K. R.

Chowdhary

Fundamentals
of Artificial
Intelligence
Fundamentals of Artificial Intelligence
K. R. Chowdhary

Fundamentals of Artificial
Intelligence

123
K. R. Chowdhary
Department of Computer Science
and Engineering
Jodhpur Institute of Engineering
and Technology
Jodhpur, Rajasthan, India

ISBN 978-81-322-3970-3 ISBN 978-81-322-3972-7 (eBook)

https://doi.org/10.1007/978-81-322-3972-7
© Springer Nature India Private Limited 2020
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature India Private Limited
The registered company address is: 7th Floor, Vijaya Building, 17 Barakhamba Road, New Delhi
110 001, India
To my wife
Suman Lata
Preface

The course of Artificial Intelligence (AI), is one of the standard course in all
computer science curricula at undergraduate programs, in many of masters pro-
grams, as well as taught a course in many of multi-disciplinary majors like Masters
in Mathematics, Masters in Information Technology, and in non-CS Programs like
Electrical Engineering, Electronics and Communication Engineering, and in some
Institutions it is taught by the name Intelligent Systems. The concepts and theories
of AI are often applied in many Engineering disciplines, like automation, pro-
duction, packing, optimizations, planning; in routing of information and goods (in
transporting); in message communications and message filtering; in natural lan-
guage translation, speech synthesis, speech recognition, search engines, information
retrieval, information extraction, information summarization, information filtering;
in robotics, computer/machine vision systems; expert systems in medical diagnosis,
fault diagnosis in machines, surveillance, remote sensing, automated vehicles; in
computer games, combinatorial games; in recent use in on-line sales, on-line
education, and many more.
ACM Curricula 2013 defines the scope of Artificial Intelligence as follows:1
Intelligent Systems (IS): Artificial intelligence (AI) is the study of solutions for
problems that are difficult or impractical to solve with traditional methods. It is used
pervasively in support of everyday applications such as email, word-processing and
search, as well as in the design and analysis of autonomous agents that perceive
their environment and interact rationally with the environment. The solutions rely
on a broad set of general and specialized knowledge representation schemes,
problem solving mechanisms, and learning techniques. They deal with sensing
(e.g., speech recognition, natural language understanding, computer vision),
problem-solving (e.g., search, planning), and acting (e.g., robotics) and the archi-
tectures needed to support them (e.g., agents, multi-agents). The study of Artificial

1
https://www.acm.org/binaries/content/assets/education/cs2013_web_final.pdf.

vii
viii Preface

Intelligence prepares the student to determine when an AI approach is appropriate

for a given problem, identify the appropriate representation and reasoning mech-
anism, and implement and evaluate it.

Contents

This book covers in enough details all the topics recommended in the latest ACM
“Computer Science Curricula 2013” (available at the time of writing this manu-
script). The topics as per the curricula 2013 and covered in this text are as follows:
Fundamental Issues
Basic Search Strategies
Basic Knowledge Representation and Reasoning
Basic Machine Learning
Advanced Search
Advanced Representation and Reasoning
Reasoning under Uncertainty
Agents
Natural Language Processing
Advanced Machine Learning
Perception and Computer Vision
In addition, looking to their importance and scope in AI, this text covers fol-
lowing additional topics.
Logic Programming and Prolog
Constraint Satisfaction Problems
Game Theory
Data Mining
Information Retrieval
Speech Recognition
This book is written, keeping in mind the diversity of its readership—the con-
cepts and fundamental principles are given highest significance, more and more
explanation is based on figures and solved examples. This is supplemented with
exhaustive list of exercises at the end of each chapter, which are aimed to cover
wide range of problem solving skills for a learner. Each chapter begins with
introduction to the topic of chapter, its scope, and relation to other fields of AI,
supported with list of applications to keep constant motivation of the learner from
chapter to chapter, followed with learning outcome of that chapter. To maintain a
consistent interest of the student as well for ease of understanding, each chapter is
organized to gradually pick up—from simple concepts to advanced theory and
applications. For curious and interested readers, each chapter is supplemented with
exhaustive list of original source references, in the form of recent articles published
in Journals, Transactions, and Proceedings of Conferences of AI and CS.
Preface ix

The text is designed to be self taught, concept driven, simple language, with
large number of exercises both solved and unsolved. The simplicity of English
language used helps in easy grasping by an average student, and equally suited for
non-English speaking students having medium of instruction as English. It covers
all major areas of AI presently in focus and that belong to recent developments,
including machine learning, machine vision, IR, agents, etc.
The contents of this text are based on teaching this course by the author for many
years at M.B.M. Engineering College, Jodhpur (India), as well as based on the
course of AI taught to undergraduate students at Indian Institute of Technology
Jodhpur (IITJ) from the session 2011–12 to 2015–16 as a visiting faculty, the time
during which the book came into this shape. The slides of the topics taught at IITJ
can be referred by reader in the author’s website (http://www.krchowdhary.com/).
The contents of this text is class room tested over the years, and has been refined
and improved again and again over these years of teaching; all the exercises are also
class room tested.

Intended Audience and Prerequisites

The book is intended primarily for students majoring in computer science at under-
graduate and graduate level, but will also be of interest as a foundation to researchers
in any of the discipline of AI, either for self study or reference. In general, it is suitable
for all AI curricula of undergraduate courses in Engineering, particularly, Computer
Science and Engineering, Information Technology, Electrical Engineering,
and non-engineering curricula, like Master in Computer Applications, Master of
Science. Some of the chapters, particularly, the Chaps. 13–21 may form course
contents of graduate curricula. The chapters can be read in the order they appear in this
text. Readers may correspond to the author (at mail-ids: [email protected] or
[email protected]) to point out the errors, suggestions/criticism, so as to
make improvements in the script for future editions.
The prerequisites for this course are: mathematical logic, discrete mathematical
structures, and data structures with basic level of programming. This course can be
taught at third or final year of undergraduate program by name as AI or Intelligent
Systems.
Since all the solved, as well as many unsolved problems in the book are
classroom tested, these can go in parallel to the respective chapter’s delivery.
A keen learner is recommended to refer to the exhaustive list of references provided
at the end of each chapter for more details.
x Preface

Examples, Exercises, and Learning Curve

The exercises are designed to continuously develop and strengthen the ideas pre-
sented in the book’s many proofs and examples. More than 250 exercises and
solved examples have been provided in the text to rigorously practice the material
presented, and to consolidate the principles presented to strengthen the over all
understanding of topics.

Supplementary Reading Material

Lecture notes, classroom slides, problems—both solved and unsolved—quizzes,

exercises, questions papers, and solutions of mid-semesters and end-semesters
of the AI courses taught at IITJ, and list of semester projects along with their
guidelines, is available in the author’s website http://www.krchowdhary.com, can
be used as supplementary reading material by the learner. In due course, and time to
time, more help and learner’s material shall be added by the author for this text on
this URL. The students may send their feedback to the author for
suggestions/errors/improvements in the contents of AI-related text on this website,
their efforts shall be an invaluable contribution to improve the quality of this book.

Jodhpur, India Prof. K. R. Chowdhary

Acknowledgements

First and foremost, I thankfully acknowledge to all the authors mentioned in the
bibliography/references in this text, whose publications were helpful for teaching
this course of AI, as well as in bringing this text in final shape.
One of the major contribution to be acknowledged is due to my students of
AI class over a duration of one and half decade, who interacted in many ways while
I was teaching this course—by asking interesting questions, some times as curious
observers, and some times raising their doubts—all these led me to rethink, analyse
their point of view, and to discover new ways to present the ideas. This also helped
me to understand the subject even better during the time of 1997–2013 when I taught
undergraduate batches of CSE, IT, and Master’s students of computer applications at
M.B.M Engineering College, Jodhpur.
The major part of this text came into shape during the time I taught at Indian
Institute of Technology Jodhpur (IITJ) (with course nomenclatures as CS365, CS
30002, CS 324, CS323) during the academic years from 2011–12 to 2015–16.
Many thought provoking questions and interactions by students during this time
helped in organizing and presenting the material in better and digestible form, as
well as many helpful and constructive comments in their feedback system not only
improved the delivery and quality of problems solutions and exercises, but was also
a constant force to reshape the content from academic year to academic year.
I would particularly like to mention their names, however, list is long and many
names also exist, though not explicitly mentioned here. These are: Ashish Kumar,
Praneeth A. S., Sonu Mehta, Manu Agarwal, Akansha Saran, Ruchi Toshiwal,
Sonal Gupta, Tokla Sai Teja, Mahesh, Kalpnath Rao, Rishi Mishra, Syed Navaid
Ahmad, Jitendra Singh Garhwal, Pitta Divya Shree, Rishi Mishra, Jinak Jain, Avan
Jayendra Rathod, Nithin V., Pavan V. Sukalkar, Sumit Jangid, Siddarth Singh Rao,
Banoth Surya Prasad, Saurabh Kumar Gangwar, Abhishek Bassan, Arvind Pandey,
Shivam Verma, Dhanajit Brahma, Pranjal Singh, Ujjwal Anand, Jalaj Sharma, and
Shrey Maheshwari.

xi
xii Acknowledgements

I am thankful to the former Director IITJ, Prof. P. Kalara, and the next Director
Prof. C. V. R. Murty, for providing me the opportunity to continuously teach the
courses of AI, semester after semester, at IIT Jodhpur during the academic years
2011–12 to 2015–16, as well as the thanks are due to various heads of CS
department of IITJ, at that time, Dr. Venkata Ramana Badaria and Dr. Gaurav Harit
for their constant support while teaching at CSE department of IITJ. My thanks to
IITJ librarin Ms. Kshema Prakash for providing needed library support
time-to-time.
I am thankful to my guru Prof. V. S. Bansal, former Prof. and Head, Department of
Electrical Engineering, Professor Emeritus, and former Dean of the M.B.M.
Engineering College, whose every meeting with me starts and ends with technical
discussions on AI, as well as thanks to Prof. (late) Rajesh Purohit, Prof. N. C. Barwar,
and Prof. Anil Gupta of M.B.M. for their active involvements while discussing the
advances in AI, time to time.
Completing a script for a book in general is a job of deep involvement for many
years, and particularly this one for me, which took enormous amount of time for
preparation of lecture notes, referring to original material in Journals, filtering and
putting the text in digestible form looking to the need of the audience, was naturally
a big share of time that belonged to my life partner Suman Lata, and children
Pratibha, Keerti, Prabhat, and Garvita (dauhit r i). So in many indirect ways their
abundant contribution exists in completing this script, and they are rightfully
entitled for many thanks for their hidden contribution in bringing the book in this
shape.

Jodhpur, India Prof. K. R. Chowdhary

September 2019
Contents

1 Introducing Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The Turing Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Goals of AI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Roots of AI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4.1 Philosophy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4.2 Logic and Mathematics . . . . . . . . . . . . . . . . . . . . . . 6
1.4.3 Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4.4 Psychology and Cognitive Science . . . . . . . . . . . . . . 7
1.4.5 Biology and Neuroscience . . . . . . . . . . . . . . . . . . . . 8
1.4.6 Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Artificial Consciousness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6 Techniques Used in AI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.7 Sub-fields of AI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.7.1 Speech Processing . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.7.2 Natural Language Processing . . . . . . . . . . . . . . . . . . 11
1.7.3 Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.7.4 Engineering and Expert Systems . . . . . . . . . . . . . . . 12
1.7.5 Fuzzy Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.7.6 Models of Brain and Evolution . . . . . . . . . . . . . . . . 13
1.8 Perception, Understanding, and Action . . . . . . . . . . . . . . . . . . 14
1.9 Physical Symbol System Hypothesis . . . . . . . . . . . . . . . . . . . 14
1.9.1 Formal System . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.9.2 Symbols and Physical Symbol Systems . . . . . . . . . . 15
1.9.3 Formal Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.9.4 The Stored Program Concept . . . . . . . . . . . . . . . . . . 16
1.10 Considerations for Knowledge Representation . . . . . . . . . . . . . 16
1.10.1 Defining the Knowledge . . . . . . . . . . . . . . . . . . . . . 17
1.10.2 Objective of Knowledge Representation . . . . . . . . . . 17

xiii
xiv Contents

1.10.3 Requirements of a Knowledge Representation . . . . . 18

1.10.4 Practical Aspects of Representations . . . . . . . . . . . . 18
1.10.5 Components of a Representation . . . . . . . . . . . . . . . 19
1.11 Knowledge Representation Using Natural Language . . . . . . . . 20
1.12 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2 Logic and Reasoning Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Argumentation Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3 Role of Knowledge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4 Propositional Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.1 Interpretation of Formulas . . . . . . . . . . . . . . . . . . . . 31
2.4.2 Logical Consequence . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.3 Syntax and Semantics of an Expression . . . . . . . . . . 33
2.4.4 Semantic Tableau . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5 Reasoning Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.5.1 Rule-Based Reasoning . . . . . . . . . . . . . . . . . . . . . . 38
2.5.2 Model-Based Reasoning . . . . . . . . . . . . . . . . . . . . . 38
2.6 Proof Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.6.1 Normal Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.6.2 Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.6.3 Properties of Inference Rules . . . . . . . . . . . . . . . . . . 41
2.7 Nonmonotonic Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.8 Hilbert and the Axiomatic Approach . . . . . . . . . . . . . . . . . . . 43
2.8.1 Roots and Early Stages . . . . . . . . . . . . . . . . . . . . . . 44
2.8.2 Axiomatics and Formalism . . . . . . . . . . . . . . . . . . . 45
2.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3 First Order Predicate Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2 Representation in Predicate Logic . . . . . . . . . . . . . . . . . . . . . 52
3.3 Syntax and Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4 Conversion to Clausal Form . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.5 Substitutions and Unification . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.5.1 Composition of Substitutions . . . . . . . . . . . . . . . . . . 60
3.5.2 Unification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.6 Resolution Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.6.1 Theorem Proving Formalism . . . . . . . . . . . . . . . . . . 64
3.6.2 Proof by Resolution . . . . . . . . . . . . . . . . . . . . . . . . 64
3.7 Complexity of Resolution Proof . . . . . . . . . . . . . . . . . . . . . . . 65
Contents xv

3.8 Interpretation and Inferences . . . . . . . . . . . . . . . . . . . . . . . . . 66

3.8.1 Herbrand’s Universe . . . . . . . . . . . . . . . . . . . . . . . . 68
3.8.2 Herbrand’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . 71
3.8.3 The Procedural Interpretation . . . . . . . . . . . . . . . . . . 72
3.9 Most General Unifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.9.1 Lifting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.9.2 Unification Algorithm . . . . . . . . . . . . . . . . . . . . . . . 79
3.10 Unfounded Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.11 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4 Rule Based Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.2 An Overview of RBS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.3 Forward Chaining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.3.1 Forward Chaining Algorithm . . . . . . . . . . . . . . . . . . 93
4.3.2 Conflict Resolution . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3.3 Efficiency in Rule Selection . . . . . . . . . . . . . . . . . . . 97
4.3.4 Complexity of Preconditions . . . . . . . . . . . . . . . . . . 98
4.4 Backward Chaining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.4.1 Backward Chaining Algorithm . . . . . . . . . . . . . . . . 99
4.4.2 Goal Determination . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.5 Forward Versus Backward Chaining . . . . . . . . . . . . . . . . . . . 100
4.6 Typical RB System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.7 Other Systems of Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.7.1 Model-Based Systems . . . . . . . . . . . . . . . . . . . . . . . 103
4.7.2 Case-Based Reasoning . . . . . . . . . . . . . . . . . . . . . . 104
4.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5 Logic Programming and Prolog . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.2 Logic Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.3 Interpretation of Horn Clauses in Rule-Chaining . . . . . . . . . . . 114
5.4 Logic Versus Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.4.1 Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.4.2 Procedure-Call Execution . . . . . . . . . . . . . . . . . . . . 119
5.4.3 Backward Versus Forward Reasoning . . . . . . . . . . . 120
5.4.4 Path Finding Algorithm . . . . . . . . . . . . . . . . . . . . . . 121
5.5 Expressing Control Information . . . . . . . . . . . . . . . . . . . . . . . 122
5.6 Running Simple Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.7 Some Built-In Predicates . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
xvi Contents

5.8 Recursive Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

5.9 List Manipulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.10 Arithmetic Expressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.11 Backtracking, Cuts and Negation . . . . . . . . . . . . . . . . . . . . . . 135
5.12 Efficiency Considerations for Prolog Programs . . . . . . . . . . . . 137
5.13 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6 Real-World Knowledge Representation and Reasoning . . . . . . . . . . 143
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.2 Taxonomic Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.3 Techniques for Commonsense Reasoning . . . . . . . . . . . . . . . . 147
6.4 Ontologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.5 Ontology Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.5.1 Language and Reasoning . . . . . . . . . . . . . . . . . . . . . 151
6.5.2 Levels of Ontologies . . . . . . . . . . . . . . . . . . . . . . . . 152
6.5.3 WordNet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.5.4 Axioms and First-Order Logic . . . . . . . . . . . . . . . . . 154
6.5.5 Sowa’s Ontology . . . . . . . . . . . . . . . . . . . . . . . . . . 154
6.6 Reasoning Using Ontologies . . . . . . . . . . . . . . . . . . . . . . . . . 156
6.6.1 Categories and Objects . . . . . . . . . . . . . . . . . . . . . . 156
6.6.2 Physical Decomposition of Categories . . . . . . . . . . . 157
6.6.3 Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.6.4 Object-Oriented Analysis . . . . . . . . . . . . . . . . . . . . . 157
6.7 Ontological Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
6.8 Situation Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.8.1 Action, Situation, and Objects . . . . . . . . . . . . . . . . . 159
6.8.2 Formalism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6.8.3 Formalizing the Notions of Context . . . . . . . . . . . . . 163
6.9 Nonmonotonic Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
6.10 Default Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
6.10.1 Notion of a Default . . . . . . . . . . . . . . . . . . . . . . . . . 168
6.10.2 The Syntax of Default Logic . . . . . . . . . . . . . . . . . . 169
6.10.3 Algorithm for Default Reasoning . . . . . . . . . . . . . . . 170
6.11 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
7 Networks-Based Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.2 Semantic Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
7.2.1 Syntax and Semantics of Semantics Networks . . . . . 182
7.2.2 Human Knowledge Creation . . . . . . . . . . . . . . . . . . 184
Contents xvii

7.2.3 Semantic Nets and Natural Language Processing . . . 184

7.2.4 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.3 Conceptual Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.4 Frames and Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
7.4.1 Inheritance Hierarchies . . . . . . . . . . . . . . . . . . . . . . 189
7.4.2 Slots Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 190
7.4.3 Frame Languages . . . . . . . . . . . . . . . . . . . . . . . . . . 191
7.4.4 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
7.5 Description Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
7.5.1 Definitions and Sentence Structures . . . . . . . . . . . . . 196
7.5.2 Concept Language . . . . . . . . . . . . . . . . . . . . . . . . . 197
7.5.3 Architecture for DL Knowledge
Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
7.5.4 Value Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . 202
7.5.5 Reasoning and Inferences . . . . . . . . . . . . . . . . . . . . 203
7.6 Conceptual Dependencies . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
7.6.1 The Parser . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
7.6.2 Conceptual Dependency and Inferences . . . . . . . . . . 209
7.6.3 Scripts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
7.6.4 Conceptual Dependency Versus Semantic Nets . . . . . 211
7.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
8 State Space Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
8.2 Representation of Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
8.3 Graph Search Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
8.4 Complexities of State-Space Search . . . . . . . . . . . . . . . . . . . . 220
8.5 Uninformed Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
8.5.1 Breadth-First Search . . . . . . . . . . . . . . . . . . . . . . . . 222
8.5.2 Depth-First Search . . . . . . . . . . . . . . . . . . . . . . . . . 224
8.5.3 Analysis of BFS and DFS . . . . . . . . . . . . . . . . . . . . 225
8.5.4 Depth-First Iterative Deepening Search . . . . . . . . . . 227
8.5.5 Bidirectional Search . . . . . . . . . . . . . . . . . . . . . . . . 228
8.6 Memory Requirements for Search Algorithms . . . . . . . . . . . . . 229
8.6.1 Depth-First Searches . . . . . . . . . . . . . . . . . . . . . . . . 229
8.6.2 Breadth-First Searches . . . . . . . . . . . . . . . . . . . . . . . 230
8.7 Problem Formulation for Search . . . . . . . . . . . . . . . . . . . . . . . 230
8.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
xviii Contents

9 Heuristic Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
9.2 Heuristic Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
9.3 Hill-Climbing Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
9.4 Best-First Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
9.4.1 GBFS Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 245
9.4.2 Analysis of Best-First Search . . . . . . . . . . . . . . . . . . 247
9.5 Heuristic Determination of Minimum Cost Paths . . . . . . . . . . 249
9.5.1 Search Algorithm A . . . . . . . . . . . . . . . . . . . . . . . . 249
9.5.2 The Evaluation Function . . . . . . . . . . . . . . . . . . . . . 251
9.5.3 Analysis of A Search . . . . . . . . . . . . . . . . . . . . . . . 253
9.5.4 Optimality of Algorithm A . . . . . . . . . . . . . . . . . . . 254
9.6 Comparison of Heuristics Approaches . . . . . . . . . . . . . . . . . . 254
9.7 Simulated Annealing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
9.8 Genetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
9.8.1 Exploring Different Structures . . . . . . . . . . . . . . . . . 260
9.8.2 Process of Innovation in Human . . . . . . . . . . . . . . . 261
9.8.3 Mutation Operator . . . . . . . . . . . . . . . . . . . . . . . . . 261
9.8.4 GA Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
9.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
10 Constraint Satisfaction Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 273
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
10.2 CSP Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
10.3 Representation of CSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
10.3.1 Constraints in CSP . . . . . . . . . . . . . . . . . . . . . . . . . 277
10.3.2 Variables in CSP . . . . . . . . . . . . . . . . . . . . . . . . . . 279
10.4 Solving a CSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
10.4.1 Synthesizing the Constraints . . . . . . . . . . . . . . . . . . 281
10.4.2 An Extended Theory for Synthesizing . . . . . . . . . . . 283
10.5 Solution Approaches to CSPs . . . . . . . . . . . . . . . . . . . . . . . . 285
10.6 CSP Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
10.6.1 Generate and Test . . . . . . . . . . . . . . . . . . . . . . . . . . 288
10.6.2 Backtracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
10.6.3 Efficiency Considerations . . . . . . . . . . . . . . . . . . . . 292
10.7 Propagating of Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
10.7.1 Forward Checking . . . . . . . . . . . . . . . . . . . . . . . . . 294
10.7.2 Degree of Heuristics . . . . . . . . . . . . . . . . . . . . . . . . 294
10.8 Cryptarithmetics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
10.9 Theoretical Aspects of CSPs . . . . . . . . . . . . . . . . . . . . . . . . . 298
Contents xix

10.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299

Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
11 Adversarial Search and Game Theory . . . . . . . . . . . . . . . . . . . . . . 303
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
11.2 Classification of Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
11.3 Game Playing Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
11.4 Two-Person Zero-Sum Games . . . . . . . . . . . . . . . . . . . . . . . . 307
11.5 The Prisoner’s Dilemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
11.6 Two-Player Game Strategies . . . . . . . . . . . . . . . . . . . . . . . . . 310
11.7 Games of Perfect Information . . . . . . . . . . . . . . . . . . . . . . . . 312
11.8 Games of Imperfect Information . . . . . . . . . . . . . . . . . . . . . . 312
11.9 Nash Arbitration Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
11.10 n-Person Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
11.11 Representation of Two-Player Games . . . . . . . . . . . . . . . . . . . 317
11.12 Minimax Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
11.13 Tic-tac-toe Game Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
11.14 Alpha-Beta Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
11.14.1 Complexities Analysis of Alpha-Beta . . . . . . . . . . . . 326
11.14.2 Improving the Efficiency of Alpha-Beta . . . . . . . . . . 327
11.15 Sponsored Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
11.16 Playing Chess with Computer . . . . . . . . . . . . . . . . . . . . . . . . 329
11.17 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
12 Reasoning in Uncertain Environments . . . . . . . . . . . . . . . . . . . . . . 337
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
12.2 Foundations of Probability Theory . . . . . . . . . . . . . . . . . . . . . 339
12.3 Conditional Probability and Bayes Theorem . . . . . . . . . . . . . . 340
12.4 Bayesian Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
12.4.1 Constructing a Bayesian Network . . . . . . . . . . . . . . 344
12.4.2 Bayesian Network for Chain of Variables . . . . . . . . 345
12.4.3 Independence of Variables . . . . . . . . . . . . . . . . . . . 347
12.4.4 Propagation in Bayesian Belief Networks . . . . . . . . . 348
12.4.5 Causality and Independence . . . . . . . . . . . . . . . . . . 351
12.4.6 Hidden Markov Models . . . . . . . . . . . . . . . . . . . . . 353
12.4.7 Construction Process of Bayesian Networks . . . . . . . 354
12.5 Dempster–Shafer Theory of Evidence . . . . . . . . . . . . . . . . . . . 356
12.5.1 Dempster–Shafer Rule of Combination . . . . . . . . . . 357
12.5.2 Dempster–Shafer Versus Bayes Theory . . . . . . . . . . 358
12.6 Fuzzy Sets, Fuzzy Logic, and Fuzzy Inferences . . . . . . . . . . . 361
xx Contents

12.6.1 Fuzzy Composition Relation . . . . . . . . . . . . . . . . . . 363

12.6.2 Fuzzy Rules and Fuzzy Graphs . . . . . . . . . . . . . . . . 365
12.6.3 Fuzzy Graph Operations . . . . . . . . . . . . . . . . . . . . . 367
12.6.4 Fuzzy Hybrid Systems . . . . . . . . . . . . . . . . . . . . . . 369
12.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
13 Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
13.2 Types of Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . 377
13.3 Discipline of Machine Learning . . . . . . . . . . . . . . . . . . . . . . . 379
13.4 Learning Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
13.5 Classes of Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
13.5.1 Supervised Learning . . . . . . . . . . . . . . . . . . . . . . . . 383
13.5.2 Unsupervised Learning . . . . . . . . . . . . . . . . . . . . . . 384
13.6 Inductive Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
13.6.1 Argument-Based Learning . . . . . . . . . . . . . . . . . . . . 387
13.6.2 Mutual Online Concept Learning . . . . . . . . . . . . . . . 389
13.6.3 Single-Agent Online Concept Learning . . . . . . . . . . 391
13.6.4 Propositional and Relational Learning . . . . . . . . . . . 392
13.6.5 Learning Through Decision Trees . . . . . . . . . . . . . . 393
13.7 Discovery-Based Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
13.8 Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
13.8.1 Some Functions in Reinforcement Learning . . . . . . . 399
13.8.2 Supervised Versus Reinforcement Learning . . . . . . . 400
13.9 Learning and Reasoning by Analogy . . . . . . . . . . . . . . . . . . . 401
13.10 A Framework of Symbol-Based Learning . . . . . . . . . . . . . . . . 405
13.11 Explanation-Based Learning . . . . . . . . . . . . . . . . . . . . . . . . . 406
13.12 Machine Learning Applications . . . . . . . . . . . . . . . . . . . . . . . 408
13.13 Basic Research Problems in Machines Learning . . . . . . . . . . . 409
13.14 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 410
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
14 Statistical Learning Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
14.2 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
14.3 Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . 418
14.3.1 Learning Pattern Recognition from Examples . . . . . . 419
14.3.2 Maximum Margin Training Algorithm . . . . . . . . . . . 421
14.4 Predicting Structured Objects Using SVM . . . . . . . . . . . . . . . 422
14.5 Working of Structural SVMs . . . . . . . . . . . . . . . . . . . . . . . . . 424
Contents xxi

14.6 k-Nearest Neighbor Method . . . . . . . . . . . . . . . . . . . . . . . . . . 425

14.6.1 k-NN Search Algorithm . . . . . . . . . . . . . . . . . . . . . 426
14.7 Naive Bayes Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
14.8 Artificial Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 430
14.8.1 Error-Correction Rules . . . . . . . . . . . . . . . . . . . . . . 433
14.8.2 Boltzmann Learning . . . . . . . . . . . . . . . . . . . . . . . . 434
14.8.3 Hebbian Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
14.8.4 Competitive Learning Rules . . . . . . . . . . . . . . . . . . 435
14.8.5 Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
14.9 Instance-Based Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
14.9.1 Learning Task . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
14.9.2 IBL Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
14.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
15 Automated Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
15.2 Automated Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
15.3 The Basic Planning Problem . . . . . . . . . . . . . . . . . . . . . . . . . 448
15.3.1 The Classical Planning Problem . . . . . . . . . . . . . . . . 449
15.3.2 Agent Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
15.4 Forward Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
15.5 Partial-Order Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
15.6 Planning Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
15.6.1 A General Planning Language . . . . . . . . . . . . . . . . . 456
15.6.2 The Operation of STRIPS . . . . . . . . . . . . . . . . . . . . 457
15.6.3 Search Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458
15.7 Planning with Propositional Logic . . . . . . . . . . . . . . . . . . . . . 458
15.7.1 Encoding Action Descriptions . . . . . . . . . . . . . . . . . 460
15.7.2 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
15.8 Planning Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
15.9 Hierarchical Task Network Planning . . . . . . . . . . . . . . . . . . . 462
15.10 Multiagent Planning Systems . . . . . . . . . . . . . . . . . . . . . . . . . 464
15.11 Multiagent Planning Techniques . . . . . . . . . . . . . . . . . . . . . . 465
15.11.1 Goal and Task Allocation . . . . . . . . . . . . . . . . . . . . 466
15.11.2 Goal and Task Refinement . . . . . . . . . . . . . . . . . . . 466
15.11.3 Decentralized Planning . . . . . . . . . . . . . . . . . . . . . . 466
15.11.4 Coordination After Planning . . . . . . . . . . . . . . . . . . 467
15.12 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
xxii Contents

16 Intelligent Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471

16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
16.2 Classification of Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
16.3 Multiagent Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475
16.3.1 Single-Agent Framework . . . . . . . . . . . . . . . . . . . . . 476
16.3.2 Multiagent Framework . . . . . . . . . . . . . . . . . . . . . . 476
16.3.3 Multiagent Interactions . . . . . . . . . . . . . . . . . . . . . . 477
16.4 Basic Architecture of Agent System . . . . . . . . . . . . . . . . . . . . 479
16.5 Agents’ Coordination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 480
16.5.1 Sharing Among Cooperative Agents . . . . . . . . . . . . 481
16.5.2 Static Coalition Formation . . . . . . . . . . . . . . . . . . . . 482
16.5.3 Dynamic Coalition Formation . . . . . . . . . . . . . . . . . 482
16.5.4 Iterated Prisoner’s Dilemma Coalition Model . . . . . . 483
16.5.5 Coalition Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 485
16.6 Agent-Based Approach to Software Engineering . . . . . . . . . . . 486
16.7 Agents that Buy and Sell . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
16.8 Modeling Agents as Decision Maker . . . . . . . . . . . . . . . . . . . 488
16.8.1 Issues in Mental Level Modeling . . . . . . . . . . . . . . . 489
16.8.2 Model Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
16.8.3 Preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
16.8.4 Decision Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
16.9 Agent Communication Languages . . . . . . . . . . . . . . . . . . . . . 493
16.9.1 Semantics of Agent Programs . . . . . . . . . . . . . . . . . 495
16.9.2 Description Language for Interactive Agents . . . . . . 497
16.10 Mobile Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
16.11 Social Level View of Multiagents . . . . . . . . . . . . . . . . . . . . . 500
16.12 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504
17 Data Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
17.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
17.2 Perspectives of Data Mining . . . . . . . . . . . . . . . . . . . . . . . . . 509
17.3 Goals of Data Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
17.4 Evolution of Data Mining Algorithms . . . . . . . . . . . . . . . . . . 512
17.4.1 Transactions Data . . . . . . . . . . . . . . . . . . . . . . . . . . 513
17.4.2 Data Streams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514
17.4.3 Representation of Text-Based Data . . . . . . . . . . . . . 514
17.5 Classes of Data Mining Algorithms . . . . . . . . . . . . . . . . . . . . 515
17.5.1 Prediction Methods . . . . . . . . . . . . . . . . . . . . . . . . . 515
17.5.2 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518
17.5.3 Association Rules . . . . . . . . . . . . . . . . . . . . . . . . . . 519
Contents xxiii

17.6 Data Clustering and Cluster Analysis . . . . . . . . . . . . . . . . . . . 519

17.6.1 Applications of Clustering . . . . . . . . . . . . . . . . . . . . 521
17.6.2 General Utilities of Clustering . . . . . . . . . . . . . . . . . 522
17.6.3 Traditional Clustering Methods . . . . . . . . . . . . . . . . 523
17.6.4 Clustering Process . . . . . . . . . . . . . . . . . . . . . . . . . 523
17.6.5 Pattern Representation and Feature Extraction . . . . . 525
17.7 Clustering Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
17.7.1 Similarity Measures . . . . . . . . . . . . . . . . . . . . . . . . 527
17.7.2 Nearest Neighbor Clustering . . . . . . . . . . . . . . . . . . 528
17.7.3 Partitional Algorithms . . . . . . . . . . . . . . . . . . . . . . . 529
17.8 Comparison of Clustering Techniques . . . . . . . . . . . . . . . . . . 531
17.9 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
17.10 Association Rule Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
17.11 Sequential Pattern Mining Algorithms . . . . . . . . . . . . . . . . . . 541
17.11.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . 542
17.11.2 Notations for Sequential Pattern Mining . . . . . . . . . . 542
17.11.3 Typical Sequential Pattern Mining . . . . . . . . . . . . . . 543
17.11.4 Apriori-Based Algorithm . . . . . . . . . . . . . . . . . . . . . 544
17.12 Scientific Applications in Data Mining . . . . . . . . . . . . . . . . . . 549
17.13 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 555
18 Information Retrieval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
18.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
18.2 Retrieval Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 560
18.3 Boolean Model of IR System . . . . . . . . . . . . . . . . . . . . . . . . . 561
18.4 Vector Space Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563
18.5 Indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
18.5.1 Index Construction . . . . . . . . . . . . . . . . . . . . . . . . . 565
18.5.2 Index Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . 568
18.6 Probabilistic Retrieval Model . . . . . . . . . . . . . . . . . . . . . . . . . 569
18.7 Fuzzy Logic-Based IR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 570
18.8 Concept-Based IR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574
18.8.1 Concept-Based Indexing . . . . . . . . . . . . . . . . . . . . . 575
18.8.2 Retrieval Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 578
18.9 Automatic Query Expansion in IR . . . . . . . . . . . . . . . . . . . . . 579
18.9.1 Working of AQE . . . . . . . . . . . . . . . . . . . . . . . . . . 583
18.9.2 Related Techniques for Query Processing . . . . . . . . . 585
18.10 Using Bayesian Networks for IR . . . . . . . . . . . . . . . . . . . . . . 587
18.10.1 Representation of Document and Query . . . . . . . . . . 587
18.10.2 Bayes Probabilistic Inference Model . . . . . . . . . . . . 588
xxiv Contents

18.10.3 Bayes Inference Algorithm . . . . . . . . . . . . . . . . . . . 589

18.10.4 Representing Dependent Topics . . . . . . . . . . . . . . . . 592
18.11 Semantic IR on the Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592
18.12 Distributed IR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595
18.13 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
19 Natural Language Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603
19.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603
19.2 Progress in NLP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 606
19.3 Applications of NLP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 608
19.4 Components of Natural Language Processing . . . . . . . . . . . . . 609
19.4.1 Syntax Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 609
19.4.2 Semantic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 611
19.4.3 Discourse Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 611
19.5 Grammars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612
19.5.1 Phrase Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 613
19.5.2 Phrase Structure Grammars . . . . . . . . . . . . . . . . . . . 613
19.6 Classification of Grammars . . . . . . . . . . . . . . . . . . . . . . . . . . 616
19.6.1 Chomsky Hierarchy of Grammars . . . . . . . . . . . . . . 616
19.6.2 Transformational Grammars . . . . . . . . . . . . . . . . . . 617
19.6.3 Ambiguous Grammars . . . . . . . . . . . . . . . . . . . . . . 619
19.7 Prepositions in Applications . . . . . . . . . . . . . . . . . . . . . . . . . . 620
19.8 Natural Language Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . 621
19.8.1 Parsing with CFGs . . . . . . . . . . . . . . . . . . . . . . . . . 622
19.8.2 Sentence-Level Constructions . . . . . . . . . . . . . . . . . 624
19.8.3 Top-Down Parsing . . . . . . . . . . . . . . . . . . . . . . . . . 625
19.8.4 Probabilistic Parsing . . . . . . . . . . . . . . . . . . . . . . . . 627
19.9 Information Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 630
19.9.1 Document Preprocessing . . . . . . . . . . . . . . . . . . . . . 630
19.9.2 Syntactic Parsing and Semantic Interpretation . . . . . . 631
19.9.3 Discourse Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 632
19.9.4 Output Template Generation . . . . . . . . . . . . . . . . . . 633
19.10 NL-Question Answering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633
19.10.1 Data Redundancy Based Approach . . . . . . . . . . . . . 634
19.10.2 Structured Descriptive Grammar-Based QA . . . . . . . 635
19.11 Commonsense-Based Interfaces . . . . . . . . . . . . . . . . . . . . . . . 636
19.11.1 Commonsense Thinking . . . . . . . . . . . . . . . . . . . . . 638
19.11.2 Components of Commonsense Reasoning . . . . . . . . 638
19.11.3 Representation Structures . . . . . . . . . . . . . . . . . . . . 640
Contents xxv

19.12 Tools for NLP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 642

19.12.1 NLTK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 642
19.12.2 NLTK Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
19.13 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 645
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 647
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
20 Automatic Speech Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 651
20.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 651
20.2 Automatic Speech Recognition Resources . . . . . . . . . . . . . . . . 653
20.3 Voice Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
20.4 Speech Recognition Algorithms . . . . . . . . . . . . . . . . . . . . . . . 656
20.5 Hypothesis Search in ASR . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
20.5.1 Lexicon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
20.5.2 Language Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
20.5.3 Acoustic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 659
20.6 Automatic Speech Recognition Tools . . . . . . . . . . . . . . . . . . . 662
20.6.1 Automatic Speech Recognition Engine . . . . . . . . . . . 663
20.6.2 Tools for ASR . . . . . . . . . . . . . . . . . . . . . . . . . . . . 664
20.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 666
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 667
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 668
21 Machine Vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 669
21.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 669
21.2 Machine Vision Applications . . . . . . . . . . . . . . . . . . . . . . . . . 671
21.3 Basic Principles of Vision . . . . . . . . . . . . . . . . . . . . . . . . . . . 672
21.4 Cognition and Classification . . . . . . . . . . . . . . . . . . . . . . . . . 675
21.5 From Image-to-Scene . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677
21.5.1 Inversion by Fixing Scene Parameters . . . . . . . . . . . 678
21.5.2 Inversion by Restricting the Problem Domain . . . . . . 678
21.5.3 Inversion by Acquiring Additional Images . . . . . . . . 679
21.6 Machine Vision Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . 680
21.6.1 Low-Level Vision . . . . . . . . . . . . . . . . . . . . . . . . . . 680
21.6.2 Local Edge Detection . . . . . . . . . . . . . . . . . . . . . . . 681
21.6.3 Middle-Level Vision . . . . . . . . . . . . . . . . . . . . . . . . 683
21.6.4 High-Level Vision . . . . . . . . . . . . . . . . . . . . . . . . . 685
21.7 Indexing and Geometric Hashing . . . . . . . . . . . . . . . . . . . . . . 687
21.8 Object Representation and Tracking . . . . . . . . . . . . . . . . . . . . 689
21.9 Feature Selection and Object Detection . . . . . . . . . . . . . . . . . 692
21.9.1 Object Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 694
21.10 Supervised Learning for Object Detection . . . . . . . . . . . . . . . . 696
xxvi Contents

21.11 Axioms of Vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 698

21.11.1 Mathematical Axioms . . . . . . . . . . . . . . . . . . . . . . . 699
21.11.2 Source Axioms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 699
21.11.3 Model Axioms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 700
21.11.4 Construct Axioms . . . . . . . . . . . . . . . . . . . . . . . . . . 700
21.12 Computer Vision Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 701
21.13 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 703
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 706

Further Readings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 707

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 709
About the Author

Prof. K. R. Chowdhary is currently a Professor of Computer Science at JGI,

Jodhpur. He is the former Professor and Head, Department of Computer Science &
Engineering, M.B.M. Engineering College, Jodhpur, India. He completed his Ph.D.
from the same university in 2004. Prof. Chowdhary has over 35 years of teaching
and research experience. He had also taught at IIT Jodhpur (2010–2017), and
served as the Director at JIET Group of Institutions, Jodhpur from 2014 to 2019. He
was honored with senior membership award of Association for Computing
Machinery in 2008 and Eminent Computer Engineer’s Award from the Institute of
Engineers, India in 2011. He has authored several papers published in national and
international journals and conference proceedings. His areas of specialization
include: Discrete Mathematical Structures, Theory of Computation, AI, Machine
Learning, Natural Language and Speech Processing.

xxvii
Acronyms

ABL Argument-based Learning

ABOX Assertion Box
ACL Agent Communication Language
ADJ Adjective
ADV Adverb
AI Artificial Intelligence
ANNs Artificial Neural networks
AQE Automatic Query Expansion
ASR Automatic Speech Recognition
ATR Automatic Target Recognition
AUX Auxiliary verb
BFS Breadth-first Search
BOW Bag-of-words
CBR Case-based Reasoning
CD Conceptual Dependency
CF Cluster Feature
CFG Context-free Grammar
CG Conceptual Graph
CGT Combinatorial Game Theory
CLIR Cross-language IR
CNF Conjunctive Normal Forms
CSP Constraint Satisfaction Problem
CYC enCYClopedia ontology
DAG Directed Acyclic Graph
DET Determiner
DFID Depth-First Iterative Deepening
DFS Depth-first Search
DL Description Logic
DNF Disjunctive Normal Forms
DST Dempster-Shafer Theory

xxix
xxx Acronyms

FIR Fuzzy Information Retrieval

FOL First Order Logic
FOPL First Order Predicate Logic
FST Finite State Transducer
GAs Genetic Algorithms
GB Gödel-Bernays
HMM Hidden Markov Model
IBL Instance-based Learning
IE Information Extraction
IR Information Retrieval
KB Knowledge Base
KIF Knowledge Interchange Format
k-NN k-Nearest Neighbor
KR Knowledge Representation
LM Language Modeling
LOOM Lexical OWL Ontology Matcher
MAS Multi-agent System
MDP Markov Decision Process
NLP Natural Language Processing
NLTK Natural Language Tool-Kit
NN Nearest Neighbor
NP Nondeterministic Polynomial
OCR Optical Character Recognition
OWL Web Ontology Language
PCA Principle Component Analysis
PD Prisoner’s Dilemma
PSSH Physical Symbol System Hypothesis
QA Question Answering
RBS Rule-based System
RGB Red Green Blue
RL Reinforcement Learning
SRC Search Results Clustering
STRIPS STanford Research Institute Problem Solver
SVM Support Vector Machine
SVN Support Vector Network
TBOX Terminology Box
TREC Text REtrieval Conference Series
TSP Traveling Salesman Problem
WAM Warren Abstract Machine
WSD Word Sense Disambiguation
XCON Expert Configurator
ZF Zermilo-Fraenkel
Chapter 1
Introducing Artificial Intelligence

Abstract The field of Artificial Intelligence (AI) has interfaces with almost every
discipline of engineering, and beyond, and all those can be benefited by the use of
AI. This chapter presents the introduction to AI, its roots, sub-domains, Turing test
to judge if the given program is intelligent, what are the goals of AI for Engineers
and Scientists, what are the basic requirements for AI, symbol system, what are the
requirements for knowledge representation, and concludes with a chapter summary,
and an exhaustive list of exercises. In addition raises many questions to ponder over,
like, consciousness, and whether it is possible to create artificial consciousness.

Keywords Artificial intelligence · Turing test · Imitation game · Philosophy ·

Logic · Computation · Cognitive science · Evolution · Artificial consciousness ·
Speech processing · Expert systems · Physical symbol system hypothesis · Formal
system · Knowledge representation

1.1 Introduction

The Artificial Intelligence (AI) is a branch of Computer Science, which is mainly

concerned with automation of Intelligent behavior. This behavior we may consider
from all domains—the human, animal world, and vegetation. A compact definition
of Intelligence is:

Intelligence = Perceive + Analyze + React.

The following are often quoted definitions, all expressing this notion of intelli-
gence but with different emphasis in each case:

• “The capacity to learn or to profit by experience.”

• “Ability to adapt oneself adequately to relatively new situations in life.”
• “A person possesses intelligence insofar as he has learned, or can learn, to adjust
himself to his environment.”
• “The ability of an organism to solve new problems.”

© Springer Nature India Private Limited 2020 1

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_1
2 1 Introducing Artificial Intelligence

• “A global concept that involves an individual’s ability to act purposefully, think

rationally, and deal effectively with the environment.”
• “Intelligence is a very general mental capability that, among other things, involves
the ability to reason, plan, solve problems, think abstractly, comprehend complex
ideas, learn quickly and learn from experience.”
The foundation materials of AI comprises—data structures, knowledge represen-
tation techniques, algorithms to apply the knowledge and language, and the program-
ming techniques to implement all these.
To get an idea of Intelligence it requires answering these and many similar ques-
tions:

• Is intelligence due to a single faculty or it is a name for a collection of distinct

unrelated faculties?
• Is it a priori existence or it can be learned? What does exactly happen when we
learn some thing, that is, in terms of information and storage structures.
• What is truly the process of creativity and intuition in human? These again, in
terms of knowledge and its structures.
• Does the intelligence require an internal mechanism, or it can be concluded from
the behavior observed?
• What is the mechanism for representing the knowledge in the living cells?
• Are the machines self aware like humans? What are the basic requirements for
creating the facility of self-awareness in machines?
• Is it that computer intelligence can be defined only when we know the intelligence
in reference to human beings?
• Would it ever be possible to achieve the intelligence in computers? Or, is it true
that achievement of intelligence is possible only when or is it that an intelligent
entity requires the richness of sensation and experience which might be found only
in a biological existence?

Partly, the aim of Artificial Intelligence (AI) is to find the answer to these questions,
through the tools provided by AI. These tools are because, the AI offers the medium,
as well as the test-bed for theories of intelligence, which can be expressed in the
form of computer programs, and can be tested as well as verified by running these
programs on computers.
Unlike Physics and Chemistry, AI is still a premature field, hence, its structure,
objectives, and procedures are less clearly defined, and not clear like those in physics
and chemistry. The AI has been more concerned about expanding limits of computers,
apart from defining itself.
Learning Outcomes of this Chapter1 :
1. Defining AI. [Familiarity]
2. Describe Turing test thought experiment. [Familiarity]

1 https://www.acm.org/binaries/content/assets/education/cs2013_web_final.pdf (pp. 121–129).

1.1 Introduction 3

3. Differentiate between the concepts of optimal reasoning/behavior and humanlike

reasoning/behavior. [Familiarity]
4. Sub-fields of AI. [Familiarity]
5. Determine the characteristics of a given problem that an intelligent system (i.e.,
AI-based system) must solve. [Assessment]

1.2 The Turing Test

In 1950, in an article “Computing Machinery and Intelligence,” Alan M. Turing

proposed an empirical test for machine intelligence, now called Turing Test (see
Fig. 1.1). It is designed to measure the performance of an intelligent machine against
humans, for its intelligent behavior. Turing called it imitation game, where machine
and human counter-part are put in different rooms, separate from a third person,
called interrogator. The interrogator is not able to see or speak directly to any of the
other two, and does not know which entity is a machine, and communicates to these
two solely by textual devices like a dumb terminal [11, 12].
The interrogator is supposed to distinguish the machine from the human solely
based on the answers received for the questions asked over the interface device,
which is a keyboard (or teletype). Even after having asked the number of questions,
if the interrogator is not able to distinguish the machine from the human, then as per
the argument of Alan Turing, the machine can be considered intelligent. Interrogator
may ask highly computation oriented questions to identify the machine, and other
questions related to general awareness, poetry, etc., to identify the human [6].
The game (with the “player machine” omitted) is often in practice under the
name of viva-voce to discover whether someone really understands something or
has “learned it parrot fashion”.
Many researchers argue that the Turing test is not sufficient to establish the pres-
ence of intelligence. Some of the arguments for and against the above test can be as
follows:

Closed Room
Interrogator

cat ..
dog..

T2
T1 T3
Computer
running an T1, T2, T3: Dumb terminals
intelligent program

Fig. 1.1 Turing test (imitation game)

4 1 Introducing Artificial Intelligence

1. It takes human being as a reference for intelligent behavior, rather than debating
over the true nature of intelligence: against.
2. The unmeasurable things are not considered, e.g., whether a computer uses inter-
nal structures, or for example, whether the machine is conscious of its actions,
which are currently not answerable: against.
3. Eliminates any bias to human oriented interaction mechanisms, since a computer
terminal is used as a communication device: for.
4. Biased towards only symbolic problem solving: against.
5. Perceptual skills or dexterity cannot be checked: against.
6. Unnecessarily constrains the machine intelligence to human intelligence: against.

Though, the number of counts of against are far more than for, there is yet no
known test which is considered better than the Turing Test.
In the part success of the Turing Test, a powerful computer has deceived humans
in the thinking process, where this machine modeled intelligence of a young boy,
to become the first machine to pass the Turing test, conducted in June 2014. In this
experiment, five machines were tested at the Royal Society in central London to
check if these machines could fool the people into thinking. The machines behaved
like humans, and the conversation was in the form of text. A computer program, by
the name “Eugene Goostman” was developed to simulate a young boy, which came
out to convince the one-third of the judges that it was human [5].

1.3 Goals of AI

AI is the area of computer science aiming for the design of intelligent computer
systems, i.e., systems that have characteristics of intelligence, like, we observe in
human behavior, for example, to understand language(s), and have abilities of learn-
ing, reasoning, and problem solving [9].
For many researchers, the goal of AI is to emulate human cognition, while to
some researchers, it is the creation of intelligence without considering any human
characteristics. To many other researchers, AI is aimed to create useful artifacts
for the comforts and needs of human, without any criteria of an abstract notion of
intelligence.
The above variation in aims is not necessarily a wrong, as each approach uncovers
new ideas and provides a base for pursuing the research in AI. However, there is a
convincing argument that due to the absence of a proper definition of AI, it is difficult
to establish as what can and what cannot be done through AI.
One of the goals for studying AI is to create intelligence in machines as a general
property—not necessarily based on any attribute of humans. When this is the goal,
it also includes the objective of creation of artifacts of human comforts and needs,
which can be the driving force of technological development. However, this goal
also requires a notion of intelligence, to start with.
1.3 Goals of AI 5

Further, the problem gets compounded, since artifacts manufacturer may say their
product is better in terms of saving labor and money, while cognitive scientists may
say that their system correctly predicted human behavior. Apart from this dispute,
without proper theoretical base of AI, it is not a wise idea to build complex system—
on which one has confidence, can be analyzed for performance and error analysis
can be carried out.
The definition of AI be such that for any system, it covers input, output, and their
relationship based on the structure of the system. There is a need of such definition to
be as general as possible so that it can be uniformly applicable. In the absence of such
a definition, one takes the AI as that exists in Chess playing, the one in automated
vehicle driving, and the one existing in medical expert system for diagnosis—all
these approaches for the definition of AI are varying from case to case.
The scientific goal of AI is to determine theories about knowledge representation,
learning, rule-based systems, and search that explain various sorts of intelligence.
The engineering goal of AI is to acquire ability in the machine so that it can
solve the real-life problems. The basic techniques used by AI for this purpose are
knowledge representation, machine learning, rule systems, and state space search.
In the past, computer scientists and engineers used to be more concerned with
the engineering goals, but the psychologists, philosophers, and cognitive scientists
were more keen on the scientific goals. In spite of these opposite concerns, there are
common techniques that the two approaches can feed to each other. Hence, we will
proceed with both the goals in mind.

1.4 Roots of AI

The field of AI does not live in isolation, and has significant roots in number of older
disciplines, particularly,

Philosophy,
Logic/Mathematics,
Computing,
Psychology/Cognitive Science,
Biology/Neuroscience, and
Evolution.

In the above domains, there is significant overlap, for example, between philos-
ophy and logic, and between mathematics and computation. By looking at each of
these in turn, we get a better understanding about their role in AI, and how these
fields have developed to play that role in AI [6, 10].
6 1 Introducing Artificial Intelligence

1.4.1 Philosophy

The evidence of philosophy goes as back as Socrates times (∼400 BC) where he
asks for an algorithm to distinguish between piety (a reverence of supreme being)
from non-piety. In around 300 BC, Aristotle formulated various types of deductive
reasoning approaches, to mechanically generate conclusions using initial premises.
One approach of deductive reasoning he used was modus ponens, now also a standard
technique of inference in predicate and propositional logic. It is stated as
If “A is True” → “B is True”, and “A holds True” then conclude that “B holds True”.

As an example of inference, “If it’s raining then you get wet, and it’s raining, then
you should have got wet”.
The philosopher Rene Descartes (1596–1650) introduced the concept of mind-
body dualism, which says that part of the mind is exempted from following the
physical laws. The conclusion of which he has drawn as free will. In the present time
also, when the AI, machine learning, and data science are dominantly used, but it
is argued that machine cannot supersede human in intelligence, as they do not have
free will, and they need to be assigned the goal by a human.
Gottfried Wilhelm Leibnitz, a German philosopher and mathematician (1646–
1716), who supported the materialist nature of mind, said that mind operates by
ordinary physical processes. In the present context, this means that mental processes
can be performed by the machines.

1.4.2 Logic and Mathematics

The logic, has history of development from the time of Greece philosophers—Plato
and Aristotle (∼300–400 BC), however, there has been more recent development at
a rapid pace. These were due to the following
• Earl Stanhope’s Logic (1777), using which Earl demonstrated a machine capable
of solving the problem using the inference rule, called syllogisms, and the numer-
ical problems of logical nature, and elementary questions based on the theory of
probability.
• George Boole (1815–1864) introduced the language-based formal logic for a draw-
ing logical inference in 1847, later became popular by name Boolean Algebra.
• Gottlob Frege (1848–1925) introduced the first-order logic that today forms the
most common knowledge representation system, called as FOPL (first-order pred-
icate logic).
• Kurt Gödel (1906–1978), in 1931, demonstrated that there are limits of logic.
Through his incompleteness theorem he showed that in any formal logic, which
is powerful enough to describe the properties of natural numbers, there exist true
statements, but their truth cannot be proved using any algorithm.
1.4 Roots of AI 7

• Roger Penrose in 1995 tried to prove that human mind has non-computable capa-
bilities.

1.4.3 Computation

In the nineteenth and twentieth-century, many scientists defined the formalism of

what is computation, basic theory of it, and that there are things that are not com-
putable irrespective of whatever are the computing resources and time provided.
In the year 1869, William Jevon constructed a Logic Machine capable of handling
Boolean Algebra and Venn Diagrams, and could solve logical problems faster than
human beings.
Alan M. Turing (1912–1954) tried to characterize exactly what are the functions
that can be computed. He used, what is now called as Turing Machine. Unfortunately,
it is difficult to give the notion of computation as a formal definition, however,
the Church-Turing thesis, due to Alonzo Church and Turing, states that a Turing
machine is capable of computing any computable function, which is now, accepted
as a sufficient definition of computability. Turing also showed that there are some
functions which no Turing machine can compute (e.g., Halting Problem)—these are
non-computable functions.
John von Neumann (1903–1957) gave, now what is called as, von Neumann
architecture—a description of a logical model of computation and computer, without
any physical realization of a computer.
In 1960s, two important concepts emerged—Intractability (the solution time of
a problem grows at least exponentially) and Reduction of complex problems into
simpler problems.

1.4.4 Psychology and Cognitive Science

Cognitive Psychology or Cognitive Science is the study about the functioning of the
mind, human behavior, and the processing of information about the human brain. An
important consequence of human intelligence is human languages. The early work
on knowledge representation in AI was about human language, and was produced
through research in linguistics.
It is humans’ quest to understand as to how our and other animals’ brains lead
to intelligent behavior, with the aim to ultimately build AI systems. On the other
hand, it is also aimed to explore the properties of artificial systems, like, computer
models/simulations to test our hypotheses concerning human systems.
Many people working in sub-fields of AI are in the process of building models
of how the human system operates, and use artificial systems for solving real-world
problems.
8 1 Introducing Artificial Intelligence

1.4.5 Biology and Neuroscience

The field of neuroscience says that human brains, which provide intelligence, are
made up of tens of billions of neurons, and each neuron is connected to hundreds or
thousands of other neurons. A neuron is an elementary processing unit, performing a
function called firing, depending on the total amount of activity feeding into it. When
a large number of neurons are connected together, it gives rise to a very powerful
computational device that can compute, as well as learn how to compute.
The concept of human brains, having the capability to compute, as well as learn,
is used to build artificial neurons in the form of electronics circuits, and connect
them as circuits (called ANN—artificial neural networks) in large quantities, to build
powerful AI systems. In addition, the ANN are used to model various human abilities.
The major difference between the functions of neurons and the process of human
reasoning is that neurons work at sub-symbolic level, whereas much of conscious
human reasoning appears to operate at a symbolic level, for example, we do most of
the reasoning in the form of thoughts, which are manipulations of sentences.
The collection of neurons in the form of programs called Artificial Neural Net-
works (ANN), perform well in executing simple tasks, and provide good models
of many human abilities. However, there are many tasks of AI that they are not so
good at ANN, and other approaches are more promising in those areas, compared
to ANN. For example, for natural language processing (NLP) and reasoning, the
symbolic logic, called predicate logic is better suited.

1.4.6 Evolution

Unlike the machines, the humans (intelligence) has a very long history of evolution,
of millions of years, compared to less than hundred years for electronic machines
and computers. The first exhaustive document of human evolution, the evolution
by natural selection is due to Charles Darwin (1809–1882). The idea is that fitter
individuals will naturally tend to live longer and produce more children (may not be
truly valid in the modern world). Hence, after many generations, a population will
automatically emerge with good innate properties [3].
Due to this evolution, the structure of the human brain, and even the knowledge, are
to a sufficient extent built-in at the time of the birth. This is an advantage over ANNs,
which have no pre-stored knowledge, hence they need to acquire the entire knowledge
by learning only. However, the present-day computers are powerful enough that even
the evolution can be simulated using them, and can evolve the AI systems. It has
now become possible to evolve the neural networks to some extent so that they are
efficient at learning. But, may still be challenging to recreate the long history of the
evolution of humans in the ANNs.
A closely related field to ANNs is genetic programming, which is concerned with
writing the programs that evolve over time, and do not need to modify them as it is
done in usual programs when the system requirement changes [4].
1.4 Roots of AI 9

1.5 Artificial Consciousness

Right from the time, automated machines like computers came into existence, it
was the quest of researchers to build machines that can compete in intelligence with
humans. Looking at the current progress rate of smart machines, like smartphones,
it is believed that in the not very far future, it may be possible to build machines
which may be, if not more, but have comparable intelligence to that of humans.
Using such machines, it may be possible to produce human like consciousness in
machines—called as artificial consciousness.
On the contrary, even a far of realization of artificial consciousness gives rise to
several philosophical nature of questions:

• Can the computers be made to think or they will just calculate?

• Is consciousness a human prerogative only, or it can be created in machines also?
• Is the consciousness due to the material comprised in the human brain, or it can
be created in silicon also (the computer hardware)?

To provide the answers to these questions is difficult as of now, mainly because it

requires combining the knowledge from the fields of computer science, neurophys-
iology, and philosophy.
On the other hand, the very talk of artificial consciousness—a possible product of
the human imagination, express human desires, and fears about future technologies—
may influence the course of progress.
At a social level, the science fiction stories simulate future scenarios that can help
prepare us for crucial transitions by predicting the consequences of such technolog-
ical advances [1].

1.6 Techniques Used in AI

The AI systems have a lot of variations, for example, the rule-based systems are based
on symbolic representations, and work on inferences. There are other extremes,
the ANN-based system work on the interface with other neurons, and connection
weights. In spite of all these, there are four common features among all of them.
Representation
All AI systems have an important feature of knowledge representation. The rule-
based systems, frame-based systems, and semantic networks make use of a sequence
of if-then rules, while the artificial neural networks make use of connections along
with connection weights.
Learning
All AI systems have capability of learning, using which they automatically build up
the knowledge from the environment, e.g., acquiring the rules for a rule-based expert
system, or determining the appropriate connection weights in an artificial neural
network.
10 1 Introducing Artificial Intelligence

Rules
The rules of an AI-based system can be implicit or explicit. When explicit, the rules
are created by a knowledge engineer, say, for an expert system, and when implicit,
they can be, for example, in the form of connection weights in a neural network.
Search
The search can be in many forms, for example, searching the sequence of states that
lead to solution faster, or searching for an optimum set of connection weights in an
ANN by minimizing the fitness function.

1.7 Sub-fields of AI

Considering AI as replication of human intelligence may be misleading, and primi-

tive. The later is because the true process of human intelligence and its sources are
still under debate. However, if AI is taken to mean the advanced computing, it means
more justified. In the past two decades, particularly after the year 2k, the AI applica-
tions have evolved and expanded, in the commercial, industrial, medicines and drug
decide, medical science, consumer products, manufacturing processes, and even in
management, to list only a few of its total domain. The use of AI techniques in every
organization has become necessary to maintain competitiveness in the market. Many
organizations keep secret of the true AI techniques they use.
AI now consists of many sub-fields, using a variety of techniques, such as the
following:

Speech Processing: To understand speech, speech generation, machine dialog,

machine user-interface.
Natural Language Processing: Information retrieval, Machine translation, Ques-
tion/Answering, summarization.
Planning: Scheduling, game playing.
Engineering and Expert Systems: Troubleshooting medical diagnosis, Decision
support systems, teaching systems.
Fuzzy Systems: For fuzzy controls.
Models of Brain and Evolutionary: Genetic algorithms, genetic programming,
Brain modeling, time series prediction, classification.
Machine Vision and Robotics: Object recognition, image understanding, Intelligent
control, autonomous exploration.
Machine Learning: Decision tree learning, version space learning.

Most of these have both engineering and scientific aspects. Many of these are
going to be discussed in this text. Following is brief Introduction to some of these
areas.
1.7 Sub-fields of AI 11

Fig. 1.2 Sound waves for the text “Hello” repeated five times by a five year child

1.7.1 Speech Processing

The speech processing and understanding of human speech has number of applica-
tions, some of them we come across quite often, like, speech recognition for dictation
systems, speech production for automated announcements, voice-activated control,
human–computer interface (HCI), and voice-activated transactions are a few exam-
ples.
One of the primary goals is, how do we get from sound waves to text streams and
vice-versa? The Fig. 1.2 is an example, showing the sound wave pattern for the text
“Hello” repeated five times.
To be precise, how should we go about segmenting the stream into words? How
can we distinguish between “Recognize speech” and “Wreck a nice beach”?

1.7.2 Natural Language Processing

Consider the machine understanding and translation of simple sentences given below.

Ram saw the boy in the park with a telescope.

Ram saw the boy in the park with a dog.
In the parse-tree in Fig. 1.3a, the sentence structure is “Ram saw, the boy in the
park, with a Telescope.” Whereas in Fig. 1.3b, it is “Ram saw, the boy in the park
with a dog.” In first, it shows the association of verb saw and telescope, i.e., someone
is seeing using telescope. In Fig. 1.3b the association of boy, and dog is shown, and
all are in the park. The further deeper contexts help in resolving this ambiguity.
Though the sentences appear simple, finding out the meaning of each using the
machine is difficult, as the parse-tree of each need to be analyzed for the meaning
associated in each, in addition, the context knowledge is important.
12 1 Introducing Artificial Intelligence

Fig. 1.3 Parse-trees with

different semantics

Some of the common applications of Natural Language Processing (NLP) are [2]:

1. Word Processing and Desktop Publishing

2. Spell check and Correction
3. Information Retrieval
4. Information Extraction
5. Information Categorization
6. Question-answering
7. Information summarization
8. Machine Translation.

1.7.3 Planning

Any planning, or specifically the robotic planning, is concerned with choosing (com-
puting) the correct sequence of actions to complete a specific task. This requires a
convenient and efficient representation of the problem domain. The plan steps are
called states defined in formal language, like, in predicate language, or in the form
of rules, depending on what type of planning has been used. A plan may be taken as
a sequence of operations that transform the initial state into the goal state ultimately,
the later is the solution. The best planning seeks to explore the best path of states for
reaching the goal state. Hence, the best path or an optimum path seeks exploration
or searching, to find out the best possible path efficiently, in terms of time and space
required for running the planning algorithm.

1.7.4 Engineering and Expert Systems

These are primarily based on symbolic processing—which is the main stream of

AI. The fundamental problems such as, how to represent knowledge, is the con-
1.7 Sub-fields of AI 13

tent of the following chapters. Various representation schemes are classified into two
categories: Symbolic representation and Graphical representation. The first is propo-
sitional and predicate logic based, while the other is graph-based, like—semantic
networks, frames, ontologies and conceptual dependencies.

1.7.5 Fuzzy Systems

The primary aim of fuzzy or soft-computing is to exploit the tolerance for imprecision
and uncertainty to achieve tractability, robustness, and low cost applications. Fuzzy
systems can be integrated with other techniques such as neural networks, and prob-
abilistic reasoning. In fuzzy systems, the membership are partial (fuzzy/unclear).
Examples of fuzzy sets are: old, cloudy, high speed, rainy, etc. This is in contrast to
the classical-based systems, where boundaries are crisp, like, member of computer
science class, an Indian national, chair (member of the set of chair), where all three
are representing the full membership of each category of the given set.

1.7.6 Models of Brain and Evolution

The models of the human brain and evolution correspond to two major approaches to
AI. The AI as a model of the brain corresponds to Symbolic AI, and has remained the
major field of AI. This has the property of high level of mathematical abstractions,
and considered as the macroscopic view of AI. The human psychology operates at
a symbolic level, as well as the AI programming languages, and early engineered
systems fall in this type of AI.
The other approach of AI is like human evolution, which is based on low level
biological and genetic models of living beings. The neural-based computing (artificial
neural networks), and genetic algorithms derived from these concepts from life. These
biological models of AI may not necessarily be resembling verbatim to their living
being counterparts, but the AI techniques that are based on GA (genetic algorithms)
evolve the solution like the populations of human or like other forms of the life
evolves [7].
Neural networks, similar to expert systems, are modeled on the human brain and
learn by themselves from patterns. Thus, learning can then be applied to classifica-
tion, prediction, or control applications. The GAs, are computer models based on
genetic and evolution. Their basic idea is that genetic programs work towards finding
better and better solutions to problems, just as species evolve to better adapt to their
environments. The GAs comprises three basic processes: reproduction of solution
based on their fitness, crossover of genes, and mutation for random change of genes.
A broader definition GA is evolutionary computing, which includes not only GAs
but classifier systems, genetic programming in which each solution is a computer
program, and a part of artificial life.
14 1 Introducing Artificial Intelligence

1.8 Perception, Understanding, and Action

These fields are concerned with vision, speech processing, and robotics. The basic
theme is applications that make machine sense (e.g., to see, hear, or touch), then
understand and think, and finally take action.
For example, the basic objective of machine vision may be to make the machine
“understand” the input consisting of reflected brightness values. Once this under-
standing is achieved, the results can be used for interpretation of patterns, inspecting
parts, action of robots, and so forth. Developing an understanding presents the same
difficulty in all areas of AI, including knowledge based systems—people understand
what they see by integrating an optical image with complex background knowledge.
Such background knowledge has been built over years of experiencing perceptions.
Creating this type of information processing in the machine is a challenging task;
however, some interesting applications have already appeared as the evidence to
support future progress.
The speech processing uses two major technologies. One of these areas focus
on input or speech recognition where acoustic input, like optical input in machine
vision, is difficult task to automate. People understand what they hear with complex
background knowledge. Speech recognition technology includes signal detection,
pattern recognition, and possibly semantics—a feature closely related to Natural
Language understanding. The other technology concerns to the creation of output
or text-to-speech (tts) synthesis. Speech synthesis is easier than recognition, and its
commercialization has been well established.
The field of Robotics integrate many techniques of sensing, and, is one of the AI
areas in which industrial applications have the longest and widest successful records.
The abilities of these robots are relatively limited. For example, only in a limited task
such as welding seams and installing windshields, etc [7].

1.9 Physical Symbol System Hypothesis

The symbols are the basic requirements of intelligent activity, e.g., by human the sym-
bols are basic number systems, alphabet of our languages, sign language, etc. Similar
is the case with entire computer science, the languages, commands, computations, all
have symbols as the base. When the information is processed by computers, on the
completion of the task, we measure the progress, as well as the quality of results and
efficiency of computations, only based on its symbols’ contents in the end results [8].
1.9 Physical Symbol System Hypothesis 15

1.9.1 Formal System

Basic requirement of achieving AI is formal system, which is based on physical

symbol system hypothesis. The term was coined by Allen Newell and Herbert Simon,
which states that a “physical symbol system” is a necessary requirement for AI to
function. As per this, the physical patterns, called symbols, are combined to produce
structures (i.e., expressions), and the processes act on these to manipulate to produce
new expressions [8].
This symbol system hypothesis claims that human intelligence is due to this
symbol system, which comprises all the alphabets, numerals, and other punctuation
symbols. Thus, the symbol system is a “necessary” requirement for achieving the
intelligence. Based on this argument, one can say that, if the machines are provided
with symbol system with symbol manipulation capabilities, it is “sufficient” for
achieving the intelligence in the machines.
As per the Physical Symbol System Hypothesis (PSSH), the capabilities of symbol
manipulation are the essence for human’s, as well as machines’ intelligence. Hence,
it is a necessary and sufficient tool for achieving intelligence in both the machines and
humans. There is also experimental evidence, that, in various problem solving, like,
in mathematical puzzles, planning of activities, and execution, the symbol system is
the key requirement. By the term “necessary” here means, that the system possessing
general intelligence, on analysis, will prove to be based on a physical symbol system.
And, the term “sufficient” means that the physical symbol system can be organized to
be exhibiting the general intelligence. When the problem-solving process of humans
were simulated step by step on computers by the researchers, it was found to be
simply the process of symbols’ manipulations.
Of course, various researchers have criticized this hypothesis strongly, but still,
it forms the central part of AI research. The critics argue that the symbol systems
work only for high level processes like chess, games, and puzzles, but not suitable
for low level systems like vision and speech recognition. This distinction is based on
the fact that high-level symbols directly correspond to objects, like cat, house,
hill, etc, but not to the low-level symbols that are present in the machinelike neural
networks (or ANN).

1.9.2 Symbols and Physical Symbol Systems

If we look at the entire knowledge of computer science, it is the symbols, which have
been used to explain this knowledge at the most fundamental level. The explanation
is nothing but the scientific proposition of nature, which is empirically derived over
a long period of time, through a graduate development. Hence, the symbols are at the
root of artificial intelligence, and are also the primary topic of artificial intelligence.
For all the information processed by computers in the service of finding the end
goals, the intelligence of the system (computers) is their ability to reach goals in
16 1 Introducing Artificial Intelligence

the face of difficulties and complexity of the solution, as well as the complexity
introduced by the environment. The fundamental requirements to achieve artificial
intelligence is to store and manipulate the symbols, however, there is no uniformity
and specific requirement of storage structures, as the structures vary from method to
method used for implementation of AI, which are mostly the variants of network-
based representation and predicate-based representations.
The “physical systems” used have two important characteristics: 1. Operation of
systems is governed by the laws of physics, when they are realized as engineered
systems, and made of engineered components; and 2. The “symbols” are not limited
only to the symbols used by human beings.

1.9.3 Formal Logic

The “physical symbol system” hypothesis has its root to Russel’s formalizing logic,
which states that one need to capture the basic conceptual notion of mathematics in
logic and put that notion to proof and deduction as sound base. This notion, with the
effort ultimately grew in the form of mathematical logic—the propositional logic,
predicate logic, and their variants [13].

1.9.4 The Stored Program Concept

The second generation of computers brought the concept of stored program concept
in the mid-forties, after the Eniac computer. The arrival of these computers was con-
sidered as a milestone in terms of conceptual progress, as well as practical availability
of systems. In such systems, the programs are treated as data, which, in turn, was
processed by other programs, like a compiler program processing another program
as data to generate object code. Interestingly, this capability was already verified and
existed in Turing machine, which came as early as 1936. The Turing machine is a
model of computing given by Alan M. Turing, where, in a universal Turing machine,
an algorithm (another Turing machine) and data are on the very same tape. This idea
was realized practically when machines were built with enough memory to make it
practicable to store actual programs in some internal place, along with the data on
which the program will act, as well as the data which will be produced as a result of
the execution of the programs.

1.10 Considerations for Knowledge Representation

As far as AI is concerned, the following are the aspects of knowledge representation:

1.10 Considerations for Knowledge Representation 17

• What is the meaning of Knowledge?

• How the Knowledge can be represented in the machine?
• What are the requirements of representation of knowledge, e.g., structures, meth-
ods, size, etc.
• How the practical and theoretical aspects differ for knowledge representation?
• Can it be? Or, if yes, how to represent the knowledge using Natural Language?
• Can we call the databases as a form of knowledge representation?
• What are the semantic networks, and what are the frames? How the knowledge
can be represented using these approaches?
• How the knowledge can be represented using the First-Order Predicate Logic
(FOPL)?
• What is a Rule-Based Systems?
• What is an expert system?
• Out of the many techniques, which is the best technique for knowledge represen-
tation?

1.10.1 Defining the Knowledge

As per the Webster English language dictionary, the following are the meanings of
knowledge:
1. The act or state of knowing; clear perception of fact, truth, or duty; certain appre-
hension; familiar cognizance; cognition. [1913 Webster]
Knowledge, which is the highest degree of the speculative faculties, consists in
the perception of the truth of affirmative or negative propositions—Locke. [1913
Webster]
2. That which is or may be known; the object of an act of knowing; a cognition—
Chiefly used in the plural. [1913 Webster]
3. That which is gained and preserved by knowing; instruction; acquaintance;
enlightenment; learning; scholarship; erudition. [1913 Webster]

1.10.2 Objective of Knowledge Representation

The objective of knowledge representation is to express the knowledge in computer

so that the AI programs can use it to perform reasoning and inferences using this in an
efficient way. The knowledge is represented using certain representation language,
18 1 Introducing Artificial Intelligence

for example, a predicate like language. The language has two important components
in it.
Syntax
The system of a language defines the methods using which we or the machine can
distinguish the correct structures from incorrect, i.e., it makes possible to identify
the structurally valid sentences.
Semantics
The semantics of a language defines the world, or facts in the world of the concerned
domain. And, hence defines the meaning of the sentence in reference, to the world.

1.10.3 Requirements of a Knowledge Representation

A good knowledge representation system for any particular domain should possess
the following properties.
Adequacy of representation
The representation system should be able to represent all kind of knowledge needed
in the concerned AI-based system.
Adequacy of Inference
The representation should be such that all that can be inferable by manipulating the
given knowledge structures should be inferred by the system, when needed.
Inference Efficiency
The knowledge structures in the representation are so organized that the attention of
the system, in the form of deductions, navigates in such a direction that it can reach
the goal quickly.
Efficient acquisition
It should be able to acquire the new information automatically and efficiently, as and
when needed, and also to update the knowledge regularly. In addition, there should
be a provision that knowledge engineer can update the information in the system.

1.10.4 Practical Aspects of Representations

We are aware of good and bad knowledge representation, when we consider the
knowledge representation in English or any other natural language. These are due to
factors, like, syntax, semantics, partial versus full knowledge on any subject, depth
and breadth of knowledge, etc.
1.10 Considerations for Knowledge Representation 19

There are many theoretical requirements for good knowledge representations,

which can be met by dealing with a number of practical aspects, as follows:
• The representations should be complete, so that everything needs to be represented,
can easily be represented.
• The representations should be simple and clear, so that one can easily understand
what is being communicated by the representation,
• The important objects and their relations should be explicit and accessible, so that
it becomes easy to see what is going on, and how the components of knowledge
interact with each other.
• The irrelevant detail of the knowledge should be suppressed in the representation,
so that they do not introduce complications. However, when needed, these are still
available.
• The representation should be concise, so that information can be stored, retrieved
and manipulated rapidly.
• The representation should be such that the overall system is fast.
• They must be computable and implementable with standard computing procedures.

To realize the above, a lot depends on algorithms used, representation structures,

hardware, as well as dissemination of knowledge before it is represented.

1.10.5 Components of a Representation

To carry out the analysis of any representation system, it is useful to break the entire
representation into smaller (smallest) components, which are in the most fundamental
form. Accordingly, the components of an AI representation are divided into the four
fundamental components:
Lexical components
The lexical components of knowledge representation are the symbols and words of
the vocabulary used for representation.
Syntactic/Structural components
It describes how the symbols can be arranged systematically to create meaningful
sentences. These structures are the grammar of the language used for representation.
Semantic components
They help is associating real-world meaning to objects and entities.
Procedural components
These procedures are used for creating and modifying the representations, and also
for answering the questions using these procedures.
20 1 Introducing Artificial Intelligence

1.11 Knowledge Representation Using Natural Language

We humans are intelligent beings, who make use of the knowledge represented
in the form of natural language (like, English, Hindi, Chinese, etc.), we update
that knowledge (i.e., acquisition), and do the reasoning and inferences using this
representation. Of course, there are many other types of knowledge representation
and inferencing with the humans, which are not symbolic-based, like those acquired
through smell, touch, hearing, and taste (through tongue). Hence, why not use the
natural language for knowledge representation for machines also? The following are
the trade-offs for representation using natural language.
Advantages
There are very strong advantages in favor of using natural language for knowledge
representation.
• The natural language is strong at expressiveness, using which we can represent
almost everything (real-world situations, pictures, symbols, ideas, emotions) and
can carry out the reasoning using that.
• It is the most abundantly used source for knowledge representation by humans,
for example, can we list the name of textbooks not written in natural language? It
is hard to reply!

Disadvantages
In spite of strong points in favor of natural language-based representation, there are
serious difficulties in realizing such representation for machines, due to the following
reasons:
• The syntax and semantics of the natural language are very complex, which are not
so easily understood. Hence, it becomes challenging and risky if solely depended
on machines.
• The uniformity in representation is lacking—the sentences carrying the identical
meaning can be represented in many different syntax (structures).
• There is a lot of ambiguity in the natural language, a sentence/word may have
many different meanings, and the meanings are context-dependent. Hence, it is
overly risky to try these for machines, unless the machines are having intelligence
at par with the human.

1.12 Summary

Intelligence is defined as:

Intelligence = Perceive + Analyze + React

1.12 Summary 21

AI has its inter-related goals for scientific, as well as engineering areas. Its roots
are in several historical disciplines, which include, philosophy, logic, computation,
psychology, cognitive science, neuroscience, biology, and evolution.
The major sub-fields of AI now include: neural networks, machine learning, evo-
lutionary computation, speech recognition, text-to-speech translation, fuzzy logic,
genetic algorithms, vision systems and robotics, expert systems, natural language
processing, and planning. Many of these domains have dependency and are inter-
related, for example, neural network is one of the techniques for machine learning.
The common techniques used across these sub-fields are: knowledge representation,
search, and information manipulations.
Human brain and evolution are also the areas of AI modeling.
The study of logic and computers have demonstrated that intelligence lies in the
physical symbol system (PSS)—a collection of patterns and processes. The PSS needs
the capability to manipulate the patterns, i.e., it should be able to create the patterns,
modify the patterns, and should be able to destroy the patterns. The patterns have
important properties, that they can designate objects, processes, and other patterns.
When the patterns designate processes, the later can be interpreted, i.e., to perform
the process steps. The two significant classes of symbol systems we are familiar with
those that are used by human beings, and those by computers. The later uses binary
strings or patterns.
The PSSH (physical symbol system hypothesis) says that to achieve the intelli-
gence, it is sufficient to have three things,
i. a representation system, using which anything can be represented,
ii. a manipulation system, using which the symbols can be manipulated, and
iii. search using which the solution can be searched.
For the above, it is in fact, not important as whether the medium of storage is the
human brain (neurons) or the electronic memory of computer systems.

Various approaches for knowledge representations (KR) are:

i. Natural languages versus databases.
ii. Frame versus semantic network-based representation.
iii. Propositional and predicate logic-based representation.
iv. Rule-based representation.
Knowledge representation helps to know the object or the concerned concept.
Various characteristics of KR are:
i. KR has syntax and semantics.
ii. Requirements for knowledge representation are: adequacy of representation and
inferencing, and efficiency of inference and acquisition.
iii. Its practical aspects are: complete, computable, and suppression of irrelevant
data.
iv. Components for KR are: lexical, structural, semantic, and procedural.
22 1 Introducing Artificial Intelligence

Exercises

1. Try to analyze your own learning behavior, and list the goals of learning, in the
order of their difficulty level of learning.
2. List the living beings—human, dog, cow, camel, elephant, cat, birds, insects, in
the order of their intelligence levels. Also, justify your argument.
3. Suggest some method to combine large number of human beings, in a group,
and formulate a method/algorithm to perform the pipeline or any fast computing
work.
4. Write an essay, describing the various anticipated theories as to how the human
processes the information?
5. Consider a task requiring knowledge like baking a cake. Out of your imagination,
suggest what are the knowledge requirements to complete this task.
6. Out of your reasoning, explain the distinction between knowledge and belief.
7. What is the basic difference between neural network level processing and pro-
cessing carried out for human reasoning?
8. What are the major advantages of humans over modern computers?
9. List the examples where PSSH is not sufficient or not the basis of achieving
Intelligence. Justify your claims.
10. Write an essay, describing how the things (objects/activities) are memorized by
a. Plants.
b. Birds.
c. Sea animals.
d. Land animals.
e. Human.
What can be the size of memories in each of the above cases?
11. Discuss the potential uses of AI in the following applications:
a. Word processing systems.
b. Smartphones.
c. Web-based auction sites.
d. Scanner machines.
e. Facebook.
f. Twitter.
g. Linkedin.
h. Amazon and Flipkart.
12. How the artificial neural networks (ANN) and genetic algorithms differ from
each other in respect of leaning to be used for problem solution? (Note: You
need to explore the AI resources to answer this question).
Exercises 23

13. It is an accepted scientific base that physical characteristics of life are genetically
transferred. Do you believe that information and knowledge are also genetically
transferred? Justify for yes/no?
14. Are the beliefs of rebirth and reincarnation are also the goals of AI research?
How and how not?

References

1. Buttazzo G (2001) Artificial consciousness: utopia or real possibility? IEEE Comput 34(7):24–
30
2. Church KW, Lisa FR (1995) Commercial applications of natural language processing. Commun
ACM 38(11)
3. Forrest S (1993) Genetic algorithms: principles of natural selection applied to computation.
Science 261(13):872–878
4. Goldberg DE (1994) Genetic and evolutionary algorithms coming of age. Commun ACM
37(3):113–119
5. http://www.theguardian.com/technology/2014/jun/08/super-computer-simulates-13-year-
old-boy-passes-turing-test. Accessed 19 Dec 2017
6. Ludger GF (2009) Artificial intelligence - structures and strategies for complex problem solv-
ing, 5th edn. Pearson, New Delhi
7. Munakata T (1994) Commercial and industrial applications of AI. Commun ACM 37(03)
8. Newell A, Herbert AS (1976) Computer science as an empirical inquiry: symbols and search.
Commun ACM 19(3):113–126
9. Russell SJ (1997) Rationality and intelligence. Artif Intell 94:57–77
10. Russell SJ, Norvig P (2005) Artificial intelligence, a modern approach, 2nd edn. Pearson, New
Delhi
11. Stanford Encyclopedia of Philosophy. http://plato.stanford.edu/entries/turing-test/. Accessed
19 Dec 2017
12. Turing AM (1950) Computing machinery and intelligence. MIND - Q Rev Psychol Philos
Lix(236):433
13. Whitehead AN, Russell B (1910) Principia mathematica, vol 1 (Part I. Mathematical logic).
Merchant Books. ISBN 978-1-60386-182-3
Chapter 2
Logic and Reasoning Patterns

Abstract Logic is the foundation of AI, and the majority of AI’s principles are based
on logical or deductive reasoning. The chapter presents: contributions of pioneers of
logic, the argumentation theory, which is based on logic and with its roots in propo-
sitional logic, the process of validating the propositional formulas, their syntax and
semantics, interpretation of a logical expression through semantic tableau, followed
with presents the basic reasoning patterns used by human, and their formal notations.
In addition, presents the normal forms of propositional formulas and application of
resolution principle on these for inference. The nonmonotonic reasoning and its sig-
nificance is briefly described. At the end, the chapter presents the axiomatic system
due to Hilbert and its limitations, and concludes with chapter summary.

Keywords Logic · Propositional logic · Deductive reasoning · Argumentation

theory · Syntax and semantics of propositional formulas · Nonmonotonic
reasoning · Hilbert’s axiomatic system

2.1 Introduction

The ancient Greeks are the source of modern logic, their education system empha-
sized the competence in rhetoric (proficient in language) and philosophy; the
words axioms and theorem are from Greek. The logic was used to formalize the
deductions—the derivation of true conclusions—from true premises. Later it was
formalized as a set theory by the mathematician George Boole. Till the arrival of the
nineteenth century, the logic remained more of a philosophical nature, rather than a
mathematical and scientific tool. Later, since complex things could not be reasoned
through logic, the logic became part of mathematics, where mathematical deduction
became justifiable through formalizing a system of logic, and resulted in one very
important breakthrough. This was, about the set of true statements, stated as “the set
of provable statements are only those that are true statements.” This is because some
proof exists for those due to some other true statements.
At the beginning of nineteenth century, the mathematician David Hilbert intro-
duced the logic, as well as theories of the nature of logic-a far more generalization

© Springer Nature India Private Limited 2020 25

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_2
26 2 Logic and Reasoning Patterns

of the logic. But, this generalization received a blow when another mathematician
Kurt Gödel showed in 1931 that there are true statements of arithmetics that are not
provable, through his incompleteness theorem.
Now, though mathematical logic remains the branch of pure mathematics, it is
extensively applied to computer science and artificial intelligence in the form of
propositional logic and predicate logic (first-order predicate logic (FOPL)).
As per the Newell’s and Simons’s Physical Symbol System Hypothesis (PSSH),
discussed in the previous chapter, the knowledge representation is the first require-
ment of achieving intelligence. This chapter presents the knowledge representation
using propositional logic, introduces first-order predicate logic (FOPL), and drawing
of inferences using propositional logic.
Logic is a formal method for reasoning, using its concepts can be translated into
symbolic representation, which closely approximate the meaning of these concepts.
The symbolic structures can be manipulated using computer programs to deduce
facts to carry out the form of automated reasoning [9].
The aim of logic is to learn principles of valid reasoning as well as to discern
good reasoning from bad reasoning, identifying invalid arguments, distinguishing
inductive versus deductive arguments, identifying fallacies as well as avoiding the
fallacies.
The Objective of logic is to equip oneself with various tools and techniques, i.e.,
decision procedures for validating given arguments, detecting and avoiding fallacies
of a given deductive or inductive argument.
We study the logic because of the following reasons:

• Logic deals with what follows from what? For example, Logical consequence,
inference pattern, and validating such patterns,
• We want the computer to understand our language and does some intelligent tasks
for us (Knowledge representation),
• To engage in debates, solving puzzles, game like situation,
• Identify which one is a fallacious argument and what is a type of fallacy?
• Proving theorems through deduction. To find out whether whatever proved is
correct, or whatever obviously true has a proof, and
• To solve some problems concerning the foundations of mathematics.

Learning Outcomes of This Chapter:

1. Convert logical statements from informal language to propositional logic expres-
sions. [Usage]
2. Apply formal methods of symbolic propositional such as calculating the validity
of formula and computing normal forms. [Usage]
3. Use the rules of inference to construct proofs in propositional. [Usage]
4. Describe how symbolic logic can be used to model real-life situations or appli-
cations, including those arising in computing contexts such as software analysis
(e.g., program correctness), database queries, and algorithms. [Usage]
2.1 Introduction 27

5. Apply formal logic proofs and/or informal, but rigorous, logical reasoning to real
problems, such as predicting the behavior of software or solving problems such
as puzzles. [Usage]
6. Explain the difference between rule-based and model-based reasoning techniques.
[Familiarity]
7. Describe the strengths and limitations of propositional logic. [Familiarity]

2.2 Argumentation Theory

The Argumentation theory is the study of how conclusions can be reached through
logical reasoning, that is, whether the claims are soundly based on premises or not
(Fig. 2.1). It includes the arts and sciences of civil debate, dialog, conversation, and
persuasion. It includes the studies of rules of inference, logic, and procedural rules
in both artificial and real world settings.
An argumentation system comprises debate and negotiations, aimed at reaching
to a mutually agreeable conclusion. The argumentation may also consist of erroneous
dialogs, where victory over an opponent is the only goal, without consideration of
the truth. The argumentation theory is an art, as well as science, using these people
protect their self-interests and beliefs using rational dialogs at commonplaces of their
meeting points, and during the process of argumentation.
People make use of argumentation theory in law also, for example, in preparing an
argument to be presented before the court of law, in debate in the court of law, in trials,
and in testing the validity of certain kinds of evidence. Scholars of Argumentation
theory study the post-hoc rationalizations by which organizational actors try to justify
decisions even made irrationally.
The simple block diagram for logical reasoning shown in Fig. 2.1 has internal
structure, comprising the following:

1. a set of assumptions or premises (or antecedents),

2. a method of reasoning or deduction, and
3. a conclusion or consequence.

If the premises are P1 , P2 , . . . , Pn , then they are conjuncted and their conjunction
imply the conclusion C, i.e., P1 ∧ P2 ∧ · · · ∧ Pn → C.
An argument must have at least one premise and one conclusion. Often, classical
logic is used as the method of reasoning so that the conclusion follows logically
from the assumptions or support. One challenge is that if the set of assumptions is

Fig. 2.1 Inference process Set of

premises Method of Conclusion
reasoning
28 2 Logic and Reasoning Patterns

inconsistent then anything can follow logically from inconsistency. Therefore, it is

common to insist that the set of assumptions be consistent. It is the practice to have
a minimal set. Such argumentation has been applied to the field of medicine also.
The second school of argumentation investigates abstract arguments, where “argu-
ment” is considered a primitive term, so no internal structure of arguments is taken
on the account.

2.3 Role of Knowledge

We have discussed in the above section about knowledge and logic. The Logic needs
a base of knowledge to infer or conclude new knowledge. The knowledge is also
used for learning, retrieval and reasoning. The learning is not only adding new facts
into an existing knowledge base, but before the new data are put into the storage,
they need to be classified for ease of retrieval. The interaction and inference with
existing facts avoid the redundancy and duplication of knowledge in the knowledge
base. In addition, the learning updates the existing facts.
Having stored the knowledge in the process of learning, one important objective of
that is, retrieval. The representation scheme used in the knowledge base has critical
effect on the efficiency of the retrieval system. As humans, we are very good in
retrieval from our knowledge (memories), and many AI systems have used that for
modeling AI learnings.
The knowledge is also used for reasoning process, i.e., to infer new facts from the
existing facts in the knowledge. For example, observing many birds flying, to infer
that all the birds fly, as well for solving a complex problem, say, based on sufficient
facts, to infer that a customer financed by a bank, will be able to repay the loan the
bank has financed to him.

2.4 Propositional Logic

The propositional logic deals with individual Propositions, which are viewed as
atoms, i.e., these cannot be further broken down into smaller constituents. For build-
ing propositional logic, first we describe the logic with the help of a formula called
Well-Formed Formulas (wff, read as woofs). A formula is a syntactic concept, which
means whether or not a string of symbols is a formula not. It can be determined
solely based on its formal construction, i.e., whether it can be built according to
its construction rules. Therefore, we are in a position to verify that a sequence of
symbols is a formula or not, as per the specified rules. This function of verification,
in a compiler, is done by a parser—to verify whether the formula belongs to the
particular programming language or not. A parser also constructs a parse-tree of the
given formula through which it tells how the formula is constructed [2].
2.4 Propositional Logic 29

The meaning (semantics) is associated with each formula by defining its inter-
pretation, which assign a value true (T) or false (F) to every formula. The syntax is
also used to define the concept of proof—the symbolic manipulations of formulas
to deduce the given theorem. The important thing we should note is that provable
formulas are only those which are always true.
We start the propositional logic with the individual propositional variables. These
variables themselves are formulas, which cannot be further analyzed. We represent
these by English alphabets and subscripted alphabets p, q, r, s, t, p1 , p2 , q1 , q2 , . . . ,
etc. These formulas may have smaller constituents but it is not the role of propositional
logic to go into the details of their constructions. The use of letters to represent
propositions is not in true sense variables, they simply represent the propositions or
statements in a symbolic form, and they are not the variables in the sense used in
predicate logic (to be discussed later), or in high-level languages like C or Fortran,
where a variable stands for a domain of values. For example, an integer variable
in a Fortran program stands for any integer number as per the specifications of the
language.
The other symbols of propositional logic are operators as follows:
∧ conjunction operator,
∨ disjunction operator,
¬ not or inverting operator,
→ implication, i.e., if . . . than . . . rule, and
⊥ contradiction (false).
Let following be the propositions:
p = Sun is star.
q = Moon is satellite.
We can construct the following formulas using the above propositions:
p ∧ q = Sun is star and Moon is satellite.
p ∨ q = Sun is star or Moon is satellite tennis.
¬ p ∨ q = Sun is not star or Moon is satellite.
¬ p → q = if Sun is not star then Moon is satellite.
A formula in propositional logic can be recursively defined as follows:
(i) Each propositional variable and null are formulas, therefore, p, q, φ are formu-
las,
(ii) If p, q are formulas, then p ∧ q, p ∨ q, ¬ p, p → q, ( p), are also formulas,
(iii) A string of symbols is a formula only as determined by finitely many applica-
tions of above (i) and (ii), and
(iv) nothing else is propositional formula.
This recursive form of the definition can be expressed using B N F (Backups-Naur
Form) notation as follows:
30 2 Logic and Reasoning Patterns

1. f or mula := atomic f or mula | f or mula ∧ f or mula | f or mula ∨ f or mula

| f or mula → f or mula | ¬ f or mula | ( f or mula)
2. atomic f or mula := ⊥ | p | q | r | p0 | p1 | p2 | . . .
(2.1)

In the above notation, the symbols— f or mula and atomic f or mula, that appears
to the left-hand are called non-terminals and represent grammatical classes. The
p, q, r, ⊥, p1 , etc, that appear only to the right-hand side, are called terminals, and
represent the symbols of the language.
A sentence in the propositional language is obtained through a derivation that
starts with a non-terminal, and repeatedly applied the substitution rules from the
BNF notations, until the terminals are reached [8].

Example 2.1 Derivation for p ∧ q → r .

The sequence of substitutions rules to derive this formula, i.e., to establish that it is
syntactically correct, are as follows:

f or mula ⇒ f or mula → f or mula

⇒ f or mula ∧ f or mula → f or mula
⇒ atomic ∧ f or mula → f or mula
⇒ p ∧ f or mula → f or mula
⇒ p ∧ atomic → f or mula
⇒ p ∧ q → f or mula
⇒ p ∧ q → atomic
⇒ p ∧ q → r.

The symbol atomic stands for atomic formula and the symbol “⇒” stands for
“implies”, i.e., the expression to right to this is implied by the expression to left of
“⇒”.
The derivation can also be represented by a derivation-tree (parse-tree), shown
in Fig. 2.2. From the derivation-tree, we can obtain another tree shown in Fig. 2.3,
called syntax-tree or formation-tree, by replacing each non-terminal by the child that
is an operator under that. There is always unique syntax-tree for every formula. 
Considering two propositions p, q, the interpretation (semantics) of the formulas
constructed when they are joined using binary operators (∨, ∧, →) are shown in the
truth-table Table 2.1.
The Material conditional ‘→’ joins two simpler propositions, e.g., p → q, read
as “if p then q”. The proposition to the left of the arrow is called the antecedent and
to the right is consequent. There is no such designation for conjunction or disjunction
operators because they are commutative operations. The p → q expresses that q is
true whenever p is true. Thus it is true in every case in Table 2.1, except in row
three, because this is the only case when p is true but q is not. Using “if p then q”,
2.4 Propositional Logic 31

Fig. 2.2 Parse-tree for the

expression p ∧ q → r

Fig. 2.3 Syntax-tree for the →

expression p ∧ q → r

∧ r

p q

Table 2.1 Interpretation of propositional formulas

p q p∨q p∧q p→q
F F F F T
F T T F T
T F T F F
T T T T T

we can express that “if it is raining outside then there is a cold over Kashmir”.
The material conditional is often confused with physical causation. The material
conditional, however, only relates two propositions by their truth values—which is
not the relation of cause and effect. It is contentious in the literature whether the
material implication represents logical causation.

2.4.1 Interpretation of Formulas

The interpretation of formula is assigning truth value to that formula. As discussed

earlier, a formula can be atomic or in may be complex, i.e., joining or atomic formulas.
The following are some definitions related to the interpretation of formulas [1].
32 2 Logic and Reasoning Patterns

Definition 2.1 (Satisfied, model, valid, and tautology) A propositional formula A is

satisfied iff I (A) = T r ue for some interpretation I . A satisfying interpretation is
called model for A. The formula A is called valid, denoted by |= A, iff I (A) = T r ue
for all interpretations I . A valid propositional formula is also called tautology.
A propositional formula is unsatisfiable (also called contradiction, ⊥), iff it is
not satisfiable, i.e., I (A) = False, for all interpretations I . If I (A) = False for
some interpretation I , then A is called non-valid or falsifiable, and denoted by |= A.
Definition 2.2 (Simultaneously satisfiable) A set of formulas S = {A1 , A2 , . . . , An }
is simultaneously satisfiable iff there exists an interpretation I such that I (Ai ) =
T r ue for all i. The S is unsatisfiable iff for every interpretation I there exits an i
such that I (Ai ) = False.

2.4.2 Logical Consequence

The logical consequence or logically follows is the central concept in the foundations
of logic. It is much more interesting to assume that a set of formulas is true and then
to investigates the consequences of these assumptions [1].
Assume that θ and ψ are formulas (sentences) of a set P, and I is an interpre-
tation of P. The sentence θ of propositional logic is true under an interpretation I
iff I assigns the truth value T to that sentence. The θ is false under an interpretation
I iff θ is not true under I .
Definition 2.3 (Logical consequence) A sentence ψ of propositional logic is a logi-
cal consequence of a sentence (or set of sentences) θ , represented as θ |= ψ, if every
interpretation I that satisfy θ also satisfy ψ.
In fact, ψ need not be true in every possible interpretation, only in those interpre-
tations which satisfy θ , i.e, those interpretations which satisfy every formula in θ .
In the formula (( p → q) ∧ p)  q, the q is logical consequence of (( p → q) ∧ p).
The sign ‘’, is sign of deduction, and S  q is read as S deduces q, where S is a set
of formulas and q is the formula.
A sentence of propositional logic is consistent iff it is true under at least one
interpretation. It is inconsistent if it is not consistent.
Example 2.2 Determine the logical consequence of ψ = ( p ∨ r ) ∧ (¬q ∨ ¬r ) from
θ = { p, ¬q}, i.e., find θ |= ψ, and validity for ψ.
Here ψ is logical consequence of θ , denoted by θ |= ψ, because ψ is true under all the
interpretations such that I ( p) = T r ue, and I (q) = False, is the interpretation,
for which θ is satisfied.
However, ψ is not valid, since it is not true under the interpretation I ( p) =
F, I (q) = T, I (r ) = T .
Further note that θ  ψ is a valid statement because the expression θ  ψ is
always true. 
2.4 Propositional Logic 33

2.4.3 Syntax and Semantics of an Expression

Syntax is name given to a correct structure of a statement. It is the meaning asso-

ciated with the expression. It is mapping to the real-world situation is semantics.
The semantics of a language defines the truth of each sentence with respect to each
possible world. For example, the usual semantics for interpretation of the statement
( p ∨ q) ∧ r is true in a world where either p or q or both are true and r is true.
Different worlds can be all the possible sets of truth values of p, q, r , which is total
8. The truth values are simply the assignment to these variables, and not necessarily
the values which are only true. For example, I ( p) = F, I (q) = F, I (r ) = T ;
and I ( p) = T, I (q) = F, I (r ) = T are the possible worlds for the expression
( p ∨ q) ∧ r .

2.4.4 Semantic Tableau

Semantic tableau is relatively efficient method for deciding satisfiability for the for-
mula of propositional calculus. The method (or algorithm) systematically searches
for a model for a formula. If it is found, the formula is satisfiable, else not satisfi-
able. We start with the definition of some terms, and then analyze some formulas to
motivate us for the construction of semantic tableau [1].

Definition 2.4 (Literal and complementary pair) A literal is an atom or negation of

an atom. For any atom p, the set { p, ¬ p} is called complementary pair of literals.
For any formula A, {A, ¬A} is complementary pair of formulas.

Example 2.3 Analysis of the satisfiability of a formula.

Consider that a formula A = p ∧ (¬q ∨ ¬ p), has an arbitrary interpretation I .

Given this, I (A) = T iff I ( p) = T and I (¬q ∨ ¬ p) = T . Hence, I (A) = T
iff either,

1. I ( p) = T and I (¬q) = T , or
2. I ( p) = T and I (¬ p) = T .

Hence A is satisfiable if either (1) interpretation holds or (2) holds. But (2) is not
feasible. So, A is satisfiable when the interpretation of (1) holds true. Note that the
satisfiability of a formula is reduced to the satisfiability of literals.
It is clear that a set of literals is satisfied if and only if it does not contain com-
plementary pair of literals. In the above case, the pair of literals { p, ¬ p} in case
(2) is complementary pair, hence the formula is unsatisfied for this interpretation.
However, the first set { p, ¬q} is not the complementary pair, hence it is satisfiable.
From the above discussion, we have trivially constructed a model for the formula
A by assigning True to positive literals and False to negative literals. Hence, p =
34 2 Logic and Reasoning Patterns

Fig. 2.4 Tree for semantic p ∧ (¬q ∨ ¬p)

tableau

p, (¬q ∨ ¬p)

p, ¬q p, ¬p

True, and q = False makes the set in (1) true, hence { p = T, q = F} is a model for
formula A.
The above is a search process, and can be represented by a tree shown in Fig. 2.4.
The leaves in the tree represent a set of literals that must be satisfied. A leaf containing
complementary pair of literals is marked closed by ×, while the satisfying leaf is
marked as open by .
The construction process of the tree can be represented as an algorithm, to find
out if some model exists for a formula, and what is that model. 
Definition 2.5 (Semantic Tableau) Semantic Tableau is a tree, each node of which
will be labeled with a set of formulas, and these formulas are inductively expanded
to leaves such that each leaf is marked as open by or closed by ×.
Definition 2.6 (Completed tableau) A semantic tableau whose construction is ter-
minated is called completed tableau . A completed tableau is closed if all the leaves
are marked closed. Otherwise, it is open i.e., some leaves are open.
Definition 2.7 (Unsatisfiable formula) Any formula A is unsatisfiable if its com-
pleted tableau T is closed.
Corollary 2.1 (Method for semantic tableau) A formula A is satisfied if its tableau
T is open. Thus a method for semantic tableau is an algorithm for the validity of a
propositional calculus formula.
Example 2.4 Find out whether ( p ∨ q) ∧ (¬ p ∧ ¬q) is satisfiable, using tableau
method.
Let A = ( p ∨ q) ∧ (¬ p ∧ ¬q). For the satisfaction of A, I (A) = T r ue for some
assignments. That is, I ( p ∨ q) = T r ue and I (¬ p ∧ ¬q) = T r ue. Thus, I (A)
is True if either,
• I ( p) = T, I (¬ p) = T r ue, I (¬q) = T r ue, or
• I (q) = T r ue, I (¬ p) = T r ue, I (¬q) = T r ue.
So that, two sets of literals are,

( p, ¬ p, ¬q) and (q, ¬ p, ¬q).

Since both contain complementary pairs, hence neither of the literals is satisfiable.
So it is impossible to find a model for A, and A is unsatisfiable.
2.5 Reasoning Patterns 35

2.5 Reasoning Patterns

How can we reason about solving any problem? To a certain extent, it depends on
the chosen knowledge representation. The followings are the methods in broad about
how the reasoning is performed by humans [2].
Deductive Reasoning
It is a process by which general premises are used to obtain the inferences, which
are specific. For example, we may have the following premises and conclusion:
Premise-I: I do shopping when the weather is good on weekends.
Premise-II: Today is Saturday and the sky is clear.
Conclusion: Therefore, I will go for shopping Today.
To perform the deductive reasoning, the problem is first formulated in the way as
we did in the above example. Having done this, the conclusions must be valid when the
premises are true. Beginning with a small set of axioms, postulates, and definitions,
the Greek mathematician Euclid proved a total of 465 geometric propositions as the
logical consequences of the input assumptions.
One of the most fundamental rules of inference is modus ponens rule. We have
the following example for modus ponens.
Premise-I: All the men are mortal.
Premise-II: Socrates is man.
Conclusion: Therefore, Socrates is mortal.
The new knowledge, “Socrates is mortal” has been deduced from the earlier two
sentences.
The enumeration table of all possible worlds for modus ponens are shown in
Table 2.2. We note that it is a valid inference, as the sentence (( p → q) ∧ p) → q,
with q as the inference implied, is true in all the rows.
Other deductive reasoning approaches are : modus tollens and syllogism, and
abduction. The Table 2.3 shows the formulas for these rules.
Abduction is deductive type logic, which provides only a “plausible inference.”
For example, given that: “smoking causes lung cancer” and “Sam died due to lung
cancer”, through abduction one would infer that “Sam was smoker”. However, this

Table 2.2 Modus ponens is valid inference

p q p→q ( p → q) ∧ p (( p →
q) ∧ p) → q
F F T F T
F T T F T
T F F F T
T T T T T
36 2 Logic and Reasoning Patterns

Table 2.3 Inference rules

Rule Formula Description
Modus ponens (( p → q) ∧ p)  q If p then q; p; therefore q
Modus tollens (( p → q) ∧ ¬q)  ¬ p If p then q; not q; therefore
not p
Abduction ( p → q) ∧ q  p if p then q; q; therefore q
Hypothetical syllogism (( p → q) ∧ (q → r )) if p then q;  ( p → r ) if q
then r ; therefore, if p then r
Disjunctive syllogism (( p ∨ q) ∧ ¬ p)  q Either p or q, or both; not p;
therefore, q

conclusion is not necessarily true, because there are other reasons also for lung cancer,
which are not due to smoking. When statistics and probability theory are used along
with abduction, it may result in most probable inferences out of the many likely
inferences. To illustrate how the abduction based reasoning works, we consider a
logical system comprising a general rule and one specific proposition.
All successful enterprising industrialists are rich (general rule). Rajan is a rich person (specific
proposition). Therefore, a plausible inference can be that Rajan is a successful, enterprising
industrialist.

However, this conclusion can be false also, because there are many other paths
to richness, such as a lottery, inherited property, coming across a treasure, and so
on. If we have a table of all the riches and how they became rich, we may draw the
probability of abduction for richness to be true in this case.
Inductive Reasoning
The inductive reasoning arrives at a conclusion about all members of a class. It
is based on examination of only a few members of the class and based on that it
generalizes for the entire class. It is broadly reasoning from a specific to the general.
For example, the traffic police comes to know about following situation on a particular
day about nature of road accidents:
1st accident was due to wrong side drive,
2nd accident was due to wrong side drive,
3rd accident was due to wrong side drive.
One would logically infer that all the accidents are due to wrong side driving.
Another example is about the birds for their flying attribute.
Crow fly,
peacock fly,
pigeon fly.
Thus, we conclude that all the birds fly.
Another example is about the progressive sum of 1st n odd integers:
2.5 Reasoning Patterns 37

1 = 12
1 + 3 = 22
1 + 3 + 5 = 32
1 + 3 + 5 + 7 = 42

Thus, by induction we prove that, the sum of n successive odd integers is n 2 .

The outcome of the inductive reasoning process will frequently contain some
measures of uncertainty because including all possible facts in the premises are
usually impossible.
We know that the inference of an accident’s example is not always true, and also
of “all birds fly” is not true, because, ostrich and penguins do not fly. However, for
1st odd integers sum, it is true.
The deductive or inductive approaches are used in logic, rule-based systems, and
in frames.
Analogical Reasoning
The analogical reasoning assumes that when question is asked, the answer can be
derived by analogy, as in the case of following example.
Premise: All the 100 m racers get 5% additional in their merit score.
Question: How much one 400 m racer will get additional in academic score?
Conclusion: Because, 400 m is a race, and an sports activity like 100 m, so it will
also benefit one with 5% in final scores.
Analogical reasoning is a type of verbalization of an internalized learning process.
An individual uses processes that require the ability to recognize previously encoun-
tered experiences. This approach is not very common in AI, however, the case-based
reasoning, semantic networks, and frames use this analogical reasoning approach.
Formal Reasoning
It uses the process of syntactic manipulation of data structures to deduce new facts.
A typical example is the mathematical logic used in proving theorems in geometry.
For example, proof by resolution.
Procedural and Numeric Reasoning
It uses mathematical models or simulation to solve the problems. The model-based
reasoning is an example of this approach.
Generalization and Abstraction
The approaches of generalization and abstraction, both can be used with the logical
and semantic representation of knowledge.
Meta-level Reasoning
The meta-level reasoning involves the knowledge about what you, how much you
know about so and so. Also, which approach to use, how successful the inference will
38 2 Logic and Reasoning Patterns

be, depends on a great extent on which knowledge representation method is used.

For example, reasoning by analogy can be more successful with semantic networks
than with frames.

2.5.1 Rule-Based Reasoning

The rule-based reasoning is also called pattern matching, and uses forward and
backward chaining. The implementation of rule-based system makes use of modus
ponens and other approaches. Consider the rule:
Rule 1: If export rises the prosperity increases.
Using the modus ponens, if the premises, e.g., “The export rises” is true, the
conclusion of the rule is accepted as true. We call this accepting the rule as “rule
fires”. The firing of a rule occurs when all its premises are satisfied, whether all are
true or some are false. On firing, the resulting conclusion is stored in the assertion
base, to use for further firing of the rules and generate the assertions. When a premise
is not available as an assertion, it can be obtained by querying the user, or by firing
other rules. Testing of a rule premise or conclusion is as simple as matching a symbol
pattern.
Every rule in the knowledge base can be checked to see if its premises or conclu-
sion can be satisfied by previously made assertions. This process of matching, if done
using forward chaining, i.e., premises to conclusions. If it is done from conclusions
to premises, it is called backward chaining.

2.5.2 Model-Based Reasoning

A reasoning within a context is important in any reasoning system. In real-life situa-

tions, one often provides a lot of missing contexts or out of context information when
answering certain queries. This situation can be correctly modeled by supplementing
the existing knowledge about the world, with additional context-specific information.
When it is supplemented by context information, reasoning within context becomes
a deduction process.
The added information may act as constrain to the existing information in the
system, as in the absence of this additional information the deduction process has
more paths of freedom in the reasoning process. But, due to the availability of this
added context information the reasoning task becomes easier because the domain
in which reasoning takes place gets restricted (constrained) due to having lesser
flexibility of deduction paths to be navigated. This task can be formalized as a task
of varying contexts.
The knowledge that comprises the information for reasoning in the model-based
system is in the form of a set of models of the world. These models satisfy the
2.5 Reasoning Patterns 39

assignments and examples of the world. This is, in contrast, to the use of only the
formulas in the first-order predicate logic to describe the world. The other difference
is that the model-based approach is motivated from a cognitive point of view – the
forerunners of this approach of reasoning are cognitive psychologists who support
the “reasoning by examples.” When a model-based reasoning system is presented
with a query, the reasoning is performed by evaluating the query on these models.
Let us suppose that model-based knowledge base representation Γ , and a query
α are both given, and it is required to find out if Γ implies α (i.e., Γ |= α)? This
we can determine in two steps: 1) evaluate α on all the models in the representation,
2) If there is a model of Γ that does not satisfy α, the Γ does not model the alpha
(i.e., Γ |= α), otherwise we conclude that Γ |= α. This means if the model-based
representation contains all the models of Γ , then by definition, this approach verifies
the implication correctly, and produces the correct deduction.
However, there is a problem—the representation of Γ , such that it explicitly holds
all the models, is not a plausible solution. The model-based approach is feasible only
if Γ can be replaced by small model-based representation, and after that also it should
correctly support the deduction.
Various topics in reasoning are as follows:
Monotonic versus nonmonotonic reasoning,
Reasoning with uncertainty,
Shallow and deep representation of knowledge,
Semantic networks,
Blackboard approach,
Inheritance approach,
Pattern matching,
Conflict resolution.
These are discussed in current, and the following chapters, in details.

2.6 Proof Methods

There are two different methods, one is through model checking and other is deduction
based. The first comprises enumeration of truth-tables, and is always exponential in
n, where n is the size of the set of propositional symbols. The other, i.e., deduction
based approach is repeated application of inference rules. The inference rules are
used as operators in the standard search algorithm. In fact, the application of the
inference approach to proof is called searching for solution. Proper selection of
search directions is important here, as these will eliminate many unnecessary paths
that are not likely to result in the goal. Consequently, the proof-based approach for
reasoning is considered better and efficient compared to model enumeration/checking
based method. The later is exhaustive and exponential in n, where n is the size of the
set of propositional symbols.
40 2 Logic and Reasoning Patterns

The property, the logical system follows, is the fundamental property of mono-
tonicity. As per this, if S  α, and β is additional assertion, then S ∧ β  α.
Thereby, the application of inference rules is legitimate (sound) rule, which helps
in the generation of new knowledge from the existing. If a search algorithm like DFS
(depth first search) is used, it will always be possible to find the proof, as it will
search the goal, whatever the depth may it be. Hence, the inference method in this
case is complete also [7].
Before the inference rules are applied on the knowledge base, the existing sen-
tences in the knowledge base (KB) needs to be converted into some normal form.

2.6.1 Normal Forms

A logical expression can be represented as sum-of-product terms or product-of-sum

terms. If a given logical expression is represented as sums of elementary products,
then this form is called disjunctive normal form (DNF), and if it is represented as
product of elementary sums, it is called conjunctive normal form (CNF). In DNF, the
elementary product terms are called minterms, while in a CNF elementary sum terms
are called maxterms. For a given formula, an equivalent disjunctive normal form with
only disjunctions of minterms is called principle disjunctive normal form or sum-
of-products canonical form. Similarly, an equivalent CNF with only conjunctions of
maxterms is called principle conjunctive normal form or product-of-sums canonical
form [2].
One technique to get a CNF expression for a given DNF expression, say, ¬a¬bc +
¬ab¬c + ¬abc + a¬bc is given in steps as follows:
1. Considering a DNF expression of three variable a, b, c, write down all the
minterms: ¬a¬b¬c, ¬a¬bc, ¬ab¬c, ¬abc, a¬b¬c, a¬bc, ab¬c, abc.
2. Cross out all combinations in the original DNF. We are left with ¬a¬b¬c,
a¬b¬c, ab¬c, abc.
3. Next, write the expression in CNF by inverting each subset of three variables
and ORing as (a + b + c)(¬a + b + c)(¬a + ¬b + c)(¬a + ¬b + ¬c) in the
form of CNF.
Obtaining DNF from CNF is just the reverse process.

2.6.2 Resolution

The resolution rule is an inference which uses deduction approach. It is used in

theorem proving. If two disjunctions have complementary literals, then a resultant
inference of these is disjunction of these expressions, with complementary terms
removed. If p = p1 ∨ p2 ∨ c and q = q1 ∨ ¬c are two formulas, then resolution of
2.6 Proof Methods 41

p and q results to dropping of c and ¬c and disjunction is performed of the remaining

propositions of p and q, as follows:

( p1 ∨ p2 ∨ c), (q1 ∨ ¬c)

(2.2)
p1 ∨ p2 ∨ q 1

The necessary condition for the above is that C should not be a function of any
of the p1 , p2 , q1 .

Example 2.5 Show by resolution that ( p → q) → [(r ∧ p) → (r ∧ q)] is a

tautology:

⇒ ¬(¬ p ∨ q) ∨ [¬(r ∧ p) ∨ (r ∧ q)]

⇒ ( p ∧ ¬q) ∨ [(¬r ∨ ¬ p) ∨ (r ∧ q)]
⇒ ( p ∧ ¬q) ∨ [((¬r ∨ ¬ p) ∨ r ) ∧ ((¬r ∨ ¬ p) ∨ q)]
⇒ ( p ∧ ¬q) ∨ [(r ∨ ¬r ∨ ¬ p) ∧ (q ∨ ¬r ∨ ¬ p)]
⇒ ( p ∧ ¬q) ∨ [(q ∨ ¬r ∨ ¬ p)]
⇒ (q ∨ ¬r ∨ ¬ p ∨ p) ∧ (q ∨ ¬r ∨ ¬ p ∨ ¬q)
⇒T ∧T
⇒ T.

2.6.3 Properties of Inference Rules

An inference rule is a mechanical process of producing new facts from the existing
facts and rules. The semantics of predicate logic provides a basis for a formal theory
of logical inference. It allows the creation of new facts from the existing facts and
rules [5, 7].
An interpretation of a predicate statement means the assignment of true or false
value to that statement. An interpretation that makes a sentence true is said to satisfy
a sentence. An interpretation that satisfies every member of a set is said to satisfy the
set.

Definition 2.8 (Logically follows) If every interpretation that satisfies S also satisfy
X , then we say the expression X logically follows from a set of expressions S (the
knowledge base). In other words, the knowledge base S entails sentence X if and
only if X is true in all worlds where knowledge base is true. If a sentence X logically
follows S, we represent it as S |= X .

The term logically follows simply means that X is true for every, potentially
infinite interpretations that satisfy S. However, it is not a practical way of interpre-
tations. In fact, inference rules provide a computationally feasible way to determine
the expression X , when it logically follows a set of premises S.
42 2 Logic and Reasoning Patterns

An example of an inference rule is Modus Ponens:

[(P → Q) ∧ P] → Q (2.3)

which is a valid statement (a tautology). Here, the Q also logically follows (entails)
from (P → Q) ∧ P. That is, [(P → Q) ∧ P] |= Q.

Definition 2.9 ‘Sound’ inference system.

When every inference X deduced from S also logically follows S, then the inference
system is sound. This is expressed by,

S  x ⇒ S |= x. (2.4)

The ‘’ is sign of ‘deduction’.

Soundness means that you cannot prove anything that is wrong. 

Definition 2.10 (A Complete inference system) If every X which logically follows S

can also be deducted (inferred), then the inference rule is complete. This is expressed
by
S |= x ⇒ S  x. (2.5)

Completeness means that you can prove anything that is right.

Another rule of inference is Modus tollens, specified as,

[(P → Q) ∧ ¬Q] → ¬P (2.6)

is sound and complete.

The reader may verify whether the inference rule of modus tollens is sound or
complete or both or none?

2.7 Nonmonotonic Reasoning

The classical logic or FOPL (first order predicate logic) discussed far, is not all time
sufficient to model the real-world knowledge of the world we live in. The reason are:
things become false to true or vice-versa over a time, addition of new knowledge in
the knowledge base may contradict the existing knowledge (e.g., the statement “the
surface of the earth is curved” becomes false on poles), things may be partially true
instead either true or false, and some times there is a probability of being some thing
true or false, and so on. Hence, there is a requirement of an all together different
approach and method of inferencing for real world situations.
The Nonmonotonic logic is the study of those ways of inferring from given knowl-
edge that do not necessarily satisfy the monotonicity property, satisfied by all methods
2.7 Nonmonotonic Reasoning 43

Fig. 2.5 Nonmonotonic

reasoning

No. of Inferences
Size of Knoledge

based on classical logic. In classical logic, if a conclusion is warranted on the basis

of certain premises (knowledge), no additional premises will ever invalidate the con-
clusion.
In everyday life, however, it seems clear that we humans draw sensible conclusions
from what we know and that, on the face of new information we often have to take
back previous conclusions. This happens even when the new information we gathered
in no compel us to take back our previous assumptions (see Fig. 2.5).
For example, we may hold the assumption that “most birds fly”, but that “penguins
are birds that do not fly”. On learning that “Tweety is a bird”, we infer that “Tweety
flies.” However, on learning that “Tweety is a penguin,” will in no way make us
change our mind about the fact that most birds fly, and also that penguins are birds
that do not fly or the fact that Tweety is a bird. However, it should make us aban-
don our conclusion about Tweety’s flying capabilities. It is desirable that intelligent
automated systems will have to do the same kind of (nonmonotonic) inferences.
Considering that Γ is a set of sentences of propositional logic, and α is inferred
from it, i.e Γ  α. For any new propositional sentences β, if Γ ∪ {β}  α then it is
monotonic reasoning. If it is not necessary that Γ ∪ {β}  α, then it is nonmonotonic
reasoning. We note from Fig. 2.5, that some times, even when we add into knowledge
base, the number of inferences decreases instead of increasing; and, this is property
of nonmonotonic reasoning.
Some of the systems that perform such nonmonotonic inferences are—negation
as failure, circumscription, modal system, default logic, autoepistemic logic, and
inheritance systems.

2.8 Hilbert and the Axiomatic Approach

An axiomatic system comprises a set of axioms and a set of primitives, where the
primitives are object names but, these objects are left undefined. The axioms are
the sentences that make assertions about the primitives. Further, these assertions are
not provided with any justifications, so they are neither true nor false. The subse-
quent or new assertions about the primitives are called theorems, are rigorous logical
consequences of axioms and previously proved theorems.
44 2 Logic and Reasoning Patterns

In 1899 the mathematician David Hilbert published his ground-breaking research

in the form of a book. He provided a complex deductive system based on five groups
of axioms, namely:
1. Axioms of incidence,
2. Axioms of order,
3. Axioms of congruence,
4. Axioms of continuity, and
5. an axiom of parallels.
As per Hilbert’s approach, the basic concepts of geometry comprises points, lines
and planes of Euclidean geometry. However, these concepts are never explicitly
defined. Instead, they are implicitly defined by the axioms such that, points, lines,
and planes are any family of mathematical objects that satisfy the given axioms of
geometry.
Twenty years later Hilbert was considered as the chief promoter of a program
intended to provide solid foundations to arithmetic, based on purely axiomatic
methods—the mathematics that model all the computations. It was called formalist
program, and Hilbert was identified as the champion of the formalist approach to
mathematics as a whole [6].

2.8.1 Roots and Early Stages

The formal definitions in an axiomatic system serves the purpose to simplify the
things as they can be used to create new objects made of complex combinations of
primitives and previously defined terms (objects and theorems). If a definite meaning
is assigned to a primitive of an axiomatic system, called as an interpretation, the
theorems become meaningful assertions.
Following are some definitions of the axiomatic system.

Definition 2.11 (Model (for axiomatic system.)) If all the axioms are true for a
given interpretation, then everything asserted by the theorem is also true. Such an
interpretation is called a model for the axiomatic system.

Definition 2.12 (Inconsistent (axiomatic system.)) Since a contradiction can never

be true, an axiomatic system using a contradiction can arrive at a logical deduction
that it has no model. An axiomatic system with this property is called inconsistent.

Definition 2.13 (Consistent (axiomatic system.)) If an abstract axiom system does

have a model, then such system is consistent.

Definition 2.14 (Isomorphic) If two models of the same axiom system can be proved
as structurally equivalent, then they are isomorphic to each other.

An axiomatic system can have more than one model.

2.8 Hilbert and the Axiomatic Approach 45

Definition 2.15 (Categorical Axioms) If all models of an axiom system are isomor-
phic then the axiom system is categorical.

Thus, for a categorical axiom system, there exists a model—the one and only
interpretation in which its theorems are all true.
The qualities—truth, logical necessity, consistency, and uniqueness were consid-
ered as the base of classical Euclidean geometry. Till recently, it was accepted that
Euclidean geometry is the only way to think about space. Now, the axiomatic sys-
tems are taken as the basis of geometry, and later all of the mathematics including
the computational mathematics and algorithms.
Hilbert’s definition of an axiomatic system lays the foundation of theory and
verifies that this system satisfies three main properties: independence, consistency,
and completeness. He proposed that just as in geometry, this kind of axiomatic
analysis should be applied to other fields of knowledge, and in particular to physical
theories. When we study any system of axioms as per Hilbert’s perspectives, the
focus of interest remained always on the disciplines themselves rather than on the
axioms. The axioms are just a means to improve our understanding of the discipline,
and not aimed to turn mathematics into a formally axiomatized game. For example,
in the case of geometry, a set of axioms were selected in such a way that they reflected
the basic manifestations of the intuition of space [4].

2.8.2 Axiomatics and Formalism

To understand the role of axioms, we will discuss the axioms of the set, as they
are useful in reasoning and inferences. By analyzing the mathematical arguments,
logicians become convinced that the notion of “set” is the most fundamental concept
of mathematics. For example, it can be shown that the notion of an integer can be
derived from the abstract notion of a set. Thus, in our world all the objects are sets,
and we do not postulate the existence of any more primitive objects. To support
this intuition, we can think our universe as all sets which can be built by successive
collecting processes, starting from the empty set, and we allow the formation of
infinite sets.
The first set of axioms for a general set theory was given by E. Zermelo in 1908,
and later developed by A. Fraenkel, hence usually referred to as Zermelo-Fraenkel
(ZF) set theory, the one we are most concerned. Another systems of axioms, which
has only finitely many axioms, but is less natural, was developed by von Neumann,
Bernays, and Gödel. The later is usually referred to as Gödel-Bernays (GB) set theory.
Following are some of the important axioms of ZF set theory [3, 8].

1. Axioms of Extensibility.

∀x∀y(∀z(z ∈ x ↔ z ∈ y) → x = y) (2.7)
46 2 Logic and Reasoning Patterns

The above says that set is determined by its members. We can define the subsets
as follows:
x ⊆ y ↔ ∀z(z ∈ x → z ∈ y). (2.8)

Also,
x ⊂ y ↔ x ⊆ y ∧ ¬x = y. (2.9)

2. Axiom of the Null set.

∃x∀y(¬y ∈ x). (2.10)

The set defined by this axiom is the empty or null set and we denote it by φ.

3. Axiom of Unordered Pairs.

∀x∀y∃z∀w(w ∈ z ↔ w = x ∨ w = y). (2.11)

We represent the set z by {x, y}. Also, {x} is {x, x} and we put x, y =
{{x}, {x, y}}. The set x, y is called ordered pair of x and y.
Using the above we can define a function as follows: a function is a set f of ordered
pairs such that x, y, x, z ∈ f → y = z. The set of x such that x, y ∈ f is
called domain, and set of y is called range. We say, f maps in set u if the range
of f is in u.

4. Axiom of set Union. It can be expressed as:

∀x∃y∀z(z ∈ y ↔ ∃t (z ∈ t ∧ t ∈ x)). (2.12)

The above says that y is union of all sets in x. Using the axiom Eq. 2.12, we can
deduce that given x and y, there exists z, such that z = x ∪ y, that is, t ∈ z ↔
t ∈ x ∨ t ∈ y.
To motivate for the next axiom being described, if x is an integer, the successor
of x will be defined as x ∪ {x}. Then the “axiom of infinity” generates a set that
contains all the integers and thus infinite.
5. Axiom of Infinity. It can be expressed as follows, and we understand that it is the
principle of Induction.

∃x(φ ∈ x ∧ ∀(y ∈ x → y ∪ {y} ∈ x). (2.13)

6. Axiom of Power Set. This axiom states that there exists for each x the set y for all
the subsets of x.
∀x∃y∀z(z ∈ y ↔ z ⊆ x). (2.14)

If the axiom of extensionality is dropped, the resulting system may contain atoms,
i.e., sets x such that ∀y(¬y ∈ x) yet the sets x are different. Indeed, one possible
view is that integers are atoms and should not be taken as sets.
2.8 Hilbert and the Axiomatic Approach 47

The first interesting axiom is the Axiom of Infinity. If we drop it, then we can take
a model for ZF set of all finite sets which can be built from φ.
The axioms discussed above can be used to prove theorems, like, mathematical
induction, invertible functions, and in fact another theorem of set theory, as well as
the corollaries, but the same are not appropriate to cover here, and a curious reader
is encouraged to refer the literature given in the bibliography.

2.9 Summary

Logic is used for valid deductions, and it avoids fallacy reasoning. Logic is also
useful in argumentation theory—a study of how conclusions can be reached through
logical reasoning, that is, whether the claims are soundly based on premises or not.
Argumentation includes debate and negotiation, that are concerned with reaching
mutually acceptable conclusions. The logic is used in proofs, games, and puzzles
solutions. The arguments have the internal structure: comprising of premises, rea-
soning process, and consequence.
The most commonly used, Propositional logic, represents sentences using single
symbols, called atoms, which are joined using the operators ∨, ∧, ¬, → to cre-
ate compound sentences. The sign of “→” in p → q is material implication, also
called conditional join, if p then q. Propositional logic expressions are called sen-
tences/statements; these are interpreted as true or false. The sentences are called wff,
and are defined recursively. A formula is a syntactic concept, which means whether
or not a string of symbols is a formula.
The meaning (semantics) is associated with each formula by defining its inter-
pretation, which assign a value true (T) or false (F) to every formula. Interpretation
of a statement means the assignment of true values to its atoms. A set of truth values
assigned to the atoms in a statement is called its world. Assignment of truth values
to the atoms in a statement, which makes the statement true is called model of the
statement.
The model checking is the process of truth-table enumeration, and is exponential
on n, the number of atoms in a statement. The derivation can also be represented by
a derivation-tree (parse-tree).
A propositional formula A is satisfied iff v(A) = T r ue for some interpretation
v. A satisfying interpretation is called model for A. The formula A is called valid,
denoted by |= A, iff v(A) = T r ue for all interpretations v. A sentence is logically
true (valid) iff it is true under every interpretation. |= θ means that θ is valid.
A reasoning, in which addition of new knowledge may produce inconsistency in
the knowledge base, is called nonmonotonic reasoning. As per the property of mono-
tonicity, if S  α, and β is additional assertion, then S ∧ β  α. The Nonmonotonic
logic is the study of those systems that do not satisfy the monotonicity property
satisfied by all methods based on classical logic.
The reasoning pattern comprises inference methods: modus ponens, modus tollens,
syllogism; and Proof methods: resolution theorem, model checking, model checking,
48 2 Logic and Reasoning Patterns

Normal forms. Deducing new knowledge from the existing set of the knowledge
base is called inferencing. The Modus ponens, modus tolens, syllogism are inference
rules, and Sound and Complete are good properties of inference systems.
Semantic tableau is a method for deciding satisfiability for the formula of proposi-
tional calculus, which systematically searches for a model for a formula. If it is found
the formula is satisfiable, else not satisfiable. Semantic Tableau is a tree, each node
of which will be labeled with a set of formulas, and these formulas are inductively
expanded to leaves such that each leaf is marked as open by or closed by ×.
The resolution rule is an inference which uses deduction approach. It is used in
theorem proving.
If every interpretation that satisfies S also satisfy X , then we say expression X
logically follows from a set of expressions S (the knowledge base). The Soundness
means that you cannot prove anything that is wrong, and Completeness means that
you can prove anything that is right.
An axiomatic system comprises a set of axioms and primitives.

Exercises

1. Prove the following assertions:

a. α is valid only if T |= α, where T stands for True.
b. For any (not necessarily valid) α, F |= α, where F is logically False.
c. Γ |= β if and only if the sentence (Γ ⇒ β) is valid. Here, Gamma is knowl-
edge base and β is sentence.
d. α ≡ β if and only if the sentence (α ⇔ β) is valid. Here, α and β are sentences.
2. Give your argument in favor and against, that the material join p → q is not
same as the cause effect p → q.
3. Determine, which of the following formulas are valid /satisfied /contradiction?
a. (( p → q) ∧ (¬ p → r )) → (q ∨ r )
b. ( p ∨ q) → ( p ∧ q)
c. p → ¬q
d. ( p ∧ q) → ( p ∨ q).
4. Show that ( p → q) → (¬q → ¬ p) is valid (Hint: Construct truth-table, the
interpretation of this formula shall be true for all the worlds).
5. Show that formula (¬ p ∧ ¬q) ∧ ( p ∨ q) is unsatisfiable (Hint: Construct truth-
table).
6. Assume a vocabulary with only four propositions, A, B, C, and D. Find out the
number of models for each of the following sentences?
a. P ⇔ Q ⇔ R.
b. P ∧ Q,
c. (P ∧ Q) ∨ (Q ∧ R),
Exercises 49

7. If S is a set of propositional formulas, then show that S |= F iff S ∪ {¬F} is

contradiction (Hint: A set of propositional formulas is contradiction, iff every
valuation of S there is a formula p in the set such that S |= p).
8. If θ |= ψ then prove that θ ∪ {A} |= ψ for any formula A.
9. If θ |= ψ and A is valid then prove that θ − {A} |= ψ for any formula A.
10. Find out the satisfiability of the following formulas using semantic Tableau
methods:
a. p ∧ (¬ p ∨ ¬q).
b. ( p ∧ q) ∨ (¬ p ∧ ¬q).
11. Show that if S is unsatisfiable then S − {Ai } is also unsatisfiable for every i ≤
i ≤ n.
12. Establish the consistency/inconsistency of the following statements using Tableau
method.
a. ( p ∧ q) ∨ ( p → q),
b. (¬q → ¬ p) ↔ ( p → q),
c. (¬ p ∨ ¬q) ∧ ( p ∨ q).
13. Convert the following statements into CNF:
a. a¬bc + bc + ¬ab¬c,
b. a + bc + ¬b¬c.
14. Convert the following statements into DNF:
a. (a + b)(a + ¬b + c)(a + ¬c),
b. (¬a + b) ∨ (a → b → c).
15. Find out the resolvent for { p → q, ¬q ∨ ¬r, r → p}.
16. Write a recursive algorithm T RU E(x, M) that returns true if and only if propo-
sitional logic sentence x is true in the model M, where M assigns a truth value
for every symbol in x. The algorithm should run in time linear in the size of the
sentence.

References

1. Ben-Ari M (2008) Mathematical logic for computer science, 2nd edn. Springer International,
London. ISBN-978-81-8128-344-3
2. Chowdhary KR (2012) Fundamentals of discrete mathematical structures, 2nd edn. EEE PHI
India. ISBN: 978-81-203-4506-5
3. Cohen PJ (2008) Set theory and the continuum hypothesis. Dover, New York
4. Davis M, Putnam H (1960) A computing procedure for quantification theory. J ACM 7(3):201–
215. https://doi.org/10.1145/321033.321034
5. Gelder AV (1991) The well-founded semantics for general logic programs. J ACM 38(3):620–
650
6. Hilbert D (2005) The foundations of geometry. Ebook # 17384
50 2 Logic and Reasoning Patterns

7. Kaushik S (2002) Logic and prolog programming. New Age International, New Delhi
8. Puppes P (1972) Axiomatic set theory. Dover, New York
9. Shankar N (2009) Automated deduction for verification. ACM Comput Surv 41(4):20:1. https://
doi.org/10.1145/1592434.1592437
Chapter 3
First Order Predicate Logic

Abstract The first order predicate logic (FOPL) is backbone of AI, as well a method
of formal representation of Natural Language (NL) text. The Prolog language for AI
programming has its foundations in FOPL. The chapter demonstrates how to translate
NL to FOPL in the form of facts and rules, use of quantifiers and variables, syntax
and semantics of FOPL, and conversion of predicate expressions to clause forms.
This is followed with unification of predicate expressions using instantiations and
substitutions, compositions of substitutions, unification algorithm and its analysis.
The resolution principle is extended to FOPL, a simple algorithm of resolution is
presented, and use of resolution is demonstrated for theorem proving. The interpre-
tation and inferences of FOPL expressions are briefly discussed, along with the use
of Herbrand’s universe and Herbrand’s theorem. At the end, the most general unifier
(mgu) and its algorithms are presented, and chapter is concluded with summary.

Keywords First Order Predicate Logic (FOPL) · Natural language · Quantifiers ·

Syntax and semantics of FOPL · Unification · Most general unifier · Resolution
theorem · Theorem proving · Herbrand’s universe · Herbrand’s theorem

3.1 Introduction

This chapter presents a formulation of first-order logic which is best suited as a basic
theoretical instrument—a computer based theorem proving program. As per the
requirements of theory, an inference method should be sound—allows only logical
consequences of premises deducible from the premises. In addition, it should be
effective—algorithmically decidable whether a claimed application of the inference
principle is really an application of it. When the inference principle is performed
by computer, the complexity of the inference principle is not an issue. However, for
more powerful principles, usage of combinatorial information processing for single
application may become dominant.
The system described in the following is an inference principle—the resolution
principle, is a machine-oriented rather than human-oriented system. Resolution prin-
ciple is quite powerful in psychological sense also, as it obeys a single type of

© Springer Nature India Private Limited 2020 51

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_3
52 3 First Order Predicate Logic

inference, which is often beyond the ability of the human to grasp. In theoretical
sense, it is a single inference principle that forms the complete system of first-order
logic. However, this latter property is not of much significance, but it is interesting
in the sense that no any other complete system of first-order logic is based on just
one inference principle, if ever one tries to realize a device of introducing a logical
axioms, or by a schema as an inference principle. The principle advantage of using
the resolution is due to its ability that allows us to avoid any major combinatorial
obstacles to efficiency, which used to be a serious problem in earlier theorem-proving
procedures.
Learning Outcomes of this Chapter:
1. Translate a natural language (e.g., English) sentence into predicate logic state-
ment. [Usage]
2. Apply formal methods of symbolic predicate logic, such as calculating validity
of formula and computing normal forms. [Usage]
3. Use the rules of inference to construct proofs in predicate logic. [Usage]
4. Convert a logic statement into clause form. [Usage]
5. Describe the strengths and limitations of predicate logic. [Familiarity]
6. Apply resolution to a set of logic statements to answer a query. [Usage]
7. Implement a unification-based type-inference algorithm for a simple language.
[Usage]
8. Precisely specify the invariants preserved by a sound type system. [Familiarity]

3.2 Representation in Predicate Logic

The first Order Predicate Logic (FOPL) offers formal approach to reasoning that has
sound theoretical foundations. This aspect is important to mechanize the automated
reasoning process where inferences should be correct and logically sound.
The statements of FOPL are flexible enough to permit the accurate representation
of natural languages. The words—sentence or well formed formula will be indicative
of predicate statements. Following are some of the translations of English sentences
into predicate logic:

• English sentence: Ram is man and Sita is women.

Predicate form: man(Ram) ∧ woman(Sita)
• English sentence: Ram is married to Sita.
Predicate form: married (Ram, Sita)
• English sentence: Every person has a mother.
The above can be reorganized as: For all x, there exists a y, such that if x is person
then x’s mother is y.
Predicate form: ∀x∃y[person(x) ⇒ hasmother(x, y)]
3.2 Representation in Predicate Logic 53

• English sentence: If x and y are parents of a child z, and x is man, then y is not
man.
∀x∀y[[parents(x, z) ∧ parents(y, z) ∧ man(x)] ⇒ ¬man(y)]

We note that predicate language comprises constants {Ram, Sita}, variables

{x, y}, operators {⇒, ∧, ∨, ¬}, quantifiers {∃, ∀} and functions/ predicates
{married(x, y), person(x)}. Unless specifically mentioned, the letters a, b, c, . . . at
the beginning of English alphabets shall be treated as constants to indicate names of
objects and entities, and those at the end, i.e., u, v, w, . . . shall be used as variables
or identifiers for objects and entities.
To indicate that an expression is universally true, we use the universal quantifier
symbol ∀, meaning ‘for all’. Consider the sentence “any object that has a feathers
is a bird.” Its predicate formula is: ∀x[hasfeathers(x) ⇒ isbird (x)]. Then certainly,
hasfeathers(parrot) ⇒ isbird (parrot) is true. Some expressions, although not always
True, are True at least for some objects: in logic, this is indicted by ‘there exists’,
and the existential quantifier symbol ∃ is used for this. For example, ∃x[bird (x)],
when True, this expression means that there is at least one possible object, that
when substituted in the position of x, makes the expression inside the parenthesis as
True [1].
Following are some examples of representations of knowledge FOPL.

Example 3.1 Kinship Relations.

mother(namrata, priti).(That is, Namrata is mother of Preeti.)

mother(namrata, bharat).
father(rajan, priti).
father(rajan, bharat).
∀x∀y∀z[father(y, x) ∧ mother(z, x) ⇒ spouse(y, z)].
∀x∀y∀z[father(y, x) ∧ mother(z, x) ⇒ spouse(z, y)].
∀x∀y∀z[mother(z, x) ∧ mother(z, y) ⇒ sibling(x, y)].

In above, the predicate father(x, y) means x is father of y; spouse(y, z) means y

is spouse of z, and sibling(x, z) means x is sibling of y. 
Example 3.2 Family tree.
Suppose that we represent “Sam is Bill’s father” by father(sam, bill) and “Harry is
one of Bill’s ancestors” by ancestor(harry, bill). Write a wff to represent “Every
ancestor of Bill is either his father, his mother, or one of their ancestors”.

∀x∀y[ancester(y, bill) ⇒(father(y, bill) ∨ mother(y, bill))

∨ ((father(x, bill) ∧ ancester(y, x))
∨ ((mother(x, bill) ∧ (ancester(y, x))].
54 3 First Order Predicate Logic

Example 3.3 Represent the following sentences by predicate calculus wffs.

1. A computer system is intelligent if it can perform a task which, if performed by

a human, requires intelligence.

∃x[[(perform(human, x) → requires(human, intelligence))

∧ (perform(computer, x)] → intelligent(computer))]

2. A formula whose main connective is ⇒ is equivalent to a formula whose main

connective is ∨.

∀x∀y[(formula(x) ∧ mainconnective(x, ⇒ ))
∧ (formula(y) ∧ mainconnective(y, ∨))
→ x ≡ y].

3. If a program cannot be told a fact, then it cannot learn that fact. ∀x[(program(x) ∧
¬told (x, fact)) → ¬learn(x, fact)] 

Example 3.4 Blocks World.

Consider that there are physical objects, like—cuboid, cone, cylinder placed on the
table-top, with some relative positions, as shown in Fig. 3.1. There are four blocks
on the table: a, c, d are cuboid, and b is a cone. Along with these there is a robot
arm, to lift one of the object having clear top.

Fig. 3.1 Blocks world

robot arm

d
b

a
3.2 Representation in Predicate Logic 55

Following is the set of statements about the blocks world (called knowledge base):
cuboid (a).
cone(b).
cuboid (c).
cuboid (d ).
onground (a).
onground (c).
ontop(b, a).
ontop(d , c).
toplcear(b).
topclear(d ).
∀x∀y[topclear(x) ∧ toplcear(y) ∧ ¬cone(x) ⇒ puton(y, x)].
The knowledge specified in the blocks world indicate that objects a, c, d are
cuboid, b is cone, a, c are put on the ground, and b, d are on top of a, c, respectively,
and the top of b, d are clear. These are called facts in knowledge representation. At
the end, the rule says that there exists objects x and y such that both have their tops
clear and x is not a cone, then y can be put on the object x. 
Bound and Free Variables
A variable in a wff is bound if it is within the scope of a quantifier naming the variable,
otherwise the variable is free. For example, in ∀x(p(x) → q(x, y)), x is bound and y
is free; in ∀x(p(x) → q(x)) → r(x), the x in r(x) is free variable. In the latter case
it is better to rename the variable to remove the ambiguity, hence we rephrase this
statement as ∀x(p(x) → q(x)) → r(z). An expression can be evaluated only when
all the variables in that are bound.
If F1 , F2 , . . . , Fn are wffs with ∧, ¬ as operators in each of Fi , then F1 ∨ F2 ∨
· · · ∨ Fn is called DNF (disjunctive normal form). Alternatively, if operators in Fi
are ∨, ¬ then F1 ∧ F2 ∧ · · · ∧ Fn is called CNF (conjunctive normal form). The
expression Fi called a term, consists only literals. We will prefer the CNF for predicate
expression. Thus, for an inference to be carried out, it is necessary to convert a
predicate expression to CNF. For example, ∃x[p(x) ⇒ q(x)] can be converted to
¬p(a) ∨ q(a), where a is an instance of variable x. The expression ¬p(a) ∨ q(a) is a
term of CNF. A formula in CNF, comprising ∧, ∨, ¬ along with constants, variables,
and predicates, is called clausal or clause form [2].

3.3 Syntax and Semantics

Two types of semantics are defined for the programing languages: (1) operational
semantics, and, (2) fixpoint semantics. The operational semantics defines input-
output relation computed by a program in terms of the individual operations per-
formed by the program inside a machine, like, basic logical and arithmetic opera-
tions. The meaning of a program is input-output relation obtained by executing the
56 3 First Order Predicate Logic

program on a machine. The other semantics—fixpoint semantics, is machine inde-

pendent. It defines the meaning of a program to be the input-output relation which is
the minimal fixpoint of a transformation associated with the program. The Fixpoint
semantics is used to justify existing methods for proving properties of programs, and
to justify new methods of proof.
We know the distinction between the syntax and the semantics from previous
chapter as well from the study of programming languages. The Syntax deals with
the formal part of language in abstraction from its meaning. It concerns with the
definition of well-formed formulas. Syntax in its narrow sense and also deals with
the study of axioms, rules of reference and proofs, which constitute proof theory.
Semantics is concerned with the interpretation of language and includes such notions
as meaning, logical implication and truth.
It is convenient to restrict attention to predicate logic programs written in clausal
form. Such programs have an especially simple syntax but retain all the expressive
power of the full predicate logic [3].

• Atomic formula. A string of symbols consisting of a predicate symbol of degree

n ≥ 0 followed by n terms is an atomic formula.
• Clause. A clause is a disjunction L1 ∨ · · · ∨ Ln of literals Li , each of which is
atomic formula P(t1 , . . . , tm ) or the negation of atomic formulas, where P is a
predicate symbol and ti , are terms. A finite set (possibly empty) of literals is called
a clause. The empty clause is denoted by: []
• Sentence. A sentence is a finite set of clauses.
• Literals. An atomic formula is a literal; and if A is an atomic formula then ¬A is
also literal.
The atomic formulas are positive literals, and negations of atomic formulas are
negative literals.
• Term. A term is either a variable or an expression like f (ti , . . . , tn ), where f is a
function symbol, ti are terms, and constants are 0-ary function symbols. A variable
is also a term, and a string of symbols comprising a function symbol of degree
n ≥ 0 followed by n terms is again a term.
A set of clauses {C1 . . . , Cn } is interpreted as a conjunction of clauses C1 . . . Cn .
A clause C containing just the variables x1 , . . . , xn is called universally quantified.
For example,

for all x1 , . . . , xn C (3.1)

is universally quantified clause.

• Ground Literals. A literal having no variables is called Ground Literal.
• Ground clauses. A clause with every member of it as a ground literal, is called a
Ground Clause. Empty clause—[] is a Ground Clause.
• Well-formed expressions. The Terms and Literals are (the only) Well-Formed
expressions.
• Lexical Order of Well-formed expressions. This is set of all well formed expressions
ordered in lexical order. The ordering is as follows: A precedes B if A is shorter
3.3 Syntax and Semantics 57

than B. If A and B have same length, then A has the alphabetically earlier symbol
in the first symbol position, at which A and B have distinct symbols.
• Herbrand Universe. It is set of ground terms associated with any set of S of clauses.
Let F be the set of all function symbols which occur in clause set S. If F contains any
function symbols of degree 0, then the functional vocabulary of S is F, otherwise,
it is {a} ∪ F, where a is a ground term. In this case, the Herbrand universe of S is
set of all ground terms with only symbols of the functional vocabulary of S.
• Models. It is a set of ground literals having no complementary pair. If M is a
Model and S is a set of ground clauses, then M is a model of S if, for all C ∈ S,
C contains a member of M . In general, if S is any set of clauses, and H is the
Herbrand Universe of S, then M is model of H (S).
• Satisfiability. A set S is Satisfiable if there is a model of S, otherwise S is Unsat-
isfiable.

For every sentence S1 of first order predicate logic there is always a sentence S2
in Clausal Form which is satisfiable if and only if S1 is also satisfiable. In other
words, for every non-clause form sentence there is a logically equivalent clause form
sentence. Due to this, all questions concerning to the validity or satisfiability of
sentences in FOPL can be addressed to sentences in clausal form.
Procedure for obtaining clausal-form for any well-formed formula (wff ) are dis-
cussed later in this chapter. In the above we have defined part of the syntax of
predicate logic, which is concerned with the specification of well-formed formulas.
The formalism we are going to use in the next section is based on the notions of
unsatisfiability and refutation rather than upon the notions of validity and proof.
To work on the criteria of refutation and unsatisfirability, it is necessary to convert
the given wff into clausal form.
To determine whether a finite set of sentences (S) of first-order predicate is satis-
fiable, it is sufficient to assume that each sentence in S is in clause form, and there is
no existential quantifiers as the prefix to S. In addition, the matrix of each sentence in
S is assumed to be a disjunction of formulas, each of which is either atomic formula
or the negation of an atomic formula. Therefore, the syntax of S is designed such
that the syntactical unit is a finite set of sentences in this special form, called clause
form. Towards the end of conversion process, the quantifier prefix is omitted from
each sentence, since it is necessary that universal quantifiers bind each variable in
the sentence. The matrix of each sentence is simply a set of disjuncts and the order
and multiplicity of the disjuncts are not important.

3.4 Conversion to Clausal Form

Following are the steps to convert a predicate formula into clausal-form [2].

1. Eliminate all the implications symbols using the logical equivalence: p → q ≡

¬p ∨ q.
58 3 First Order Predicate Logic

2. Move the outer negative symbol into the atom, for example, replace ¬∀x p(x)
by ∃x¬p(x).
3. In an expression of nested quantifiers, existentially quantified variables not in the
scope of universal quantifiers are replaced by constants. Replace ∃x∀y(f (x) →
f (y)) by ∀y(f (a) → f (y)).
4. Rename the variables if necessary. For example, in ∀x(p(x)) → q(x), rename
second free variable x, as ∀x(p(x) → q(y)).
5. Replace existentially quantified variables with Skolem functions; then elimi-
nate corresponding quantifiers. For example, for ∀x∃y[¬p(x) ∨q(y)], we obtain
∀x[¬p(x) ∨ q(f (x)). These newly created functions are celled Skolem functions,
and the process is called Skolemization.
6. Move the universal quantifiers to the left of the equation. For example, substitute
∃x[¬p(x) ∨ ∀y q(y)] by ∃x∀y[¬p(x) ∨ q(y)]
7. Move the disjunctions down to the Literals, i.e., terms should be connected by
conjunctions only, vertically.
8. Eliminate the conjunctions.
9. Rename the variables, if necessary.
10. Drop all the universal quantifiers, and write each term in a separate line.

The resulting sentence is a CNF, and suitable for inferencing using resolution.

Example 3.5 Convert the expression ∃x∀y[∀z p(f (x), y, z) ⇒ (∃u q(x, u) ∧
∃v r(y, v))] to clausal form.

The steps discussed above are applied precisely, to get the clausal form of the pred-
icate formula.

1. Eliminate implication.

∃x∀y[¬∀z p(f (x), y, z) ∨ (∃u q(x, u) ∧ ∃v r(y, v))]

2. Move negative symbols to the atom.

∃x∀y[∃z¬p(f (x), y, z) ∨ (∃u q(x, u) ∧ ∃v r(y, v))]

3. Replace existentially quantified variables not in the scope of universal quantifier

to constants.

∀y[∃z¬p(f (a), y, z) ∨ (∃u q(a, u) ∧ ∃v r(y, v))]

4. Rename variables (not required in this example.)

5. Replace existentially quantified variables that are functions of universal quanti-
fied variables, by Skolem functions:

∀y[¬p(f (a), y, g(y) ∨ (q(a, h(y) ∧ r(y, l(y))]

3.4 Conversion to Clausal Form 59

6. Move ∀ to left is not required in this example.

7. Move disjunctions down to Literals.

∀y[(¬p(f (a), y, g(y)) ∨ (q(a, h(y))) ∧ (¬p(f (a), y, g(y)) ∨ r(y, l(y)))]

8. Eliminate conjunctions.

∀y[¬p(f (a), y, g(y)) ∨ (q(a, h(y)), (¬p(f (a), y, g(y)) ∨ r(y, l(y))]

9. Renaming variable is not required in this example.

10. Drop all universal quantifiers and write each term on separate line.

¬p(f (a), y, g(y)) ∨ (q(a, h(y)),

¬p(f (a), y, g(y)) ∨ r(y, l(y)).

Example 3.6 Convert the following wff to clause form.

(∀x)(∃y){[p(x, y) ⇒ q(y, x)] ∧ [q(y, x) ⇒ s(x, y)]}

⇒ (∃x)(∀y)[p(x, y) ⇒ s(x, y)]

For [p(x, y) ⇒ q(y, x)] ∧ [q(y, x) ⇒ s(x, y)] by application of syllogism, it can be
reduced to [p(x, y) ⇒ s(x, y)]. Thus, original expression reduces to:

= (∀x)(∃y)[p(x, y) ⇒ s(x, y)] ⇒ (∃x)(∀y)[p(x, y) ⇒ s(x, y)]

= ¬(∀x)(∃y)[p(x, y) ⇒ s(x, y)] ∨ (∃x)(∀y)[p(x, y) ⇒ s(x, y)]
= (∃x)¬(∃y)[p(x, y) ⇒ s(x, y)] ∨ (∃x)(∀y)[p(x, y) ⇒ s(x, y)]
= (∃x)(∀y)¬[p(x, y) ⇒ s(x, y)] ∨ (∃x)(∀y)[p(x, y) ⇒ s(x, y)]
= (∀y)¬[p(a, y) ⇒ s(a, y)] ∨ [p(a, y) ⇒ s(a, y)]
= [p(a, y) ∧ ¬s(a, y)] ∨ [¬p(a, y) ∨ s(a, y)]
=T

3.5 Substitutions and Unification

The following definitions are concerned with the operation of instantiation, i.e, sub-
stitutions of terms for variables in the well-formed expressions and in sets of well-
formed expressions [8].
Substitution Components
A substitution component is any expression of the form T /V , where V is any variable
and T is any term different from V . The T can be any constant, variable, function,
predicate, or expression.
60 3 First Order Predicate Logic

Substitutions
A substitution is any finite set (possibly empty) of substitution components, none
of the variables of which are same. If P is any set of terms, and the terms of
the components of the substitution θ are all in P, we say that θ is a substitu-
tion over P. We write the substitution where components are T1 /V1 , . . . , Tk /Vk as
θ = {T1 /V1 , . . . , Tk /Vk }, with the understanding that order of components is imma-
terial. We will use lowercase Greek letters θ, λ, μ denote substitutions.
Instantiations
If E is any function string of symbols, and θ = {T1 /V1 , . . . , Tk /Vk } is any substitu-
tion, then the instantiation of E by θ is the operation of replacing each occurrence of
variable Vi , 1 ≤ i ≤ k, in E by term Ti . The resulting string denoted by Eθ is called
an instance of E by θ . That is, if E is the string E0 Vi1 E1 . . . Vin En , n ≥ 0, then Eθ is
the string E0 Ti1 E1 . . . Tin En . Here, none of the substrings Ej of E contain occurrences
of variables V1 , . . . , Vk after substitution. Some of Ej are possibly null, and each Vij
is an occurrence of one of the variables V1 , . . . , Vk .

3.5.1 Composition of Substitutions

If θ = {T1 /V1 , . . . , Tk /Vk } and λ are any two substitutions, then the composition of
 
θ and λ denoted by θ λ is union θ ∪ λ , defined as follows:

The θ is set of all components Ti λ/Vi , 1 ≤ i ≤ k, such that Ti λ (λ substituted

in θ ) is different from Vi , and λ is set of all components of λ whose variables are
not among V1 , . . . , Vk .
Within a given scope, once a variable is bound, it may not be given a new binding
in future unifications and inferences. If θ and λ are two substitution sets, then the
composition of θ and λ, i.e., θ λ, is obtained by applying λ to the elements of θ and
adding the result to λ.
Following examples illustrate two different scenario of composition of substitu-
tions.

Example 3.7 Find out the composition of {x/y, w/z}, {v/x}, and {A/v, f (B)/w}.

Let us assume that θ = {x/y, w/z}, λ = {v/x} and μ = {A/v, f (B)/w}. Following
are the steps:

1. To find the composition λμ, A is is substituted for v, and v is then substituted for
x. Thus, λμ = {A/x, f (B)/w}.
2. When result of λμ is substituted in θ , we get composition θ λμ = {A/y,
f (B)/z}. 

Example 3.8 Find out the composition of θ = {g(x, y)/z}, and λ = {A/x, B/y,
C/w, D/z}.
3.5 Substitutions and Unification 61

By composition,

θ λ = {g(x, y)/z} ◦ {A/x, B/y}

= {g(A, B)/z, A/x, B/y, C/w}

The {D/z} has not been included in the resultant substitution set, because other-
wise, there will be two terms for the variable z, one g(A, B) and other D. 
One of the important property of substitution is that, if E is any string, and σ = θ λ,
then Eσ = Eθ λ. It is straight forward to verify that εθ = θ ε = θ for any substitu-
tion θ . Also, composition enjoys the associative property (θ λ)μ = θ (λμ), so we may
omit the parentheses in writing multiple compositions of substitutions. The substi-
tutions are not in general commutative; i.e., it is generally not the case that θ λ = λθ ,
because for this Eθ λ has to be equal to Eλθ , which is not guaranteed. However, the
composition has distributive property.
The point of the composition operation on substitution is that, when E is any
string, and σ = θ λ, the string Eσ is just the string Eθ λ, i.e., the instance of Eθ by λ.

3.5.2 Unification

If E is any set of well-formed expressions and θ is a substitution, then θ is said to unify

E, or to be a unifier of E, if Eθ is a singleton. Any set of well-formed expressions
which has a unifier is said to be unifiable [6].
In proving theorems using quantified variables, it is often necessary to “match”
certain subexpressions. For example, to apply the combination of modus ponens
and universal instantiation (Eq. 3.5) to produce “mortal(socrates)”, it was necessary
to find substitution {socrtaes/x} for x that makes man(x) and man(socrates) equal
(singleton).
Unification algorithm determines the substitutions needed to make two predicate
expressions match. For this, all the necessary condition is that variables must be uni-
versally quantified. Unless the variables in an expression are existentially quantified,
they are assumed to be universally quantified. This criteria allows us full freedom
choosing the substitutions. The existentially quantified variables can be eliminated
by substituting them with constants or with Skolem functions that makes the sentence
true. For example, in sentence,

∃x mother(x, jill),

we can replace x with a constant designating jill’s mother, susan, to get:

mother(susan, jill);

and write unifier as {susan/x}.

62 3 First Order Predicate Logic

For perform unification, a variable can be replaced by any term, including other
variable or function expressions of arbitrary complexity. This also includes function
expressions that themselves contain variables. For example, the function expression,
mother(joe), may be substituted for x in human(x) to get human(mother(joe)).
Example 3.9 Find out the substitution instances for foo(x, a, zoo(y)), given the sim-
ilar predicates with literal arguments.
1. foo(fred , a, zoo(z)), where fred is substituted for x and z for y, i.e., λ1 =
{fred /x, z/y}. Thus,

foo(fred , a, zoo(z)) = foo(x, a, zoo(y))λ1 .

2. foo(w, a, zoo(jack)), where λ2 = {w/x, jack/y}; hence

foo(w, a, zoo(jack)) = foo(x, a, zoo(y))λ2 .

3. foo(z, a, zoo(moo(z))), where λ3 = {z/x, moo(z)/y}, hence;

foo(z, a, zoo(moo(z))) = foo(x, a, zoo(y))λ3 .

We use the notation x/y to indicate that x is substituted for the variable y; we
also call this as bindings, so y is bound to x. A variable cannot be unified with a
term containing that variable. So p(x) cannot be substituted for x, because this would
create an infinite regression: p(p(p(...x)...).

3.6 Resolution Principle

The resolution rule can be traced back to 1960, which was introduced by Davis
and Putnam. However, this algorithm required all around ground instances for the
given formula, which resulted to a combinatorial explosion. However, the source of
combinatorial explosion was eliminated in 1965, when J. A. Robinson introduced
an algorithm for unification. The unification allowed the instantiation of the formula
during the proof “in demand”, just as per the need through the newly introduced most
general unifier (mgu) algorithm [8].
The resolution method for (propositional) logic due to J. A. Robinson (1965) is
sound and complete, and a well-known procedure for checking the unsatisfiability of
a set of clauses. The resolution is mathematical oriented rather than human oriented.
It is quite powerful both in the psychological sense that it condones single inferences
which are often beyond the ability of human to grasp, and in theoretical sense that
it alone, as sole inference principle, forms a complete system of FOPL. Because of
only one inference type, it allows to avoid the combinatorial obstacles to efficiency.
Let us refresh us with the terminology we discussed in the beginning of this
chapter. We will designate a literal by symbol L, which is either a propositional
3.6 Resolution Principle 63

Fig. 3.2 DAG for theorem {P, Q} {P, ¬Q} {¬P, Q} {¬P, ¬Q}
proving

{P } {Q}

{¬P }

{}

symbol, P, or the negation, ¬P. A finite set of literals {L1 , . . . , Lk } is a clause,

which is interpreted as a disjunction of literals, L1 ∨ · · · ∨ Lk . With k = 0, it is an
empty clause, denoted as []. A conjunction of a set of clauses Γ = {C1 , . . . , Cn } is
interpreted as C1 ∧ · · · ∧ Cn . In short, we write set of clauses as Γ = C1 , . . . , Cn .
The resolution method is a procedure for determining whether a set of clauses Γ ,
is unsatisfiable. To find out latter, the resolution method first builds a certain kind
of labeled DAG (Directed Acyclic Graph) whose terminal nodes are labeled with
clauses in Γ and the interior nodes are labeled as per the resolution rule [5].
Consider that there are any two clauses C = A ∪ {P} and D = B ∪ {¬P} (where
P is a propositional letter, P ∈ / A and ¬P ∈ / B). The resolvent of C and D is the
clause R = A ∪ B obtained by canceling out P and ¬P, and making disjunction
of the remaining part in each clause. A resolution DAG for Γ is a DAG whose
terminal nodes are labeled with clauses from Γ and such that every interior node n
has exactly two predecessors, n1 and n2 so that n is labeled with the resolvent of the
clauses labeling n1 and n2 . A resolution refutation for Γ is a resolution DAG with a
single root whose label is the empty clause. Note that the root will come at the end,
as shown in Fig. 3.2.

Example 3.10 Resolution refutation.

A resolution refutation for the set of clauses

Γ = {{P, Q}, {P, ¬Q}, {¬P, Q}, {¬P, ¬Q}}.

is shown in Fig. 3.2. 

A recursive algorithm can be given for construction of a resolution DAG using

any number of clauses, and to prove its correctness. That means, if the input set of
clauses is unsatisfiable, the output resolution DAG is a resolution refutation. This
confirms the completeness of propositional resolution constructively.
64 3 First Order Predicate Logic

3.6.1 Theorem Proving Formalism

It is a syntactic inference procedure, when applied to clauses, determines, if the

satisfied set is unsatisfiable. Proof is similar to proof by contradiction and deduce []
(i.e., null). If for example, we have set of clauses (axioms) C1 , C2 , . . . , Cn , and we
want to deduce D, i.e., D is logical consequence of of C1 , C2 , . . . , Cn . For this we add
¬D to the set {C1 , C2 , . . . , Cn }, then we show that set is unsatisfiable by deducing
contradiction [7].
The process of deduction using resolution is given in Algorithm 3.1. Given two
clauses C1 , C2 with no variables in common, and if l1 is a literal in C1 and its
complement literal l2 is a literal in C2 , then l1 , l2 can be dropped and disjunction C
is obtained from the remaining part of C1 , C2 . The C is called resolvent of C1 , C2 .
Let C1 = ¬P ∨ Q, and C2 = ¬Q ∨ R, then following can be deduced through
resolution,
P ⇒ Q, Q ⇒ R
(3.2)
P⇒R
equivalently,
(¬P ∨ Q), (¬Q ∨ R)
. (3.3)
∴ (¬P ∨ R)

It can be easily verified that (¬P ∨ Q) ∧ (¬Q ∨ R) |= (¬P ∨ R), hence (¬P ∨
Q) ∧ (¬Q ∨ R) ⇒ (¬P ∨ R) is a valid statement. Thus, ¬P ∨ R is inference or the
resolvent. Arriving to a proof by above is called proof by refutation.
Resolution says that if there are axioms of the form ¬P ∨ Q and there is another
axiom of the form ¬Q ∨ R, then ¬P ∨ R logically follows; called the resolvent. Let
us see why it is so? When ¬P ∨ Q is True, then either ¬P is True or Q is True. For
other expression, when ¬Q ∨ R is True, then either ¬Q is True or R is True. Then we
can say that ¬P ∨ R is certainly True. This can be generalized to two expressions,
when we have any number of expressions, but two must be of opposite signs.

3.6.2 Proof by Resolution

To prove a theorem, one obvious strategy is to search forward from the axioms, using
sound rules of inference. We try to prove a theorem by refutation. It requires to show
that negation of a theorem cannot be True. The steps for a proof by resolution are:
1. Assume that negation of the theorem is True.
2. Try to show that axioms and assumed negation of theorem, together are True,
which cannot be True.
3. Conclude that above leads to contradiction.
4. Conclude that theorem is True because its negation cannot be True.
3.6 Resolution Principle 65

To apply the resolution rule,

1. Find two sentences that contain the same literal, one in its positive form and one
in its negative form, like,

CNF : summer ∨ winter, ¬winter ∨ cold ,

2. use the resolution rule to eliminate the complement literals from both sentences
to get,
CNF : summer ∨ cold .

The Algorithm 3.1 is an algorithm for theorem proving through resolution-

refutation, where α is the theorem to be proved, and β is set of axioms, both of
these are input to the algorithm. All the inputs to algorithm are in the clause form.
The algorithm returns “true” if the theorem is true, else returns “False”.

Algorithm 3.1 Algorithm-Resolve(Input: α, β)

1: Γ = β ∪ {¬α}
2: while there is a resolvable pair of clauses Ci , Cj ∈ Γ do
3: C = resolve(Ci , Cj )
4: if C = NIL then
5: return “Theorem α is true”
6: end if
7: Γ = Γ ∪ {C}
8: end while
9: Report that theorem is False

3.7 Complexity of Resolution Proof

The question is, how you can be so clever to pickup the right clauses to resolve? The
answer is that you take advantage of two ideas:
1. You can be sure that every resolution involves the negated theorem, directly or
indirectly.
2. You know where you are and where you are going, hence you can compute the
difference to help you proceed with your intuition for selection of clauses.
Consider there are total n clauses, c1 . . . cn . We can try to match c1 with c2 . . . cn ,
and in next level c2 is matched with c3 . . . cn , and so on. This results to breadth first
search (BFS). Consider that resolvents generated due to this matching are c1 . . . cm .
Next all the newly generated clauses are matched with the original, and then they
are merged into the original. This process is repeated until contradiction is reached,
showing that theorem is proved. Since, the entire set of clauses are compared, the
66 3 First Order Predicate Logic

proof is bound to result, if it exists, at all. This gives completeness to the resolution
proof.
The other alternative is, nodes which are farther and farther away are matched
before those which are closer to the root. The c1 is matched with first child c2 out of
c2 . . . cn . Then c2 is matched with its first child generated, and so on, resulting to the
search process called DFS (depth first search).
However, the above both are brute-force algorithms, and are complex. The other
methods are heuristic based. In fact, there is difficulty to express your concepts
required in pure logic. One of the approaches is to use the clauses having smallest
number of literals. Another, to use negated clauses.
The resolution search strategies are subject to the exponential-explosion problem.
Due to this, those proofs which require long chains of inferences, will be exponen-
tially expensive in time.
All resolution search strategies are subject to a version of halting problem, for
search is not guaranteed to terminate unless there actually is a proof. In fact, all
complete proof procedures for the first order predicate calculus are subject to halting
problem. Complete proof procedures are said to be semi-decidable, because they are
generated to tell you whether an expression is a theorem, only if the expression is
indeed a theorem.
Theorem proving is suitable for certain problems, but not for all problems, due to
the following reasons:

1. Complete theorem proving requires search, and search is inherently exponential,

2. Theorem provers may not help you to solve practical problems, even if they do
their work instantaneously.

3.8 Interpretation and Inferences

A FOPL statement is made of predicates, arguments (constants or variables), func-

tions, operators, and quantifiers. Interpretation is process of assignment of truth
values (True/False) to subexpressions and atomic expressions, and computing the
resultant value of any expression/statement. A statement or expression in predicate
logic is also called wwf (well formed formula).
Consider the interpretation of predicate formula:

∀x[bird (x) → flies(x)]. (3.4)

To find out the satisfiability of the formula (3.4), we need to substitute (instantiate)
a value for x (an instance of x) and check if flies(x) is true. Until, that x is found,
it may require instantiation with large number of values. Similarly, to check if the
Eq. (3.4) is valid, it may require infinitely large number of values in the domain of
3.8 Interpretation and Inferences 67

x to be verified. If any one of that makes the formula false, the formula is not valid.
Thus, checking of satisfiability as well as validity of a formula is predicate logic are
complex process. The approach of truth table and tableau method we discussed in
the previous chapter are applicable here also.
Given a predicate sentences of m number of predicates each having one argument,
and domain size of all the arguments is n, in the worst case it will require total nm
substitutions to test for satisfiability, as well as for validity checking. However, a
sentence of m propositions will require in the worst case only 2m substitutions.
Hence, satisfiability checking in predicate sentences is much more complex than that
in proposition logic. It equally applies with expressions having existential quantifiers,
like, ∃x[bird (x) → flies(x)].
Thus, it is only the proof methods, using which logical deductions can be carried
out in realistic times.

Example 3.11 Given, “All men are mortal” and “Socrates is man”, infer using pred-
icate logic, that “Socrates is mortal”.

The above statement can be written in predicate logic as:

∀x[man(x) ⇒ mortal(x)],
man(socrates). (3.5)

Using a rule called universal instantiation, a variable can be instantiated by a

constant and universal quantifier can be dropped. Hence, from (3.5) we have,

man(socrates) ⇒ mortal(socrates),
man(socrates). (3.6)

Using the rule of modus ponens on (3.6) we deduce “mortal(socrates)”. It

is also logical consequence. If Γ = {[man(socrates) ⇒ mortal(socrates)] ∧ man
(socrates)}, and α = mortal(socrates), then we can say that Γ  α.
The set of formulas Γ is called knowledge base. To find out the result for the
query “Who is man?”, we must give the query

?man(X ).

in Prolog (to be discussed later), which will match (called unify or substitute) man(X )
with man(socrates) with a unification set, say, θ = {socrates/X }. The substitution
which returns man(socrates) is represented by man(X )θ . 
68 3 First Order Predicate Logic

Example 3.12 Prolog Program.

The sentence in Eq. (3.5) will appear in prolog as,

mortal(socrates) :- man(socrates).
man(socrates).

Here, the sign ‘:-’is read as ‘if’. The subexpression before the sign ‘:-’is called
‘head’or procedure name and the part after ‘:-’is called body of the rule. The sentence
(3.6) can also be written in a clause form (to be precise, in Horn clause form) as,

mortal(socrates) ∨ ¬man(socrates).
man(socrates). (3.7)

3.8.1 Herbrand’s Universe

Defining an operational semantics for a programming language is nothing but to

define an implementation independent interpreter for it. In case of predicate logic,
the proof procedure itself behaves like an interpreter. The Herbrand’s Universe and
Herbrand’s base play an important role in interpretation of predicate language. In the
following, we define the Herbrand’s Universe and Herbrand’s Base.

Definition 3.1 (Herbrand’s Universe) In a predicate logic program, a Herbrand Uni-

verse H, is a set of ground terms that use only function symbols and constants.

Definition 3.2 (Herbrand’s Base) A set of atomic formulas formed by predicate

symbols in a program, is called Herbrand’s base. The additional condition is that,
arguments of these predicate symbols are in the Herbrand Universe.

For a predicate program, the Herbrand universe and Herbrand base are countably
infinite if the predicate program contains a function symbol of positive arity. If the
arity of function symbols is zero, then both the haerbrand’s universe and base are
finite [3].
In special cases, when resolution proof is used on FOPL, it reduces the expressions
to propositional form. If the set of clauses is A, its Harbrand’s universe is set of all
the ground terms formed using only the function symbols and constants in A. For
example, if A has constants a, b, and a unary function symbol f , then the Herbrand
universe is the infinite set:

{a, b, f (a), f (b), f (f (a)), f (f (b)), f (f (f (a)), . . . }.

The Herbrand’s base of A is the set of all ground clauses cθ where c ∈ A and θ
is a substitution that assigns the variables in c to terms in the Herbrand’s universe.
3.8 Interpretation and Inferences 69

Horn clauses
The Horn clauses are terms without variables, these are constructed using constants
and function symbols that occur in the set of clauses A. These terms form the data
structures, which are manipulated by the program built-in in the clauses A. The
collection of all such terms, determined by A, is called Herbrand universe. Every
n-ary predicate symbol P occurring in A denotes an n-ary relation over the Herbrand
universe of A. We call the n-tuples which belong to such relations as input-output
tuples and the relations themselves as input-output relations.
In an inference system, the operational semantics determine a unique denotation
for a formula P such that the n-tuple (t1 , . . . , tm ) belongs to the denotation of P ∈ A,
iff A  P(t1 , . . . , tn ). That is,

D(P) = {(t1 , . . . , tn ) : A  P(t1 , . . . , tn )}. (3.8)

We first pick an arbitrary constant, say, a, and then construct the variable-free
terms. Formally, D(P) is inductively defined as follows:

1. All constants occurring in P belong to D(P); if no constant occurs in P, then

a ∈ D(P).
2. For every n-ary functional symbol p occurring in P, if t1 , t2 , . . . , tn ∈ D(P) then
p(t1 , t2 , . . . , tn ) ∈ D(P).

Here X  Y means X derives Y . For resolutions systems, if X  Y then there

exists a refutation of the sentence in clausal form with atoms as X and Y .
In goal oriented inference systems, the procedure calls are replaced by procedure
bodies. Such inference systems correspond to standard notion of operational seman-
tics. In theoretical sense, any inference system, based on predicate logic represents
an abstract machine, that generates only those derivatives which are determined by
this inference system.
For predicate logic, the corresponding programs compute the relations represented
by predicate symbols in the set of clauses A. These relations may be in the form of
predicates or functions. However, these function or predicate symbols are not treated
as functions computed by the program, but they result into some data structures,
and these data structures are actually the input and output objects of the relations /
functions being computed.

Definition 3.3 Herbrand’s Structure.

Let P be a formula in Skolem form (when a constant, say, a is substituted for x

in ∃x p(f (x), b) results to Skolem form p(f (a), b))). A structure A = (UA , IA )
suitable for p is a Herbrand structure for P if it satisfies the following conditions:

1. UA = D(P), and
2. for every n-ry function symbol p occurring in P and every t1 , t2 , . . . , tn ∈ D(P) :
f A (t1 , t2 , . . . , tn ) = f (t1 , t2 , . . . , tn ).
70 3 First Order Predicate Logic

In above, A = (UA , IA ) is model with UA as formula and IA as its

interpretation. 

Definition 3.4 (General logic program) It is a finite set of general rules, with both
positive and negative subgoals.

A general logic program comprises rules and facts. A general rule’s format is: its
head, or conclusion is to the left of the symbol “←,” (read “if”), and its subgoals
(called body) is right of the symbol “←”. Following is an example of rule, where
p(X ) is head, q(X ) is positive subgoal, and r(X ) is a negative subgoal.

p(X ) ← q(X ), ¬ r(X ). (3.9)

This rule may be read as “p(X ) if q(X ) and not r(X ).” A Horn rule has no
negative subgoals, and a Horn logic program is made of only Horn rules.
We will follow the conventions of Prolog for naming the objects: the logical
variables begin with an uppercase letter, while the constants, functions, and predicates
begin with a lowercase letter. For both the predicate and its relation, we will use the
same symbol, for example p.
Followings may be the arguments of a predicate:

1. a constant/variable is a term;
2. a function symbol with terms as arguments, is a term.

The terms may be viewed as data structures of the program, with function symbols
serving as record names. Often a constant is treated as a function symbol of arity
zero.
Following are the definitions of some important terms.

Definition 3.5 (Herbrand Instantiation) Herbrand instantiation of a general logic

program is the set of rules obtained by substituting terms in the Herbrand universe
for variables, in every possible way.

Definition 3.6 An instantiated rule is one only, whereas “uninstantiated” logic

programs are assumed to be a finite set of rules, and instantiated logic programs may
be infinite in number.

Definition 3.7 (Complement of a set) For a set of literals L its complement is a set
formed by complementing of each literal in L, represented by ¬L.

Further,
– p is said to be inconsistent with L if p ∈ ¬L,
– Sets of literals R and L are inconsistent if at least one literal in R is inconsistent
with L, i.e., when R ∩ ¬L = φ,
– A set of literal is inconsistent if it is inconsistent with itself; otherwise it is con-
sistent.
3.8 Interpretation and Inferences 71

3.8.2 Herbrand’s Theorem

Herbrands theorem is a fundamental theorem based on mathematical logic, that

permits a certain type in reduction from FOPL to propositional logic [4].
In its simplest form, the Herbrand’s theorem states that a formula of first-order
predicate logic ∃x A, where A is quantifier free, is provable if and only if there exist
ground terms M1 , . . . , Mn such that,

|= A[x := M1 ] ∨ · · · ∨ A[x := Mn ]. (3.10)

When using the classical formulation, the Herbrand’s theorem relates the validity
of a first-order formula in Skolem prenex form1 to the validity of one of its Herbrand
extensions. That means, the formula ∀x1 . . . ∀xn ψ(x1 . . . , xn ) is valid if, and only if,
 m
i ψ(ti,1 , . . . , ti,n ) is valid for some m ≥ 1 and some collection of ground Herbrand
terms ti,j .
Since it is possible that every classical first-order formula can be reduced to this
Skolem prenex form through the Skolemization while preserving its satisfiability, the
Herbrand’s theorem provides a way to reduce the question of validity of first-order
formulas to propositional logic formula.
However, the required Herbrand’s extension and the terms ti,j cannot be computed
recursively (for otherwise first-order logic would be decidable), this result is highly
useful for the automated reasoning as it gives a way to some highly efficient proof
methods such as resolution and the resolution refutation.
Theorem 3.1 A closed formula F in Skolem form is satisfiable if and only if it has
a Herbrand model.
Proof If the formula has a Herbrand model then it is satisfiable. For the other direc-
tion let A = (UA , IA ) be an arbitrary model of F. We define a Herbrand structure
B = (UB , IB ) as follows:
Universe: UB = D(F)
Functional Symbols: f B (t1 , t2 , . . . , tn ) = f (t1 , t2 , . . . , tn )
Predicate Symbols: (t1 , . . . , tn ) ∈ P B iff A (t1 ), . . . , A (tn ) ∈ P A .
Claim: B is also a model of F.
We prove a stronger assertion: For every closed form G in Skolem form such that
G ∗ only contains atomic formulas of F ∗ : if A |= G then B |= G.
By induction on the number n of universal quantifiers of G.
Basis (n = 0). Then G has no quantifiers at all.
It follows A (G) = B(G), this proves the theorem. 
To perform reasoning with the Herbrand base, the unifiers are not required, and we
have a sound and complete reasoning procedure, which is guaranteed to terminate.
The idea used in this approach is: Herbrand’s base will typically be an infinite set of
propositional clauses, but it will be finite when Herbrand’s universe is finite (there

1A string of quantifiers followed by a quantifier-free part, e.g., ∀x1 . . . ∀xn ψ(x1 . . . , xn ).

72 3 First Order Predicate Logic

is no function symbols and only finitely many constants appear in it). Sometimes
we can keep the universe finite by considering the type of the arguments (say t) and
values of functions (f ), and include a term like f (t) in the universe only if the type of
t is appropriate for the function f . For example, f (t) may be, birthday(john), which
produces a date.

3.8.3 The Procedural Interpretation

It is easy to procedurally interpret the sets of clauses, say, A, which contain at most one
positive literal per clause. However, along with this any number of negative literals
can also exist. Such sets of clauses are called Horn sentences or Horn Clauses or
simply clauses. We distinguish three kinds of Horn clauses [3].

1. ‘[]’the empty clause, containing no literals and denoting the truth value false, is
interpreted as a halt statement.
2. B̄1 ∨ · · · ∨ B̄n , a clause consisting of no positive literals and n ≥ 1 negative literals,
is interpreted as a goal statement. Note that goal statement is negated and added
into the knowledge base to obtain the proof through resolution refutation.
3. A ∨ B̄1 ∨ · · · ∨ B̄n , a clause consisting of exactly one positive literal and n ≥ 0
negative literals is interpreted as a procedure declaration (i.e., rule in Prolog
program). The positive literal A is the procedure name and the collective negative
literals are the procedure body. Each negative literal Bi , in the procedure body
is interpreted as a procedure call. When n = 0 the procedure declaration has an
empty body and interpreted as an unqualified assertion of fact.

In the procedural interpretation, a set of procedure declarations is a program. Com-

putation is initiated by an initial goal statement, which proceeds by using declared
procedures to derive new goal statements (subgoals) Bi s from old goal statements,
and terminates on the derivation of the halt statement. Such derivation of goal state-
ments is accomplished by resolution, which is interpreted as procedural invocation.
Consider that, a selected procedure call Ā1 inside the body of a goal statement as,

Ā1 ∨ · · · ∨ Āi−1 ∨ Āi ∨ Āi+1 ∨ · · · ∨ Ān (3.11)

and a procedure declaration is given as,

A ∨ B̄1 ∨ · · · ∨ B̄m , m ≥ 0. (3.12)

Suppose, the name of procedure A matches with the procedure call Ai , i,e., some
substitution θ of terms for variables makes Ai and A identical. In such a case, the
resolution derives a new goal statement by disjunction formulas (3.11) and (3.12) as
given below, subject to substitution θ .
3.8 Interpretation and Inferences 73

(Ā1 ∨ · · · ∨ Āi−1 ∨ B̄1 ∨ · · · ∨ B̄m ∨ Āi+1 ∨ · · · ∨ Ān )θ. (3.13)

In general, any derivation can be regarded as a computation, and any refutation

(i.e. derivation of []) can be regarded as a successfully terminating computation. It is
to be noted that, only goal oriented resolution derivations correspond to the standard
notion of computation.
Thus, a goal-oriented derivation, from an initial set of Horn clauses A and from
an initial goal statement (computation) C1 ∈ A, is a sequence of goal statements
C1 , . . . , Cn . So that each Ci contains a single selected procedure call and Ci+1 ,
obtained from Ci by procedure invocation relative to the selected procedure call
in Ci , using a procedure declaration in A.
For the implementation of above, one method is model elimination. Using this, the
selection of procedure calls is governed by the last-in/first-out rule: a goal statement
is treated as a stack of procedure calls. The selected procedure call must be at the
top of the stack. The new procedure calls which by procedure invocation replace the
selected procedure call are inserted at the top of the stack. This would result to a
depth-first search procedure.
The Predicate logic is a nondeterministic programming language. Consequently,
given a single goal statement, several procedure declarations can have a name which
matches the selected procedure call. Each declaration gives rise to a new subgoal
statement. A proof procedure which sequences the generation of derivations in the
search for a refutation behaves as an interpreter for the program incorporated in the
initial set of clauses.
The following example explains how to use procedural interpretation to append
two given lists.
Example 3.13 Appending two lists [3].
Let a term cons(x, y) is interpreted as a list whose first element, the head, is x and
whose tail y is the rest of the list. The constant nil denotes the empty list. The terms
u, x, y, and z are variables. The predicate append(x,y,z) denotes the relationship: z is
obtained by appending y to x.
The following two clauses constitute a program for appending two lists.

append (nil, x, x). (3.14)

append (cons(x, y), z, cons(x, u)) ∨ append (y, z, u). (3.15)

The clause in statement (3.14) represents halt statement. In (3.15) there is a

positive literal for procedure name, and negative literal(s) for the procedure body,
both together it is procedure declaration. The positive literal means, if cons(x, y)
is appended with z, it results to x appended with u such that u is, y appended
with z. The later part is indicated by the complementary (negative) term. Note that
clausal expression (3.15) is logically equivalent to the expression append (y, z, u) →
append (cons(x, y), z, cons(x, u)).
74 3 First Order Predicate Logic

Suppose it is required to compute the result of appending list cons(b, nil) to the
list cons(a, nil). Therefore, the goal statement is,

append (cons(a, nil), cons(b, nil), v), (3.16)

where v (a variable) and a, b (constants), are the “atoms” of the lists. To prove using
resolution, we add the negation of the goal,

append (cons(a, nil), cons(b, nil), v), (3.17)

into the set of clauses. The program is activated by this goal statement to carry out
the append operation. With this goal statement the program is deterministic, because
only one choice is available for matching. The following computation follows with
a goal directed theorem prover as interpreter: The goal statement,

C1 = append (cons(a, nil), cons(b, nil), v). (3.18)

matches with the clause statement (3.15) with matchings: x = a, y = nil, z = cons
(b, nil). Also, v = cons(x, u) = cons(a, u), i.e., there exists a unifier θ1 = {cons
(a, w)/v}. The variable u has been renamed as w. On unifying clauses (3.18) and
(3.15), the next computation C2 is:

C2 = append (nil, cons(b, nil), w)θ1 . (3.19)

Keeping θ1 accompanying the predicate in above is for the purpose that if C2 is

to be unified with some other predicate, the matching of the two shall be subject to
the same unifier θ1 .
As next matching, C2 can be unified with (3.14) using a new unifier θ2 =
{cons(b, nil)/w} to get next computation,

C3 = []θ2 . (3.20)

The result of the computation is value of v in the substitution, i.e.,

v = cons(a, u)
= cons(a, w)
= cons(a, cons(b, nil)).

The above result is equal to goal: append (cons(a, nil), cons(b, nil), v). 

Example 3.14 Theorem proving using resolution-refutation.

Following axioms are about the observed block relationship shown in Fig. 3.3, which
are already in clausal form.
3.8 Interpretation and Inferences 75

Fig. 3.3 Objects on table

on(cylinder, box).

on(box, table).

It is required to be shown that object cylinder is above table, i.e., above(cylinder,

table), given the the following rules:

∀x∀y[on(x, y) → above(x, y)], and

∀x∀y∀z[above(x, y) ∧ above(y, z) → above(x, z)].

After we have gone through the procedure for conversion to clausal form, the
above axioms are transformed into clause forms.

¬on(u, v) ∨ above(u, v).

¬above(x, y) ∨ ¬above(y, z) ∨ above(x, z).
The expression to be proved is “above(cylinder, table)”; its negation is
¬above(cylinder, table). Let us list all the clauses systematically.

(1) ¬on(u, v) ∨ above(u, v).

(2) ¬above(x, y) ∨ ¬above(y, z) ∨ above(x, z).
(3) on(cylinder, box).
(4) on(box, table).
(5) ¬above(cylinder, table).

Now, we manually run the Algorithm 3.1 on the clauses (1)–(5), as well as those
which would created new, to unify them according to unification Algorithm 3.2,
until we reach to a null resolvent.
First we resolve clauses (2) and (5) and bind x to ‘cylinder’and z to ‘table’.
Applying the resolution, we get resolvent (6). Unifier for this is {cylinder/x, table/z}.
76 3 First Order Predicate Logic

(2) ¬above(cylinder, y) ∨ ¬above(y, table) ∨ above(cylinder, table).

(5) ¬above(cylinder, table).
(6) ¬above(cylinder, y) ∨ ¬above(y, table).

Next, resolve clauses (1) with (6), binding u with y and v with ‘table’, we get (7).
Unifier for this is {y/u, table/v}.

(1) ¬on(y, table) ∨ above(y, table).

(6) ¬above(cylinder, y) ∨ ¬above(y, table).
(7) ¬on(y, table) ∨ ¬above(cylinder, y).

We use (1) again with (7) with u bound to cylinder and v to y. Unifier for this is
{cylinder/u, y/v}. On resolving we get (8).

(1) ¬on(cylinder, y) ∨ above(cylinder, y).

(7) ¬on(y, table) ∨ ¬above(cylinder, y).
(8) ¬on(cylinder, y) ∨ ¬on(y, table).

Next, use clause (3) and (8), binding y to box, with unifier {box/y}. We get (8) as
resolvent.

(3) on(cylinder, box).

(8) ¬on(cylinder, box) ∨ ¬on(box, table).
(9) ¬on(box, table).
Finally, the clauses (4) and (9) are resolved to get ‘[]’:

(4) on(box, table).

(9) ¬on(box, table).
(10) [].

Since we have arrived at the contradiction, it shows that negation of the theorem:
¬above(cylinder, table) must be False. Hence the theorem above(cylinder, table)
must be True. 

3.9 Most General Unifiers

The simple approach to avoid needless search in a first-order derivation is to keep

the search procedure as general as possible. Consider, for example the following two
clauses, each as a literal only.

c1 = p(g(x), f (x), z),

and
c2 = ¬p(y, f (w), a).
3.9 Most General Unifiers 77

They are unified by the substitution θ1 ,

θ1 = {b/x, g(b)/y, a/z, b/w},

and also by θ2 ,

θ2 = {f (z)/x, g(f (z))/y, a/z, f (z)/w}.

Note that a constant, or variable, or a function substitutes for a variable, and not
the other way.
We may very well be able to derive the empty clause using c1 , c2 with substitution
of θ1 , followed with application of resolution. But if we cannot, we will need to
consider other substitutions like θ2 .
The trouble is that both of these substitutions are overly specific. We can see that
any unifier must give w the same value as x, and to y the same as g(x), but we do not
need to commit yet to a value for x. The substitution,

θ3 = {g(x)/y, a/z, x/w}

unifies the two literals without making an arbitrary choice that might preclude a path
to the empty clause. The θ3 is a most general unifier (mgu).
More precisely, an mgu θ of literals ρ1 and ρ2 is a unifier that has the property
 
that for any other unifier θ , there is a further substitution θ ∗ such that θ = θ θ ∗ .
So starting with θ , you can always get to any other unifier by applying additional
substitutions. For example, given θ3 , we can get to θ1 by further applying λ = {b/x}
so that θ1 = θ3 λ. And, we can get to θ2 by μ = {f (z)/x} so that θ2 = θ3 μ. Note that
an mgu need not be unique. For example, θ4 = {g(w)/y, a/z, w/x} is also an mgu
for c1 and c2 .
The key fact about mgus is that we can limits the resolution rule to mgus without
loss of completeness. This helps immensely in the search since it dramatically reduces
the number of resolvents that can be inferred from these two input clauses.

Example 3.15 Given a unifier, obtain a more general unifier.

Suppose you have two expressions p(x) an p(y). One way to unify these is to substitute
any constant expression for x and y: S = {fred /x, fred /y}. But this is not the most
general unifier, because if we substitute any variable for x and y, we get a more
general unifier: G = {z/x, z/y}. The first unifier is a valid unifier, but it would lessen
the generality of inferences that we might want to make.

Let E = {p(x), p(y)},

S = {fred /x, fred /y},
G = {z/x, z/y}.

Now let S = {fred /z}
78 3 First Order Predicate Logic

Then ES = {p(fred ), p(fred )}


and GS = {fred /x, fred /y}

and therefore EGS = {p(fred ), p(fred )} = ES.

So, given a unifier, you can always create a more general unifier. When both of
these unifiers are composed and instantiate the original expression E, you get the
same instance as it was obtained with the earlier unifier.

3.9.1 Lifting

It is necessary to show that the general resolution principle is sound and complete.
However, a technical difficulty is the completeness of the proof. Using the Herbrand’s
theorem and semantic trees, we can prove that there is a ground resolution refutation
of an unsatisfiable set of clauses. But, this cannot be generalized as a proof for
general resolution, because the concept of semantic trees cannot be generalized.
Why it cannot be generalized, is due to the variables, which give rise to potentially
infinite number of elements in the Herbrand’s base, as we will show it shortly.
Fortunately, there is a technique, called, “Lifting”, to prove completeness of a
theorem. Following are the steps for lifting:

1. first prove the completeness of the system for a set of ground classes, then,
2. as a second step, lift the proof to non-ground case.

Example 3.16 Infinite inferences.

Let us assume that there are two non-ground clauses: 1. p(u, a) ∨ q1 (u) and, 2.
¬p(v, w) ∨ q2 (v, w). If the signature pattern contains function symbols, then these
clauses have infinite set of instances, as follows:

{p(r, a) ∨ q1 (r) | r is ground}.

{¬p(s, t) ∨ q2 (s, t) | s, t are ground}.

We can resolve above instances if and only if r = s and t = a. Then we can

apply the resolution refutation and obtain the inference given in the denominator of
Eq. (3.21), which are infinite, due to variable s.
p(s, a) ∨ q1 (s), ¬p(s, a) ∨ q2 (s, a)
(3.21)
q1 (s) ∨ q2 (s, a)


The above difficulty can be overcome by taking a ground resolution refutation
and “lifting” it to a more abstract general form.
The lifting is an idea to represent infinite number of ground inferences of the form
given in Eq. (3.21) by a single non-ground inferences:
3.9 Most General Unifiers 79

p(u, a) ∨ q1 (u), ¬p(v, w) ∨ q2 (v, w)

q1 (v) ∨ q2 (v, a)

This lifting can be done using most general unifier, we will be discussing shortly.
Example 3.17 Find out the Lifting for following clauses:
C1 = p(u) ∨ p(f (v)) ∨ p(f (w)) ∨ q(u)
C2 = ¬p(f (x)) ∨ ¬p(z) ∨ r(x)
Using the substitution θ = {f (a)/u, a/v, a/w, a/x, f (a)/z}, the above clauses
become C1 = p(f (a)) ∨ q(f (a)), and C2 = ¬p(f (a)) ∨ r(a). Using C1 and C2 , it
resolves to C  = q(f (a)) ∨ r(a). The lifting claims that there is a clause C =
q(f (x)) ∨ r(x) which is resolvent for clauses C1 and C2 , such that clause C  is ground
instance of C. This can be realized using the unification algorithm to obtain a most
general unifier (mgu) of clauses C1 and C2 , the latter two clauses resolves to C, as
{f (x)/u, x/v, x/w, f (x)/z}.

3.9.2 Unification Algorithm

A unification algorithm is central to most of the theorem-proving systems. This

algorithm receives as input a pair of expressions, and returns as output a set of
substitutions (assignments) that make the two expressions look identical.
The unification algorithm recursively compares the structures of the clauses to be
matched, working across element by element. The criteria is that,
1. the matching individuals, functions, and predicates must have the same names,
2. the matching functions and predicates must have the same number of arguments,
and
3. all bindings of variables to values must be consistent throughout the whole match.
To unify two atomic formulas in an expression A, we need to understand the
disagreement set.
Definition 3.8 Disagreement Set.
If A is any set of well-formed expressions, we call the set D the disagreement set
of A, whenever D is the set of all well-formed subexpressions of the well-formed
expressions in A, which begin at the first symbol position at which not all well-formed
expressions in A have the same symbol. 
Example 3.18 Find out the disagreement set for given set of atoms.
Let the string is, A = {p(x, h(x, y), y), p(x, k(y), y), p(x, a, b)}, having three predi-
cate expressions. The disagreement set for A is,

D = {h(x, y), k(y), a}. (3.22)

80 3 First Order Predicate Logic

Once the disagreement is resolved through unification for this this symbol position,
there is no disagreement at this position. The process is repeated for the new first
symbol position at which all wffs in A do not have same symbol, and so on, until A
becomes a singleton.
Evidently, if A is nonempty and is not a singleton (a set with exactly one element),
then the disagreement set of A is not a singleton and nonempty. Also, if θ unifies A,
and A is not singleton, the θ unifies the disagreement set A. 
For A to be a finite nonempty set of well-formed expressions for which the sub-
stitution Algorithm 3.2 terminates with “return σA ”, the substitution σA available as
output of the unification algorithm is called the most general unifier (mgu) of A, and
A is said to be most generally unifiable [8, 9].

Algorithm 3.2 Unification-Algorithm (Input: A, Output: σA )

1: Set σ0 = ε, k = 0
2: while true do
3: if Aσk is a singleton then
4: Set σA = σk
5: terminate
6: end if
7: Let Uk be the earliest and Vk be the next earliest element in the disagreement set Dk of Aσk
(see Eq. 3.22)
8: if Vk is a variable, and does not occur in Uk then
9: set σk+1 = σk {Uk /Vk },
10: k =k +1
11: else
12: (A is not unifiable)
13: exit.
14: end if
15: end while

Through manually running the Algorithm 3.2 for the disagreement set in (3.22),
stepwise computation for σk is as follows:
For k = 0, and σ0 = ε,

σk+1 = σk {k(y)/h(x, y)}

⇒ σ1 = {k(y)/h(x, y)}.

which, in the next iteration becomes,

σ2 = σ1 {a/k(y)}
= {k(y)/h(x, y)}{a/k(y)}.
3.9 Most General Unifiers 81

The same process is repeated for the disagreement set of 3rd argument in A, which
results to substitution set as {b/y}.

σ3 = σ2 {b/y}
= {k(y)/k(x, y)}{a/k(y)}{b/y}.

On substituting these, we have,

A = {p(x, a, b), p(x, a, b), p(x, a, b}.

which is a singleton, and σ3 is mgu.

For obtaining the unifier σk , the necessary relation required between Uk and Vk
is, Vk has to be a variable, and Uk can be a constant, variable, function, or predicate.
Vk may even be a predicate or function with variable.
The Algorithm 3.2 always terminates for finite nonempty set of well-formed
expressions, otherwise it would generate an infinite sequence of A, Aσ1 , Aσ2 , . . . ,
each of which is a finite nonempty sets of well-formed expressions, with the property
that each successive set contains one less variable than its predecessor. However, this
is impossible because A contains only finitely many distinct variables.
The Algorithm 3.2 runs in O(n2 ) time on the length of the terms, and an even
better. However, there exists more complex, but linear time algorithms for same.
Because, most general unifiers (mgus) greatly reduce the search, and can be calculated
efficiently, almost all Resolution-based systems implementations are based on the
concept of mgus.

3.10 Unfounded Sets

In the well-founded semantics, the unfounded sets provide the basis for negative
conclusions. Let there is a program P (set of rules and facts in FOPL), its associated
Herbrand base is H , and suppose its partial interpretation is I . Then, some A ⊆ H
is called an unfounded set of P with respect to the interpretation I , with following
condition: for each instantiated rule R ∈ P, at least one of the following holds: (In
the rules P , we assume that p is a head, and qi are the corresponding subgoals.)

1. Some positive / negative subgoal qi of the body of the rule is false in the interpre-
tation I ,
2. Some positive subgoals qi of the body occurs in the unfounded set A.

For rule R with respect to I , a literal that makes conditions 1 or 2 above true is
called witness of unusability.
Intuitively, the interpretation I is intended model of P. The rules that satisfy
condition 1 cannot be used for further derivations because their hypotheses are already
known to be false.
82 3 First Order Predicate Logic

The condition 2 in above, called unfoundedness condition, states that all the rules
which might still be usable to derive something in A, should have an atom (i.e., a
fact) in A as true. In other words, there is no single atom in A, that can be established
to be true by the rules of P (as per the knowledge of interpretation I ). Therefore, if
we infer that some or all atoms in A are false, there is no way available later, using
that we could infer that an atom is true [4].
Hence, the well-founded semantics uses conditions 1 and 2 to draw negative
conclusions, and simultaneously infers all atoms in A to be false. The following
example demonstrates the construction of unfounded set from the set of rules and
facts.
Example 3.19 Unfounded set.
Assume that we have a program in predicate logic with instantiated atoms.

p(c).
q(a) ← p(d ).
p(a) ← p(c), ¬ p(b).
q(b) ← q(a).
p(b) ← ¬ p(a).
p(d ) ← q(a), ¬ q(b).
p(d ) ← q(b), ¬ q(c).
p(e) ← ¬ p(d ).

From above rules, we see that A = {p(d ), q(a), q(b), q(c)} is an unfounded set
with respect to φ (null set). Since A is unfounded, its subsets are also unfounded. The
component, {q(c)} is unfounded due to condition (1), because there is no rule available
to establish its truth. The set {p(d ), q(a), q(b)} is unfounded due to condition (2) (their
subgoals or body appear in unfounded set.
There is no way available to establish p(d ) without first establishing q(a) or q(b).
In other words, whether we can establish ¬q(b) to support the first rule for p(d ) is
irrelevant as far as determination of unfoundedness is the concern.
Interestingly, there is no way available to establish q(a) in the absence of first
establishing p(d ), and also there is no way available to establish q(b) without first
establishing q(a). Further, q(c) can never be proven. We note that among p(d ), q(a),
and q(b) as goals, none can be proved without the other two or their negation as
subgoals.
The pair p(a), p(b), even though they depend on each other, but does not form
an unfounded set due to the reason that the only dependence is through negation.
Hence, it can be concluded that the any attempt for proof of p(a) and p(b) will fail,
but this claim is faulty.
The difference between sets {p(d ), q(a), q(b)} and {p(a), p(b)} is as follows:
declaring any of p(d ), q(a), or q(b) false (unfounded), does not create a proof that
any other element of the set is true.
3.10 Unfounded Sets 83

Finally, consider the set {p(a), p(b)}: If any of the elements p(a) or p(b) is taken
false, it becomes possible to prove that the other is true. And, if both are declared
false together, there is an inconsistency. 

3.11 Summary

First-order logic is best suited as a basic theoretical instrument of a computer theorem

proving program. From the theoretical point of view, an inference principle need
only be sound (i.e., allow only logical consequences of premises to be deduced from
them) and effective (i.e., it must be algorithmically decidable whether an alleged
application of the inference principle is indeed an application of it). The resolution
principle satisfies both.
Two types of semantics, namely, operational and fixpoint, have been defined
for programing languages. The operational semantics defines input-output relation
computed by a program in terms of the individual operations performed by the
program inside the machine. Meaning of a program is nothing but the input-output
relation obtained due to executing it in a machine.
A machine independent alternative to semantics, called fixpoint semantics, defines
the meaning of a program as input-output relation which is the minimal fixpoint of
a transformation associated with the program.
A FOPL statement is made of predicates, arguments (constants or variables),
functions, operators, and quantifiers. Interpretation is process of assignment of truth
values (True/False) to subexpressions and atomic expressions, and computing the
resultant value of any expression/ statement.
It is easy to procedurally interpret the sets of clauses which contain at most one
positive literal per clause. However, along with this any number of negative literals
can also exist. Such sets of clauses are called Horn sentences or Horn Clauses or
simply clauses.
The Predicate logic is a nondeterministic programming language. Consequently,
given a single goal statement, several procedure declarations can have a name which
matches the selected procedure call. Each declaration gives rise to a new subgoal
statement.
A proof procedure which sequences the generation of derivations in the search for
a refutation behaves as an interpreter for the program incorporated in the initial set
of clauses. Defining an operational semantics for a programming language means
to define an implementation independent interpreter for it. For predicate logic, the
proof procedure behaves as such an interpreter.
The Herbrand universe is the set of ground terms which use the function symbols
and constants that appear in the predicate logic program. The Herbrand base is
defined as the set of atomic formulas formed by predicate symbols in the program,
whose arguments are in the Herbrand universe.
Herbrands theorem is a fundamental theorem of mathematical logic, which allows
a certain type of reduction of first-order logic to propositional logic.
84 3 First Order Predicate Logic

A substitution component is any expression of the form T /V , where V is any

variable and T is any term different from V , is called unifier. The V is called variable
of component T /V , and T is called term of the component. A most general unifier
(mgu) (i. e., simplest one) θ of literals ρ1 and ρ2 is a unifier that has the property that
 
for any other unifier θ , there is a further substitution θ ∗ such that θ = θ θ ∗ .
The backbone of most theorem-proving systems is a unification algorithm. This
algorithm returns a set of substitutions for a pair of input expressions. These substitu-
tions may be assignments to variables or expressions, which make the two expressions
(or variables or functions) identical or equivalent. To prove a theorem, one obvious
strategy is to search forward from the axioms, using sound rules of inference. We
try to prove a theorem by refutation. It requires to show that negation of a theorem
cannot be True.

Exercises

1. Apply the Resolution theorem to prove:

“Socrates is mortal”, given that
All men are mortal, and
Socrates is man.
2. What are the other methods for automated theorem proving? Explain any three
in brief.
3. Convert the following into clause form:

∀x[p(x) ∧ q(x)] ⇒ [R(x, I ) ∧ ∃y(∃z r(y, z)

⇒ S(x, y))] ∨ ∀x T (x).

4. Show that a formula in CNF is valid if and only if each of its disjunctions contains
a pair of complementary literals P and ¬P.
5. Prove or disprove the followings:
a. If S is a first-order formula, then S is valid iff S → ⊥ is contradiction.
b. If S is a first-order formula and x is a variable, then S is contradiction iff
∃xS is a contradiction.
6. Using the resolution principle prove the validity of following formula:

∀x∃y(p(f (f (x)), y) ∧ ∀z(p(f (x), z)

→ p(x, g(x, z)))) → ∀x∀y p(x, y).

7. Is the predicate logic deterministic or nondetermnistic programming language?

justify for yes / no.
Exercises 85

8. Consider a set of statements of FOPL that uses two 1-place predicates: Large
and Small. The set of object constants are a, b. Find out all possible models for
this program. For each of the following sentences find out the models in which
each of the sentence becomes true.
a. ∀x Large(x).
b. ∀x ¬Large(x).
c. ∃x Large(x).
d. ∃x ¬Large(x).
e. Large(a) ∧ Large(b).
f. Large(a) ∨ Large(b).
g. ∀x [Large(x) ∧ Small(x)].
h. ∀x [Large(x) ∨ Small(x)].
i. ∀x [Large(x) ⇒ ¬Small(x)].
9. Find out the clauses for the following FOPL formulas.
a. ∃x∀y∃z(P(x) ⇒ (Q(y) ⇒ R(z))).
b. ∀x∀y((P(x) ∧ Q(y)) ⇒ ∃zR(x, y, z)).
10. Define the required predicates and represent the following sentences in FOPL.
a. Some students opted Sanskrit in fall 2015.
b. Every student who opts Sanskrit passes it.
c. Only one student opted Tamil in fall 2015.
d. The best score in Sanskrit is always higher than the best score in Tamil.
e. There is a barber in a village who shaves every one in the village who does
not shave himself / herself.
f. A person born in country X , each of whose parents is a citizen of X or a
resident of X , is also a resident of X .
11. Determine whether the expression p and q unify with each other in each of the
following cases. If so, give the mgu, if not justify it. The lowercase letters are
variables, and upper are predicate, functions, and literals.
a. p = f (x1 , g(x2 , x3 ), x2 , b); q = f (g(h(a, x5 ), x2 ), x1 , h(a, x4 ), x4 ).
b. p = f (x, f (u, x)); q = f (f (y, a), f (z, f (b, z))).
c. p = f (g(v), h(u, v)); q = f (w, j(x, y)).
12. What can be the strategies for combination of clauses in resolution proof? For
example, if there are N clauses, in how many ways they can be combined?
13. Why resolution based inference is more efficient compared modus-ponens?
14. Let Γ is knowledge base and α is inference from Γ . Give a comparison among
the following inferences, in terms of their performances:
a. Proof by Resolution, i.e., Γ  α,
b. Proof by Modus poenes, i.e., Γ  α,
c. Proof by Resolution Refutation, i.e., Γ ∪ {¬α}  φ.
86 3 First Order Predicate Logic

15. Given n number of clauses, draw a resolution proof tree to demonstrate combin-
ing them. Suggest any two strategies.
16. Given the knowledge base in clausal form, is it possible to extract answers from
that making use of resolution principle? For example, finding an answer like,
“Where is Tajmahal located?”
17. Represent the following set of statements in predicate logic, convert them to
clause from, then apply the resolution proof to answer the question : Did Ranjana
kill Lekhi?
“Rajan owns a pat. Every pat owner is an animal lover. No animal lover ever
kills an animal. Either Rajan or Ranjana killed a pat, called Lekhi.”
18. Explain:
a. Unification
b. Skolemization
c. Resolution principle versus resolution theorem proving.
19. Use resolution to show that the following set of clauses is unsatisfiable.

{p(a, z), ¬p(f (f (a)), a), ¬p(x, g(y)) ∨ p(f (x), y)}.

20. Derive ⊥ from the following set of clauses using the resolution principle.

{p(a) ∨ p(b), ¬p(a) ∨ p(b), p(a) ∨ ¬p(b), ¬p(a) ∨ ¬p(b)}.

21. Give resolution proofs for the inconsistency ∀x shaves(Barber, x) →

¬shaves(x, x), where Barber is a constant.
22. Consider ab locks-world described by facts and rules:

Facts:
ontable(a), ontable(c), on(d , c), on(b, a), heavy(b),
cleartop(e), cleartop(d ), heavy(d ), wooden(b), on(e, b).
Rules:
All blocks with clear top are black.
All wooden blocks are black.
Every heavy and wooden block is big.
Every big and black block is on a green block.
Making use of resolution theorem find out the block that is on the green block.
23. Given the following knowledge base:
If x is on top of y then y supports x.
If x is above y and they are touching each other then x is on top of y.
A phone is above a book.
A phone is touching a book.
Exercises 87

Translate the above knowledge base into clause form, and use resolution to show
that the predicate “supports(book, phone)” is true.
24. How resolution can be used to show that a sentence is:
a. Valid?
b. Unsatisfiable?
25. “The application of resolution principle for theorem proving is a non-
deterministic approach.” justify this statement.
26. a. Use Herbrand’s method to show that formula,

∀x shaves(barber, x) → ¬shaves(x, x)

is unsatisfiable?
b. What is Herband’s universe for S = {P(a), ¬P(f (x)) ∨ P(g(x))}?

27. Prove that ∀x¬p(x) and ¬∃x p(x) are equivalent statements.
28. Let S and T be unification problems. Also, let σ be a most general unifier for S
and θ be a most general unifier for σ (T ). Show that θ σ is a most general unifier
for S ∪ T .
29. Write the axioms describing predicates: grandchild, grandfather, grandmother,
soninlaw, fatherinlaw, brother, daughter, aunt, uncle, brotherinlaw, and first-
cousin.
30. For each pair of atomic sentences in the following, find out the most general
unifier.
a. knows(father(y), y) and knows(x, x).
b. {f (x, g(x)) = y, h(y) = h(v), v = f (g(z), w)}.
c. p(a, b, b) and p(x, y, z).
d. q(y, g(a, b)) and q(g(x, x), y).
e. older(father(y), y) and older(father(x), ram).
31. Explain what is wrong with the below given definition of set membership pred-
icate ∈:

∀x, s : x ∈ {x | s}
∀x, s : x ∈ s ⇒ ∀y : x ∈ {y | s}.

32. Consider the following riddle: “Brothers and sisters have I none, but that man’s
father is my father’s son”. Use the rules of kinship relations to show who that
man is?
33. Let the following be a set of facts and rules:
Rita, Sat, Bill, and Eden are the only members of a club.
Rita is married to Sat.
Bill is Eden’s brother.
Spouse of every married person in the club is also in the club.
88 3 First Order Predicate Logic

a. Represent the above facts and rules using predicate logic.

b. Show that they do not conclude “Eden is not married.”
c. Add some some more facts, and show that now the augmented set conclude
that Eden is not married.

References

1. Chowdhary KR (2015) Fundamentals of discrete mathematical structures, 3rd edn. EEE, PHI
India
2. Davis M, Putnam H (1960) A computing procedure for quantification theory. J ACM 7(3):201–
215. https://doi.org/10.1145/321033.321034
3. Emden V, Kowalki RA (1976) The semantics of predicate logic as a programming language. J
ACM 23(4):733–742
4. Av Gelder et al (1991) The well-founded semantics for general logic programs. J ACM
38(3):620–650
5. Luckham D, Nilsson NJ (1971) Extracting information from resolution trees. Artif Intell 2:27–54
6. Nilsson NJ (1980) Principles of artificial intelligence, 3rd edn. Narosa, New Delhi
7. Robinson JA (1963) Theorem-proving on the computer. J ACM 10(2):163–174
8. Robinson JA (1965) A machine-oriented logic, based on the resolution principle. J ACM
12(1):23–41
9. Stickel ME (1981) A unification algorithm for associative-commutative functions. J ACM
28(3):423–434
Chapter 4
Rule Based Reasoning

Abstract The popularity of rules-based systems (RBSs) is due to their naturalness.

This chapter presents the potential applications of RBSs, the working of RBS, forward
and backward chaining RBSs, their Algorithms, and inferencing using these systems.
The analysis of complexity of preconditions, and efficiency of rule selection are
introduced to sufficient depth, as well the cofmparison between the two types of RBSs
are presented. A typical RBS, and other methods—model-based and case-based
approaches are also discussed. In addition, number of solved, as well exhaustive list
of exercises are provided at the end of the chapter for practice. The chapter concludes
with its summary.

Keywords Rule-based systems (RBSs) · Forward chaining · Backward chaining ·

Forward chaining Algorithm · Backward chaining Algorithm · Model-based
reasoning · Case-based reasoning · Conflict resolution

4.1 Introduction

Symbolic rules are popular for knowledge representation and reasoning. Their pop-
ularity stems mainly from their naturalness, which facilitates comprehension of the
represented knowledge. The basic form of a rule is,

i f <conditions> then <conclusion>

where <conditions> represent the conditions or premises of a rule, and the

<conclusion> represent its conclusion or consequence. The conditions of a rule
are connected between each other with logical connectives such as AND/OR thus
forming a logical function. When sufficient conditions of a rule are satisfied, the
conclusion is derived and the rule is said to fire (or trigger). Rules represent general
knowledge regarding a domain.

© Springer Nature India Private Limited 2020 89

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_4
90 4 Rule Based Reasoning

Table 4.1 Application areas of rule based systems

Problem System functions
Troubleshooting Analyzing situations, suggesting measures,
and prescribing preventative actions
Process control Identifying problematic data and remedies
of inequalities
Quality assurance Assessment of tasks, proposing the practices,
and enforcing the requirements
Equipment maintenance Diagnosing various faults and recommending
the repairs
Component selection Specifying requirements and matching parts from
an electronics catalog
Computer operation Analyzing requirements to be fulfilled, selecting
and operating a software
Product configuration Specifying preferences, and identifying parts
that satisfy constraints

In this chapter we will discuss the use of easily stated if-then rule to solve prob-
lems. In particular you will learn the forward-chaining from the assertions and back-
ward chaining from hypotheses. The illustrative examples will explain the rule firing
sequences and graph search. The complexities of Algorithms for both forward and
backward rule chaining are also analyzed and presented. In addition, the complexi-
ties of preconditions of rules, and other popular types of expert systems, like—Case
based reasoning, and model based reasoning are presented.
Rule-based systems (RBSs) constitute the one of the most common and simple to
implement means for codifying the problem-solving know-how of human experts.
Experts tend to express most of their problem-solving techniques in terms of a set of
situation-action rules, and this suggests that RBSs should be the method of choice
for building knowledge-intensive expert systems. The RBSs share following key
properties [3]:
1. They incorporate practical human knowledge in conditional if-then rules,
2. Their skill increases at a rate proportional to the enlargement of their knowledge
bases,
3. By selecting the relevant rules and combining them appropriately, it is possible
to solve a large range of problems, of varying complexities.
4. They can be designed to adaptively decide the best sequence of rules to execute,
and
5. The RBS can retrace the reasoning steps and explain the justification for the
conclusions drawn. In this retracing, it can show each rule executed, using natural
language.
Table 4.1 lists the major application areas addressed by RBS.
4.1 Introduction 91

Learning Outcomes of this Chapter:

1. Explain the difference between rule-based, case-based and model-based reasoning
techniques. [Familiarity]
2. Inferences in rule-based systems. [Usage]
3. Complexities of rule-based systems. [Assessment]

4.2 An Overview of RBS

Roughly speaking, a Rule Based System comprises a knowledge base and an infer-
ence engine (see Fig. 4.1). The knowledge base is collection of facts and rules. Facts
are simple statements with constant or variable object lists, while the Rules always
express a conditional sentence, with one or more premises and a consequent com-
ponent. The rules needs to be interpreted, which means, if a premise can be satisfied
the consequent also be satisfied. The consequent can be a conclusion or an action.
When it is an action, the effect of satisfying the premises is to schedule that action
for execution, and when the consequent defines a conclusion, the effect is to infer
the conclusion [2].
Several key techniques for organizing RBSs have emerged. Rules can be used to
express deductive knowledge, such as logical relationships, and thereby to support
inference, verification, or evaluation tasks. Conversely, rules can be used to express
goal-oriented knowledge that an RBS can apply in seeking problem solutions and
cite in justifying its own goal-seeking behavior. The rules can also be used to express
causal relationships, which an RBS can use to answer “what-if” questions, or to
determine possible causes for specified events.

Fig. 4.1 Inference cycle of a

Kowledge Working
forward-chaining RBS base memory

Match

Conflict set

Inference Engine
Select

Execute
92 4 Rule Based Reasoning

In a rule based system, each if pattern may match to one or more of assertions in
a collections of assertions. The collections of assertions is called working-memory
(Fig. 4.1). The assertions collectively may match to premises of one or more rules in
the knowledge base. This is done by “match” block of the inference-engine. All the
rules matching with the assertions in the working memory are put in the conflict set,
from where one of the rule is “selected” based on some conflict resolving criteria set
for, and then the selected rule is executed. The resulting consequences/conclusions
(the then patterns) are put in the working memory to form new assertions, and the
process continues till desired result (goal) is not reached.
A system like this is called deduction system, as it deduces new inferences from
the rules and assertions. A realistic example of a rule is:

R1: if 1. stiff neck,

2. high temperature, and
3. impairment of conciousness occur togetehr,
then
meningitis is suspected.

In the above, meningitis is a disease related to“Inflation of membrane of spinal

chord and brain” will be suspected by this rule if the “if” patterns 1, 2, 3 are found true.
RBS can be applied in judiciary systems also. Following is an example of a more
complex a rule.

If the plaintiff received an eye injury

and it was one eye injured
and treatment for eye required surgery
and recovery from the injury was almost complete
and visual acuity was slightly reduced by the injury
and there is fixed condition,
then increase the injury trauma factor by $5,000.

Facts is the kind of data in a knowledge base that express assertions about prop-
erties, relations, propositions, etc. In contrast to rules, which the RBS interprets
as imperatives, facts are usually static and inactive implicitly. Also, a fact is silent
regarding the pragmatic value and dynamic utilization of its knowledge. Although
in many contexts facts and rules are logically interchangeable, in the RBSs they are
quite distinct.
In addition to its static memory for facts and rules, an RBS uses a working-
memory to store temporary assertions. These assertions record earlier rule-based
inferences. We can describe the contents of working memory as problem-solving
state information. Ordinarily, the data in working memory adhere to the syntactic
conventions of facts. Temporary assertions thus correspond to dynamic facts.
The basic function of an RBS is to produce results. The primary output may be—
a problem solution, an answer to a question, or result of an analysis of some data.
Whatever the case, an RBS employs several key processing determining its overall
4.2 An Overview of RBS 93

activity. A world manager maintains information in working memory, and a built-in

control procedure defines the basic high-level loop; if the built-in control provides
for programmable specialized control, an additional process manages branches to
and returns from special control blocks.
Some times, the patterns specify the actions rather than assertions, e.g., “to put
them on the table”. In such case the rule based system is called reaction system.
In both the deduction systems and reaction systems, forward chaining is the pro-
cess of moving from the if patterns to then patterns, where if patterns identifies the
appropriate situation for deductions of new assertion, and performance of an action
in the case of reaction system.

4.3 Forward Chaining

In a forward chaining system, we start with the initial facts, and use the rules to
draw new conclusions (or take certain actions), given those facts. Forward chaining
systems are primarily data-driven. Whenever an if pattern is observed to match an
assertion, the antecedent is satisfied. When all the if patterns of a rule are satisfied,
the rule is triggered. When a triggered rule establishes a new assertion or performs
an action, it is fired. This procedure is repeated until no more rules are applicable
(Fig. 4.1).
The selection process is carried out in two steps:
1. Pre-selection: Determining the set of all the matching rules, also called the
conflict-set.
2. Selection: Selection of a rule from the conflict set by means of a conflict resolving
strategy.

4.3.1 Forward Chaining Algorithm

A simple forward chaining Algorithm 4.1 starts from the known facts in the knowl-
edge base, and triggers all the rules whose premises are the known facts, then adds
the consequent of each into the knowledge base. This process is repeated until the
query is answered or until there is no conclusion generated to be added into the
knowledge base. We will use symbols θ, λ, γ to represent substitutions. The unify
is a function unifies the newly generated assertion q  and the query α, and returns a
unifying substitution λ if they are unified, else returns null.
Let α be the goal. The forward-chaining Algorithm 4.1 picks up any sentence
s ∈ Γ , where Γ is knowledge base, and checks all possible substitutions θ for s.
Let, the predicate form of s is p1 ∧ p2 ∧ · · · ∧ pn → q. On substituting θ , say, it
results to ( p1 ∧ p2 ∧ . . . pn )θ = ( p1 ∧ p2 ∧ . . . pn ), such that pi ’s ∈ Γ . Let ( p1 ∧
p2 ∧ . . . pn ) → q  , such that q  = qθ . This q  is added into new inference. If the
94 4 Rule Based Reasoning

inference q  and goal α unify, then their unifier λ is returned as the solution, else the
Algorithm continues.

Algorithm 4.1 Forward-chaining(Input: Γ, α) // α is a query, Γ is knowledge base

1: while True do
2: new = {}
3: for each sentence s ∈ Γ do
4: Convert s into the format p1 ∧ p2 ∧ · · · ∧ pn → q
5: for each substitution θ such that ( p1 ∧ p2 ∧ . . . , ∧ pn ) ← ( p1 ∧ p2 ∧ . . . , ∧ pn )θ for
some pi s ∈ Γ do
6: q  ← qθ
7: new ← new ∪ {q  }
8: λ ← uni f y(q  , α)
9: if λ is not null then
10: return λ
11: end if
12: end for
13: Γ ← Γ ∪ new // add new inferences in knowledge base
14: end for
15: end while
16: Return Fail

The Algorithm may fail to terminate in the case when the raised query has no
answer. For example, every natural number can be generated by recursively applying
successor operation on a natural number, and assuming that 0 is natural number. This
will lead to indefinite loop for very large numbers.

naturalnum(0).
∀x[naturalnum(x) → naturalnum(succ(x)).]

The following example demonstrates the manual run of forward chaining Algo-
rithm.

Example 4.1 Produce the inference, given the knowledge base Γ = {man(x) →
mor tal(x), man(socrates)} and query α = mor tal(w), i.e., “Who is mortal?”

We follow the Algorithm 4.1 manually and note that p1 ∧ · · · ∧ pn = p1 = man(x);

p1 = man(socrates), substitution θ = {socrates/x}, and q = mor tal(x). Also,

q  = qθ
= mor tal(x){socrates/x}
= mor tal(socrates).

Also, new = {} ∪ {q  } = mor tal(socrates).

On unification of α (i.e, mor tal(w)), and q  , the unifier λ obtained is,
4.3 Forward Chaining 95

λ = uni f y(q  , α)
= uni f y(mor tal(socrates), mor tal(w))
= {socrates/w}.

Hence, the answer for query is w = socrates. 

Complexity Issues
The complexity of the forward chaining Algorithm is determined by the inner loop in
the Algorithm 4.1. It finds all the possible premises such that the premises unify with
certain set of facts in the knowledge base Γ . The process is called pattern matching,
and tried for every rule, for every substitution θ . Thus, if set of rules are P, there are
n k substitutions for each rule, assuming that in worst case k arguments and n number
of literals’ assignments for each argument in the P. This makes the complexity of
above Algorithm as,
|P|n k (4.1)

which is exponential. However, since the size and arities are constant in the real-
world, the complexity of expression (4.1) is a polynomial in nature. To search the
predicates for matching, they are indexed and a hash function is generated for quick
search.

4.3.2 Conflict Resolution

When we are doing data-directed reasoning, we may not like to fire all of the rules
in case more than one rule is applicable. In cases where we want to eliminate some
applicable rules, some kind of conflict resolution is necessary for arriving at the
most appropriate rule(s) to fire. In a deduction system all rules generally fire, but in
a reaction system, when more than one rule are triggered at the same time, only one
of the possible actions is desired [4].
The conflict resolving strategy is used to avoid superfluous rules. The most obvious
strategy is to choose a rule at random, out of those that are applicable. Following are
some common approaches for deciding the preferences for the rule to fire.
Selection by order
Selection by order may either on the order in which the rule appears in the knowledge
base, or the order may be based on how recent the rule was used.
Order
Pick the first applicable rule in the order they have been presented. This is the type
of strategy used by Prolog, and is one of the most common ones. Production system
programmers would take this strategy into account when formulating rule sets.
96 4 Rule Based Reasoning

Recency
Select an applicable rule based on how recently it has been used. There are different
versions of this strategy, ranging from firing the rule that matches on the most recently
created (or modified) wff, to firing the rule that has been least recently used. The
former could be used to make sure that a problem solver stays focused on what it
was just doing (typical of depth-first search); the latter would ensure that every rule
gets a fair chance to influence the outcome (typical of breadth-first search).
Selection according to syntactic structure of the rule
There are many ways one can consider the syntax, e.g., it may be conditions and their
specificity, or the size of the rule: largest, smallest, or in increasing order of size.
Specificity Criteria
Select the applicable rule whose conditions are most specific. A set of conditions is
taken as more specific than other if the set of wffs that satisfy it is a subset of those
that satisfy the other. For example, consider the following three rules:

i) if bird(x) then add(canfly(x))

ii) if bird(x) ∧ weight(y, x) ∧ gt(y, 5kg) then
add(cannotfly(x))
iii) if bird(x) ∧ penguin(x) then add(cannotfly(x))

Here, the second and third rules are both more specific than the first, because the
condition in (ii) and (iii) are more specific than in (i) (condition of (i) is very general).
If we have a bird that has weight greater than 5 kg, or it is a penguin, then the first
rule applies, because the condition of “bird” is satisfied. However, in this case the
other two rules should take precedence, as they are more specific. Note that if the bird
is a penguin and heavy, neither second nor third is applicable, and another conflict
resolution criteria must be applied to decide between the second and third rules.
Largest Rule First
Apply the syntactically largest rule, i.e., the rule which contains the most proposi-
tions.
Incremental in Size
Solutions to subproblems are constructed incrementally, and are cached to avoid the
recompilation.

Selection by Means of Supplementary Knowledge

The supplementary knowledge about the rules may be in the form of priority of the
rules, or as meta-knowledge.
Highest Priority First
For this purpose each rule must be given a priority, which may be represented, e.g.,
by a numeric value.
4.3 Forward Chaining 97

Apart from the priority criteria, there are additional rules, called meta-rules, con-
trol the selection.
The simple way of implementing the forward-chained interpreter is to check the
preconditions of all the rules, sort the applicable rules in an agenda (i.e., an ordered
list) according to the criteria given by the conflict-resolving strategy, and then perform
the action of the leading rule in the agenda, followed with this previous agenda is
deleted and the process begin as a new.

4.3.3 Efficiency in Rule Selection

Following are the criteria for selection of rules for implementation of reasoning
process.
Similarity Between Rules
The first improvement over the brute-force approach is based on the similarity
between preconditions of different rules. Using this, rules with a common proposi-
tions may be eliminated all together if these propositions do not hold. For achieving
this, the rules are structured in a tangled hierarchy by constructing a network of inter-
connecting trees in which the internal decision nodes are formed by the propositions
and the leaves of the rules. The path from the root of the tree to a rule includes all
the propositions of the precondition of this rule.
Check Applicability of New Rules
This improvement is based on the fact that a complete new calculation of an agenda
requires far more effort than a modification on the basis of changes. From one cycle
to next, the knowledge base is altered only by the action of the selected rule, which
may add or delete portions of knowledge. Thus, it needs to be tested that whether the
deleted knowledge destroy the applicability of any rule in the old agenda and whether
the added knowledge make any new rules applicable. The later test is performed by
running through the rule network with new knowledge as starting point.
Indexing
One way of improving over the brute-force Algorithm is to index all the rules so that
each parameter of the predicate produces a reference to its rule. When a parameter
takes on a new value or changes its value, the reference enables the rules concerned
to be found and checked efficiently.
Consider the following rule, for which many conjunct ordering may be possible,

∀x∀y[ p(x)q(x, y) → r (x, a)]. (4.2)

Here, all the p(x) may be ordered, next find q(x, y) for each x and y. Alterna-
tively, for each x find p(x), then find all q(x, y) for different y. Which approach is
better? Finding this is solution of a conjective-ordering problem, which is NP-hard.
A heuristic can use most constrained, the one having fewest values. This shows
98 4 Rule Based Reasoning

a close relationship between CSP (constraint satisfaction problem), discussed in

next chapters, and a pattern matching problem. Due to these complexity issues the
forward-chaining lgorithm 4.1 is NP-hard in its inner loop.

4.3.4 Complexity of Preconditions

Different formalisms have differing expressive power of describing the preconditions

and the rules. The differences are those between propositional and predicate logic.
The count of number of rules is not a sufficient criteria for the indication of the size of
an expert system. Ideally each relevant situation of the domain should be covered by
just one rule. Description of a situation by different rules will impair the modularity
of the knowledge base since these rules should be meaningful only as group. On the
other hand, more expressive power of the rule formalism require a more complex and
often more inefficient rule interpreter so that one must check how much complexity
is necessary for the domain [4].
The simplest method of evaluating preconditions is lookup in the knowledge
base. For example, in the case of query: “Is the pain intensifying synchronous with
respiration?”, it is only necessary to check whether the value—“synchronous with
respiration”, is stored under the object - “pain intensifying”. With the logical connec-
tives, and, or, not, this situation can be described. However, in the case of numerical
or temporal relationships, the looking alone is not sufficient, and additional, calcula-
tions are required. For example, for the query “Did shortening of breath occur before
throat pain”. here, the preposition “before” has reference to time. Does it mean 1
second or a minute, or a day, because all these satisfy the criteria of before ! There is
need of comparison, using the relations operators, like, less than, greater than, equal,
which are needed to be applied along with the time information. Also, depending on
the situation, magnitude of precede be available in the knowledge base.
One of the drawback of forward chaining is that it makes all allowable inferencing
based on the facts and rules available, irrespective of whether they are relevant to
deriving goals or not. This makes the forward chaining, not a very efficient approach
for reasoning. To avoid this problem, we work for selected subclass of rules. The other
approach, the backward chaining, arrives to only those premises which certainly lead
to goal, hence the system is called goal driven.

4.4 Backward Chaining

While forward chaining allows conclusions to be drawn only from a given knowledge
base, a backward-chained rule interpreter is suitable for requesting still unknown
facts. In a backward chaining system, you start with some hypothesis (goal) you are
trying to prove, and keep looking for rules that would allow you to conclude that
hypothesis, perhaps setting new sub-goals to prove as you go.
4.4 Backward Chaining 99

For this, the goal expression is initially placed in the working memory, followed
with this, the system performs two operations:
1. matches the rule’s “consequent” with the goal. For example, q in p1 ∧ p2 ∧ · · · ∧
pn → q,
2. selects the matched rule p1 ∧ p2 ∧ . . . pn → q and places its premises ( p1 ∧
· · · ∧ pn ) in the working memory.
The second step in above corresponds to the transformation of the problem into
subgoals. The process continues in the next iteration, with these premises becoming
the new goals to match against the consequent of other rules. Thus, the backward
chaining system, in the human sense are hypothesis testing.
Backward chaining uses stack. First, the goal is put in the stack, then all the rules
which results to this goal are searched and put on the stack. These becomes the sub-
goals, and the process goes on till all the facts are proved or are available as literals
(ground clauses). Every time the goal and premises are decided, the unification is
performed.

4.4.1 Backward Chaining Algorithm

The Algorithm for backward chaining returns the set of substitutions (unifier) which
makes the goal true. These are initially empty. The input to the Algorithm is knowl-
edge base Γ , goals α, and current substitution θ (initially empty). The Algorithm
returns the substitution set λ for which the goal is inferred. The Algorithm 4.2 is the
backward-chaining Algorithm.

Algorithm 4.2 Backward-chaining(Input: Γ, α, θ ) // α is a query, Γ is knowledge

base, θ current substitution (initially empty), λ represent substitution set for the query
to be satisfied (initially empty)
1: θ = {}, λ = {}
2: q  ← αθ
3: for each sentence s ∈ Γ , where s = p1 ∧ p2 ∧ . . . , ∧ pn → q and γ ← uni f y(q, q  ) = null
do
4: αnew ← ( p1 ∧ p2 ∧ · · · ∧ pn )
5: θ ← θγ
6: λ ←backward-chaining(Γ, αnew , θ) ∪ λ
7: end for
8: return λ

Example 4.2 Apply the backward-chaining Algorithm for inferencing from a given
knowledge base.

Let Γ = {man(x) → mor tal(x), man(socrates)}. Assume that it is required to

infer answer for “Who is mortal?” That is, goal α = mor tal(w), find w. The loop
iterations in the backward-chaining Algorithm 4.2are as follows:
100 4 Rule Based Reasoning

1st Iteration: Initially θ is empty, hence, q  = αθ = mor tal(w). From Algorithm

and knowledge base Γ , the sentence is, s = (man(x) → mor tal(x)). Next, γ =
uni f y(mor tal(x), mor tal(w)) = {w/x}. Also, the new goal, αnew = man(x). The
new value of current substitution is, θ ← θ γ = {}{w/x} = {w/x}. Next, compute
λ = backwar d-chaining(Γ, man(x), {w/x}) ∪ λ, as a recursive call.
Recursive call: We apply the Algorithm in a recursive mode, and get q  = man(x)
{w/x} = man(w). Next, the subgoal man(w) (i.e., q  ) matches with other subgoal
(it is a fact) man(socrates) ∈ Γ . Hence, γ = uni f y(man(socrates), man(w)) =
{socrates/w}. Next, the new goal is, αnew = man(socrates) and new substitution is:

θ = θγ
= {w/x}{socr taes/w}
= {socrates/x}.

In the next call of recursion at step 5, q  = αθ = man(socrates) {socrates/x},

the substitution fails, as there is no x where “socrates” can be substituted. Hence,
q  = null. Since γ is null, the for loop at step 3 terminates, and returned value of λ
is {socrates/w}, i.e., w = socrates, or mor tal(socr taes). 

4.4.2 Goal Determination

If the goal is not known in the knowledge base, the rule interpreter first decides
whether it can be derived or must be requested from user. For the derivation of goal,
all the rules which includes the goal in their consequent part are processed. These
rules can be identified efficiently if they are indexed according to their consequent
parts.
Backward chaining therefore contains implicitly a dialogue control, where order of
the questions decides the order of the rules for deriving a parameter. This dependency
on the order reduces the modularity of the rule-based system. The efficiency of the
backward-chained rule interpreter is determined by the formulation of the goal: the
more precise the goal, the smaller is the search tree of rules to be checked and
questions to be asked. For example, in medicine, a query may be: “What is name of
disease?” as against an alternate query of “Is X the name of disease?”.
The examples of back-ward chained rule interpreter are MYCIN, and PROLOG.

4.5 Forward Versus Backward Chaining

Whenever the rules are such that typical set of facts relate to large number of con-
clusions, i.e., the fan-out is larger than fan-in, it is better to use backward chaining.
This is to stop in explosive growth in search space. On the other hand, when the rules
4.5 Forward Versus Backward Chaining 101

Table 4.2 Knowledge base S. no. Rule

for Animal-Kingdom
1 sponge(x) → animal(x)
2 ar thopod(x) → animal(x)
3 ver tebrate(x) → animal(x)
4 f ish(x) → ver tebrate(x)
5 mammal(x) → ver tebrate(x)
6 car nivor e(x) → mammal(x)
7 dog(x) → car nivor e(x)
8 cat (x) → car nivor e(x)

are such that a typical hypothesis can lead to many facts, i.e., fan-in is larger than
fan-out, it is better to use forward chaining. The reason being that if we go backward
in a case where fan-in is high, it would lead to exponentially growing search space,
so we prefer forward chaining in such knowledge.
If there is a situation that all the required facts are known, and we want to get all
the possible conclusions from these, the forward chaining is better. However, if the
rules are such that facts are incompletely known, we cannot proceed from the facts
to conclusions, and thus goal driven strategy should be used.
Forward chaining is often preferable in cases when there are many rules with the
same conclusions, because we choose the satisfying premises. In backward chaining
it may lead to non-satisfying premises. The following examples demonstrate this.

Example 4.3 Given a knowledge base for an animal kingdom, infer animal(bruno)
after adding of dog(bruno). The taxonomy of the animal kingdom includes such rules
as shown in Table 4.2.

It is required to show that, K B + dog(br uno) → animal(br uno).

1. Forward Chaining. We start with rule 7, and unify dog(br uno) with dog(x).
Next, we will successively infer and add carnivore(bruno), mammal(bruno), ver-
tebrate(bruno), and animal(bruno). The query will then succeed immediately.
The total work is proportional to the height of the hierarchy of this taxonomy,
which is 4.
2. Backward-chaining. Alternatively, if we use backward chaining, the query
animal(br uno) will unify with the first rule above and generate the sub-query
sponge (br uno), which will initiate a search for br uno through all the subdi-
visions of sponges. Not finding, it tries with 2nd rule, but orthopod is not in
consequent. Next, with rule 3, the goal driven chain is: “animal → vertebrate →
fish”, which fails. The successful invocation rule sequence is “3 → 5 → 6 →
7.” Thus, it searches the entire taxonomy of animals looking for dog(br uno).
We note that work done in the background chaining much more than forward
chaining. However, this is not necessarily true always. 
102 4 Rule Based Reasoning

In some cases, it is desirable to combine both forward and backward reasoning,

and due to the merits of individual rules they are identified as forward / backward.
However, this process results to a far more complex mechanism.

4.6 Typical RB System

The R1 (later called as XCON—for expert configurator) was production rule-based

expert system, which was written in OPS5 language by John P. McDermtt in 1978
to help in ordering DEC’s VAX computer system by automatically selecting the
computer systems components, based on the requirements of customers. When fully
developed it had 2500 rules, which provided over 95% accuracy in configuration of
systems as per customers needs [5].
The XCON differed from other systems mainly in its use of “match” rather than
“generate-and-test” as its central problem solving strategy. It exploits knowledge
of its task domain to generate a single acceptable solution. The input to XCON
is customer’s order and output is a set of diagrams showing spatial relationships
among the components of the order. These diagrams act as guidance to the technician
who assembles the system. For example two inter-dependent activities needs to be
performed for configuring a VAX mini-computer system:
1. It is determined that the customer’s order is complete; if it is not, whatever are
the components in shortfall as per the standard order, must be added.
2. Next, the spatial relationships among all of the components is to be determined
(including those that are added).
The criterion of success whether a configuration is complete or not, cannot be deter-
mined by any simple test, but involves instead particular knowledge about all the
individual components and their relationships. The criterion of successful spatial
arrangement is that it should be geometrically compact, and in addition, it involves
particular knowledge on a component by component basis. Thus, the task accom-
plishment is defined by a large set of constraints comprising a large amount of
knowledge.

4.7 Other Systems of Reasoning

In addition to the methods discussed above, there exists reasoning methods, like,
model based reasoning, case based reasoning, and some hybrid combinations of
these.
4.7 Other Systems of Reasoning 103

4.7.1 Model-Based Systems

Consider a situation of applying the goal driven strategy of trouble-shooting elec-

tronic circuits, like, TV, or electron-microscope. On confirming that the system does
not function (i.e., given the goal), we try to find out systems modules (i.e., sub-goals)
and check which out of these are working. For the unit which is not working, we try
to find out its sub units, and carry on the check like this, till we reach to component
levels, like resisters, capacitors, etc. In addition, we also need to look at the history
of failure rate of the components. This, in fact becomes a very exhaustive work.
Instead, it would be better if we take the physical model of the system, with
mathematical equations describing the interactions and relationships between them.
It would then base the diagnosis on the signal reading at various points in the circuit.
A Model based reasoning comprises following details:
1. A description of each component in the device. These description can simulate
the behavior of the component.
2. A description of the device’s internal structure: A representation of components
and their interconnections, with ability of simulation of interactions.
3. Information about the input output relationships and magnitudes of values at input
and outputs.
To have a glimpse of a model based reasoning, consider a model shown in Fig. 4.2.
The input and output relations are functionally defined, expected and actual values
are shown in the output. Given this, we can determine the faulty component.
This model comprises adders and multiplications units which deliver the output
as per the functions they perform. In the multiplier-2 the actual output is 56, but
expected is 63. Hence the possibility is, either the Mult-2 is faulty or its inputs are
faulty. Since the actual output of Mult-1 and its expected are matching, that confirms
that, Mult-1 and its two inputs are correct. Thus possibility left is, either first input
to Mult-2 is wrongly delivered, or add-2 is malfunctioning, or input D is incorrect.
Assuming that connections are not faulty, then either Mult-2 is faulty, or Add-3 is
faulty, or the input D is faulty, which can be systematically checked to locate the
fault.

Fig. 4.2 Model to be

A =2
troubleshooted Add-1

Expected O/P =35

B=3 Mult-1
Actual O/P = 35

Add-2

Mult-2 Expected O/P = 63

C=4
Actual O/p = 56
Add-3
D=5
104 4 Rule Based Reasoning

4.7.2 Case-Based Reasoning

The case-based reasoning (CBR) uses specific database for problem solving. Con-
sider the case of medical practitioner, who uses medical history of failed as well as
successful “cases” for diagnosis of future patients. If a new patient’s case appears
similar to one of the earlier patient’s successful case treatment case, the knowledge
history of old case is useful. At the same time if some old case has failed, that will
not be tried again due to available history.
Figure 4.3 shows the functional block diagram of case based reasoning. Often, the
researchers use the combination of above cases for reasoning due to their advantages.
The similar is the situation in the case of legal practitioners. Lawyers, are allowed
by law, to quote the old case histories in their legal proceedings, if that case matches
in the present situations. They can also give a different interpretation for the old case
to deal with new situations, even if they do not fully match [1].
The above are the examples of case based reasoning.
Many other cases can be counted in this category: programmers often reuse the
old code to fit in the new situation with partial modification, architects reuse the
design of older successful architectures, the wars winners learn from the history of
wars won, all of us often follow the success stories of others in life, and so on. Many a
times, the case where we need to apply the reasoning is not identical to the previous,
hence some transformations are required to fit into the current situation. Thus, this
reasoning is a a kind of learning, from the analogy, called analogical leaning.
For each new problem, the case based reasoning researchers share the following
common operations:
1. Retrieve appropriate cases from the storage (case database).
2. Modify the retrieved case to fit into the current situations.
3. Apply the transferred rule.
4. Save the solution along with the case details, for success or failure for future use.

Input Elaborate
the problem

Indexing Problem construct

target case
Retrieve
similar case
Case retrieval
Cases Historical cases
data base
Rreuse old
Save
Revise
Confirmed solution: adapted cases
new case

Fig. 4.3 Case-based Reasoning

4.8 Summary 105

4.8 Summary

Some of the applications of rule based expert systems are: component selection,
equipment maintenance, product configuration, computer operation & troubleshoot-
ing, process control, and quality assurance. The basic form of a rule is,

i f <conditions> then <conclusion>

where <conditions> represent the conditions or premises of a rule, and

<conclusion> represent its conclusion or consequence.
A Rule Based System (RBS) consists of a knowledge base and an inference engine.
In addition to its static memory for facts and rules, an RBS uses a working-memory
to store temporary assertions. These assertions record earlier rule-based inferences.
Some times, the patterns specify the actions rather than assertions, e.g., “to put them
on the table”. In such case the rule based system is called reaction system.
The rule based systems are two types: forward chaining (data driven) and back-
ward chaining (goal) driven. In gaol driven system, one starts from the hypothesis
(goal) and tries to prove it by finding the subgoals, then these subgoal becomes the
goals, and so on. In data driven system, we search from the data and search the goal. In
both the cases one needs to perform the rule chaining. While forward chaining allows
conclusions to be drawn only from a given knowledge base, a backward-chained rule
interpreter is suitable for requesting still unknown facts. The other types of reasoning
systems are: Case based system (CBS) and Model based systems (MBS).
When we are doing data-directed reasoning, we may not like to fire all of the rules
in case more than one rule is applicable. For this, some kind of conflict resolution is
necessary for arriving at the most appropriate rule(s) to fire. Some conflict resolving
strategies may be: Order of listing, recency, specificity, syntactic structures, etc.
The criteria for reasoning process can be: similarity between rules, applicability
of new rules, and Indexing.
Whenever the rules are such that typical set of facts relate to large number of
conclusions, i.e., the fan-out is larger than fan-in, it is better to use backward chaining.
On the other hand, when the rules are such that when a typical hypothesis can lead
to many facts, i.e., fan-in is larger than fan-out, it is better to use forward chaining.
The case based reasoning uses specific database for problem solving. A medi-
cal practitioner, who uses medical history of failed as well as successful cases for
diagnosis of future patients. If a new patient’s case appears similar to one of the
earlier patient’s successful case treatment case, the knowledge history of old case is
useful. At the same time if some old case has failed, that will not be tried again due
to available history.
106 4 Rule Based Reasoning

Exercises

1. List the various components of a Production System, and explain each.

2. In the above context, explain the importance of “binding”, “matching” and “con-
flict resolution”.
3. Suppose there is a production system with four initial facts: A, B, C, D and
following three rules:

R1 : i f A then E
R2 : i f B ∧ F then G
R3 : i f C ∧ E then F

a. Using these rules and facts, explain what is meant by “backward chaining”
and show explicitly how it can be used to determine the truth of G?
b. Explain what is meant by “forward chaining”, and show explicitly how it can
be used in this case to determine new facts.
4. Consider the following possibilities, suggest the solution strategy to be adopted
if the system is implemented as a rule based system:
a. A subgoal literal is generated such that the higher goal is a subset of the
subgoal.
b. A subgoal literal is generated whose negation unifies with the higher-goal.
c. A subgoal S literal is generated that is equal to another goal G, and G is
neither higher nor lower to S.
5. Translate the following knowledge base into predicate form.
If x is on top of y then y supports x.
If x is above y and they are touching each other then x is on top of y.
A phone is above a book.
A phone is touching a book.

a. Use forward-chaining to show that the predicate “supports(book, phone)”

is true.
b. Use forward-chaining to show that the predicate “supports(book, phone)”
is true.

Count the number of triggering and rule firing in each of the above cases.
6. Given the propositions A, B, C, D and rules R1 to R4 ,

R1 = i f A ∧ X ∧ Y then Z
R2 = i f B ∧ V then Y
R3 = i f C ∧ V then X
R4 = i f D then C
Exercises 107

a. Use the forward-chaining to determine if Z can be inferred from the above

knowledge base.
b. Use the backward-chaining to determine if goal Z succeeds from the
above knowledge base.
7. What are the conflict resolving strategies in case more than one rule matches with
the assertions? Discuss the merits and demerits of each strategy for selection of
a rule to fire.
8. Briefly justify each of the following for conflict resolution. Also, give examples
in each case.
a. Specificity criteria.
b. Syntactically longest rule first.
c. Ordering rules.
d. Most recent first.
e. In order of increasing size of the rule.
9. List the following formulas in oder of specificity, and construct a tree such that
those having same specificity will stand at the same level in the tree.
a. b∨c
b. b∧c
c. a∨c
d. a∧c
e. a∧b
f. a∧b∧c
g. b
h. c
10. How the assignment of priority to rules can be implemented? Consider the
Table 4.2 to demonstrate the assignment of suitable priorities to rules listed
in this table. Suggest the structure as how the priorities are to be stored and rules
be invoked based on these priorities.
11. Explain what structures can be used and how they will operate, for knowledge
based systems, for the following functionalities?
a. If a precondition of one or more rules is false, the system should exclude
these rules from participation in the inference process.
b. If firing of some rules infer new knowledge, this should be added into the
existing knowledge.
c. To maintain and use index file for rules.
12. For the rule (4.2), if domain size of p is | p| = m, domain size of variable x is n,
and of y is l,
a. Find out the worst-case time complexity of the rule to be selected,
b. In how many ways ordering can be done?
c. Justify that it is NP-hard problem as a general case.
108 4 Rule Based Reasoning

13. Which type of rule chaining you will prefer in the following situations? Also,
identify, what is the goal in each case.
a. To diagnose the case of malaria infection, based on the symptoms of
cough, soar throat, regular fever, shivering.
b. Identify a thief, based on the nature of theft, finger prints, and goods
stolen.

14. ] Suggest an approach to combine the forward and backward chaining for the
knowledge base shown in Table 4.2 (Hint: Try backward chaining to begin with,
if it is not doing efficiently, then switch over to forward chaining.)
15. To prove a theorem of geometry using rule-based systems, represent the follow-
ing statements as production rules:
a. Corresponding sides of two congruent triangles are also congruent.
b. Corresponding angles of two congruent triangles are also congruent.
c. If corresponding sides of two triangles are congruent then the triangles
are congruent.
d. If corresponding sides and the angle covered by them are equal then the
triangles are congruent.
e. Base angles of an isosceles triangle are congruent.
16. In what order the rules should be fired for inference efficiency? Discuss the merits
and demerits of choosing a particular order of rule firing.
17. What are the criteria of selection of premises for firing the rules? Discuss this
based on specificity and generality of the premises.
18. Consider a network of applicability of rules. Imagine that due to firing of a
rule, a new inference is generated. This inference may contradict some fact or
goal. Explain, what should be the structure of dependency network so that any
inconsistency caused due to new inference is taken care of.
19. Suggest the strategy, as how a tree like rules structure can be used for reasoning
in forward direction. Explain or suggest your strategy for following:
a. Whether the breadth-first or depth-first search is better for rule searching?
b. Whether the top-down or bottom-up search is better for firing of rule
sequences?
c. How to eliminate one or both the rules, say R1 and R2 , for consideration,
if one of the precondition is false, which is common in both R1 and R2 ?
d. In the begin of a reasoning process whether you would prefer to choose
to fire a rule with large number of preconditions or small?
20. Analyze the Algorithm 4.1 and justify that it is NP-hard as a general case, how-
ever, polynomial in real world situation.
21. Analyze the complexity of preconditions in the forward reasoning process.
22. How the following situations are implemented in knowledge base for rule-
chaining? Explain.
a. “Is the pain intensifying with respirations?”
Exercises 109

b. “Did shortening of breadth occur before throat pain?”

23. Given the knowledge base of rule-chaining of Table 4.2, show an analysis as, in
general, which type, forward or backward search strategy is better? Also, justify
your claim.
24. Why, a rule-based system having combination of forward and backward rule-
chaining is more complex than either of these both?

References

1. Allen BP (1994) Case-based reasoning: business applications. Commun ACM 37(3):40–42

2. Patterson DW (2001) Introduction to artificial intelligence and expert systems. PHI India
3. Hayes-Roth F (1985) Rule-based systems. Commun ACM 28(9):921–922
4. Puppe F (1993) Systematic introduction to expert systems. Springer-Verlag
5. McDermott JP (1980) R1: A rule-based configurator of computer systems—technical report
CMU-CS-80-119, 1–56
Chapter 5
Logic Programming and Prolog

Abstract Prolog is logic programming languages for AI, based on predicate logic.
This chapter discusses the structure, syntax, and semantics of Prolog language, pro-
vides comparison with procedural language like C, interpretation of predicate logic
and that of Prolog, both formally as well through worked out examples, and explain
how the recursion is definition as well solution of a problem, and explains with sim-
ple examples as how the control sequencing takes place in Prolog. Use of two open
source compilers of prolog using simple worked out examples is demonstrated. Each
concept of Prolog and logic programming is explained with the help of worked out
examples. At the end, a brief summary gives glimpse of the chapter, followed with
number of exercises to strengthen the learning.

Keywords Logic programming · Prolog · Predicate logic · Prolog compiler ·

Control sequencing · Knowledge base · Query · Horn clause · Recursion ·
Rule-chaining · Backward chaining · Forward chaining · List · Cut · Fail

5.1 Introduction

This chapter presents the basic concepts of logic programming, and Prolog language.
Prolog is a logic programming language, implemented in two parts:
1. Logic, which describes the problem, and
2. Control, provides the solution method.

This is in contrast to procedural programming languages, where description and

solution go together, and are hardly distinguishable. This, feature of prolog helps in
separate developments for each part, one by the programmer and other by imple-
menter.
PROLOG is a simple, yet powerful programming language, based on the principles
of first-order predicate logic. The name of the language is an acronym for the French
‘PROgrammation en LOGique’(programming in logic). PROLOG was designed by
A. Colmerauer and P. Roussel at the University of Marseille (Canada), around 1970.
The PROLOG has remained connected with a new programming style, known as

© Springer Nature India Private Limited 2020 111

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_5
112 5 Logic Programming and Prolog

logic programming. Prolog is useful in problem areas, such as artificial intelligence,

natural language processing, databases, etc., but pretty useless in others, such as
graphics or numerical computations. The main applications of the language can be
found in the area of Artificial Intelligence; but PROLOG is being used in other areas
in which symbol manipulation is of prime importance. Following are the application
areas.

• Computer algebra.
• Design of parallel computer architectures.
• Database systems.
• Natural-language processing.
• Compiler construction.
• Design of expert systems.

The popular compilers of prolog are swi-prolog and gnu prolog; both are available
in open source, and runs on Windows and Linux platforms.
Learning Outcomes of this Chapter:

1. Prolog versus procedure-oriented languages. [Assessment]

2. Working with Prolog, using gprolog and swi-prolog compilers. [Usage]
3. Writing small prolog programs and running. [Usage]
4. Translating predicate logic into Prolog. [Usage]
5. Prolog Syntax and semantics. [Familiarity]
6. Forward-chaining versus back-ward chaining. [Assessment]
7. Using backward-chaining for reasoning and inference in Prolog. [Assessment]

5.2 Logic Programming

In conventional languages, like, C, C++, and Java, a program is a description of

sequence of instructions to be executed one-by-one by a target machine to solve
the given problem. The description of the problem is implicitly incorporated in this
instruction sequence, where usually it is not possible to clearly distinguish between
the description (i.e., logic) of the problem, and the method (i.e., control) used for its
solution.
In logic programming language, like Prolog, the description of the problem and
the method for its solution are explicitly separated from each other. Hence, in an
algorithm of a Prolog program, these two parts are distinctly visible, and can be
expressed by [6]:

Algorithm = Logic + Control (5.1)

In the above equation, term ‘logic’ represents the descriptive part of an algorithm,
and ‘control’ represents the solution part, which takes the description as the point
5.2 Logic Programming 113

Fig. 5.1 Logic

programming algorithm = logic + control

what how(search)

Horn clauses resolution

prolog database prolog interpreter

of departure. In other words, the logic component defines what the algorithm is
supposed to do, and the control component indicates how it should be done.
For solution through logic program, a problem is always described in terms of
relevant objects and relations between them. These relations are represented in a
clausal form of logic—a restricted form of first-order predicate logic. The logic
component for a problem is called a logic program, while the control component
comprises method for logical deduction (reasoning) for deriving new facts from the
logic program. This results to solving a problem through deduction. The deduction
method is designed as a general program, such that it is capable of dealing with any
logic program that follows the clausal form of syntax.
There are number of advantages of splitting of an algorithm into a logic component
and a control component:
• We can develop these two components of the program independent of each other.
For example, when writing the description of the problem as logic, we do not have
to be familiar with how the control component operates for that problem; thus
knowledge of the declarative reading of the problem specification suffices.
• A logic component may be developed using a method of stepwise refinement; we
have only to watch over the correctness of the specification.
• Changes to the control component affect (under certain conditions) only the effi-
ciency of the algorithm; they do not influence the solutions produced.
The implementation of Logic programming is explained in Fig. 5.1.
When all facts and rules have been identified, then a specific problem may be
looked upon as a query concerning the objects and their interrelationships. In sum-
mary, to provide specification of a logic program amounts to following jobs:
• specify the facts about the objects and relations between them for the problem
being solved;
• specify the rules about the objects and their interrelationships;
• specify the queries to be posed concerning the objects and relations.
An algorithm for logic program can be shown to be decomposed into the compo-
nents shown in Fig. 5.2.
114 5 Logic Programming and Prolog

Fig. 5.2 Components of

logic program

5.3 Interpretation of Horn Clauses in Rule-Chaining

Horn clause is restricted form of a predicate logic sentence. A typical representation

of a problem in Horn clause form is:

1. a set of clauses defining a problem domain and,

2. a theorem consisting of: (a) hypotheses represented by assertions A1 ←, . . . ,
An ← and (b) a conclusion in negated form and represented by a denial ←
B1 , . . . , Bm .

The reasoning process can be carried out as back-ward reasoning, or forward rea-
soning. In backward-chaining, reasoning is performed backwards from the conclu-
sion, which repeatedly reduces the goals to subgoals until ultimately all the subgoals
are solved directly in the form of original assertions.
In the case of problem-solving using forward-chaining approach, we reason for-
wards from the hypotheses, and repeatedly derive new assertions from old ones until
eventually the original goal is solved directly by derived assertions [6].
For our reasoning using forward and backward chaining, we consider the family-
tree of Mauryan Dynasty (India) as shown in Fig. 5.3 [5].
The problem of showing that chandragupta is a grandparent of ashoka can be
solved either backward-chaining or forward-chaining. In forward-chaining, we start
with the following assertions:
father(bindusara, asoka) ←
father(chandragupta, bindusara) ←
Also, we shall use the clause par ent (x, y) ← f ather (x, y) to derive new
assertions,
5.3 Interpretation of Horn Clauses in Rule-Chaining 115

Fig. 5.3 Mauryan dynasty

family-tree

par ent (chandragupta, bindusara) ←

par ent (bindusara, ashoka) ←
Continuing forward-chaining we derive, from the definition of grandparent, the
new assertion,
grandpar ent (chandragupta, ashoka) ←
which matches the original goal.
Reasoning using backward-chaining, we start with the original goal, which shows
that chandragupta is a grandparent of ashoka,

← grandpar ent (chandragupta, ashoka)

and use the definition of grandparent to derive two new subgoals,

← par ent (chandragupta, z), par ent (z, ashoka),

by denying that any z is both a child of chandragupta and a parent of ashoka. Contin-
uing backward-chaining and considering both subgoals (either one at a time or both
simultaneously), we use the clause,

par ent (x, y) ← f ather (x, y)

to replace the subproblem par ent (chandragupta, z) by f ather (chandragupta, z)

and the subproblem par ent (z, ashoka) by f ather (z, ashoka). The symbol “←”
is read as “if”. The newly derived subproblems are solved compatibly by assertions
which determine “bindusara” as the desired value of z.
In both the backward-chaining and forward-chaining solutions of the grandparent
problem, we have mentioned the derivation of only those clauses which directly
contribute to the ultimate solution. In addition to the derivation of relevant clauses, it
is often unavoidable, during the course of searching for a solution, to derive assertions
or subgoals which do not contribute to the solution. For example, in the forward-
chaining search for a solution to the grandparent problem, it is possible to derive the
irrelevant assertions as,
116 5 Logic Programming and Prolog

par ent (dur dhara, bindusara) ←

male(chandragupta) ←
Also, in backward-chaining search it is possible to replace the subproblem,
par ent (chandragupta, z)
by the unsolvable and irrelevant subproblem,
mother (chandragupta, z).
There are proof procedures which understand logic in backward-chaining, e.g., model
elimination, resolution, and interconnectivity graphs. These proof procedures operate
with the clausal form of predicate logic and deal with both Horn clauses and non-
Horn clauses. Among clausal proof procedures, the connection graph procedure is
able to mix backward and forward reasoning.
The terminology we used here—the backward-chaining, is also called “top-
down”. Given a grammar formulated in clausal form, top-down parsing algorithm
generates a sentence to its original form, i.e., the assertions. The forward-chaining
is also called “bottom-up”, where we start from the assertions and try to reach to the
goals.

5.4 Logic Versus Control

Different control strategies for the same logical representation generate different
behaviors. Also, the information about a problem-domain can be represented using
logic in different ways. Alternative representations can have a more significant impact
on the efficiency of an algorithm compared to alternative control strategies for the
same representation.
Consider the problem of sorting a list x and obtaining the list y. In our represen-
tation, we can have a definition with an assertion consisting of two arguments: “y is
permutation of x”, and “y is ordered”, i.e.,

sorting x gives y ← y is a permutation of x, y is ordered.

Here “←” is read as ‘if’ and ‘,’ is logical AND operator. The first argument generates
permutations of x and then it is tested whether they are ordered. Executing procedure
calls as coroutines, the procedure generates permutations, one element at a time.
Whenever a new element is generated, the generation of other elements is suspended
while it is determined whether the new element preserves the orderedness of the
existing partial permutation.
A program consists of a logic component and a control component. The logic
component defines the part of the algorithm that is problem specific, which not
only determines the meaning of algorithm but also decides the way the algorithm
behaves. For systematic development of well-structured programs using successive
refinements, the logic component needs to be defined before the control component.
5.4 Logic Versus Control 117

Due to these two components of a logic program, the efficiency of an algorithm can
be improved through two different approaches: 1. by improving the efficiency of
logic component, 2. through the control component. Note, that both improvements
are additive, and not as alternative choices.
In a logic programs, the specification of control component is subordinate to logic
component. The control part can be explicitly specified by the programmer through a
separate control specification language, or the system itself can determine the control.
When logic is used like in relational calculus to specify queries (i.e., higher level
of language) to knowledge base, the control component is entirely determined by
the system. Hence, for higher level of programming language, like queries, lesser
effort is required for programming the control part, because in that case the system
assumes more responsibility about efficiency, as well as to exercise control over the
use of given information.
Usually, a separate control-specifying language is preferred by advanced pro-
grammers to exercise the control with higher precision. A higher system efficiency
is possible if programmer can communicate to system a more precise information to
have finer control. Such information may be a relation, for example, F(x, y), where y
can also be a function of x. This function could be used by a backtracking interpreter
to avoid looking for another solution to the first goal in the goal statement. This can
be expressed, for example by,
← F(A, y), G(y),
when the second goal fails. Other example of such information can be that one
procedure,
S←T
may be more likely to solve problem S than another procedure,
S ← R.
This kind of information of cause-effect relation is common in fault diagnosis
where, on the basis of past experience, it might be possible to estimate that symptom
S is more likely have been caused by T rather than by R.
In the above examples, the control information is problem-specific. However, if
the control information is correct, and the interpreter is correctly implemented, then
the control information should not affect the meaning of the algorithm, which is
decided by the corresponding logic component of the program.

5.4.1 Data Structures

In a well-structured program it is desirable to keep data structures separate from pro-

cedure that interact with the data structure. Such separation means data structures can
be manipulated without altering the procedure. Usually, there is need of alteration
of data structures—some times due to the need of change of requirements and other
118 5 Logic Programming and Prolog

times with an objective to to improve the algorithm by replacing data structure by

a more efficient one. In large and complex programs, often the information demand
made on the data structures are only determinable in the final stages of the program
design. If data structures are separated from the procedures, the higher level proce-
dures can be written before the data structures are finalized, and these procedures
can be altered conveniently later any time without effecting the data structure [6].
In Prolog, the data structures of a program are already included in the logic
component of the program. Consider, for example, the data structure “Lists”, which
can be represented by following terms:

• nil; empty list,

• cons(x, y); list with first element x, and y is another list.

Hence, the following example names a three-element list consisting of individuals

as 2, 1, 3 in that order.
cons(2, cons(1, cons(3, nil)))

Example 5.1 A logic program for quick-sort.

In quick-sort, the predicates empty, first, rest, partitioning, and appending, can be
defined independently from the definition of sorting (see Eqs. 5.2 and 5.3). For this
definition we assume that, partitioning of a list x2 by element x1 gives a list u com-
prising the elements of x2 that are less or equal to x1 , and a list v of elements of x2
that are greater than x1 .

sor ting x gives y ← x is empt y, y is empt y (5.2)

sor ting x gives y ← f ir st element o f x is x1 , rest is x2 ,

partitioning x2 by x1 gives u and v,
sor ting u gives u  ,
sor ting v gives v  ,
appending w to u  gives y,
f ir st element o f w is x1 ,
r est o f w is v  . (5.3)

The data-structure-free definition of quicksort interacts with the data structure of

lists through the following definitions:
nil is empty ←
first element of cons(x, y)isx ←
rest of cons(x, y)isy ←
If the predicates: empty, first, rest, are dropped from the definition of quick-sort,
and instead the preliminary forward-chaining/forward-chaining deduction is used,
then the original data-structure-free definition of quick-sort can be replaced by a
definition that mixes the data structures with the procedures,
5.4 Logic Versus Control 119

sor ting nil gives nil ←

sor ting cons(x1 , x2 ) gives y ← partitioning x2 by x1 gives u, v,

sor ting u gives u  ,
sor ting v gives v  ,
appending list cons(x1 , v  ) to u  gives y.


There is another advantage of data-structure-independent definition: with well-
chosen interface procedure names, data-structure-independent programs are virtually
self-documenting. In conventional programs that mix procedures and data structures,
the programmer needs to provide separate documentation for data structures.
On the other hand, in-spite of the arguments in support for separating procedures
and data structures, the programmers usually mix them together for the sake of run-
time efficiency.

5.4.2 Procedure-Call Execution

In a simplest backward reasoning based execution, the procedure calls are executed
one at a time, in the order they have been written. Typically, an algorithm can be
made to run faster by executing the same procedure calls in the form of coroutines
or as communicating parallel-processes. Consider an algorithm A1 ,

A1 = L + C 1 (5.4)

where logic component is L and control component is C1 . Assume that from A1 we

have obtained a new algorithm A2 , having logic component L and control component
C2 ,
A2 = L + C 2 (5.5)

where we replaced control strategy C1 by new control strategy C2 , and the logic
L of the algorithm remains unchanged. For example, executing procedure calls in
sequence, the procedure,

sor ting x gives y ← y is a per mutation o f x, y is or der ed, (5.6)

first generates permutations of x and then tests whether they are ordered. By executing
procedure calls as coroutines, the procedure generates permutations, one element at
a time. Whenever a new element is generated, the generation of other elements is
suspended while it is determined whether the new element preserves the orderedness
of the existing partial permutation.
120 5 Logic Programming and Prolog

In the Similar way to procedure (5.6), procedure calls in the body of quick-sort
can be executed (either as coroutines or as parallel/sequential processes). When they
are executed in parallel, partitioning the rest of x can be initiated as soon as the
first element of the rest are generated (see procedure (5.3) on p. 118). Note that the
sorting for u and v can take place in parallel, the moment first elements of u and v
are generated. And, the appending of lists can start soon after the first elements of
u  , and the sorted output of u, are made available.
The high level language SIMULA provides the facility of writing both the usual
sequential algorithms, as well as algorithms with coroutines. Here, the program-
mer can provide the controls about: When the coroutines should be suspended and
resumed? However, in this language, the logic and control are inextricably inter-
twined in the program text, like in other procedure oriented languages. Hence, the
control of an algorithm cannot be modified without rewriting the entire program.
In one way, the concept of separating logic and controls is like separating data
structures and procedures. When a procedure is kept separate from a data structure,
we are able to distinguish clearly as what functions are fulfilled by which data struc-
tures. On the other hand, when logic is separated from control, it becomes possible
to distinguish, what the program (i.e., logic) does, and how it does it (i.e., control-
ling takes place). In both conditions it becomes obvious as what the program does,
and hence it can be more easily determined whether it correctly does the job it was
intended for.

5.4.3 Backward Versus Forward Reasoning

Recursive definitions are common in mathematics, where they are more likely to
be understood as forward-chaining rather than backward-chaining. Consider, for
example, the definition of factorial, given below.

f actorial o f 0 is 1 ← (5.7)

f actorial o f x is u ←y plus 1 is x,
f actorial o f y is v,
x times v is u. (5.8)

The mathematician is likely to understand such a definition forward-chaining,

generating the sequence of assertions as follows:

f actorial o f 0 is 1 ←
f actorial o f 1 is 1 ←
f actorial o f 2 is 2 ←
f actorial o f 3 is 6 ←
and so on.
5.4 Logic Versus Control 121

Conventional programming language implementations understand recursions as

backward-chaining. Programmers, accordingly, tend to identify recursive definitions
with backward-chaining execution. However, there is one exception, and that is
Fibonacci series, which is efficient when interpreted as forward reasoning. It is left
as an exercises for the students to verify the same.

5.4.4 Path Finding Algorithm

Consider an algorithm A, with C1 , C2 control components, and L 1 , L 2 as logic com-

ponents, which can often be analyzed in different ways [6].

A = L 1 + C1 = L 2 + C2 . (5.9)

Some of the behavior determined by C1 in one analysis might be determined by the

logic component L 2 in another analysis. This has significance for understanding the
relationship between programming style and execution facilities. In the short term
sophisticated behavior can be obtained by employing simple execution strategies
and by writing complicated programs. In the longer term the same behavior may be
obtained from simpler programs by using more sophisticated execution strategies.
A path-finding problem illustrates a situation in which the same algorithm can be
analyzed in different ways. Consider the problem of finding a path from vertex a to
vertex z in the directed graph shown in Fig. 5.4.
In one analysis, we can employ a predicate go(x) which states that it is possible
to go to x. The problem of going from a to z is then represented by two clauses. One
asserts that it is possible to go to a. The other denies that it is possible to go to z.
The arc directed from a to b is represented by a clause which states that it is possible
to go to b if it is possible to go to a. Different control strategies determine different
path-finding algorithms. Forward search from the initial node a is forward-chaining
based reasoning from the initial assertion go(a) ← (see Table 5.1). Backward search
from the goal node z is backward reasoning from the initial goal statement ← go(z)
(see Table 5.2).

Fig. 5.4 Directed graph path d

finding
e x
b z
u
y
a

c v
122 5 Logic Programming and Prolog

Table 5.1 Forward-chaining go(a) ←

go(b) ← go(a)
go(c) ← go(a)
go(d) ← go(b)
go(e) ← go(b)
go(x) ← go(e)
go(z) ← go(x)
...

Table 5.2 Backward- ← go(z)

chaining
go(z) ← go(x)
go(z) ← go(y)
go(x) ← go(e)
go(y) ← go(u)
go(y) ← go(v)
go(e) ← go(b)
go(b) ← go(a)
...

Carrying out a bidirectional search from both the start node and the goal node
results to a combination of backward and forward reasoning. In that case, whether a
path-finding algorithm investigates one path at a time (in depth-first) or develops all
paths together (in breadth-first) will depend on search strategy used.

5.5 Expressing Control Information

The distinction between backward-chaining and forward-chaining based execution

can be expressed in a graphical notation using arrows to indicate the flow of control.
The same notation can be used to represent different combinations of backward-
chaining and forward-chaining based execution. The notation does not, however,
aim to provide a complete language for expressing useful control information.
The arrows are attached to atoms in clauses to indicate the direction of transmission
of processing activity from one clause to other clause. For every pair of matching
atoms in the initial set of clauses (i.e., one atom in the conclusion of a clause and
the other among the conditions of a clause), and there is an arrow directed from one
atom to the other. For backward-chained reasoning, arrows are directed from goals
to assertions. For the grandparent problem, we have the graph shown in Fig. 5.5.
5.5 Expressing Control Information 123

← grandparent(chandragupta, ashoka)

grandparent(x, y) ← parent(x, z), parent(z, y)

parent(x, y) ← f ather(x, y) parent(x, y) ← mother(x, y)

f ather(chandragupta, bindusara) ←
f ather(bindusara, ashoka) ←
mother(durdhara, bindusara) ←

Fig. 5.5 Control-flow for backward-chaining

← grandparent(chandragupta, ashoka)

grandparent(x, y) ← parent(x, z), parent(z, y)

parent(x, y) ← f ather(x, y) parent(x, y) ← mother(x, y)

f ather(chandragupta, bindusara) ←
f ather(bindusara, ashoka) ←
mother(durdhara, bindusara) ←

Fig. 5.6 Control-flow for forward-chaining

A processing activity in backward-chaining in this figure is shown as starting

with the initial goal statement, and transmits activity to the body of the grandparent
procedure, whose procedure calls, in turn, activate the parenthood definitions. At
the end, the assertions provided in the knowledge base passively accepts processing
activity, and does not further transmit it to other clauses.
For forward-chaining execution, arrows are directed from assertions to goals (see
Fig. 5.6). The processing activity originates with the database of initial assertions
(i.e., bottom of the graph in this figure). They transmit activity to the parenthood
definitions, which, in turn, activate the grandparent definition. Processing terminates
when it reaches all the conditions in the passive initial goal statement, at the top of
the graph [6].
The grandparent definition can be used in a combination of backward-chaining
and forward-chaining methods. Using numbers to indicate sequencing, we can
124 5 Logic Programming and Prolog

Fig. 5.7 Combination of 3

logic and control-I

grandparent(x, y) ← parent(x, z), parent(z, y)

1 2

Fig. 5.8 Combination of 1

logic and control-II

grandparent(x, y) ← parent(x, z), parent(z, y)

2 3

represent different combinations of backward-chaining and forward-chaining rea-

soning. For the sake of simplicity, we only show the control notation associated with
the grandparent definition.
The combination of logic and control indicated in Fig. 5.7 represents the algorithm,
with markers 1–3, having following sequence of operation:

1. indicates that the algorithm waits until x is asserted to be the parent of z, then
2. indicates that the algorithm finds a child y of z, and finally
3. indicates that the algorithm asserts that x is grandparent of y.

The combination indicted by Fig. 5.8, which represents the algorithm, has marking
1–3 as events, with following meanings:

1. this event waits until x is asserted to be parent of z, then

2. this event waits until it is given the problem of showing that x is grandparent of
y,
3. this event attempts to solve by showing that z is parent of y.

In the similar way the the algorithms takes care of rest of the controls of backward-
chaining and forward-chaining.

5.6 Running Simple Programs

Program files can be compiled using the predicate consult. The argument has to be
a Prolog atom denoting the particular program file. For example, to compile the file
socrates.pl, submit the following query to swi-Prolog [8]:

?- consult(socrates.pl).
5.6 Running Simple Programs 125

If the compilation is successful, Prolog will reply with ‘Yes’. Otherwise a list of
errors is displayed. The ‘swi-prolog’ can also run in GUI environment in Windows.
The gnu-prolog running on linux can run a prolog program ’socrates.pl’ as
follows [3]:

$ gprolog <enter>
GNU Prolog 1.3.0
Copyright ....
| ?-[socrates]. % filename without extensions.

Let us demonstrate to run a prolog program and perform interpretation of the

clauses by backward-chaining. For this, we consider our single old problem of
“socrates” and application of inference rule of modus ponens.

Example 5.2 Demonstrating Backtracking.

All men are mortal.

Socrates is a man.
Therefore, Socrates is mortal.

In terms of Prolog, the first statement corresponds to the rule: X is mortal, if X is

a man (i.e., for every X ). The second statement corresponds to the fact: ‘Socrates is
a man’. Note that ‘socrates’ is constant (literal), and X is a variable. The above rule
and fact can be written in Prolog language syntax as,

mortal(X) :- man(X).
man(socrates).

The conclusion of the argument is: “Socrates is mortal,” which can be expressed in
predicate as ‘mortal(socrates)’. After we have compiled, we run the above program,
and query it, as follows:

?- mortal(socrates).
Yes

We notice that Prolog agrees with our own logical reasoning. But how did it come
to its conclusion? Let’s follow the goal execution step-by-step [7].

1. The query mor tal(socrates) is designated the initial goal.

2. Scanning through the clauses of this program, Prolog tries to match
mor tal(socrates) with the first possible fact or head of rule. It finds mor tal(X )—
head of the first (and only) rule. When matching the two ‘socrates‘is bound to X ,
with unifier {socrates/ X }.
3. The variable binding is extended to the body of the rule, i.e. man(X ) becomes
man(socrates).
4. The newly instantiated body becomes our new subgoal: man(socrates).
126 5 Logic Programming and Prolog

5. Prolog executes the new subgoal by again trying to match it with a rule-head or
a fact.
6. Obviously, new subgoal man(socrates) matches the fact man(socrates), and
current sub-goal succeeds.
7. This, means that the initial goal succeeds, and prolog responds with ’YES’. 

We can observe the trace of sequences operations of interpretations by running it.

Prolog is a declarative (i.e., descriptive) language. Programming in Prolog means
describing the world. Using such programs means asking questions about the previ-
ously described world. The simplest way of describing the world is by stating facts,
like “train is bigger than bus”, as,

bigger(train, bus).

The following example demonstrates this [2].

Example 5.3 Knowledge base about sizes of transports.

Let’s add a few more facts to vehicles of transport as:

bigger(train, bus).
bigger(bus, car).
bigger(car, bicycle).
bigger(car, motorbike).

This is a syntactically correct program, and after having compiled it, we can ask
the Prolog system questions (or queries) about it.

?- bigger(car, bicycle). <enter>

Yes

The query ‘bigger(car, bicycle)’ (i.e. the question “Is a car bigger than a bicycle?”)
succeeds, because the fact ‘bigger(car, bicycle)’ was previously communicated to the
Prolog system. Our next query is, “is a motorbike bigger than an train?”

?- bigger(train, motorbike).
No

The reply by Prolog is “No”. The reason being that, the program says nothing about
the relationship between train and motorbike. However, we note that, the program
says—‘trains are bigger than bus’, and ‘buses are bigger than cars’, which in turn
are bigger than motorbike. Thus, trains are also to be bigger than motorbikes. In
mathematical terms, the bigger-relation is transitive. But it also not been defined
in our program. The correct interpretation of the negative answer Prolog is that:
“from the information communicated to the program it cannot be proved that an
5.6 Running Simple Programs 127

train is bigger than a motorbike”. As an exercise, we can try the proof by resolution
refutation, but it cannot be proved because it is not possible to verify the statements.
Solution would be to define a new relation, which we will call isbigger, as the
transitive closure. Animal X is bigger than Y , if this has been stated as a fact.
Otherwise, there is an animal Z , for which it has been stated as a fact that animal X
is bigger than animal Z , and it can be shown that animal Z is bigger than animal Y .
In Prolog such statements are called rules and are implemented as follows:

isbigger(X, Y) :- bigger(X, Y). %rule1

isbigger(X, Y) :- bigger(X, Z), isbigger(Z, Y). %rule2

where ‘:-’ stands for ‘if’ and comma (,) between ‘bigger(X, Z)’ and ‘isbigger(Z,Y)’
stands for ‘AND’, and a semicolon (;) for ‘OR’. If from now on we use ‘isbigger’
instead of ‘bigger’ in our queries, the program will work as intended.

?- isbigger(train, motorbike).
Yes

In the rule1 above, the predicate ‘isbigger(X, Y)’ is called goal, and ‘bigger(X,
Y)’ is called sub-goal. In the rule2 ‘isbigger(X,Y)’ is goal and the expressions after
the sign ‘:-’ are called sub-goals. The goal is also called head of the rule, and the
expressions after sign ’:-’ is called body of the rule statement.
In fact, the rule1 above corresponds to the predicate,

i f bigger (X, Y ) then isbigger (X, Y ),

or
bigger (X, Y ) → isbigger (X, Y ).

Similarly, predicate expression for rule2 is

bigger (X, Z ) ∧ isbigger (Z , Y ) → isbigger (X, Y ).

The prolog expressions which are not conditionals, i.e., like,

bigger(train, bus).
bigger(bus, car).
bigger(car, bicycle).
bigger(car, motorbike).

are called facts(or assertions). The facts and rules, together, make the knowledge
base (KB) in a program.
For the query ‘isbigger(train, motorbike)’ the Prolog still cannot find the fact
‘bigger(train, motorbike)’ in its database, so it tries to use the second rule instead.
This is done by matching the query with the head of the rule, which is ‘isbigger
128 5 Logic Programming and Prolog

(X, Y)’. When doing so, the two variables get bound: X = train, and Y = motorbike.
The rule says that in order to prove the goal ‘isbigger(X,Y)’ (with the variable
bindings that’s equivalent to isbigger(train, motorbike)), Prolog needs to prove the
two subgoals ‘bigger(X, Z)’ and ‘isbigger(Z, Y)’, with the same variable bindings.
Hence, the rule2 gets transformed to:

isbigger (train, motor bike) : −bigger (train, Z ),

isbigger (Z , motor bike).

By repeating the process recursively, the facts that make up the chain between
train and motorbike are found and the query ultimately succeeds. Our earlier Fig. 5.5
demonstrated the similar chain of actions.
Of course, we can do slightly more exciting job than just asking yes/no-questions.
Suppose we want to know, what animals are bigger than a car? The corresponding
query would be:

?- isbigger(X, car).

We could also have chosen any other name in place of X for it as long as it starts
with an uppercase letter, which makes it a variable. The Prolog interpreter replies as
follows:

X = bus; % press here ’;’ to get another match

X = train ; if exists.
No

There are many more ways of querying the Prolog system about the contents of
its database.
For example, try to find out the answer for:

?- bigger(Who, Whom).

You will get many answers! The Prolog treats the arguments Who and Whom as
variables.
As an example we ask whether there is an animal X that is both smaller than a
car and bigger than a motorbike:

?- isbigger(car, X), isbigger(X, motorbike).

No

The following example explains the execution sequence of prolog statements.

Example 5.4 Give the trace of execution of query “isbigger(bus, motorbike)”, sub-
mitted to the animal world.
The trace of the above query is shown below.
5.6 Running Simple Programs 129

? isbigger(bus, motorbike).
1. Call: isbigger(bus, motorbike)?
2. Call: bigger(bus, motorbike)?
2. Fail: bigger(bus, motorbike)?
2. Call: bigger(bus, _80)?
2. Exit: bigger(bus, car)?
2. Call: isbigger(car, motorbike)?
3. Call: bigger(car, motorbike)?
3. Exit: bigger(car, motorbike)?
2. Exit: isbigger(car, motorbike)?
1. Exit: isbigger(bus, motorbike)?
True?
Yes
{Trace}

The trace can be verified to be performing as per the Rule1 and Rule2 discussed
above. 
Many a times, when started from goal, it may not be possible to reach to facts
available. This shows that prolog is incomplete in theorem proving even for definite
clauses, as it fails to prove facts that can be concluded from knowledge base.

5.7 Some Built-In Predicates

The built-ins can be used in a similar way as user-defined predicates. The important
difference between the two is that a built-in predicate is not allowed to appear as the
principal function in a fact or in the head of a rule. This must be so, because using
them in such a position would effectively mean changing their definition [1].
Equality. We write X = Y . Such a goal succeeds, if the terms X and Y can be
matched.
Output. Besides Prolog’s replies to queries, if you wish your program to have
further output, you can use the write predicate. The argument can be any valid Prolog
term. In the case of a variable argument, its value will be printed. Execution of the
predicate causes the system to skip a line, as in the following cases.

?- write(Hello World!), nl.

Hello World!
Yes

?- X = train, write(X), nl.

train
X = train
Yes
130 5 Logic Programming and Prolog

read(N).
write(’the number is’), write(N), nl.
the number is 5
N = 5
Yes

Matchings. Following are the examples for matchings. If two expressions matches,
the output is ’Yes’ otherwise it is ’No’. The query, to prolog shows that the two
expressions cannot be matched.

?- p(X, 2, 2) = p(1, Y, X).

No

Sometimes there is more than one way of satisfying the current goal. Prolog
chooses the first possibility (as determined by the order of clauses in a program),
but the fact that there exists alternatives, is recorded. If at some point, Prolog fails to
prove a certain subgoal, the system can go back and try an alternative left behind in
the of executing of the goal. This process is known as backtracking. The following
example demonstrates backtracking.

5.8 Recursive Programming

Using recursive programs, we can provide recursive definition of functions. We know

that the factorial n! of a natural number n is defined as the product of all natural
numbers from 1 to n. Here’s a more formal, recursive definition (also known as an
inductive definition), and the code in prolog [4].
Example 5.5 Factorial Program.
Recall the definition of factorial in Eqs. (5.7) and (5.8) in Sect. 5.4.3.
0! = 1, (base case) n! = (n1)! ∗ n, for n ≥ 1 (Recursion rule)

%finding factorial.
fact(0, 1). % base case

fact(N, R) :- N >= 1, % recursion step

N1 is N - 1,
fact(N1, R1),
R is R1 * N.

For a recursive program to test the membership of an element in a set, if the
element is not as head of the list, then it is in the tail. The process is recursively
called. The membership algorithm is built-in feature of prolog, as well as it can be
user-defined.
5.8 Recursive Programming 131

Fig. 5.9 Towers of Hanoi

problem

A I B

A recursive algorithm for GCD (greatest common divisor) based on Euclid’s

Algorithm can be constructed as follows.

Example 5.6 Program for Greatest Common Divisor (GCD).

%gcd
gcd(X, X, X).
gcd(X, Y, Z) :- X > Y, D is X - Y, gcd(D, Y, Z).
gcd(X, Y, Z) :- X < Y, D is Y - X, gcd(X, D, Z).


Example 5.7 Towers of Hanoi Problem.

Given three stacks A (source), B (destination), and I (intermediate), the towers of

Hanoi problem is to move N number of disks from stack A to B using I as temporary
stack. The disks are originally on stack A such that larger diameter disks are below
the smaller diameter disks, and no two disks have equal diameters. The movement is
to be done following the rules of this game, which states that only one disk is to be
moved at a time, and at no time the bigger diameter disk shall come over a smaller
diameter disk (Fig. 5.9).

move(A,B):- nl,
write(’move top from ’),
write(A),
write(’ to ’),
write(B).
transfer(1,A,B,I) :- move(A,B).
transfer(N, A, B, I):- N > 1,
M is N -1,
transfer(M, A, I, B),
move(A, B),
transfer(M, I, B, A).

The algorithm uses the strategy: move n − 1 disks from A to I , then move a single
disk from A to B, finally move n − 1 disks from I to B. For n − 1, it recursively
calls the algorithm. The predicate nl stands for new-line. 
132 5 Logic Programming and Prolog

5.9 List Manipulation

Prolog represents the lists contained in square brackets with the elements being
separated by commas. Here is an example:

[train, bus, car, bicycle]

Elements of lists could be any valid Prolog terms, i.e. atoms, numbers, variables,
or compound terms. A term may also be other list. The empty list is denoted by ‘[]’.
The following is another example for a (slightly more complex) list:

[train, [], X, parent(X, tom), [a, b, c], f(22)]

Internally, lists are represented as compound terms using the function. (dot). The
empty list ‘[]’ is an atom and elements are added one by one. The list [a, b, c], for
example, corresponds to the following term:

.(a, .(b, .(c, [])))

We discussed in Sect. 5.4.1 about lists. A list is a recursive definition, consisting

of a head and a tail. The tail also comprises of head and rest of the elements as tail,
and so on, until the tail is empty list.

Example 5.8 Membership Program.

% membership built-in
?-member(2, [a, b, c, 2, 4, 900]).
Yes.

% membership program
ismember(X, [X|R]). % matches with 1st element
ismember(X, [Y|R]) :- ismember(X, R). % try for
% next element


The built-in program append, appends two lists.

?append([1, 2, 3], [a, b, c], X). % This is buit-in

X=[1, 2, 3, a, b, c]

Example 5.9 Appending of lists.

append([], L , L).
append([X |L1], L2, [X |L3]) : −append(L1, L2, L3).
5.9 List Manipulation 133

append([1, 2], [3, 4], Z)

θ1 = {1/X, [2]/L1, [3, 4]/L2, append([X|L1], L2, [X|L3])

[1|L3]/Z}

append(L1, L2, L3)

= append([2], [3, 4], L3)

θ2 = {2/Y, [3, 4]/L5, [2, L4]/L3} append([Y |[]], [3, 4], [Y |L4])

append([], [3, 4], L4)

Fig. 5.10 Prolog search for appending two lists

For a query, we write,

? append([1, 2], [3, 4], Z)

Z = [1, 2, 3, 4]
Yes
?-

The search, along with unifications for appending two lists is shown in Fig. 5.10.
The goal search shows alternate cycles of unification and calling of sub-goals. As a
result of the recursion, the append operation can be realized as follows. The terminal
node is matched with the fact: append([], L , L). Consequently, L4 = L5 = [3, 4].
On back substitution,

L3 = [2|L4]
= [2|[3, 4]]
= [2, 3, 4].

Z = [1|L3]
= [1|[2, 3, 4]]
= [1, 2, 3, 4].

Head and Tail
The first element of a list is called its head and the remaining list is called the tail.
An empty list does not have a head. A list containing a single element has a head
(namely that particular single element) and its tail is the empty list. A variant of the
134 5 Logic Programming and Prolog

list notation allows for convenient addressing of both head and tail of a list. This is
done by using the separator | (bar) [1].

?- [11, 12, 13, 14, 15] = [Head | Tail].

Head = 11
Tail = [12, 13, 14, 15]
Yes

Notice that the H ead and T ail are just names for variables.1 We could have used
X and Y or something else instead with the same result. We can access 2nd element
as well.
?- [bovi, jovi, kv, licet, quod, non, ] = [_, X | _].
X = jovi

The more examples are as follows, which are self explanatory.

?- append([1, 2, 3], [d, e, f, g], X).

X = [1, 2, 3, d, e, f, g]
Yes

?- append(U, V, [a, b, c, d]).

U = []
V = [a, b, c, d] ;
U = [a]
V = [b, c, d] ;
U = [a, b]
V = [c, d] ;
U = [a, b, c]
V = [d] ;
U = [a, b, c, d]
V = [] ;
No

?- length([train, [], [1, 2, 3, 4]], Length).

Length = 3
Yes

?- member(bicycle, [train, bus, car, bicycle, motorbike]).

Yes

?- reverse([1, 2, 3, 4, 5], X).

X = [5, 4, 3, 2, 1]
Yes

1A Prolog variable starts with uppercase letter.

5.10 Arithmetic Expressions 135

5.10 Arithmetic Expressions

Prolog is not designed to handle arithmetics efficiently. Hence, it handles expressions

and assignment operations in some different way [1].

?3 + 5 = 8.
No

?X is 3 + 5.
X = 8
Yes

The terms 3 + 5 and 8 do not match as the former is a compound term, whereas
the latter is a number.
The following are arithmetical relational predicates:

X > Y
X < Y
X >= Y
X =< Y
X \= Y
X = Y

The last two predicates express inequality and equality, respectively.

5.11 Backtracking, Cuts and Negation

The Prolog language has number of predicates to explicitly control the backtracking
behavior of its interpreter. This way the Prolog deviates from the logic programming
idea. For example, the predicate True takes no arguments, and it always succeeds.
Some of the other explicit predicates of Prolog as discussed below.
Fail Predicate
The predicate Fail also has no arguments, the condition fail never succeeds. The
general application of the predicate fail is to enforce backtracking, as shown in the
following clause:
a(X ) : - b(X ), f ail.

When the query a(X ) is entered, the PROLOG interpreter first tries to find a
match for b(X ). Let us suppose that such a match is found, and that the variable X is
instantiated to some term. Then, in the next step f ail, as a consequence of its failure,
enforces the interpreter to look for an alternative instantiation to X . If it succeeds
136 5 Logic Programming and Prolog

in finding another instantiation for X , then again f ail will be executed. This entire
process is repeated until no further instantiations can be found. This way all possible
instantiations for X will be found. Note that if no side-effects are employed to record
the instantiations of X in some way, the successive instantiations leave no trace. It
will be evident that in the end the query a(X ) will be answered by no. But, we have
been successful in backtracking, i.e., going back and trying all possible instantiations
for X , which helps in searching all the values.
The negation in prolog is taken as failure as shown in the following program.
Example 5.10 Negation as failure.

bachelor(P) :- male(P), not(married(P)).

male(rajan).
male(rajam).
married(dicken).

When run, the queries responded as obvious. In the third case, married(Who) suc-
ceeds, so the negation of goal fails.

?bachelor(rajan).
yes
?bachelor(dicken).
no
bachelor(Who).
Who = dicken
no

Cut Predicate
Some times it is desirable to selectively turn off backtracking. This is done by cut (!).
The cut, denoted by !, is a predicate without any arguments. It is used as a condition
which can be confirmed only once by the PROLOG interpreter: on backtracking it is
not possible to confirm a cut for the second time. Moreover, the cut has a significant
side effect on the remainder of the backtracking process: it enforces the interpreter
to reject the clause containing the cut, and also to ignore all other alternatives for the
procedure call which led to the execution of the particular clause.

Example 5.11 Backtracking.

a :- b,c,d.
c :- p,q,!,r,s.
c.

Suppose that upon executing the call a, the successive procedure calls b, p, q, the
cut and r have succeeded (the cut by definition always succeeds on first encounter).
Furthermore, assume that no match can be found for the procedure call s. Then as
usual, the interpreter tries to find an alternative match for the procedure call r . For
5.11 Backtracking, Cuts and Negation 137

each alternative match for r , it again tries to find a match for condition s. If no
alternatives for r can be found, or similarly if all alternative matches have been tried,
the interpreter normally would try to find an alternative match for q. However, since
we have specified a cut between the procedure calls q and r , the interpreter will not
look for alternative matches for the procedure calls preceding r in the specific clause.
In addition, the interpreter will not try any alternatives for the procedure call c; so,
clause 3 is ignored. Its first action after encountering the cut during backtracking is
to look for alternative matches for the condition preceding the call c, i.e., for b. 

5.12 Efficiency Considerations for Prolog Programs

For a given goal, prolog explores the premises for rules in the knowledge base,
making the goal true. If there are premises p1 ∧ p2 ∧ · · · ∧ pn , it fully explores the
premise (called choice point) pi before proceeding to pi+1 .
The solution through prolog is unification, and binding of variables, pushing and
retrieving the stack, associated with backtracking. When a search fails, prolog will
backtrack to the previous choice point, followed with possibly unbounding of some
of the variables. It always keeps track of all the bound variables at any moment, and
those kept in the stack. In addition, it has to manage the index for fast searching of
predicates. This is called trail. Accordingly, even the most efficient prolog interpreters
consume thousands of machine instructions for even the simple unifications and
matching.
For the huge task, and due to nature of computing, required, the normal proces-
sors give very poor performance to prolog programs. Hence, the prolog programs
are compiled into intermediate programs, called WAM (Warren Abstract Machine).
WAM helps prolog running faster as well as making it parallel.
Prolog may some times lead to incomplete loops.
The true version of prolog is called pure Prolog. It is obtained from a variation of
the backward chaining algorithm that allows Horn clauses with the following rules
and conventions:

• The Selection Rule is to select the leftmost literals in the goal.

• The Search Rule is to consider the clauses in the order they appear in the current
list of clauses, from top to bottom.
• Negation as Failure, that is, Prolog assumes that a literal L is proven if it is unable
to prove ¬L.
• Terms can be set equal to variables but not in general to other terms. For example,
we can say that x = A, x = F(B) but we cannot say that A = F(B).

These rules makes fast processing. But, unfortunately, the Pure Prolog inference
Procedure is Sound but not Complete. This can be seen by the following example.
Using this we are unable to derive in Prolog P(a, c) because we get caught in an
ever deepening depth-first search.
138 5 Logic Programming and Prolog

P(a, b).
P(b, c).
P(Y, X ) : - P(X, Y ).
P(X, Z ) : - P(X, Y ), P(Y, Z ).

Actual Prolog
Actual Prolog differs from pure Prolog in three major respects:
• There are additional functionalities besides theorem proving, such as functions to
assert statements, functions to do arithmetic, functions to do I/O.
• The “cut” operator allows the user to prune branches of the search tree.
• The unification routine is not quite correct, in that it does not check for circular
bindings e.g. X → Y, Y → f (X ).

5.13 Summary

Prolog is a logic programming language, implemented in two parts:

1. Logic, which describes the problem, and
2. Control, provides the solution method.
This is in contrast to other programming languages, where description and solu-
tion go together, and they are hardly distinguishable. This, feature of prolog helps
in separate developments for each part, one by the programmer and other by imple-
menter.
The Prolog is being used in many areas where symbol manipulation is of prime
importance; however, the main applications of this language is in the area of Artificial
Intelligence.
In a well-structured program it is desirable to have data structures separate from the
procedures which interrogate and manipulate them. This separation means that the
data structures can be altered without altering the higher level procedures. Typically,
an algorithm, whcih is separate from data structures, can be made more efficient by
executing the same procedure calls either as coroutines or as communicating parallel
processes.
A Prolog program is declaration of facts and rules, called knowledge base, a
searching of this knowledge base in prolog is DFS (depth first search).
Running a prolog program is querying that program. Inferencing process in prolog
is goal driven.
Due to long processing of unification, binding, searching, use of stack, a prolog
program runs very slow. To run it faster, a prolog program is converted into a virtual
machine code, which is executed by the WAM (warren abstract Machine). Efficiency
of a Prolog program can be improved through two different approaches, either by
improving the logic component or by leaving the logic component unchanged and
improving the control over its use.
Exercises 139

Exercises

1. Determine, in which of the following lists cases the unification succeeds and
where it fails? If it succeeds, write down the unifier. (Note: Uppercase are vari-
ables.)

[a, d, z, c] and [H |T ]
[apple, pear, grape] and [A, pear |Rest]
[a|Rest] and [a, b, c]
[a, []] and [A, B|Rest]
[One] and [two| []]
[one] and [T wo]
[a, b, X ] and [a, b, c, d]

2. Give Prolog predicates for natural-number and plus that implement the Peano
axioms. Using plus give a Prolog program times for multiplication. Further, using
times give a Prolog program exp for exponentiation.
3. Given the following knowledge base for prolog, find a female descendant of
‘george’, by manually running the program.

parent(george,sam).
parent(george,andy).
parent(andy,mary).
male(george).
male(sam).
male(andy).
female(mary).
ancestor(X,Z) :- parent(X, Z).
ancestor(X,Z) :- parent(X,Y), ancestor(Y,Z).

4. What is response of Prolog interpreter for following queries?

?[a,b,c,d] = [a,b,c,d|[]].
?[a,b,c,d] = [a,b,c,[d]].
?[a,b,c,d] = [a,b,[c,d]].
?[a,b,c,d] = [a|[b,c,d]].
?[a,b,c,d] = [a,b|[c,d]].
?[a,b,c,d] = [a,b,c,d,[]].
?[a,b,c,d] = [a,b,c|[d]].
?[a,b,c,d] = [a,[b,c,d]].
140 5 Logic Programming and Prolog

5. Which of the following lists are syntactically correct for Prolog language? Find
out the number of elements in the lists that are correct.

[1, 2, 3, 4| []]
[1| [2| [3| [4]]]]
[1, 2, 3| []]
[[] | []]
[[1, 2] |4]
[1| [2, 3, 4]]
[[1, 2] , [3, 4] | [5, 6, 7]] [1|2, 3, 4]

6. Write a predicate second(S, List), that checks whether S is the second element
of List.
7. Consider the knowledge base comprising the the following facts:

tran(ek, one).
tran(do, two).
tran(teen, thr ee).
tran(char, f our ).
tran( panch, f ive).
tran(cha, si x).
tran(saat, seven).
tran(aat, eight).
tran(no, nine).

Write a predicate listtran(H, E) that translates a list of Hindi number words

to the corresponding list of English number words. For example, for a list X ,
listtran([ek, teen, chaar ], X ), should give response:

X = [one, thr ee, f our ].

8. Draw the search trees for the following prolog queries:

a. ?- member (x, [a, b, c]).
b. ?- member (a, [c, b, a, y]).
c. ?- member (X, [a, b, c]).
9. Write a program that takes a grammar represented as a list of rules and given a
query of as sentence, and returns whether the sentence is grammatically correct.
10. Run the following programs in trace mode with single step, and describe the
observed behavior, as why it is so?
a. Factorial Program (Example 5.5).
b. GCD Program (Example 5.6).
c. Mortal men program (Example 5.2).
Exercises 141

11. Given the following facts and rules about a blocks world, represent them in rules
forms, then translate the rules into prolog, and find out “what block is on black
block?”

Facts:
A is on table.
B is on table.
E is on B.
C is on A.
C is heavy.
D has top clear.
E has top clear.
E is heavy.
C is iron made.
D is on C.

Rules:
Every big, black block is on a red block.
Every heavy, iron block is big.
All blocks with top clear are black.
All iron made blocks are black.

References

1. Clocksin WF, Mellish CS (2009) Programming in polog, 3rd edn. Narosa, New Delhi
2. http://www.dtic.upf.edu/~rramirez/lc/Prolog.pdf. Cited 19 Dec 2017
3. http://www.gprolog.org/. Cited 19 Dec 2017
4. Ivan B (2007) PROLOG programming for artificial intelligence, 3rd edn. Pearson Education
5. Strong John S (1989) The Legend of King Asoka—a study and translation of Asokavadana.
Princeton University Press, Princeton, N.J.
6. Kowalski RA (1979) Algorithm = Logic + Control. Commun ACM 22(7):424–436
7. Van Emden MH, Kowalski RA (1976) The semantics of predicate logic as a programming
language. J ACM 23(4):733–742
8. http://www.swi-prolog.org/. Cited 19 Dec 2017
Chapter 6
Real-World Knowledge Representation
and Reasoning

Abstract Apart from its theoretical significance, the AI must represent real-world
knowledge, and produce reasoning using that. The real-world things are collections
of entities in different classes. This chapter presents the representations structures for
such knowledge, e.g., taxonomies and reasoning based on that. Other phenomena
in real-world, that are presented are, action and change, commonsense reasoning,
ontology structures for different domains, like, language, and world. The Sowa’s
ontology for objects, and processes, both concrete and abstract, is explained. The
situation calculus is presented in its formal details, along with worked exercises.
The more prevalent real-world reasoning, like nonmonotonic and default reasoning
are also treated in sufficient details, along with supporting worked exercises. This is
followed, with summary of the chapter, and exhaustive list of practice exercises.

Keywords Real-world knowledge · Ontologies · Ontological reasoning · Sowa’s

ontologies · Situation calculus · Nonmonotonics reasoning · Default reasoning

6.1 Introduction

Who is taller, the father or his son held in his hands by the father? Can you make tea
out of salt? If you push a needle into a potato, does the needle make a hole in the
potato or in the needle? We may find these questions as absurd, but many tasks, such
as machine vision, natural language processing, planning, and reasoning in general,
requires the same kind of real-world knowledge and capabilities of reasoning. As
another example, if a six feet tall person is holding a two feet baby in his arms, and
it is also given that they are father and son, we take no time to conclude which one
is father and which one is son.
The use of real-world knowledge for natural language processing (NLP), partic-
ularly for all kinds of sense disambiguation, and for machine translation is a chal-
lenging task. Some of the ambiguities are resolved using the simple rules, while the
others can be resolved using rich understanding of the world. Consider the following
two examples:

© Springer Nature India Private Limited 2020 143

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_6
144 6 Real-World Knowledge Representation and Reasoning

“The Municipal council of the city refused permission to demonstrators because

they feared violence,” versus,
“The Municipal council of the city refused permission to demonstrators because
they advocated violence.”
In the first sentence it is required to determine that “they” refers to “Municipal
council of the city” and not to “demonstrators.” But, in the second sentence it is
required to find out that, “they” refers to “demonstrators.” The above determination
requires the knowledge about the characteristic relations of “Municipal council of
city” and “demonstrators” with the verbs “fear” and “advocated.” In NLP, such
ambiguities in the language can be resolved by true understanding of the text, which
requires bringing of real-world knowledge in the system.
Considering computer vision, we are able to recognize the objects because of
context. For example, the bowl, banana, knife in kitchen are for container, for eating,
and to cut vegetables, respectively. But, the same items in a departmental store are for
sale. So, it is relation with other objects/environment, like kitchen and shop, explain
the meaning of objects [1].
Learning Outcomes of this Chapter:

1. Compare and contrast the most common models used for structured knowledge
representation, highlighting their strengths and weaknesses. [Assessment]
2. Identify the components of non-monotonic reasoning and its usefulness as a rep-
resentational mechanism for belief systems. [Familiarity]
3. Compare and contrast the basic techniques for representing uncertainty.
[Assessment]
4. Compare and contrast the basic techniques for qualitative representation.
[Assessment]
5. Apply situation and event calculus to problems of action and change. [Usage]
6. Explain the distinction between temporal and spatial reasoning, and how they
interrelate. [Familiarity]

6.2 Taxonomic Reasoning

Taxonomy is a name given to collection of individuals, categories, and the relations

between their pairs. The taxonomies are also referred as semantic networks. As an
example, Fig. 6.1 shows a taxonomy of some categories of animals, individuals, and
the relations between categories and individuals, as well between category and its
subcategories.
There are three basic relations types [2]:
• A category is a subset of other category. For example, the categories dog and cat
are subsets of another category mammal;
6.2 Taxonomic Reasoning 145

Animal

Mammal Furry Bird CanFly

Dog Cat Robin Penguin

Bruno Tweety

Node types: Link types:

Category Subcategory

Instance
Individual
Propertyof

Property Cancel

Fig. 6.1 Taxonomy of animal world

• An individual can be an instance of a category. For example, an individual called

Bruno is an instance of a category dog; robin and penguin are subsets (subcate-
gories) of bird;
• Two categories are always disjoint. For example, dog and cat are disjoint sets.
A property can be used a tag for a category. For example a category Mammal can
be tagged with property furry, and another category bird can be tagged with property
Canfly.
One form of inference from taxonomical structure is through the relation of tran-
sitivity. In the taxonomy shown in Fig. 6.1, Bruno is shown as an instance of dog,
and dog is a subset of mammals, then it can be reasoned that Bruno is an instance of
mammal.
The other form of inference—the inheritance, works as follows:
Bruno is an instance of dog,
Dog is subset of mammal,
Mammal has property furry,
Therefore, it can be inferred through inheritance that dog and Bruno have property
furry.
146 6 Real-World Knowledge Representation and Reasoning

A variant of this inheritance is default inheritance, which have following charac-

teristics:
1. A category can be marked with a characteristic but cannot be marked as a universal
property, and
2. A subcategory or instance will always inherit property from higher order category
unless otherwise it is specifically negated.
For instance, there is a category bird, with property Canfly. And, there are subcate-
gories Robin and Penguin. The property Canfly should be inherited by sub-category
Robin, but not by Penguin. This is implemented by negating the Canfly property
for Penguin, as shown in the simple and standard taxonomy of the animal world in
Fig. 6.1, where every two categories are related in such a way that, they are either
disjoint or one is subcategory of other.
Other taxonomies structures are less straight forward. Consider a semantic net-
work for categories of people, the individual say Dr. C. V. Raman, is having following
relations with other categories, as an instance of those categories. He is simultane-
ously a Physicist, a Nobel-laureate, a ScientistIndian, and NativeofMadras. Also,
there is an overlap, and it is some times not clear as which are the properties, and
which are taxonomic categories. Hence, in taxonomizing more abstract categories,
choosing and delimiting the categories become a challenging task.
Number of specialized taxonomies have been developed in the domains such
as medicine, geonomics, languages, and other fields. The more sophisticated tax-
onomies are based on description logic, which provides the tractable constructs for
describing concepts and the relation between them. These have been applied in
semantic OWL (Web Ontology Language).
Temporal Reasoning
It provides the representation of knowledge and automating reasoning about time,
duration, and time intervals. For example, in the sentence, “After landing at Frankfurt,
it took a long time, before I could board the connecting flight to New Delhi” versus
“It for the two countries a long time before they could arrive to trade treaty.” In the
first sentence, “long time” may mean few hours, while in the second same phrase
may mean many years, or even decades. Integrating such reasoning, with natural
language (NL) interpretation, has been problematic. Many temporal relations are not
explicitly stated in text, and they are complex and text dependent.
Action and Change
Theory of action and change is other area of commonsense reasoning, and has been
well understood. Following are the constraints of representation and reasoning for
the domains of action and change:
• Events are atomic, that is, at a time only one event occurs. The reasoning will take
place based on the state of the world at the beginning, and at the end of the event.
The intermediate states while the event is in execution are not accounted for.
• Each and every change of the world is due to occurrence of some event.
6.2 Taxonomic Reasoning 147

• All the events are deterministic in nature, i.e., state of the world at the end of every
event can be fully determined using the state before that event, plus this event
itself.
The problem of representation, and the form of reasoning such as prediction
and planning, have been largely understood for the domain that satisfy the above
mentioned three constraints. However, there are other extreme domains: continu-
ous domains, simultaneous events, probabilistic events, and imperfect knowledge
domains, where, the representation and reasoning are not so well understood yet.
Some of these domains will be discussed in the following text.

6.3 Techniques for Commonsense Reasoning

The field of commonsense reasoning can be largely divided into two areas:
• knowledge based approaches, and
• machine learning approaches that cover large data corpora, which are almost
always of the text corpora types.
However, there is a very limited interaction between these approaches. There
are other approaches, e.g., crowd-sourcing based approaches, which try to construct
knowledge base as combination of other knowledge bases and participation of many
non-experts [2].
The knowledge based approaches are further divided into categories that are based
on following domains (see Fig. 6.2).
1. Mathematical logic or some other mathematical formalism,
2. Informal approaches, which are in contrary to mathematical formalism,
3. Based on theories from cognition, and
4. There are large-scale other approaches, which are usually mathematical or infor-
mal, but mainly targeted at collecting a large amount of knowledge.

Commonsense
Reasoning

Web Mining Knowledge- Crowd-

based sourcing
KnowItAll ConceptNet
OpenMind

Mathematical Informal Large-scale

Situation calculus. Scripts, Frames, CYC

Case-based reasoning

Fig. 6.2 Taxonomy of approaches to commonsense reasoning

148 6 Real-World Knowledge Representation and Reasoning

One of the successful form of a commonsense reasoning, which has mathematical

base is qualitative reasoning (QR). The QR helps in automating the reasoning and
problem solving about the physical world around us. It has also generated techniques
and systems that are being commonly used in application domains, like autonomous
support for spacecraft, on-board diagnosis of vehicular systems and their failure anal-
ysis, automated generation of control software for photocopiers, and intelligent aids
for learning about thermodynamic cycles. There are number of prominent features
that are typical for QR systems: ontologies, causality, and inference of behavior from
structure.
The other type of reasoning is quantitative reasoning, which deals with quantities,
and reasons on the criteria of high/low, big/small, long/short etc.
The commonsense reasoning targets a number of different objectives, some of
these are as follows:
• Plausible inference. It is used for drawing provisional/uncertain conclusions.
• Reasoning architecture. The reasoning architecture is a general-purpose data struc-
tures to encode knowledge and algorithms, which are helpful for carrying out the
reasoning. For example, to represent the meaning/senses of natural language sen-
tences.
• Basic domains. While performing the commonsense reasoning, human can do
complex reasoning about basic domains like space, time, basic physics, and basic
psychology. The knowledge we as human draw from these domains is largely not
fully documented and the reasoning processes are largely unavailable for intro-
spection. An automated reasoning should comprise all these reasoning capabilities.
• Reasoning modes. The commonsense reasoning system should incorporate a
variety of modes of inference, like, explanation-based reasoning, generalization,
abstraction, analogical-based reasoning, as well be able to perform the simulation.
• Independence of experts. Hand-coding of a large knowledge base for commonsense
reasoning is expensive and slow process. Hence, assembling it either automatically
or building it by drawing the knowledge of non-experts, e.g., through crowd-
sourcing, is considered as better choice.
• Breadth. To perform a powerful commonsense reasoning through machines, will
require a large body of knowledge.
• Cognitive modeling. These are the theories of automated commonsense reasoning
that describe the commonsense reasoning in people.
• Applications. To be useful, a commonsense reasoner must interface with applica-
tions smoothly, and must serve the needs of applications.

6.4 Ontologies

Ontology is a particular theory about the nature of being or the kinds of existent.
The task of intelligent systems in computer science is to formally represent these
existents. A body of formally represented knowledge is based on conceptualization.
6.4 Ontologies 149

A concept is some thing we can form a mental image of. Conceptualization consists
of a set of objects, concepts, and other entities about which knowledge is being
expressed and of relationships that hold among them. Every knowledge model is
committed to some conceptualization, implicitly or explicitly [3].
The major applications of ontologies are in natural language processing, infor-
mation retrieval, modeling and simulation. The CYC (enCYClopedia) ontology, for
example, is used for a CYC natural language system, built for the purpose of trans-
lating natural language text into cycL [4].
An ontology is a systematic arrangement of all of the important categories of
objects or concepts that exist in some field of discourse, that shows the relations
between these concepts/objects. In a completed form, an ontology is a categorization
of all of the concepts in some field of knowledge, that includes all the objects,
their properties, relations, and functions needed to define the objects, and specify
their actions. A simplified ontology may contain only a hierarchical classification
of objects and processes (called taxonomy), which shows the subsumption relations
between concepts in the specific domain of knowledge [5].
An ontology may be visualized as an abstract graph with nodes representing the
objects and labeled arcs representing the relations between them, as shown in Fig. 6.3,
which is the upper level of the CYC (from enCYClopedia) project hierarchy.
The CYC Project is an example of ontology. CYC contains more than 10,000
concept types used in the rules and facts encoded in the knowledge base. Its goal
was to build a commonsense knowledge base containing millions of facts and their
interrelationships. It was the efforts to build the knowledge in it that is seldom written
down—the knowledge for example, the reader of an encyclopedia is assumed to pos-
sess to understand the contents of encyclopedia. This kind of knowledge is required
as built-in in a system, which are required to read the restricted natural language.
A concept is some thing we can form a mental image of. At the top of the hierarchy
is the Thing “concept”, which does not have any properties of its own. The hierarchy
under Thing is quite tangled. Not all the subcategories are exclusive. In general, Thing
is partitioned in three ways: First is Represented Thing versus Internal Machine
Thing. Every CYC category must be an instance of one and only one of these sets.
Internal-Machine-Thing is anything that is local to the platform CYC is running on
(strings, numbers, and so on). Represented-Thing is everything else.
The concepts used in an ontology and hierarchical ordering are to some extent
arbitrary, and largely depends on the purpose the ontology is created. This is because,
same objects are of varying importance when used for different purposes, and one or
other set of properties of objects are chosen as the criteria for classification of these
objects. Apart from this, different degrees of aggregation of concepts may be used,
and the importance for one purpose may be of no concern for a other purpose. For
example, class is collection of students, and when all of them are playing in ground
we call it a team. As other example, when both parents are teaching they are called
teachers, when traveling together they are passengers, and at home they are spouse.
150 6 Real-World Knowledge Representation and Reasoning

Thing

Individual Represented
Intangible thing
object

Event Stuff Intangible Collection

object
Relationship

Attribute Slot
Process Intangible
value
stuff
Atrribute
Something
existing
Internal machine
thing

Tangible
Intelligence
object

Tangible
stuff

Composit tangible &

Intangible Object

Fig. 6.3 World ontology

6.5 Ontology Structures

When compared with First-Order Predicate Logic (FOPL) for knowledge represen-
tation and reasoning, in the ontologies we are particular for knowledge organization
as well as contents. This approach has generalization, combined with exceptions.
For example, “all birds fly” can be a rule. But, there are some birds which do not fly,
for example, penguin. This is an exception. We should be able to add this concepts
of knowledge, not as exception, but as extension. This requires categorization of
objects.
Ontologies provide explicit representations of domain concepts busing a structure
around which knowledge base is built. Each ontology is a structure of concepts
and relations between them, such that all concepts are defined as well interpreted
using a declarative approach. The ontology system also defines the vocabulary of the
problem domain, and a set of constructs as how the terms can be combined to model
the domain. The Fig. 6.3 shows a general structure of ontology.
6.5 Ontology Structures 151

Fig. 6.4 Language, logic, Language Common-sense

and ontology

Ontology Type-cheching

Logic Reasoning

6.5.1 Language and Reasoning

The Fig. 6.4 shows the interrelationship of language and logic with ontology, and
how each of these are related to commonsense, reasoning, and type checking. The
language, ontology, and logic are representations, where as the corresponding blocks,
i.e., common-sense, type-checking, reasoning, are respectively, the applications.
Consider the object “ball”, we identify it as a member in the collection of balls
by its type checking; we say “plants have life”, they belong to the category of all
the living beings (type-check). At lowest level in Fig. 6.4 is logic, making use of
binary operators. This logic helps in reasoning process. The common-sense is con-
text level knowledge, which is essential for language understanding. For example,
the phrase—“a good play”, while in a cricket ground is understood as good shot,
while in a theater, it is taken as performance by an artist. We resolve the phrase “a
good play” by context.
There is a close relationship between natural language understanding (NLU) and
knowledge representation and reasoning. In other words, the relationship between
language and knowledge. In fact, understanding is reasoning paradigm, as it has
become quite clear that understanding natural language is, for the most part, a com-
monsense reasoning process at the pragmatic level. As an example, to illustrate this
strong interplay between language understanding and commonsense reasoning, con-
sider the following:
(a) “Jack defended a jailed activist in every state.”
(b) “Jack knows a jailed activist in every state.”
Through commonsense, we have no difficulty in inferring for (a) that Jack supports
for the same activist in every state. Perhaps jack might have traveled to different states
for campaign. Where as in (b), we hardly conceive of the same. It may mean, “there
is a jailed activist in every state”, or an activist being jailed in every state. Such
inferences lie beyond syntactic and semantic explanations, and are in fact depend on
our commonsense knowledge of the world (where we actually live).
As another example, consider resolving of the noun He in the following:
(c) Rajan shot a policeman. He instantly
(i) ran away.
(ii) collapsed.
152 6 Real-World Knowledge Representation and Reasoning

It is obvious that the above references can only be be resolved through common-
sense knowledge. For example, typically, when shot (x, y) holds between some x
and some y, the x is the more likely a “subject” which would run away. So, in (c(i)),
most likely its is Rajan who ran away. In c(ii), y is the more likely to collapse, that
is the policeman. Note, however, that such inferences must always be considered
defeasible, since quite often additional information might result in the retraction of
previously made inferences. For example, (c(ii)) might be describing a situation in
which Rajan, a ten-year old who was shooting for practice, fell down. Similarly, and
(c(i)) might actually be describing a situation in which the policeman, upon being
slightly injured, tried to run away, may be to escape further injuries!
There are number of challenges, in term of computational complexities, in rea-
soning with uncommitted (or underspecified) logical forms. However, even bigger
challenge is to make available of large body of commonsense knowledge, with com-
putationally effective reasoning engines.

6.5.2 Levels of Ontologies

In comparing various ontologies, they can be viewed at three different levels:

1. is-a taxonomy of concepts,
2. the internal concept structure and relation between concepts, and
3. the presence or absence of explicit axioms.
The Fig. 6.5 shows various concepts related by is-a hierarchy relation (a mem-
ber of relation), also subset relation. A member can be an instance of a category.
Taxonomy is central part of most ontologies, and its organization can vary greatly.
For example, all concepts can be in one large taxonomy, or there can be number of
smaller hierarchies, or there can be no explicit taxonomy at all.
Although all general-purpose ontologies try to categorize the same world, they
are very different at the top level. They also differ in their treatment of basic parts:
things, processes, and relations (see Fig. 6.3).
The next level of comparison is internal concept structure, which can be realized
by properties and roles. The internal concept relations are the relations, for example,
between bird versus canfly, wings, and feathers (see also Fig.6.5). Concepts in some
ontologies are atomic (axioms) and might not have any properties or roles or any
other internal structure associated with them.
An important test for any ontology is the practical applications it is used for.
These can be the applications in natural language processing, information retrieval,
simulation and modeling, and so on, that use knowledge represented in the ontology.
6.5 Ontology Structures 153

Fig. 6.5 Top level ontology for animal

6.5.3 WordNet

The WordNet is a well developed lexical ontology—a manually constructed online

lexical database system, provided with open access and semantically organized lex-
ical objects. Its structure is provided with capability to distinguish between various
parts-of-speech: like nouns, adjectives, verbs and adverbs. The basic object of Word-
Net is synset, which is a set of synonyms. Hence, if a word has more than one senses
(e.g., the word “bank” for “money”, and “bank” for “river bank”), it will appear
in more than one synset. The synsets are organized in a hierarchy as super-class
(hypernyms) and subclass (hyponyms). The Fig. 6.6 shows part of wordnet hierar-
chy of tangible things, with braces enclosing concepts in same synset. Note that in
the synset {living thing, organization}, each of its subordinate item has two parts,
e.g., {plant, flora}, but in case of {nonliving thing, object} there is only one item (the
object name) in each of the subordinate, because the non-living thing is not classified
thing [5, 6].
The Wordnet is a taxonomical architecture, which does not have more elementary
concepts or axioms. For each concept (i.e., synset) provided in the Wordnet, there
is a pointer to nouns representing its parts. For example, the parts of concept “bird”
may be feathers, beak, and tail. There are provisions in the Wordnet for other types
of pointers, for example, from noun to verb, to represent functions (actions), or from
noun to adjective to represent properties of nouns.
154 6 Real-World Knowledge Representation and Reasoning

{Thing, Entityt}

{non-living thing, object} {living thing, organization}

{natural object} {substance} {plant, flora} {animal, fauna}

{food} {person, human being}

{artifact}

Fig. 6.6 WordNet ontology representing subclass relation among synsets

6.5.4 Axioms and First-Order Logic

Structurally, an ontology is collection of terms, their definitions, and axiom relating

them. The terms are typically organized as taxonomy. In some cases axioms are
central to ontology design, while in other cases the axiomatic language is associated
with central concepts.
Apart from the taxonomy and structure of concepts, axioms are the base for
representing more information about categories and their relations. In addition, the
axioms specify constraints on properties and role values for each category. Some
times, axioms are specifically specified, while other times ontology consists of cat-
egories and corresponding frames, and every thing else in hidden in the application
code. There is a fine-line difference between internal concept structure and axioms.
The category is represented using a frame formalism, and roles and properties
are slot of the frame. The same facts can also be represented using axioms, i.e., a
taxonomy can also be represented using axiomatic notations. In this notation, the
axioms are represented using first-order predicate logic. For example, “all persons
are living being and all living being are things”, is represented Fig. 6.6, as well by
the following Eq. 6.1.

∀x∀y[( per son(x) → livingbeing(x)) ∧ (livingbeing(y) → thing(y))] (6.1)

The CYC project, for example, comprised among the largest number of axioms,
i.e., in the order of 106 .

6.5.5 Sowa’s Ontology

John Sowa (1997, 1995) stated his fundamental principles for ontology design as
“distinctions, combinations, and constraints” (see Fig. 6.7). He uses philosophical
motivation as the basis for his categorization. There are three top-level distinctions:
6.5 Ontology Structures 155

Fig. 6.7 Sowa’s ontology Top Level

Object Process

Contrete Abstract

Physical Informatio
Process Object

Physical Information
Object Process

Recipe Recipe Recipe Computer

preparation Dpcument Program

First is Physical versus Information, or Concrete versus Abstract. This is a disjoint

partition of all the categories in the ontology [5].
The second principle for ontology design is based on combinations. The combi-
nations classify the objects into firstness versus secondness versus thirdness, or Form
versus Role versus Mediation. These categories are not mutually exclusive. For exam-
ple, Woman is considered to be a form (firstness) because it can be defined without
considering anything outside a person. As a mother, a teacher, or an employee, the
same individual would be an example of a role (secondness). These roles represent
an individual in relation to another type (a child, a student, an employer). Marriage
is a mediation (thirdness) category because it relates several (in this case, two) types
together, lecture class of AI is mediation as it relates many students as a type together.
The third principle is based on constraints. It is, continuant versus occurrent,or
object versus Process. Continuants are objects that retain their identity over some
period of time; occurrents are processes “whose form is in the state of flux”. For
example, Avalanche is a process, and Glacier is an object. Note that this distinction
depends on the time scale. On a grand time scale of centuries, Glacier is also a
process. These distinctions are combined to generate new categories (Fig. 6.7). At
a lower level, for example, Script (a computer program, or a baking recipe) is a
form that represents sequences and is thus defined as Abstract, Form, Process. Also,
History (an execution of a computer program), which is a proposition that describes
a sequence of processes, is then Abstract, Form, or object [7].
In the John Sowa’s ontology, top level category is for every possible combination
of distinctions. Constraints are used at the lower levels to rule out categories that
cannot exist, as well it avoids too many combinations of categories. For example,
logical constraints, would rule out triangles of more than three sides, and empirical
constraints would rule out birds that talk or think. Representation of constraints is
in the form of axioms, inherited through the ontology hierarchy to the lower levels.
156 6 Real-World Knowledge Representation and Reasoning

6.6 Reasoning Using Ontologies

We have discussed in the above that, ontologies are taxonomical structures of cate-
gories, subcategories, and members of concepts. As human, we identify the objects,
whether physical or abstract, due to this classification. For example, we can distin-
guish an oranges from cricket balls, because in the first the surface property belong
to softness, and spongy, though their size and shape are quite similar. The Structures,
e.g., a University System, a building, a computer network, operating system, or say,
even office chair, are decomposed into parts, so that we identify the object as a whole
as well the relationship of the parts remain explicit in this representation. Hence, it is
due to these associated categorizations, and sub-categorizations, along with restric-
tions we are able to identify the real world objects. The reasoning can be done using
categories and objects, using physical decomposition of categories, using measuring
criteria, or using object oriented analysis.

6.6.1 Categories and Objects

The taxonomical hierarchies are common in government, military, and other organi-
zations. Classifying objects into categories is important for knowledge organization.
Of course interactions take place at individuals levels but relations are created at the
level of categories. For example, the sentence, “Basketball team has won”, does not
refer to a single ball and nor a single player, but a relationship among team members
as a group.
The Category also helps in prediction of objects once they are classified. One refers
the objects from perceptual inputs, infers category of membership from perceived
properties of the objects, then uses category information to predict about object. For
example, from its yellow colour, size, shape, we identify that an object is a ‘Mango’.
In FOPL there are two choices for categories: predicates and objects. Following
are the examples:
basketball(b);
member (b, basketballs);
b ∈ basketballs; (the object b is a member of category basketball)
subset (basketballs, balls); (category is subclass of another category)
basketballs ⊂ balls.
dogs ∈ domesticatedanimals. (categories are members)
The category is defined by members and subset relation. We also note that cate-
gories organize the knowledge as well as they help to inherit. Thus, ‘mango’∈ ‘man-
goes’, and ‘mangoes’⊂ ‘fruits’. Thus, mango inherits the properties of ‘taste’from
fruits (fruits have ‘tastes’). Similarly, we have categories: animals, birds, foods,
institutions. A “set of institutions” ⊂ “institutions”, and MBM ∈ “institutions”. The
subclass organize the categories into a taxonomic hierarchy.
6.6 Reasoning Using Ontologies 157

Following are the examples of reasoning through taxonomies:

(x ∈ basketballs) ⇒ r ound(x); (all members of a category have same property
of roundedness)
color (x, br own) ∧ r ound(x) ∧ dia(x) = 9” ∧ x ∈ balls ⇒ x ∈ basketballs;
(the member category can be recognized by some property)
Having represented in taxonomy, when a top element is identified having some
property, those its subordinates can be easily identified, as they possess similar prop-
erties, unless there is an exception.

6.6.2 Physical Decomposition of Categories

An object can be part of another, for example, nose, eyes, hands, are part of body;
steering is part of car, wheels are part of wheel-assembly, and wheel-assembly is part
of car. This can be represented by relations of physical decompositions,
par to f (wheel, wheelassembly).
par to f (wheelassembly, car ).
par to f (x, y) ∧ par to f (y, z) ⇒ par to f (x, z).
Note that taxonomies are transitive relations. The relation of reflexivity also holds
on individual objects, e.g., par to f (x, x), as x has one part, and that part is itself.
The categories of composite objects is defined by structural relations between parts
and assemblies.

6.6.3 Measurements

The physical objects have height, weights, mass, length, and other physical measure-
ments. To characterize a physical object, values need to be assigned to the objects in
the form of their properties. Following are the examples:
length(i pad) = inches(5) or length(i pad) = centimeter s(12.5).
list price(basketball) = r upees(500).
height (x) > height (y) ⇒ taller (x, y).

6.6.4 Object-Oriented Analysis

The ontology development processes are similar to that of object-oriented design and
analysis. In both, domain vocabulary is collected in the beginning, which is often
out of domain’s generic nouns, verbs, and adjectives. The result of object-oriented
158 6 Real-World Knowledge Representation and Reasoning

analysis is in the form of a draft document for domain ontology and relevant to
the application. However, analysts do not call the result as ontology. A designer in
an object-oriented system defines the things like, objects, classes, interface func-
tions, hierarchies, and the system behavior. But, an ontological engineer make use of
intermediate representations such as graphs, tables, and semantic networks, to design
hierarchies and other concept relationships, which hold true for object oriented design
also.
In object oriented analysis, the classes act as templates. Both types of specialists
make use of templates to specify product details. The classes are then merged or
refined, as with ontologies.

6.7 Ontological Engineering

The knowledge engineering is process of knowledge representation—identification

of task, assemble relevant knowledge, decide vocabulary of predicates, functions and
constructs, encode knowledge about domain, encode description of specific problem
instance, pose queries to procedures and get answers and debug knowledge base.
The ontological engineering can be defined on the same line of knowledge engi-
neering. It is related to the creating representation of general concepts—actions, time,
physical objects and beliefs.
Ontological engineering comprise a set of activities conducted during the time of:
conceptualization, design, implementation and deployment of ontologies. Ontologi-
cal engineering covers topics of philosophy, metaphysics, knowledge representation
formalisms, development methodology, knowledge sharing and reuse, knowledge
management, business process modeling, common-sense knowledge, systematiza-
tion of domain knowledge, information retrieval from the Internet, standardization,
and evaluation. It also gives us design rationale of a knowledge base, helps define
the essential concepts of the world of interest, allows for a more disciplined design
of a knowledge base, and enables knowledge accumulation. In practice, knowledge
of above mentioned disciplines helps to:
• Organize the knowledge acquisition process;
• Specify the ontology’s primary objective, purpose, granularity, and scope; and
• Build its initial vocabulary and organize taxonomy in an informal or semi-formal
way, possibly using an intermediate representation.
Special-purpose languages/tools used for implementing ontologies, such as
Ontolingua, CycL, and LOOM (lexical OWL ontology matcher), use a frame-based
formalism, a logic-based formalism, or combination. Ontolingua is a frame-based
language that uses KIF (Knowledge Interchange Format). It is a language for publica-
tion and knowledge communication, with notation and semantics like some extended
form of first-order predicate logic. Ontolingua enables writing knowledge-level spec-
ifications independent of particular data or programming languages, and can translate
a knowledge base from one representation language into another.
6.7 Ontological Engineering 159

The LOOM takes two ontologies represented in OWL and produces a pair of
relatedconcepts from the ontologies. In order to identify the corresponding concepts,
LOOM compares the preferred names and symbols of the concepts in both ontologies.
It identifies these concepts as similar, if and only if, their preferred names or synonyms
are equivalent based on modified string functions. The string comparison removes the
delimiters from both strings, then uses approximate matching techniques to compare
them, allowing for mismatch of at most one character in a specified length.
Some languages/tools used in building ontologies are based on translation
approaches. Using these approaches it is possible to build ontologies directly at
knowledge level, and knowledge level is translated into the implementation level,
thus eliminating the need to master the implementation languages [8].

6.8 Situation Calculus

The concept of action arises in at least two major subareas of artificial intelligence,
(1) natural language processing and (2) problem solving. For the most part, the
formalisms that have been suggested in each sub-area are independent of each other
and difficult to compare. However, so far there does not exists a computational
theory of actions, which is powerful enough to capture the range of the meanings
and distinctions expressible in any natural language. The primary goal of situation
calculus is to provide a formalism which is expressive enough for representation of
actions and to explore its use in defining the meanings of English language sentences
that describe actions and events [9].
The requirement on the formalism for representation of actions is that it should
be a useful representation for action reasoning (i.e., problem solving). It has a very
useful application, i.e., to describe, how this representation could be used for planning
(of actions) or for plan recognition system. This is essential to the natural language
processing as well, because the natural language understanding is aimed to ultimately
result to problem-solving capability.
The situation calculus is also the language of choice for investigations of various
technical problems that arise in theorizing about actions and their effects. It is being
taken as a foundation for practical work in planning, control, simulation, database
updates, agent programming and robotics. In parallel with these developments of
its applications, there have emerged axiomatizations for the situation calculus, and
explorations of some of their mathematical and computational properties [10].

6.8.1 Action, Situation, and Objects

The situation calculus is a logic formalism for representing and reasoning about
dynamical domains. The concepts in the situation calculus are situations, actions
and fluents. Number of objects are typically involved in the description of the world.
160 6 Real-World Knowledge Representation and Reasoning

The situation calculus is based on a sorted domain with three sorts: actions, situations,
and objects, where the objects include everything that is not an action or a situation.
Hence, domain D is,
D = A, S, O (6.2)

where A is set of actions, S is set of situations, and O is set of objects, with variables
of all sort can be used. The actions, situations, and objects are elements of the domain,
but fluents are modeled either as predicates or as functions. The phrase, e.g., sorted
domain of actions means, that set of actions are to be carried out in a certain (sorted),
and similar meaning exists for sorted situations, and sorted fluents.
A situation is a kind of state (ako), however it is a result of chain of earlier
situations. Actions are what make the dynamic world change from one situation to
another when performed by agents. A fluent is a condition that can change over time.
Fluents are situation-dependent functions used to describe the effects of actions.
There are two kinds of them: relational fluents and functional fluents. The former
have only two values: True or False, while the latter can take a range of values.
For instance, one may have a relational fluent called handempty which is true in a
situation if the robot’s hand is not holding anything. We may need a relation like this
in a robot domain. One may also have a functional fluent, e.g., battery-level, whose
value in a situation is an integer of value between 0 and 100 denoting the total battery
power remaining in one’s notebook computer.
The set of actions form sort of domain. The action can also use variables, and
they can be quantified, for example in a robot world, possible action terms would
be move(x, y) to model a robot moving to a new location (x, y), and pickup(o) to
model the robot picking up an object o, and so on.

6.8.2 Formalism

A formalism used in situation calculus is aimed to implement the commonsense

reasoning, such that the information used by humans can be expressed in sentences
in logical order, then it is made part of the database (knowledge base). As next step, a
goal oriented program will consult these databases for the required facts to ultimately
achieve the goal. The database stores facts about what are the effects for any given
actions, for example, a set of facts and effects concerning to a robot which moves
objects from one location to another location. The situation calculus is designed
to derive the actions (effects) for a given set of facts, independent of any specific
problem.
The major part of the database are facts about effect which resulted due to actions,
e.g., action of a robotic arm. These consequences of actions in situation calculus are
of general nature, and not bound to specific problem, hence they can be used for
solution of variety of problems, without creating new databases for new instances of
problems.
6.8 Situation Calculus 161

In the situation calculus, a dynamic world is modeled to progress through a series

of situations as a result of various actions being performed within this world. A
situation is a consequence of sequence of action’s occurrences. The situation available
before any actions are performed is generally denoted by S0 , called initial situation.
A new situation resulting from the performance of an action is denoted using the
function symbol result or do. This function symbol has a ‘situation’and ‘action’as
arguments, and a new situation as a result due to performing the given action in the
given situation. We shall make use of do, called binary function symbol expressed
as,
do : action × situation → situation. (6.3)

The intended interpretation is that do(a, s) (or r esult (a, s)) denotes the successor
situation resulting from performing action a in situation s. Accordingly, the basic
formalism of the situation calculus is represented by:

s
= do(e, s) (6.4)

which asserts that s

is the resulting situation when event e (action) occurs in situation
s.
The basic ontology of situation calculus consists of:
1. situations, which corresponds to snapshots of universe or an instant of time, and,
2. actions or events, which change the world from one state to another.
It is a sorted language with sorting (order) for situation and actions.
The fluent “iscarr ying(o, s)” can be used to indicate that the robot is carrying a
particular object ‘o’in a particular situation ‘s’. If the robot initially carries nothing,
“iscarr ying(Ball, S0 )” is false while, the fluent

iscarr ying(Ball, do( pickup(Ball), S0 ))

is true. The location of the robot can be modeled using a functional fluent location,
which returns the location coordinates ‘(x, y)’of the robot in a particular situation.
To describe a dynamic domain using the situation calculus, one has to decide on
the set of actions available for the agents to perform, and the set of fluents needed
to describe the changes these actions will have on the world. For example, consider
the blocks world where some blocks of equal size can be arranged into some set of
towers on a table. The set of actions in this domain depends on what the imaginary
agent can do. If we imagine an agent to be a robot-arm that can be directed to grasp
any block that is on the top of a tower, and either add it to the top of another tower
or put it down on the table to make a new tower, then we can have the following
actions:
stack(x, y)—put block x on block y, provided that robot is holding x, and y’s
top is clear. Being action, it is read “stack x on y”, and shall not be confused with
predicate.
162 6 Real-World Knowledge Representation and Reasoning

unstack(x, y)—pick up block x from block y, provided that robot’s hand is empty,
x is on y, and x has top clear.
putdown(x)—put block x down on the table, provided that robot is holding x.
pickup(x)—pick up block x from the table, provided the robot’s hand is empty,
x is on the table and top is clear.
move(x, y)—move block x to position y.
To describe the effects of these actions, we can use the following relational fluents:
handempt y—True in a situation if the robot’s hand is empty.
holding(x)—True in a situation if the robot’s hand is holding the block x.
on(x, y)—True in a situation if block x is on block y.
ontable(x)—True in a situation if block x is on the table.
clear (x)—True in a situation if block x has top clear.
For action stack(x, y) to be performed in a situation, the fluent,
– holding(x) must be true, and
– clear (y) must be true.
Also, after stack(x, y) is performed and it results to a new situation, the fluent,
– on(x, y) will be true,
– handempt y will be true,
– holding(x) will be false, and
– clear (y) will be false.
The action stack(x, y), along with present value of fluents, and resulting action
with new value of fluents can be formally expressed as an axiom,

∀x∀y[[((holding(x) ∧ clear (y)) → (stack(x, y)] → on(x, y)

∧ handempt y ∧ ¬holding(x) ∧ ¬clear (y)).] (6.5)

Now, for example, we can say that for action stack(x, y) to be performed in
a situation, holding(x) and clear (y) must be true, and that after stack(x, y) is
performed, in the resulting new situation, on(x, y) and handempt y both will be
True, and holding(x) and clear (y) will no longer be True.
We only need the action move(x, y) to take place, to move block x to position y.
This can happen if the agent (robot) in its world moves a block from a clear position
to another clear position. The block x is clear to move means that x is on the table
with top clear, or it is top most position in a stake of blocks. Accordingly, move(x, y)
action can be expressed by axiom,

∀x∃y[(clear (x) ∧ clear (y)) → move(x, y)] → on(x, y) ∧ clear (x)]. (6.6)

For describing the effects of this action two fluents are sufficient: clear (x), and
on(x, y). The action move(x, y) can be performed only when the situation “clear (x),
clear (y), x = y” is True. The consequence of this action is that x is no longer at the
place where it was earlier, but instead at location y in the resulting new situation.
6.8 Situation Calculus 163

6.8.3 Formalizing the Notions of Context

The axiomatic system we discussed above is justified only in some specific context.
With a little creative thought it is possible to construct a more general context in an
axiomatic system. However, in such a generalized context the exact form of axioms
will not hold. We can appreciate this argument if we think of human reasoning
apparent in natural language, for example, consider axiomatizing of the adjective
(concept) “on”, to draw correct inferences from the information presented in a English
language sentence, in a world of space-craft on a Mars mission (as context), where
flight crew answers a query of other crew as,

“The book is on the table.”

As critic may propose to suggest about the precise meaning of ‘on’, causing
difficulties about what can be between the book and the table or about how much
gravity there has to be in a spacecraft in order to use the word “on” and whether
centrifugal force counts. If the space-raft is orbiting the planet Mars! Thus, we
encounter a Socratic puzzles over the issue of what the concepts mean in complete
generality and come across examples that never arise in reality—“there simply is not
a most general context”!
On the other hand, if a system is axiomatized at sufficiently higher level (i.e.,
more general), the axioms will be too long to be suitable even in special situations.
Thus, we find it convenient to say, “The book is on the table,” and omit the reference
to exact time when it is on the table (global or local), its precise location on the
table (on which part or corner it is located), and what is location of table in reference
to global (GPS) coordinates, or universal coordinates, etc. Hence, we conclude that
position of book in above example is sufficiently general. How much general the
specification be? This depends on whether that general/commonsense knowledge
has been expressed in the logic used for reasoning, in the corresponding program or
in other other formalism.
A possible solution to the question—“how much general?”, is, formalize the idea
of context, and combine it with circumscription procedure of monotonic reasoning.
This is done by adding context parameters to predicates and functions in the axioms,
because each axiom makes an assertion about certain context. Apart from this, the
axioms should express that, facts are inherited using more restricted context unless
exceptions are specified. Each assertion is assumed to apply nonmonotonically in any
particular more general context, but for that also there are exceptions. For example,
there is rule that says, birds fly. In this rule, we implicitly assume that there is
an atmosphere available. However, in a more general context, this (existence of
atmosphere) is not required to be assumed. Further, it still remains to be determined
as how the inheritance to more general contexts differs from that of more specific
contexts.
There are a some predicates about contexts as well as dealing with time. One of
the most important predicate is holds, which asserts that a property holds (i.e., is
true) during a time interval. Thus, the sentence,
164 6 Real-World Knowledge Representation and Reasoning

holds( p, t)
is true if and only if property p holds during time t. A subsequent axiom will state,
this is intended to mean that p holds at every subinterval of t as well. Note that
if we had introduced holds as a “modal operator” we would not need to introduce
properties into our ontology.
Suppose that, whenever a sentence p is present in the memory of a computer, we
consider it as in a particular context and as an abbreviation for the sentence,
holds( p, C)
where C is the name of a context. We create a relation about the generality of a
context as follows: a relation (≤), e.g., c1 ≤ c2 means context c2 is more general
than context c1 . Using this relation, we allow sentences like,
holds(c1 ≤ c2 , C)
which means that statements relating contexts (c1 ≤ c2 ) can have contexts C.
A logical system using contexts might provide operations of entering and leaving
a context yielding what we might call unnatural deduction allowing a sequence of
reasoning as given in the followings.

holds( p, C)
ENT ER C
p
...
...
q
L E AV E C

This resembles the usual logical natural deduction systems.

Example 6.1 Represent the following sentences in the formalism of situation cal-
culus.

“A man called Rex is standing at the gate of Ghana Bird Sanctuary wearing a black
T-shirt. Rex loads his toy gun, waits for few seconds, and shoots at a migratory
crane.”
The fluents, which are either true or false, in this sentence are:
standing(place): whether Rex is standing at a given place or not?
black: whether Rex is wearing black T-shirt or not?
loaded: whether the gun is loaded or not?
6.8 Situation Calculus 165

Following are the actions in above sentence:

load: Rex is loading the gun.

wait: Rex waits for few seconds.
shoot: Rex shoots his gun.
Now, let us find out what holds at initial situation S0 .
holds(standing(gate), S0 )
holds(black, S0 )
holds(alive, S0 )
¬holds(loaded, S0 )
Now we try to relate actions with situations. In other words, which fluents will
hold after performing a certain action in a given situation?
1. If Rex shoots the gun and gun is loaded, crane is not alive.
2. If Rex shoots the gun, gun will become unloaded.
3. If Rex loads the gun, gun will be loaded.
4. If Rex waits on an loaded gun, the gun remain loaded.
5. If Rex waits on an unloaded gun, the gun remain unloaded.
Our first method for writing above sentences in formal way is as follows. Consider
a general sentence “ f 2 (fluent) will be true by performing action a2 in state s if
(provided that) f 1 is true in s.”

∀s[holds( f 1 , s) → holds( f 2 , do(a2 , s))]

Now consider the first sentence: “if Rex shoots the gun and the gun is loaded,
crane is not alive,” which is written as:
1. ∀s[holds(loaded, s) → ¬holds(alive, do(shoots, s))]
2. The remaining four sentences can be expressed in situation calculus as follows:
3. ∀s[¬holds(loaded, do(shoot, s))]
4. ∀s[holds(loaded, do(load, s))]
5. ∀s[holds(loaded, s) → holds(loaded, do(wait, s))]
6. ∀s[¬holds(loaded, s) → ¬holds(loaded, do(wait, s))]
Having represented in this form of FOPL, the reasoning is done using conventional
methods used for predicate logic reasoning, e.g., resolution based theorem proving.

6.9 Nonmonotonic Reasoning

The most logic we have studied so far falls in the category of monotonic logic, and
the corresponding reasoning as monotonic reasoning. The property of monotonicity
is satisfied by all methods that are based on the classical (mathematical) logic. As
per this property, if a conclusion is warranted on the basis of certain premises, no
166 6 Real-World Knowledge Representation and Reasoning

additional premises should ever invalidate the conclusion. The Nonmonotonic logic
is those ways of reasoning to infer additional information, that do not satisfy the
monotonicity property of classical logic, that is, on the face of additional information,
some of the earlier conclusions drawn may even become invalid!
In everyday life, however, we often draw sensible conclusions from what we
know/information available to us. On receiving new/updated information, many a
times we take back our previous conclusions, even when the new information we have
collected has in no way indicated that it is contradicting the previous conclusions.
Consider the following example: we have assumption that most birds fly, and that
penguin are the birds which do not fly. On learning or receiving a new information
that “Tweety is a bird”, we infer that it flies. Further learning that “Tweety is a
penguin”, still does not effect our old inference that most birds fly. But, it (addition
of new information) will require to abandon on inference of “Tweety flies.” It is quite
appropriate to say that intelligent automated systems must have the capabilities of
this kind of inference, call as non-monotonic inference.
Many systems perform such nonmonotonic inferences. The most common are :
negation as failure, circumscription, the modal system, default logic, autoepistemic
logic and inheritance systems. Each of those systems is worth studying by itself, but a
general framework in which those many examples could be compared and classified
is necessary. The following section presents a general framework, concentrating on
properties, which are important families of nonmonotonic reasoning systems.

6.10 Default Reasoning

One of the features of commonsense reasoning that makes it different from traditional
mathematical proofs is the use of defaults. A default is a proposition that is postulated
to be true in the absence of information to the contrary. For instance, an intelligent
agent may assume by default that his observation correctly reflects the state of the
world, and be prepared to retract this assumption in the face of evidence that be is in
error [11].
The logic systems we have often come across, like, classical logic, intuitionistic
logic, and modal logic are monotonic in nature. The name monotonic indicates that
on adding a new fact in these systems never results in retraction of a conclusion
derived before addition of that new fact. In fact, after addition of new knowledge in
these logic systems, the inference ability of these systems monotonically increases.
The Default logic is called nonmonotonic, because the new fact may be an exception
to one of the defaults included in the set, and on the face of that the inference system
should withdraw the conclusions drawn already.
In an intelligent system (either computer-based or human), when it tries to solve
a problem, it may rely on complete information about the problem, and its main
task will be to draw correct conclusions using classical reasoning. The classical
(predicate) logic may be sufficient in such cases. But, in many situations, the system
may have only incomplete information at hand, and it is required to draw inferences
6.10 Default Reasoning 167

with only these available information. The nonavailability of the information may
be because of the need to respond quickly, and acquiring these more information
will take some time, but the system cannot afford to wait for the collecting of all the
relevant data. For example, the case evacuation required in a nuclear disaster, where
we cannot wait to gather the complete information, and then decide to evacuate or
not, as otherwise it would have caused havoc. Some times, the information available
is unreliable/inaccurate, and it needs to be authenticated, which will require more
time, and we cannot prolong the decision.
The classical logic possess the quality to represent and reason with certain aspects
of incomplete information. In special conditions, additional information needs to
be “filled in” to overcome the incompleteness, as it is important to make certain
decisions. In such special circumstances, the system need to make some plausible
conjectures, which are based on rules of thumb, called defaults, and this modified
logic is called default logic, and the reasoning is called default reasoning. As an
example, a doctor has to make some conjectures during a situation of emergency,
about some most probable causes of the symptoms observed. Obviously, in such
conditions it would be not appropriate to wait for the results of laboratory tests, as
possibly extensive and time-consuming tests results may arrive too late to complete
the diagnosis based on those and begin the treatment.
However, when a medical diagnosis (i.e., decision) is based on assumptions, it
may turn out to be wrong on the face of availability of new information, and may
require a modified diagnosis! The phenomenon of having to take back some previous
conclusions is called nonmonotonic reasoning or non-monotonicity, which can be
stated as: if a statement ϕ follows from a set of premises M, i.e., M |= ϕ and M ⊆ M
,
then ϕ does not necessarily follow from M
(see Example 6.2). Default Logic provides
formal methods to support this kind of reasoning.
Example 6.2 Let

M = {∀x(bir d(x) ⇒ f ly(x)), ∀y( penguin(y) ⇒ bir d(y)),

∀z( penguin(z) ⇒ ¬ f ly(z)), bir d(tweet y)}.

Given these, we have M |= f ly(tweet y), and M

= M ∪ { penguin(tweet y)} is
inconsistent. We would expect the inference

M ∪ { penguin(tweet y)} |= ¬ f ly(tweet y), (6.7)

making f ly(tweet y) a defeasible consequent. 

The default Logic is the most commonly used method for nonmonotonic reason-
ing, due to the simplicity of the notion of default used in this, and because the defaults
occur quite common in real-life situations.
168 6 Real-World Knowledge Representation and Reasoning

6.10.1 Notion of a Default

A rule used by football organizers in Kashmir valley might be: “A football game shall
take place, unless there is snow in the stadium.” This rule of thumb is represented
by the default
f ootball : ¬snow
(6.8)
takes Place
Consider the following example to understand the default reasoning: In the
absence of information that there is going to be snowfall in the stadium, it is reason-
able to assume ¬snow, and also to conclude that game will take place. Accordingly,
the preparations for the game can go on. But actually if there is a heavy snowfall
during the night before the game is scheduled, then this assumption is wrong. This
is because, when we are certain, based on definite information that there is snow, we
cannot assume ¬snow, and therefore the default cannot be applied. In this case we
need to refrain from the previous conclusion (i.e., the game will take place), so the
reasoning is nonmonotonic, as we are going to withdraw the previous conclusion [11].
Before proceeding with more examples, let us first explain why classical logic is
not appropriate to model this situation. Of course, we could use the rule:

f ootball ∧ ¬snow → takes Place. (6.9)

The problem with this rule is that we have to definitively establish that there will
be no snow in the stadium before applying the rule. But that would mean that no
game could be scheduled in the winter, as it would not be possible to comment about
“no snow” on previous night of game, for in advance to decide whether to proceed for
preparations or not! And, if we wait for the previous night of match to make decision
for match, then there is no time left to do preparations. It is necessary to analyze
the difference between two statements: having “to know that it will not snow”, and
being able to “assume that it will not snow.” The defaults works on the second, and
it supports drawing conclusions based on assumptions.
The defaults can be used to model prototypical reasoning, i.e., the most instances
of a concept carry some property. One example is the statement “Typically, children
have (living) parents”; this statement can be expressed using the default logic as,

child(X ) : has Par ents(X )

(6.10)
has Par ents(X )

A no-risk reasoning is another form of default logic based reasoning, which is

concerned to situations where we need to draw a conclusion even if it is not the most
probable one, as another decision may be more disastrous. One such example is the
6.10 Default Reasoning 169

commonly used principle of awarding justice in the courts of Law: “In the absence
of evidence to the contrary, assume that the accused is innocent.” In default form we
write this as:
accused(X ) : innocent (X )
(6.11)
innocent (X )

The interpretation of rule (6.11) is that if accused(X ) is known, and there is no

evidence that innocent (X ) (at numerator) is false, then innocent (X ) (at denomina-
tor) can be inferred.

6.10.2 The Syntax of Default Logic

A default theory T is a pair W, D , where W is set of first-order predicate logic

formulas (called the facts/axioms or belief set of T ) and a countable set D of default
rules. A default rule δ is of the form
ϕ : ψ1 , . . . , ψn
δ= (6.12)
χ

here ϕ, ψ1 , . . . , ψn , χ , are one or more closed formulas in predicate logic. The for-
mula ϕ is called prerequisite of the inference rule (6.12), ψ1 , . . . , ψn are called
justifications, and χ is consequent or inference of this rule δ.
It is important to note that the formulae in a default must be a ground clause. For
example the formula,

bir d(X ) : f lies(X )

(6.13)
f lies(X )

is not a default according to the definition: it should have all clauses as ground clauses.
Let us call this rule of inference as open defaults rule. An open default is interpreted
as a default schema, and it may represents a set of defaults, which may be infinitely
large in numbers.
A default schema is very much like a default, with the difference that ϕ, ψ1 , . . . ,
ψn , χ in default are arbitrary predicate formulas (i.e., they may contain free vari-
ables). Where as, a default schema defines a set of defaults as,

ϕσ : ψ1 σ, . . . , ψn σ
(6.14)
χσ

for all ground substitutions σ that assign values to all free variables occurring in
the schema. That means free variables are interpreted as being universally quantified
over the whole default schema. Given a default schema
170 6 Real-World Knowledge Representation and Reasoning

bir d(X ) : f lies(X )

(6.15)
f lies(X )

and the facts bir d(tweet y) and bir d(sam), the default theory is represented as

T =W, D
={bir d(tweet y), bir d(sam)},

bir d(tweet y) : f lies(tweet y)
,
f lies(tweet y)

bir d(sam) : f lies(sam)
.
f lies(sam)

The Default Logic is a simple and commonly used method of knowledge repre-
sentation and reasoning in real-world situations. It has following characteristics:
• The most important aspect is that it supports reasoning with incomplete informa-
tion.
• Defaults inferences are found naturally in many applications, like in, medical diag-
nostics, reasoning related to legal issues, information retrieval, preparing specifi-
cations of systems, and as a logic in software.
• Default Logic is commonly used to model the reasoning with incomplete infor-
mation, which was the original motivation, as well as a formalism that enables
compact representation of information.
• Important prerequisites for the development of successful applications in these
domains are,
– basic concepts’ understanding, and
– there exists a powerful implementation of default logic.

6.10.3 Algorithm for Default Reasoning

The Algorithm 6.1 for default reasoning provides extensions to the formula set W .
Let T = W, D be a closed default theory (unspecified variables are false) with a
finite set of default rules D and formula W . Let P be the set of all permutations of
elements of default set D. If P = {}, i.e., D = {}, or if W is inconsistent, then return
default theory T h(W ) as the only extension of T . The set of justifications and beliefs
are indicated by variables JUST and BELIEF, respectively.
6.10 Default Reasoning 171

Algorithm 6.1 Algorithm for Default Reasoning

1: P = All permutations of elements of D
2: while P = {} do
3: Take a permutation, per m = {d1 , . . . , dn } ∈ P
4: P = P − { per m}
5: ;Initialization
6: B E L I E F = W, J U ST = {}
7: ;Application of defaults and consistency test
8: for i = 1 to n do
9: ;[assume di = ACi :Bi
i
]
10: if (B E L I E F  Ai ) ∧ (B E L I E F  ¬Bi ) then
11: B E L I E F = B E L I E F ∪ {Ci }
12: J U ST = J U ST ∪ {Bi }
13: if ∃A(A ∈ J U ST and B E L I E F  ¬A) then
14: exit the algorithm
15: end if
16: end if
17: end for
18: end while
19: Return Th(BELIEF)
20: End

Example 6.3 Find extensions of following default theory:

 
R(x) : ¬P(x) Q(x) : P(x)
T = W, D = {R(n) ∧ Q(n)}, ,
¬P(x) P(x)

First consider the permutation of D as,

 
R(x) : ¬P(x) Q(x) : P(x)
(d1 , d2 ) = , .
¬P(x) P(x)

At the begin, we initialize B E L I E F = {R(n) ∧ Q(n)}, J U ST = {}. As per

Algorithm 6.1, d1 = AC1 :B1
1
= R(x):¬P(x)
¬P(x)
. From algorithm, we note that B E L I E F 
R(n), and B E L I E F  ¬(¬P(n)). Thus, we add Ci , i.e., ¬P(n) in the B E L I E F.
Also, add ¬P(n) into J U ST . We also note that, the justification does not negate the
B E L I E F, hence it is consistent.
Next, we repeat the loop of Algorithm 6.1, for i = 2: (d2 ) = AC2 :B 2
2
= Q(x):P(x)
P(x)
.
Before this, the updated B E L I E F = {R(n) ∧ Q(n), ¬P(n)}. We note that
B E L I E F  A2 (i.e., Q(n)), but B E L I E F  ¬B2 (i.e., ¬P(n)) does not hold.
Also, Q(n) ∈ J U ST holds, but B E L I E F  ¬Q(n) does not hold, so algorithm
continues. This conclude that belief is stable (see Table 6.1).
When above is repeated for other permutation, i.e., (d2 , d1 ), we have,
 
Q(x) : P(x) R(x) : ¬P(x)
(d2 , d1 ) = , ,
P(x) ¬P(x)
172 6 Real-World Knowledge Representation and Reasoning

Table 6.1 Belief computation-1

Iteration BELIEFS JUSTIFICATIONS Consistency test
Initial R(n) ∧ Q(n) {}
i = 1, d1 ¬P(n) ¬P(n) OK
i = 2, d2 Stable

Table 6.2 Belief computation-2

Iteration BELIEFS JUSTIFICATIONS Consistency test
Initial R(n) ∧ Q(n) {}
i = 1, d2 P(n) P(n) OK
i = 2, d1 Stable

we get the extended belief set as shown in Table 6.2.

In conclusion, the T has two different extensions: T h({R(n) ∧ Q(n), ¬P(n)}),
and T h({R(n) ∧ Q(n), P(n)}), but obviously, not both at th same time.

6.11 Summary

Ontology is a theory about the nature of being or the kinds of existent. It is a sys-
tematic arrangement of all the important categories of objects or concepts that exist
in some field of discourse, such that the arrangement shows the relations between
them (objects/concepts). When compared with First-Order Predicate Logic (FOPL),
in the ontologies we are particular for knowledge organization as well as contents,
where as in predicate logic, the emphasis is on knowledge contents only.
The language, logic, ontology are closely related to commonsense, reasoning, and
type checking, respectively. The language, ontology, and logic are representations,
where as common-sense, type-checking, reasoning, are respectively, the applications.
The natural language ‘understanding is reasoning’paradigm. In comparing various
ontologies, they can be viewed at three different levels: is-a taxonomy of concepts,
(2) internal concept structure and relation between concepts, and (3) the presence or
absence of explicit axioms.
John Sowa stated his fundamental top-level principles for ontology design as
“distinctions, combinations, and constraints”. The distinctions is Physical versus
Information, combinations classify the objects into firstness versus secondness versus
thirdness, or Form versus Role versus Mediation. The constraint is, continuant versus
occurrent, or object versus Process.
The knowledge engineering is process of knowledge representation comprising of:
identification of task, assemble relevant knowledge, decide vocabulary of predicates,
functions and constructs, encode knowledge about domain, and deploy it. The onto-
logical engineering is aimed to create representation of general concepts—actions,
time, physical objects and beliefs.
6.11 Summary 173

Some of the languages and tools for implementing ontologies are: Ontolingua,
CycL, and LOOM, which use either a frame-based formalism, or a logic-based for-
malism, or both.
Classifying objects into categories is important for knowledge organization and
reasoning. For example, the sentence, “Basketball team has won”, does not refer to
a single ball and nor a single player, but a relationship among team members as a
group. The Category helps in reasoning, e.g., prediction of objects once they are
classified.
The concept of action arises in two major subareas of artificial intelligence, (1)
natural language processing and (2) problem solving. However, the formalisms that
have been suggested in each sub-area are independent of each other and difficult
to compare. The requirement on the formalism for representation of actions is that
it should be a useful representation for action reasoning (i.e., problem solving). It
has an important application to describe how this representation could be used for
planning (of actions) or for plan recognition system.
A situation is a consequence of sequence of action’s occurrences. We shall make
use of do (a binary function symbol) expressed as,

do : action × situation → situation.

which results to a new situation given an action and a present situation.

Nonmonotonic logic is the study of those ways of inferring additional information
from given information that do not satisfy the monotonicity property. The following
systems perform nonmonotonic reasoning:
• Negation as failure,
• Circumscription,
• Modal logic system,
• Default logic,
• Autoepistemic logic, and
• Inheritance systems.
One important features of commonsense reasoning that makes it different from
traditional classical reasoning is the use of defaults. A default is a proposition that
is assumed to be true in the absence of information to the contrary, i.e., unless the
contrary information is made available, it is taken as true. In addition, the default
logic is the most commonly used method for nonmonotonic reasoning due to its
simplicity and due to the notion of a default.

Exercises

1. Give the top level ontology of following structures, represent the concepts using
relations, and attributes.
a. University system—consisting of faculties, departments, teachers, classes,
students, courses, etc.
174 6 Real-World Knowledge Representation and Reasoning

b. Organizational ontology of a manufacturing firm.

c. Organizational ontology of a project based software company.
d. Government system ontology with various bodies and their responsibilities.
2. Suggest some applications of Sowa’s ontology.
3. Compare and contrast the Wordnet and Sowa’s ontology, and explain the rea-
soning performed in both with small examples.
4. Represent the ontologies of the following worlds, and explain, how you will
perform the question-answering using each of these ontology?
a. Ontology of Shirt.
b. Ontology of Dining table.
c. Ontology of University system.
5. Suggest the approach, how you will generate e-learning exercises using the
ontology.
6. What is unnatural deduction in reference to situation calculus? Give examples
to justify your claims.
7. How the inheritance works in a world of contexts? For example, in space-craft,
on earth, and when context changes from one to other?
8. Show some similarities between “contexts” and “properties” in expressing situ-
ation calculus axioms.
9. Consider a robotic-hand which can move between several bins, pickup an object
from the bin if the hand is above the bin and the hand is empty. The hand can
drop an object into a bin if the hand is holding an object and the hand is above
the bin. Moving of hand from any bin to any other bin is always possible, it does
not require any preconditions. The actions are:
dr op(x, y) (drop object x into bin y)
move(y) (move hand to be above the bin y)
grab(x, y) (pickup object x from bin y).
The fluents are:
holding(x, s) (the hand is holding x in situation s)
over (y, s) (the hand is over bin y in situation s)
in(x, y, s) (object x is in the bin y in a situation s).

a. Write the axioms for move, dr op and grab actions.

b. Write the successor state axioms for all the f luents.

10. A robot is to pickup n (n = 10) cuboid lying on the table and drop one-by-one in
a bucket, available nearby the table. Write the statements for sequential calculus.
What are the situations, actions, and fluents here?
11. Consider the eight puzzle shown in Fig. 6.8 with initial and goal states (situa-
tions).
The objective is to go from a initial situation to the goal situation. We are allowed
to move a tile into the empty space if that tile is adjacent to the empty space (e.g.
Exercises 175

Fig. 6.8 Initial and final Initial situation Goal situation

situation of 8-puzzle
1 2 3 1 2 3

4 5 6 8 4

7 8 7 6 5

in the initial situation tile 6 and 8 are adjacent to empty space). The locations are
numbered as 1-9, as shown in initial situation, with number 9 as empty tile. The
tiles are numbered 1-8. There is a single action move(t, l) which indicates mov-
ing tile t to location l. Assume a predicate ad jacent (l1 , l2 ), which is true when
location l2 is one move from l1 . It is only possible to do a move action if a tile is
in a location adjacent to an empty location. The only fluent is location(t, l, s)
meaning tile t is in location l in situation s. Given this, write down:
• initial conditions,
• effect axioms,
• precondition axioms, and
• from the effect axioms derive the successor state axioms.
12. Given the following set of facts and default rules:
a. People typically live in the same city where they work (default: d1 )
b. People typically live in the same city where their spouses are (default: d2 )
c. John works in New Delhi (fact: f 1 )
d. John’s spouse works in Mumbai (fact: f 2 )

Answer these questions:

a. Where does John live according to default logic?
b. Where does John live according to your intuition?
13. Formalize these set of facts and default rules:
a. Bob usually speaks the truth (d1 ).
b. John usually speaks the truth (d2 ).
c. Bob says that the suspect stabbed the victim to death ( f 1 ).
d. John says that the suspect shot the victim to death ( f 2 ).
e. Nobody can be both stabbed and shot to death ( f 3 ).
f. Stabbing or shooting to death is killing ( f 4 ).
Answer these questions:
a. Did the suspect kill the victim according to default logic?
b. Did the suspect kill the victim according to your intuitions?
14. Is the formula (6.10) sufficient and correct form of default reasoning, with X as
variable? Justify your answer.
176 6 Real-World Knowledge Representation and Reasoning

15. Translate the following into first-order predicate logic, and check whether the
given conclusion follow from it?
Typically, the computer science students like computers. Female students who
like computers are typically interested in cognitive science. The computer sci-
ence students are typically female: for example Anita, Babita, Cathy; but Dorthy
is an exception to this rule. Conclusion: Anita, Babita, Cathy are interested in
cognitive science; Dorthy is not interested in cognitive science.
16. Compute the default extensions of following theories T = (M, D):
a. M = {a}, D = { a:¬b
c
, :¬c
d
, :¬d
e
}
b. M = {a → c, b → c}, D = { :¬b a
, :¬a
b
, :¬d
e
}
:¬b :¬a :¬d
c. M = {}, D = { a , b , d }
d. M = { p ∧ q}, D = { b:a
a
, :¬a
a
, :¬a
¬c
, :¬q
b
, :¬q p }
17. Compute the default extensions of T = M, D , where

M = {∀x[mynah(x) → ¬nests(x)],
∀x[ penguin(x) → ¬ f lies(x)],
∀x[bir ds(x) ≡ mynah(x) ∨ penguin(x) ∨ canar y(x)],
bir d(T weet y)},
bir d(x) : nests(x) bir d(x) : f lies(x)
D={ , }
nest (x) f lies(x)

18. Find the extensions of the following default theories:

a. T = {}, { :¬pp , p∨q:¬
¬p
p
}
:¬Rain
b. T = {¬Sun-shining ∧ Summer }, { Summer
Sun-shining
}
:∃x P(x) :r ∧¬ p(x)
c. T = {}, { r ∃x p(x)
, r ∧¬ p(x) }
d. T = { p ∨ q}, { :¬pp , p∨q:¬ p
¬p
}
19. Assume that D, W be a propositional default theory, and D
be a set of normal
defaults such that D ⊆ D
. If E is an extension of D, W , then show that there
exists an extension E
of D
, W such that E ⊆ E
.

References

1. Chowdhary KR (2004) Natural language processing for word-sense disambiguation and infor-
mation extraction. PhD thesis, Department of Computer Science and Engineering, J.N.V. Uni-
versity, Jodhpur (India)
2. Davis E, Marcus G (2015) Commonsense reasoning and commonsense knowledge in artificial
intelligence. Commun ACM 58(9):92–103
3. https://protege.stanford.edu/ . Cited 19 Dec 2017
4. Douglas BL (1995) CYC: a large-scale investment in knowledge infrastructure. Commun ACM
38(11):33–38
References 177

5. Fridman-Noy N, Hafner CD (1997) The state of the art in ontology design: a survey and
comparative review. AI Mag 53–74
6. http://clarity.princeton.edu/pub/wordnet/ . Cited 19 Dec 2017
7. Sowa JF (1995) Distinctions, combinations, and constraints. In: Proceedings of the workshop
on basic ontological issues in knowledge sharing. Montreal, Canada
8. Devedzic V (2002) Understanding ontological engineering. Commun ACM 45(4):136–144
9. Pinto JA (1994) Temporal reasoning in the situation calculus. PhD Dissertation, Submitted to
the Graduate department of Computer Science, University of Toronto
10. Thielscher M (1999) From situation calculus to fluent calculus: State update axioms as a
solution to the inferential frame problem. Artif Intell 111:277–299
11. Antoniou G (1999) A tutorial on default logics. ACM Comput Surv 31(3)
Chapter 7
Networks-Based Representation

Abstract Network-based method is another approach for knowledge representation

and reasoning. They have particularly the advantage that, using the network one can
navigate through the knowledge represented, and can perform the inferences. This
chapter presents the semantic networks, conceptual graphs, frames, and conceptual
dependencies, as well as their syntax and semantics. The DL (description logic)—a
modified predicate logic for real-world applications is treated in detail, with exam-
ples of its language—the concept language for inferencing. Conceptual dependency
(CD) is a language-independent representation and reasoning framework, such that
whatever may be the natural language used, as long as its meaning is the same, the
CD will be the same. The script language for representation and reasoning along
with its syntax, semantics, and reasoning for CD is presented, followed with chapter
summary, and an exhaustive list of exercises.

Keywords Network-based representation · Semantic networks · Conceptual

graph · Frames · Description Logic (DL) · Conceptual dependencies · Scripts

7.1 Introduction

Semantic Networks were developed with the goal of characterizing the knowledge
and the reasoning of a system by means of network-shaped cognitive structures. The
similar goals were later achieved through frame-based systems, which depend on the
notion of a “frame” as a prototype, and have the capability to express the relationship
between the frames. The frames and semantic networks are quite different from
each other, but both have cognitive features, and have the capability that allows
navigation in the structures. Due to this, both of them can be classified as network-
based structures, where the network is used for representing individuals and the
relationship between them.
Due to their human-oriented origins, the network-based systems are more appeal-
ing and effective from the practical point of view than the logical systems, that are
based on predicate logic and its variants. However, these network-based systems
were not accepted as a complete solution, due to their lack of semantic properties.

© Springer Nature India Private Limited 2020 179

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_7
180 7 Networks-Based Representation

Due to that, every system behaved differently from the others, despite their almost
identical-looking components, as well as identical-looking relationships. Hence, the
need was felt to represent semantic characteristics in these structures. The semantic
characteristics in network-based systems could be achieved by introducing the notion
of hierarchical structures. Using hierarchical structures in semantics networks and
frames, one could gain both in terms of ease of representation, and also in terms of
efficiency of reasoning.
Learning Outcomes of this Chapter:
1. Identify all of the data, information, and knowledge elements and related organi-
zations, for a computational science application. [Assessment]
2. Describe how to represent data and information for processing. [Familiarity]
3. Compare and contrast the most common models used for structured knowledge
representation, highlighting their strengths and weaknesses. [Assessment]
4. Identify the components of non-monotonic reasoning and its usefulness as a rep-
resentational mechanism for belief systems. [Familiarity]
5. Compare and contrast the basic techniques for qualitative representation. [Assess-
ment]

7.2 Semantic Networks

A network-based representation provides means of structuring and exhibiting the

structure of knowledge. In a network, the pieces of knowledge are clustered together
into coherent semantic groups. It provides a natural way of mapping knowledge
between the natural language and these networks. In addition, the network represen-
tation provides a pictorial representation of knowledge objects, their attributes, and
relationship between them [2].
The basic difference between ontologies we studied earlier and the seman-
tic networks is that, ontologies are hierarchies, which may have multiple inheri-
tances, providing knowledge organization of world. Whereas, semantic networks
are not necessarily be hierarchy and they follow the lattice structure for knowledge
representation.
There is a need of built-in feature of natural language understanding in knowledge
representation, so that it becomes possible to carry out the inference through this
representation. However, when we make use of general theorem proving framework,
like resolution, these actually desired inferences are lost due to a wide range of
inferences carried out using resolution-refutation method (see Example 3.14, p. 74).
The Semantic networks not only represent information but facilitate the retrieval
of relevant facts. For instance, all the facts about an object “Rajan” are stored with
a pointer directly to one node representing Rajan. Another advantage of semantic
networks is about the inheritance of properties. If a semantic network represents the
knowledge: “All canaries are of yellow color”, and “Tweety is canary”, the network
would be able to infer that “Tweety is of yellow color.” This inference is performed
7.2 Semantic Networks 181

Fig. 7.1 A semantic has-property has-parts

fly bird wings
network

ako
parrot has-colour green

through network mathcer or retriever. The most advanced system of inference has
Inference Engine, which can perform specialized inferences tailored to treat certain
functions, predicates, and constant symbols differently than others. This is achieved
by building into the inference engine certain true sentences, which involve these
symbols, and control is provided to handle these sentences. The inference engine
is able to recognize the special conditions, on which it makes use of specialized
machinery. It becomes possible by coupling the specialized knowledge to the form
of situations that it can deal with.
The semantic networks also called the Associative Networks, model the semantics
and words of the English language. In a system developed by their inventor Quillian,
evidenced that meanings were found between the words by the path connecting them.
These models carry an intuitive appeal: the related information is always clustered
and bound together through relational links, and the knowledge required to perform
a certain task is typically contained in a narrow domain or in the vicinity of the
concerned task. This type or knowledge organization, in some way, resembles the
way knowledge is stored and retrieved in human brain [11].
Semantic networks includes a lattice of concept types [6]. Originally they included
a different correlated nets, which were based on 56 relations, like: subtypes, instances,
case-relations, part-whole, kinship relations, and various types of attributes. A simple
example of semantic network is shown in Fig. 7.1, where ako(a-kind-of), has-parts,
color, and has-property, are binary relations.
There are several benefits of using Semantic Networks for representing knowl-
edge:

• Real-world meanings (semantics) are clearly identifiable.

• Reflects the structure of the part of the world being represented in the knowledge
structuring.
• The representation due to “is-a” and “is-partof” relations help in organizing the
inheritance based hierarchies, which are useful for inheritance-based inferences.
• Accommodates a hierarchy to be useful for default reasoning (e.g., we can assume
the height of an adult as 175 cm, but if the person is a basketball player, then we
take it as 190 cm).
• The semantic networks are useful in representing events and natural language
sentences, whose meanings can be very precise. However, the concept of semantic
networks is very general. This causes a problem, unless we are clear about the
syntax and semantics in each case.
182 7 Networks-Based Representation

The Semantic networks have been used for knowledge representation in various
applications like, natural language understanding, information retrieval, deductive
databases, learning systems, computer visions, and speech generation systems.

7.2.1 Syntax and Semantics of Semantics Networks

Unlike the predicate logic, there is no well accepted syntax and semantic for semantic
networks. A syntax for any given system is determined based on the objects and
relation primitives chosen and the rules used for connection of the objects. However,
there are some primitives which are quite established. We can define a Semantic
Network by specifying its fundamental components:

1. Lexical part:

a. Nodes denote the objects.

b. Edges or links denote the relations between objects.
c. Labels denoting the particular objects and relations between them.

2. Structural part: The nodes and the edges connecting them form directed graphs,
and the labels are placed on the edges, which represent the relation between nodes.
3. Semantic part: Meanings (semantics) are associated with the edges and node
labels, whose details depend on the application domains.
4. Procedural part: The constructors are part of the procedural part, they allow for
the creation of new edges (links) and nodes. The destructors allow the deletion of
edges and nodes, the writers allow the creation and alteration of labels, and the
readers can extract answers to questions. Clearly, there is plenty of flexibility in
creating these representations.

The word-symbols used for the representation are those which represent object
constants and n-ary relation constants. The network nodes usually represent nouns
(objects) and the arcs represent the relationships between objects. The direction of the
arrow is taken from the first to the second objects, as they represent in the relations.
The Fig. 7.2 shows a is-a hierarchy representing a semantic network. In set theory
terms, is-a corresponds to the sub-set relation ‘⊆’, and an instance corresponds to
the membership relation ‘∈’ (an object class relation) [3].
The commonly used relations are: Member-of, Subset-of, ako (a-kind of), has-
parts, instance-of, agent, attributes, shaped-like, etc. The ‘is-a’ relationship occurs
quite often, like, in sentences: “Rajan is a Professor”, “Bill is a student”, “cat is a
pet animal”, “Tree is a plant”, “German shepherd is a dog”, etc. The ‘is-a’ relation
is most often used to state that an object is of a certain type, or to state that an object
is a subtype of another, or an object is an instance of a class.
Figure 7.2 shows some important features of semantic networks. The representa-
tion makes it easy to retrieve the properties of any object efficiently due to hierarchy
of relations. These networks implement property of inheritance (a form of inference).
7.2 Semantic Networks 183

Living being
isa isa
isa isa
human vertiberate Animal Plant
isa isa isa isa isa isa isa
isa
Afro Indian pigeon crow Pet Livestock Tree Bush

isa isa isa

Dog Cat Horse
isa isa
Germal Shephered Pomeranian

Fig. 7.2 Semantic network showing “Is-a” Hierarchy

Fig. 7.3 A semantic Living being

network showing
contradiction to an inherited isa isa
property has-property
fly vertiberate Animal
isa isa
isa isa
pigeon Ostrich Pet Livestock

has-property

cannot fly

The nodes which are members or subsets of other nodes may inherit the properties
from their higher level ancestor nodes. For example, we can infer from Fig. 7.2 that
dogs are animals and pigeons are vertebrates, and both of them are living beings.
The property inherited like this is recognized as default reasoning. It is assumed
that unless there is an information to the contradictory, it is reasonable to inherit the
information from the ancestor nodes. In Fig. 7.3, Pigeon inherits the property of “can
fly” from the vertebrates, while Ostrich has a locally installed attribute of “cannot
fly”, hence the property ’fly’ will not be inherited by it.
The inference procedures for semantic networks can also be in parallel to those in
propositional and predicate logic. For example, if a class A of objects have property P,
and ‘a’ is a member of A, we can infer that ‘a’ has property P. The inferences in these
networks can be defined as per those in predicate logic, making use of unification,
chaining, modus ponens, and resolution, however, the reasoning is default type [13].
184 7 Networks-Based Representation

7.2.2 Human Knowledge Creation

The semantic networks are based on the associationist theory, which defines the
meaning of an object in terms of a network of associations with other objects. When
human perceives and reasons about an object, that perception is first mapped into
a concept. This concept is part of our entire knowledge about the world and is
connected through appropriate relationships to other concepts. These relationships
form an understanding of the properties and behavior of objects such as snow. For
example, through associations, we associate the snow with other concepts like cold,
white, snowman, slippery, and ice. Our understanding of snow and truth of statements
such as “snow is white” and “the snowman is white” manifests out of this network
of associations.
There is experimental evidence that, in addition to associate concepts, humans
also organize their knowledge hierarchically, as per that information is kept at the
highest levels, to be inherited by other concepts. The evidence have shown that if
the concepts in the networks were far off in the hierarchy, it took a relatively longer
time by humans to understand this relation compared to the concepts which were in
proximity to each other. For example, with reference to Fig. 7.2, the human response
will be faster to infer that “dog is an animal”, compared to the statement “Pomeranian
is a living being”. This time difference is argued due to the fact that humans store
the information in a hierarchical way. The fastest recall was for the cases where the
traits were specific to the object. For example, the negative response for “Can Ostrich
fly?”, will be faster than the positive response for “Can pigeons fly?”. This is because
in the reasoning for the first path: “Ostrich → cannot fly”, is faster that for “pigeon
→ vertebrate → fly”, due to path length difference [1].

7.2.3 Semantic Nets and Natural Language Processing

The inheritance based system allows to store the knowledge at the highest level of
abstraction. This results to reduction in size of the knowledge base, as well as it helps
in updating the inconsistencies. The graphs can explicitly represent the relations using
arcs and nodes, which helps in formalizing the knowledge of semantic networks.
These networks can be used to answer various queries related to the knowledge base
stored, using inferences.
Much of the research in network representation has been done in the field of
Natural Language Understanding (NLU). Often, the natural language understanding
requires the understanding of the common sense, the ways in which the physical
objects behave, the interactions that occur between humans, and the ways in which
the human institutions are organized. A natural language understanding the program
must understand the intentions, beliefs, hypothetical reasoning, plans, and goals
embedded in the natural language text. Due to these, the language understanding has
been the driving force for knowledge representation.
7.2 Semantic Networks 185

The NLU programs define the words of the English language in terms of other
words, like the English dictionary does, rather than defining in terms of primitive
words or axioms. Thus, to understand the meaning of a word we traverse a network
of words until we understand the meaning of the required word.

7.2.4 Performance

The network structures used to provide intuitive and useful representations for mod-
eling semantic knowledge. These models are required to fulfill the following four
objectives:

1. Is is possible to organize the human semantic knowledge using the general struc-
tural principles which are characteristics of semantic networks?
2. Is it possible that the human performance of semantic processing can be emulated
in terms of general processes operating on semantic networks?
3. Is it possible to emulate the human processes of semantic retrieval and search on
general structures of semantic networks?
4. Can the semantic networks emulate the human processes of semantic acquisition
and development of semantics in humans?

For all the above questions, the required answer is not only in terms of yes/no, but
to what degree it is achievable. Also, the answer depends on the applications, which
exploit the features of these networks, as well as there is yet lot to be found out.
The best example of a semantic network is semantic web, which enables the
people to access the documents and services on the Internet. The interface to service
is represented in web pages written in natural language, which must be understood
and acted on by humans. The existing web is augmented by the semantic web with
formalized knowledge and data, to be processed by computers.

7.3 Conceptual Graphs

Although there is no accepted standard for semantic network representations, but

something which is very close to the goal is Conceptual Graphs. It is the portrayal of
mental perception which consists of basic primitive concepts and relationships which
exist between them. The conceptual graphs may be regarded as formal building blocks
of Semantic networks. When they are linked together, they form a more complex
and useful network [4].
Simmons [12] suggested primitives to represent standard relationships, by using
the case structure of English verbs. In the verb oriented approach, links define the
roles played by nouns and noun phrases inaction of the sentence. Case relationships
includes: agent, object, instrument, location, and time [12].
186 7 Networks-Based Representation

PERSON: Rajan agent Eat object FOOD: Noodles

Inst

Fork

Fig. 7.4 Conceptual Graph-I

A sentence in the semantic network is represented with a verb node, and various
case links to this node represent other participants in carrying out the action. The
complete structure formed is called case-frame. While this sentence is parsed, the
algorithm identifies the verb node, and retrieves the complete case-frame from the
knowledge base. As a next step, the algorithm binds the values of an agent, object,
etc., to appropriate nodes in the case-frame.

Example 7.1 Represent the sentence: “Rajan eats noodles with fork” as a conceptual
graph.

The given sentence is represented by the conceptual graph shown in Fig. 7.4, and the
corresponding predicate formula is given as Eq. 7.1.

∃x∃y(eat(x) ∧ person(Rajan) ∧ food (Noodles) ∧ fork(y)∧

agent(x, Rajan) ∧ object(x, Noodles) ∧ inst(x, y)) (7.1)

The concept we have represented is called typed or sorted version of logic, each of
the four concepts (i.e., Rajan, noodles, eat, and fork) have type labels, that represent
the type of entity the concept refers. The two concepts, Rajan and Noodles, which
have the names, identify the referent, e.g., the concept [PERSON:Rajan] has type
PERSON and referent Rajan. Three concepts have type labels: Agent, Instrument,
Object. The Conceptual Graph (CG) as a whole indicates the semantic that a person
“Rajan” is an agent of some instance of eating the food, noodles are an object, and
the fork is an instrument. Eat and Fork has no fields to refer them by name, as these
are generic concepts.

[PERSON : Rajan] ← (AGENT ) ← [EAT ]−

→ (OBJECT ) → [FOOD : noodles]
← (INSTRU MENT ) ← [FORK] (7.2)

The concept symbols may represent entities, actions, properties, or events. For
the Fig. 7.4, a linear conceptual graph form, which is easier to represent as text is
given in the Eq. 7.2. 
7.3 Conceptual Graphs 187

We note that the representation of an English language sentence in the language

of CG in Eq. 7.2, captures much of the deep structures of the natural language, such
as the relationship between the verb and its subject (called the agent relation), and
that between verb and object. When this sentence is parsed, the built-in relationship
indicates that “Rajan” is a person, who is eating, and the fork is used for eating
noodles (as object). These linguistic relationships are stored independent of the actual
sentence, and are independent of the language of the sentence.
In 1976, John Sowa developed the concept of the conceptual graph, as an inter-
mediate language, that mapped natural language questions and assertions to a rela-
tional database. The concepts were represented using symbols of rectangles, and
the circles represented as conceptual relations. An arc that pointed to a circle,
marked the first argument of the relation, and an arc pointing away from the cir-
cle marked the last argument. The relation is expressed in mathematical form as,
(relation(arg1 , arg2 , . . . , argn ). If there is only one argument, the arrow-head is
omitted, while for relation with two or more arguments, the arrow-heads are replaced
by integers, 1, 2, . . . , n.

Example 7.2 Represent the sentence “Rajan is going to Mumbai by bus”, using a
Conceptual Graph.

The Fig. 7.5 shows a CG for the sentence: “Rajan is going to Mumbai by bus.”
In the CG all the four concepts, Person, Go, Mumbai, and Bus, have type label, for
the type of the entity referred by the concept. Two concepts, Person and Destination
have names. The verb “Go” is related to the remaining three concepts, by relations of
agent, destination, and Instrument. The complete CG indicates that the person Rajan
is an agent of some instance “Going”, the city of Mumbai is the destination, and the
bus is the instrument.

(∃x)(∃y)(Go(x)Person(Rajan)City(Mumbai)Bus(y)
Agent(x, Rajan)Dest(x, Mumbai)Inst(x, y)) (7.3)

The Fig. 7.5 can be translated into the predicate formula (7.3). 

Person: Rajan Agent Go Dest City: Mumbai

Inst

Bus

Fig. 7.5 Conceptual Graph-II

188 7 Networks-Based Representation

The only logical operators used in Fig. 7.5 are conjunction and the existential
quantifier, as expressed in the Eq. 7.3, which are the most common operators in
translations from natural languages, to first-order predicate logic [14].

7.4 Frames and Reasoning

The frames are structure-based knowledge representation technique, and are similar
to semantic networks. The latter is based on the concept of human associative mem-
ory, but may simply be thought of as data structures of nodes—“concepts”—and
links—“associations”—between them [3].
The concept of a frame was proposed in the 1970s by Minsky. As per Minsky,
when we encounter a new situation, or there is substantial change in the present
context, we select from memory a new structure, called Frame. This is a previously
remembered framework, which is adapted to fit into the current set of things as
required for the new situation, with necessary changes in the frame. The frame is a
data structure to represent a stereotype situation, like a certain kind of living room,
or a frame of a picnic, or of the classroom, etc. Each frame is attached with several
kinds of information, where some information may also be about how the frame is to
be used, while other information may be about what one may expect to be happening
next, and some information may be about what is to be done if these expectations
are not met, and so on [8].
Consider representing the sentence: “Car #12 is red”
Approach 1: red(car12). With this representation, it is easy to ask “what is red”, but
we cannot ask “what is the color of car12?”
Approach 2: color(car12, red). In this approach, it is easy to ask “What is red?”
Also, we can ask, “What is the color of car12?” But, we cannot ask “What property
of car12 has value red?”
Approach 3: property(car12, color, red). With this, it is easy to ask all the above
questions.
We call this representation as, object-property-value representation, and have for-
mat property(Object, Property, Value). To get the object-centered representation, we
merge many properties of the object of the same type into one structure, as follows:
property(Object, Property1, Value1)
property(Object, Property2, Value2)
…
property(Object, Property-n, Value-n)

The representation is called Frame, as shown in Fig. 7.6.

It is important to note that objects enable grouping of procedures for determining
the properties of objects, their parts, and interaction with parts. The first step in
structuring is to collect together all propositions concerning a particular object in
a data structure, like records in PASCAL, property lists in LISP, or relations in a
7.4 Frames and Reasoning 189

Fig. 7.6 A frame object Object

Property 1
Property 2
...
Property n

Fig. 7.7 A frame of Object Property Values

“Elephant”
Elephant is-a: mammal
color: grey
has: trunk
size: huze
habitate: India/Africa

database. Figure 7.7 shows an example of a structured representation of facts by

collecting together all the properties of an object in the data structure.
There are two types of frames: 1. The Individual frames represent a single object,
e.g., a person, part of a trip; 2. Other type, the Generic frames, represent categories of
objects, e.g., students. An example of a generic frame is, “Indian city”, and individual
frames are “Delhi”, “Mumbai”. An individual frame is a named list of buckets, also
called slots. What goes in the bucket is called a filler of the slot.
(frame − name
< slot − name1 filler1 >
< slot − name2 filler2 > . . . )

7.4.1 Inheritance Hierarchies

A frame is a network of nodes and relations, whose “top levels” are fixed, and
represent the things that are always true concerning the supposed situation. The lower
levels of this frame-based network comprise many terminals or “slots”, which must
be filled in by specific instances of the data. The conditions must also be specified
for each terminal, under which the assignment be made. The assignments are usually
smaller frames, called “sub-frames” (see Fig. 7.8). Simple conditions are indicated
by markers, which might require a terminal assignment to be a person, an object
of sufficient value, or a pointer to a sub-frame of a certain type. It is possible to
specify relations among the things assigned to several terminals using more complex
conditions.
The inheritance hierarchies serve for economic data conservation. Instead of stor-
ing all the properties of each object, all the objects are structured in a hierarchy,
and only the individual properties are stored in the object itself, while the general
properties are attached to the predecessors and inherited by all the successors.
190 7 Networks-Based Representation

Fig. 7.8 A frame

representation of hotel room

In object-centered representations, an object is a natural way to organize the

knowledge about some physical objects, like “a desk has a surface-material, number
of drawers, width, length, height, color, procedure for unlocking, etc.” Some varia-
tions can be “no drawers, multi-level surface”. Alternatively, an object may describe
a situation, e.g., for a lecture-hall the complete set of situation is: hall, students,
teacher, day, time, seating arrangement, lighting, grading, etc. Or, it can be about
a trip with slot values as: origin (of trip), destination, procedures for buying ticket,
transport, getting through customs, reserving a hotel room, locating a car rental, etc.

7.4.2 Slots Terminology

Every frame is identified by its individual name—the frame name; a frame consists of
a set attributes associated with it (see Fig. 7.8). For example, in the frame “Person”,
slots may be: name, weight, height, and age; for the frame “Computer”, the slots
may be: model, processor, memory, and price. There is also a value attached to each
attribute or slot. A frame-based approach provides a natural way of structured and
compact knowledge representation. The knowledge is organized in slots that describe
various attributes or properties of any object. This approach is appropriately suited
for object-oriented programming of expert systems.
Following is the typical Information included in a Slot:
1. Relationship: A frame provides the relationship to the other frames. The frame
Hotel room (Fig. 7.8) can be a member of other frame class Room, which in turn
can belong to the class Housing, thus providing the relationship with these other
room types.
2. Slot value: The value of a slot can be numeric, symbolic, or Boolean (True/False).
For example, a slot identified as ‘Person’ is symbolic, with slot names as ‘Age’,
and ‘Height’, both having float values. The slot’s values can be dynamically
assigned during a session with the expert system, or they can be static, or can be
initialized in the beginning while the slot is created.
7.4 Frames and Reasoning 191

3. Default value of slot: A slot may contain default value when the true value is not
available, and there is no evidence that the value chosen is in no way providing
any contradiction. For example, in a frame named as Car when slot values are
not provided, default values of the slots: wheels-count and Engine-count can be
taken as 4 and 1, respectively.
4. Slot value’s range: The range of a slot’s value is useful in checking the bounds
of the slot value—whether the provided value of a slot is within the prescribed
limit? For example, a pressure range of a car tire may range 30–50 psi (pounds
per inch).
5. Slot Procedure: A slot has a procedure attached to it; when this procedure is called
it may read a value from the given slot, or it can update the value of the slot (write
the value in it).
6. Facets: The facets provide an extension to slot value structure in a frame-based
expert system. A facet provides extended knowledge about the attribute of a frame.
It can be used for establishing the attributed value of a frame, it can control end-
user queries, and can direct the inference engine as how to process the attributes.

7.4.3 Frame Languages

Frame-based languages are knowledge representation languages, where a frame des-

ignates a concept or a concrete entity. A concept is represented by a generic frame,
called class, and concrete entity illustrates one or more classes. These classes are
also represented by a specific frame, called instance. A frame, whether it is a class
or instance, is a data structure composed of slots which carry the properties of the
entity described by the frame or relations between frames.
The slots themselves are described using facets, which contribute to the description
semantics of the slot. The facets are two types: 1. descriptive (passive), and 2. active
(reflexes). The passive faces represent domain, vales, or defaults. The active facets
(also called daemons or procedural facets) are concerned to actions. They start with
keywords like, If-needed, If-created, which gets triggered after the value of a slot
is manipulated. The slots are characterized either by vales introduced by the value
facet, or default values, which reintroduced by the default facets.
The classes are organized as a hierarchy, and relations called structural links,
that connect the frames. A commonly used link, ako (a-kind-of) connect the classes
within the hierarchy, whereas “Is-a” link allows to connect instances to classes to
which they belong. A class connected by a link ako to another class, called mother-
class, inherits all the properties possessed by the mother-class. These properties are
represented using slots. The presence of a value within a class slot means this value
is true for this slot, as well as it is true for all the sub-classes and the instances of that
class, where the value has been declared. Specifying a default value in a class means
this value is generally true for that slot, also for the sub-classes, and the instances of
that class where the default is declared. However, the default can be overridden by
declaring an exception to the default, at any of the sub-class(es).
192 7 Networks-Based Representation

It is assumed that there is no multiple inheritance conflicts among these slots. The
frames of the hierarchy communicate by means of messages. The frames make use of
methods (processes), which gets executed when a message is received by the frame
in which this message is declared or when the message is received by its sub-classes
or instances.
A domain knowledge base builder uses the frame languages to describe the type
of objects to be modeled by the system, and this representation can be provided
efficiently by the frame languages. The object description of certain type may contain
a prototype description of objects of that type. These objects are used for creating a
default description of an object when its type becomes known in the model.
The frame-based languages are good for providing facilities for describing object
attributes. For example, a frame representing a car might include descriptions for
the length of the car, number of seats, size of the engine, engine’s horse-power, etc.
These attributes can be used to include partial descriptions of the attribute values,
and are helpful in preserving the semantic integrity of a system’s knowledge base by
restricting the number and range of permitted attribute values. The frame languages
do not provide a facility for describing the declarative behavior, however, they have
facilities for attaching procedural information expressed in some other language,
e.g., LISP. The procedural capability enables behavioral models of objects, and as an
expert in an application domain. It also provides a powerful tool for object-oriented
programming, where frames are treated as objects, and they respond to messages.
There are two standard forms for attaching the procedures: 1. methods, and 2.
active values. In languages, like LISP and Prolog, the methods are used as procedures,
that are attached to the frames, and respond to messages sent to the frames. Just like
the attributes, the methods are also stored as values in slots, and they are recognized
as responders to the messages. The messages sent to the frames carry the information
about target message-responder slot, these messages contain the arguments needed by
the methods stored in the slot. The “actives values” in the slots are either procedures
or collection of production rules attached to the slot. These (procedure or rules) are
invoked when the slot’s values are accessed or new values are stored in the slots.
Thus, these slots behave like “daemons”, and monitor the change and usages of the
values.
The active values can also be used to dynamically compute values on a “when-
needed” basis. The methods and active values are so written that they apply to any
member of a class of objects, and are included by the knowledge base designer in
the class description as a part of the prototype description of a class member.

7.4.4 Case Study

Many systems make use of the approach discussed above to control the reasoning,
where functions or rule classes behave like daemons, are attached to the slots of the
frames. The attachments get invoked when the value of a slot changed. That way, they
behave like sensors or alarms, or monitors. For example, an expert system is used for
7.4 Frames and Reasoning 193

Fig. 7.9 A satellite South Solar North Solar

diagnostic system Panel Panel

South Panel Power

Controller Regulation Ordinance
Controller
North panel Unit
Controller

Baterry Battery Battery

#1 #2 #3
Electric Power distribution System

Command Battery BCR OFF True False Why

an intelligent alarm facility, and this calls a user supplied function only when a slot’s
value crosses the threshold. The user may establish an alarm by providing critical
boundaries, and an alarm function. The system stores the boundaries and function
as facets of the slots, and may attach generic active value for check of boundary
crossing, whenever a value of the slot changes.
An example of the above phenomenon can be found in the knowledge system (see
Fig. 7.9), which is a Satellite diagnostic system . The system is meant to serve as an
intelligent facility for human operators to perform diagnostic and trouble-shooting
of satellite malfunctions. Such systems can also be used as a simulator to train the
operators and diagnostic experts. An important requirement of the diagnosis task is
that it requires detailed analysis of the satellite system through a diverse set of domain
experts. The software architecture of the prototype, which can be built using an expert
system, will make use of daemons attached to the slots to respond to messages sent
to objects. The prototypes of experts have the property that they can be instantiated,
as well as deleted dynamically, as per the need, during the operation of the system.
Significant events can be controlled. This system makes elaborate and effec-
tive use of frames, including prototype expert frames, which also, like a daemon,
can be instantiated and deleted dynamically as per the requirement, during opera-
tion of the system. The use of frame-based approach to build the system’s model
allowed the designers to organize the frames in such a way, that knowledge could
easily be accessible and comprehensible to diagnostic experts, as well as the satellite
operators [9].
A satellite diagnostic, e.g., STAR-PLAN, is designed with several requirements,
that resulted in an architecture based on the integration of frames and production
rules, with the following important features:
1. First, the system’s knowledge is made accessible and comprehensible to the satel-
lite operators as well as to diagnostic experts.
2. The system is incrementally and progressively built-up, as the descriptions of
additional satellite modules become ready, and as the needed experts became
194 7 Networks-Based Representation

available. Accordingly, the system’s knowledge is required to be partitioned into

a limited number of experts, for example, knowledge about a particular type of
malfunction, or about a particular module of the satellite.
3. In the case of STAR-PLAN, it was known to the designers that eventually, the
system will become very large, and it would be operating in real-time. To meet
the speed requirements, the system was so designed that only part of the system
shall be awakened—the one which is required at a certain moment of time, and
other shall be put to sleep.
Role of Frame Language
The designers of STAR-PLAN used frame language to build the taxonomy that
described parts of a typical communication satellite. The daemons and methods were
later associated with the prototype in the taxonomy. This maintained the relationship
between various parts, and defined the behavior of each part, which resulted in a kind
of object-oriented programming style, built for creating simulation behavior.
To model the diagnostic part, two separate taxonomies were constructed, such
that each class in the first taxonomy represented experts who were assigned the task
of “watching over” to some particular components of the satellite. The members of
these classes were called Guardians. In the second taxonomy, each class represented
experts responsible for responding to a particular type of problem that may occur in
the satellite. The members of the class in this second taxonomy were called Monitors.
For each component of the satellite, guardians were created and initialized at the
start of the satellite system. When an initialize message is sent to a guardian, it places
an intelligent alarm in the system’s model of the satellite. These alarms would wake-
up their guardian by sending a message when a problem has occurred in the satellite.
Hence, the guardian is active only when the satellite demands.
Alarms/Inferencing
The methods of guardian system respond to messages from the daemons. This is done
by invoking a class of diagnostic rules that determine what kind of problem is occur-
ring. For example, to find out all the possible consequences of an anomalous situation
that might have tripped the alarms, the set of rules is applied in forward-chaining
manner. In the process, all the rules are applied that have conditions matching some
aspects of the anomalous situations, or it matches a conclusion of the already applied
rule. This process of rules’ application continuous until no match is left.
The total number of rules associated with this system are small in number, typi-
cally, 10–20, so that modularization provides a small expert system of closely related
rules to be focused to develop a guardian.
As soon as a guardian comes to know about the occurrence of a problem, it
creates amonitor and initializes it. This represents an expert for the problem. The
monitor’s task is to watch the problem right from its evolution, and to make the
recommendations to the satellite operator. Once the initialization is done, the mon-
itor may create its own daemons and put itself to sleep for a fixed duration. When
the monitor wakes up itself or it is awakened by a message from one of its dae-
mons, it invokes a set of rules to analyze the status of the satellite. If the monitor
7.4 Frames and Reasoning 195

is waking-up itself after a fixed amount of time, the rules will be invoked by a
backward-chaining rule interpreter that tests a specific hypothesis (goal) about the
problem. The backward-chaining system attempts to find a sequence of rule applica-
tions that should conclude the hypothesis. When such a sequence is found, the rules
are applied so that the hypothesis is added into the knowledge base.
Based on the conclusions arrived-at making use of rules, the monitor will either
put itself to sleep again or make recommendations to the operator. When the monitor
concludes that the problem has been solved, it removes its daemons and then removes
(deletes) itself, and frees the memory occupied.
Due to the creation and deletion of monitors in a dynamic way, the satellite
problems by the STAR-PLAN system model are actually handled like by human
operators and experts. The situation is like in real-world: when a problem is identified,
a suitable expert is called in, which works with the team until the problem is resolved,
and then the expert leaves! The monitor’s rule sets are organized for problem specific
knowledge base about the problems of the satellite. The module-based system and
its organization structure also makes it easier for a (human) domain expert to create
and debug the knowledge base of rules.

7.5 Description Logic

Approaches to knowledge representation are sometimes divided roughly into two

categories: logic-based formalisms, which evolved out of the intuition, that predicate
calculus could be used unambiguously to capture facts about the world; and other,
non-logic-based representations. The latter was often developed by building on more
cognitive notions—for example, network structures, and rule-based representations
derived from experiments on recall from human memory and human execution of
tasks like mathematical puzzle solving. Even though such approaches were often
developed for specific representational chores, the resulting formalisms were usually
expected to serve in general use [5].
Since first-order predicate logic (FOPL) provides very powerful and general
machinery, logic-based approaches were more general purpose from the very start.
In a logic-based approach, the representation language is usually a variant of the
first-order predicate calculus, and reasoning amounts to verifying logical conse-
quence. In the non-logical approaches, often based on the use of graphical interfaces,
knowledge is represented by means of some ad hoc data structures, and reasoning is
accomplished by similarly ad hoc procedures that manipulate the structures. Among
these specialized representations we find semantic networks and frames. However,
frames and semantic networks lack formal semantics. Description Logic (DL ) was
first introduced into Knowledge Representation (KR) systems to overcome these
deficiencies of semantic networks and frames. The DL makes it easier to describe
definitions and properties of categories. The DL evolved from semantic networks
to formalize the network representation while retaining the emphasis on taxonomic
structures as an organizing principle.
196 7 Networks-Based Representation

A Description Logic models concepts, roles and individuals, and their relation-
ships. The fundamental modeling concept of a DL is the axiom: “a logical statement
relating roles and/or concepts”. DL is a family of formal knowledge representation
languages, which is more expressive than propositional logic and has more efficient
decision properties than first-order predicate logic. It is used in formal reasoning on
the concepts of an application domain (known as terminological knowledge). It is
used for providing a logical formalism for ontologies and the Semantic Web. Modern
ontology languages are based on DL , such as OWL (Ontology Web Language).

7.5.1 Definitions and Sentence Structures

A DL describes the domain in term of the following:

• Individuals—are the things in the world that are being described. (For example a
house, book, ram, john, rita, etc, all starting with lowercase letters).
• Classes/Categories/Roles—are sets of individuals. It is a ako (a kind of) concept.
A class is a set of all real or potential things that would be in the class. For example,
Hunter, Teenager, etc.
• Properties/Relations—are used to describe individuals. It is ako Roles or relational
nouns, and used to describe objects that are parts or attributes or properties of other
objects. Examples are: Child, Mother, Age, etc.
Two different sets of symbols—logical symbols (with a fixed meaning) and non-
logical symbols (domain-dependent) are used in the Description Logic.
Following classes of Logical symbols are used in DL :
Punctuation: (,), [,]
Positive integers
Concept-forming operators: ∀, ∃, FILLs, AND.
.
Connectives: , =, →,
, .
Non-logical symbols:
• Constants: john, rajanShaw (camel casing, but starting with uncapitalized
letter.
• Atomic concepts: Person, FatherOfOnlyGirls, Hunter, Teenager
(camel casing, first letter capital).
• Roles: :Height, :Age, :FatherOf, :Child, Mother (same as concepts,
but precede by colons).
Concepts
In terms of semantics, the concepts are given set-theoretic interpretation, where a
concept is interpreted as a set of individuals, and roles are a set of pairs of indi-
viduals. An interpretation domain can be chosen arbitrarily, which can be infinite
also. The infinite domain and the open-world assumptions are distinctive features of
Description Logic.
7.5 Description Logic 197

7.5.2 Concept Language

Atomic concepts are thus interpreted as subsets of the interpretation domain, while
the semantics of the other constructs is then specified by defining the set of individuals
denoted by each construct. For example, the concept C  D is the set of individuals
obtained by intersecting the sets of individuals denoted by C and D, respectively.
For example, Female  Teacher. Similarly, the interpretation of ∀R.C is the set of
individuals that are in the relationship R with individuals belonging to the set denoted
by the concept C. For example, ∀ResidentsOfJodhpur.Students represents all the
students who are residents of Jodhpur.
There exists some ambiguity also, due to the natural language being the source,
where many nouns can be used to refer as a category as well as relations. For example,
child can be used as a category, i.e., a very young person, it can also be used to
represent a relation, which stands for the inverse of parent.
An important feature of Description Logic is to define complex concepts in
terms of simpler ones. This is achieved by means of concept-forming operators:
∃, ∀, AND, FILLs. A complex concepts is defined as [4, 10]:
Every atomic concept is a concept;
If R is a role and C is a concept, then ∀R.C is a concept;
If R is a role and n is a positive integer, then ∃n.R is a concept;
If R is a role and C is a constant, then FILLs R.C is a concept; and
If C1 . . . Cn are concepts, then AND C1 , . . . , Cn is a concept.
The symbol ∃ stands for the class of individuals in the domain that are related by
relation R to at least n other individuals. The following can be created as complex
concepts:
∃ 1.Child : All the individuals (the class of the individuals) that have at least one
child.
∃ 2.HasCar: All the individuals that have at least two cars.
∃ 6.HasW heels: All the individuals that have at least six wheels.
The FILLs R.C stands for those individuals that are related (R-related) to the
individual identified by C. For example, “All the individuals that have the car with
plate RJC12 is represented by FILLs HasCar.RJC12.
The ∀ R.C stands for those individuals that are R-related only to individuals of
class C. For example, ∀ BeingInThisRoom.PHDStudents represents “All the individ-
uals that are in this room and are Ph.D. students.”
198 7 Networks-Based Representation

The full syntax of a concept in DL is:

Concept : −Thing|conceptName
|AND(concept1 , concept2 , ...)
|∀RoleName.concept
| ≤ ingtger.RoleName
| ≥ integer.RoleName
|FillsRoleName.IndividualName
|SameAs(Path, Path)
|∃IndividualName.concept||⊥
Path : −[RoleName, ...]

The family of concept languages is called Attribute Language (A L ) , which is

minimal language that is of practical importance. Given that Person and Female are
both atomic concepts, then Person  Female and Person  ¬Female are
A L -concepts which intuitively describe, persons that are female, and those that
are not female. Suppose that hasChild is atomic role, then Person  ∃hasChild .,
and Person  ∀ hasChild .Female denote those persons that have a child, and all
those whose children are female. Opposite to the top, there is a bottom concept (⊥),
using which we can describe the persons without a child: Person  ∀hasChild .⊥.
Sentences
A knowledge base in a Description Logic is collection of sentences like,
If C1 and C2 are concepts, then (C1 C2 ) is a sentence;
.
If C1 and C2 are concepts, then (C1 = C2 ) is a sentence;
If C is a constant and D is a concept, then (C → D) is a sentence;
For example,
PhDStudent Student, i.e., Every Ph.D. student is also a student (not vice versa).
.
C1 = C2 , i.e., concept C1 is equivalent to concept C2 , i.e. the individuals that
satisfy C1 are precisely those that satisfy C2 .
.
PhDStudent = AND(Student, Graduated , HasFunding), i.e., a Ph.D. student is
a student that is already graduated, and that has some funding.
C → D, i.e., the individual denoted by C satisfies the description expressed by
concept d . For example, rajan → PostDoc, i.e. ”Rajan is a Post Doc.“
When compared with FOPL, the FOPL focuses on sentences, and it does not
help you with reasoning on complex categories. For example, we can say that X is
a hunter by a 1-ary predicate Hunter(X ). Similarly, there is 1-ry predicate, we can
say Shooter(X ). What if we want to say is that X is both a hunter and a shooter. In
predicate logic, it is
7.5 Description Logic 199

Hunter(X ) ∧ Shooter(X ),

whereas in DL it is a 2-ary relation

Hunter&Shooter(X ).

or
AND(Hunter, Shooter).

In the DL , intersection of concepts, which is denoted C  D, is used to restrict the

set of individuals under consideration to those that belong to both C and D. In the
syntax of DL , concept expressions are variable-free. In fact, a concept expression
denotes the set of all individuals satisfying the properties specified in the expression.
Therefore, C  D can be regarded as the first-order logic sentence, C(x) ∧ D(x),
where the variable ranges over all individuals in the interpretation domain and C(x)
is true for those individuals that belong to the concept C.
We can represent the concept of “persons that are not female” and the concept of
“individuals that are female or male” by the expressions:

Person  ¬Female

and
Female  Male.

The key characteristic features of DL lies in the constructs that are helpful for
creating relationships between concepts. The most common and elementary is the
value restriction, written as ∀R.C, which means all the individuals that are having
relationship R with the concept being described, belong to concept C. Similarly,
∃R.C is a value restriction for some individuals.

Example 7.3 Represent the concepts of semantic network form in Fig. 7.10, using
DL .

Fig. 7.10 Semantic network Person

hierarchy
v/r
isa
isa
isa
hasChild Female
(1, NIL) Parent isa

isa Women
isa
isa

Mother
200 7 Networks-Based Representation

Assume that the atomic concepts are: Female, Person, and Woman. And,
hasChild, hasFemaleRelative are atomic roles. They use the operators inter-
section, union and complement of concepts, interpreted as set operations.
The Description Logic has a characteristic feature as their ability to represent
other kinds of relationships that can hold between concepts, this is beyond the IS-A
relationships. For example, in Fig. 7.10, the concept of Parent has a property that is
usually called a “role,” and expressed by a link from the concept to a node for the
role, labeled hasChild. The role has a “value restriction,” denoted by the label
v/r, which expresses a limitation on the range of types of objects that can fill that
role. In addition, the node has a number of restrictions expressed as (1, NIL), where
the first number is a lower bound on the number of children and the second element
is the upper bound, and NIL denotes no restriction on the upper limit. Overall, the
representation of the concept of Parent here can be read as “A parent is a person
having at least one child, and all of his/her children are persons”, thus the role link
between Parent and Person in Fig. 7.10 can be expressed as a concept expression
DL as,

∃1.hasChild.Parent  ∀hasChild.Person.

Existential quantification and value restrictions are thus meant to characterize

relationships between concepts. Such an expression therefore characterizes the con-
cept of Parent as the set of individuals having at least one filler of the role hasChild
belonging to the concept Person; moreover, every filler of the role hasChild must
be a person.
Relationships are inherited from concepts to their subconcepts. For example, the
concept Mother, i.e., a female parent, is a more specific descendant of both the
concepts Female and Parent, and as a result inherits from Parent the link to
Person through the role hasChild; in other words, Mother inherits the restric-
tion on its hasChild role from Parent. The other relations are translated as:
.
Woman = Person  Female.
.
Mother = Woman  Female  Parent.
∀Female.Person
∀Parent.Person.

we observe that there may be implicit relationships between concepts. For example, if
we define Woman as the concept of a female person, it is the case that every Mother
is a Woman. It is the task of the knowledge representation system to find implicit
relationships such as these (many are more complex than this one). Typically, such
inferences have been characterized in terms of properties of the network. In this case,
one might observe that both Mother and Woman are connected to both Female
and Person, but the path from Mother to Person includes a node Parent,
7.5 Description Logic 201

which is more specific then Person, thus enabling us to conclude that Mother is
more specific than Person. 

7.5.3 Architecture for DL Knowledge Representation

A knowledge representation system based on Description Logic provides facilities

to set up knowledge base, to reason about their concepts, and manipulate them. The
Fig. 7.11 shows the related blocks and their interactions for this purpose.
The KB comprises two components, the TBOX (terminology Box) introduces the
terminology, i.e., the vocabulary of an application domain; and the ABOX (assertion
box) contains assertions about the named individuals in terms of vocabulary.
The vocabulary consists of concepts, which denotes the individuals, and the roles
which denote the binary relationship between the individuals. In addition to these,
DL system allows the users to build a complex description of concepts and roles.
The TBOX can be used to assign names to complex descriptions. The description
language has model theoretic semantics. Consequently, the semantics in ABOX and
TBOX are FOPL formulas or its extensions.
The DL system provides the services for reasoning using using KB, typically,
to reason if the terminology is satisfiable. The reasoning process checks that the
assertions are consistent. With subsumption testing, it is easy to organize the concepts
of terminology in the hierarchy.
In any application, the KR system is embedded into a large environment. The
other components interact system through queries to KB and by modifying it, i.e.,
by adding or retracting concepts, roles, and assertions.
The basic form of declaration in a TBox is a concept definition, that is, the definition
of a new concept in terms of other previously defined concepts. For example, a woman
can be defined as a female person by writing this declaration:
.
Woman = Person  Female

Fig. 7.11 DL based

knowledge representation T BOX
architecture
Description Reasoning
Language
ABOX

KB

Rules Application
programs
202 7 Networks-Based Representation

There are some important common assumptions usually made about DL termi-
nologies:
• Only one definition for a concept name is allowed;
• Definitions are acyclic in the sense that concepts are neither defined in terms of
themselves nor in terms of other concepts that indirectly refer to them.
The ABox comprises extended knowledge about the domain of interest, which
are, assertions about individuals, called membership assertions. For example,

Female  Person(sita)

show that the individual named as sita is a female person. Given this definition of
woman, one can derive from this assertion that sita is an instance of the concept
Woman. Similarly,
hasChild(sita, luv)

indicates that sita has luv as a child. Assertions of the first category are also called
concept assertions, while of the second is called role assertions.

7.5.4 Value Restrictions

Let us now turn our attention to role restrictions by looking first at the quantified
role restrictions and, subsequently, at what we call “number restrictions.” Most lan-
guages provide (full) existential quantification and value restriction that allows one
to describe, for example, the concept of “individuals having a female child” as
∃hasChild.Female, and to describe the concept of “individuals all of whose
children are female” by the concept expression ∀hasChild.Female. In order to
distinguish the function of each concept in the relationship, the individual object that
corresponds to the second argument of the role viewed as a binary predicate is called
a role filler. In the above expressions, which describe the properties of Parents
having female children, individual objects belonging to the concept Female are the
fillers of the role hasChild.
Another important kind of role restriction is given by number restrictions, which
restrict the cardinality of the sets of fillers of roles. For instance, the concept

(≥ 3 hasChild)  (≤ 2 hasFemaleRelative)

represents the concept of “individuals having at least three children and at most
two female relatives.” Number restrictions are sometimes viewed as a distinguishing
feature of Description Logics, although one can find some similar constructs in some
database modeling languages (notably Entity-Relationship models).
Beyond the constructs to form concept expressions, Description Logics provide
constructs for roles, which can, for example, establish role hierarchies. However,
7.5 Description Logic 203

the use of role expressions is generally limited to expressing relationships between

concepts.
Intersection of roles is an example of a role-forming construct. Intuitively,
hasChild  hasFemaleRelative yields the role “has-daughter,” so that the
concept expression

Woman  ≤ 2 (hasChild  hasFemaleRelative)

denotes the concept of “a woman having at most 2 daughters”.

Example 7.4 Represent the following statement in Description Logic: A cheese

pizza is defined as a pizza having toping and having a pizza base. The topping is a
cheese topping, while the base is a pizzabase. A cheese topping is a topping.

CheesePizza = Pizza
 (∃hasTopping.CheeseTopping)
 (∃hasPizzabase.PizzaBase).
cheeseTopping Topping. 

7.5.5 Reasoning and Inferences

The basic inference on concept expressions in Description Logics is subsumption,

typically written as C D (read as “C is subsumed by D”). Sometimes, this axiom
type is also referred to as is-a relationship, inspired by the often chosen wording for
this type of statement (e.g. “a cat is a mammal” would be a typical verbalization
of Cat Mammal). Determining subsumption is the problem of checking whether
the concept denoted by D (the subsumer) is considered more general than the one
denoted by C (the subsumee). In other words, subsumption checks whether the first
concept always denotes a subset of the set denoted by the second one.
The principle inferences of DL are subsumtion—checking if one category is a
subset of another category, and classification checking, whether an object belongs to
a category.
For example, one might be interested in knowing whether Woman Mother. In
order to verify this kind of relationship, one has, in general, to take into account the
relationships defined in the terminology in Fig. 7.10.
Given a knowledge base expressed as a set S of sentences:
“Does a constant c satisfies concept d ?”
“Is a concept c subsumed by a concept d ?”
204 7 Networks-Based Representation

Answering to these questions amount to compute the entailment. For example,

representation for “A Ph.D. student is, a student that already graduated, and that has
some funding.” is:
.
PhDStudent = AND(Student, Graduated, HasFunding).
As another example, to say that “Bachelors are unmarried adult males”, we write
in DL as
.
Bachelor = Unmarried  Adult  Male
The most important aspect of DL is its emphasis on tractability of inference. A
problem instance is solved by designing it and then asking if it is subsumed by one
of several possible solution categories. The complexity of DL is far simpler than
FOPL. The DL usually also lacks the negation and disjunction operators.
The main application domains of description logic are: software engineering,
configuration of large software, digital libraries, Web-based information systems,
Planning, and Data Mining.

7.6 Conceptual Dependencies

The Conceptual Dependency (CD) framework is a simplified linguistic system, aimed

to provide a computational theory of simulated performance. In Conceptual Depen-
dency terms, the linguistic process is a mapping into and out of some mental rep-
resentation. This mental representation consists of concepts related to each other
by various meaning-contingent dependency links. Each concept in the interlingual
network may be associated with some word that is its realization on a sentential level.
A representation using conceptual dependency is a linked network, which charac-
terizes the conceptualization inherently present in a piece of language, with reference
to a real-world scenarios. A simple rule for representing concepts as dependent on
other concepts is, to check whether the dependent concept further explains its gov-
ernor, and this concept cannot make sense without its governor. For example, in the
sentence,

“The big man steals the red book from the girl.”,

the analysis of the sentence is as follows: The article, ‘The’ stands for connecting
sentences in paragraphs, i.e., ‘The’ also specifies that ‘man’ might have been used
previously. The adjective, ‘Big’ refers to the concept ‘big’, which cannot stand inde-
pendently. The concept ‘man’, however, can stand alone, but in the above sentence it
is conceptually modified by ‘big’, and in the network, it is realized as governor with
its dependent. The verb ‘steals’ is an action, which is dependent on the concept of
doing the acting. A conceptualization (a statement about a conceptual actor) cannot
be complete without a concept acting (or an attribute statement). Thus, to com-
plete a two-way dependency, a link must exist between the ‘man’ and the ‘steal’.
7.6 Conceptual Dependencies 205

Fig. 7.12 CD for “The big

Dependencies
man steals the red book from
the girl”
from
man steals book girl

big red

Which indicates that they are dependent on each other, and also govern each other
as shown in Fig. 7.12.
It is mandatory that every conceptualization has a two-way dependency link. In
the Fig. 7.12, the concept ‘Book’ governs the concept ‘red’ attributively (color is an
attribute), and the whole entity is placed as objectively dependent on ‘steals’. The
construction “from the girl” is realized as being dependent on the action through the
conceptual object. This is prepositional type of dependency (denoted by ⇐). There
are different forms of this prepositional dependency, each of which is expressed by
writing the preposition over the link, that indicates the prepositional relationship.
A language may use inflections, or nothing may be used instead of prepositions
to indicate prepositional dependency. Here we will discuss a language-free system,
which represents the relation of the parts conceptually.
The CDs are intended to be used in reasoning with natural language constructs,
independent of any language or the phrases in the language. This has resulted in a
small number of primitive actions, about 10–12, and a set of dependencies which con-
nect the primitive actions with each other and with their actions, objects, instruments,
etc.
The CDs have two main objectives:
1. If the meaning of any two sentences is the same, they should be represented the
same CD, regardless of the particular words are used in these.
2. Any information, which may be present implicitly in the sentence, should be
represented explicitly in the CD.
For item 1 above, the examples are ‘get’ and ‘receive’, both will have the same
CD representation. For item 2, the machine must extract the implicit part from the
sentence.
The CDs have:
1. a set of primitive actions,
2. a set of states for representation and result of the action,
3. a set of dependencies, or conceptual relationships, which could exist between
primitives, states, and objects.
The representation of English sentences could be constructed by joining together
the building blocks to form a CD graph. The CD provides four conceptualization
primitives using which the world of meaning is built. These are:
206 7 Networks-Based Representation

ACTs: Actions
PPs: Picture producers(objects)
AAs: Action Aiders (they modify the actions)
PAs: Picture Aiders (they modify the objects)

All the actions are assumed to reduce to one or more of the primitive ACTs, out
of the following actions only [7]:
PROPEL: Apply physical pressure to an object (push)
MOVE: Move body parts by owner
GRASP: Grab an object by actor(grasp)
ATRANS: Transfer of relationship (give)
PTRANS: Transfer of physical location of an object (go)
INGEST: Ingest an object by an animal (eat)
EXPEL: Expel from an animal’s body (cry)
MTRANS: Transfer mental information(tell)
MBUILD: Mentally make a new information(decide)
ATTEND: Focus sense organ
CONC: Conceptualize or think about an idea (think)
SPEAK: Produce sound (say)
At the conceptual level, a CD framework is responsible for representing the mean-
ing of a piece of written language in language-free terms. The conceptualization is
written on a straight line, in a conceptual dependency analysis. The dependents writ-
ten perpendicular to the line are attributes of their governor, except when they are
part of another conceptualization line. The whole of conceptualizations can relate to
other conceptualizations as actors or attributes.
Each primitive comprises a set of slots associated with it, from the set of conceptual
dependencies. Associated with each slot are restrictions as to what sorts of objects
could appear in that slot. For example, the following are slots for PTRANS.
ACTOR: It is either human or animate object, that initiates the PTRANS
OBJECT: It is a physical object, that is moved (PTRANSed)
FROM: The PTRANS begins at this location
TO: PTRANS ends at this location
Figure 7.13 shows the basic roles of conceptualization primitives of CD, and
Fig. 7.14 shows a general case of primitives along with an example.
The inference rules are written based on these primitives to make explicit the
information which is implicitly presented in the English or any other language text.
For example, using the primitive PTRANS, it is possible to make inference about
the OBJECT that was PTRANSed (physically transferred), which was initially at the
FROM location, and after the PTRANS is carried out, it is at the TO location. The
same inferences shall be made no matter what type of PTRANS was present, like,
flying, driving, walking, falling, etc.
7.6 Conceptual Dependencies 207

PP ACT : Indicates that an actor Acts

PP PA : Indicates that an object has attribute

ACT PP : An object has an action

PP
R : Recipient and doner of an object are in action
ACT
PP

Rajan PROPEL ball : Rajan threw the ball

P
Rajan PTRAN : Rajan ran.

dog
poss-by : Rajan’s dog.
Rajan

Fig. 7.13 Conceptual dependency representations

Fig. 7.14 A CD graph

7.6.1 The Parser

The CD framework can be used for natural language parsing. A CD-based system
analyzes sentences into their conceptual representation by operating on pieces of
every sentence, and lookup for potential conceptual cases. All the conceptualiza-
tions are checked against a list of expressions to see if that part of the concept has
occurred before. Those concepts which never occurred, are meaningless in the sys-
tem. Consider the sentence: “tall boy went to the park with the dog.” Part of the
parser output is as shown in Fig. 7.15.
In this sentence, the problem is, where to attach the concept “with dog”? Is it to the
“park” or to the “tall boy”? The problem can be solved by conceptual semantics. The
semantics for ‘go’ contains a list of conceptual prepositions. Under the preposition
208 7 Networks-Based Representation

Fig. 7.15 Parser output for P to

“tall boy went to the park” boy go park

tall

Fig. 7.16 CD for parsing a

sentence: tall boy went...
P to with
(1) boy go park dog

tall
P to
(2) boy go park

tall with
dog

‘with’ there is a description: “any movable physical object”, and since the dog is a
physical object, the dependency is applicable. As per the sentence, the parse tree (1),
in Fig. 7.16 is allowed, while (2) is rejected.
The parser tries to analyze a sentence in a way analogous to human method.
It handles input one word at a time as it is encountered, checks potential linkages
with its own knowledge of the world and past experience, and puts its output into a
language-free formalism that can be acted on.
The CD parser is a conceptual analyzer rather than syntactic parser.
Consider a sentence: “big boy gives apples to pig”. The input sentence is processed
word by word, and parsed into CD as shown with steps in Fig. 7.17.

(1) boy (2) boy give (3) give apples

big
to
(4) give pig

Assembled sentence:
to
boy give apples pig

big

Fig. 7.17 Steps for parsing using CD

7.6 Conceptual Dependencies 209

7.6.2 Conceptual Dependency and Inferences

The representation of text in canonical form has the benefit to perform inference
using that text, because the canonical form allows to write inference rules as general
as possible. If the representation does not capture the similarities in the meaning of
the text, the rules about what can be inferred from a given text needs be duplicated
by writing one rule for every form of the representation, even though they may have
the same meaning relevant to the inference. To illustrate that inference is facilitated
by CD, the inferences are used to build a “causal chain” to connect the events like in
a story, as in the following sentence.
Simon hit Anne. Anne’s mother took Anne to the hospital. Anne’s mother called Simon’s
mother. Simon’s mother slapped Simon.

An inference system could draw many inferences through this story, e.g., 1.
Simon’s mother slapped Simon because she was angry at him for hitting Anne,
2. Anne was taken to the hospital because she was hurt, 3. Anne’s mother called
Simon’s mother because she wanted to complain later. Inferences are based on a set
of rules, organized around inference categories. The total number of categories in
this case is 16, some of these are:
1. Causative inferences: These are hypothesized as possible causes or preconditions
of actions.
2. Specification inferences: These fill in missing ‘slots’ in a CD primitive, such as
the ACTOR or INSTRUMENT of an ACTion.
3. Resultantive inferences: These are inferred as likely results of the actions.
4. Function inferences: These are inferred as likely functions of objects.
The inference rules are applied in an undirected fashion. When a system reads a
sentence, the inference rules for this are automatically applied without a goal, such
as building a causal chain of events. So, it is fortuitous (having no cause or apparent
cause): the inferences are applied to new representation in an undirected fashion,
which some times result in the confirmation of another representation.
The undirected inferences are analogous to the spontaneous nature of inferences
made by people, which some times seem uncontrollable for people. This is convincing
as one cannot learn new facts without inferring things about that. However, this
undirected behavior leads to problems: When processing a story, an expert system,
which is based on CD, would not know which inferences are most likely to lead
to building a coherent causal chain to represent the story. Hence, it would lead to a
combinatorial explosion in the number of inferences that the system needs to consider
to build the causal chain to represent the story. In other words, the expert system lacks
the commonsense knowledge about what inferences are most likely to be relevant in
a given situation. For another example,
Simon picked up the menu. He decided on fish.

As an example, consider a situation, where one visits a restaurant, and decides to

take a seat for eating (see Fig. 7.18).
210 7 Networks-Based Representation

Restaurent
Customer PTRAN Customer

table
Customer ATTEND EYES

Customer MBUILD where to sit

table
Customer PTRAN Customer

sitting position
Customer MOVE Customer

Fig. 7.18 CD for “Customer moves to Restaurant”

Once a CD with a sequence of the graph is represented, it is possible to infer many

things from the representation. Like, from the Fig. 7.18, it is possible to infer many
things, like: “Who went to a restaurant?”, “Why the customer searched some thing
in a restaurant?”, etc.

7.6.3 Scripts

The scripts represent stereotypical sequences of events, like going to a restaurant,

buying from a store, etc. The theory of scripts has an emphasis to understand and
quick access those events that always happen in a stereotypical event sequence,
without worrying about those inferences which would most likely be irrelevant.
Scripts are larger knowledge structures, used to solve problems using undirected
inference. The scripts are pre-compiled sets of likely inferences, with elements of
each set packed together so that they can be searched more efficiently, and produces
lesser number of irrelevant inferences.
A script is a set of roles (participants) involved in the script, as well as common
objects used. A script comprises scenes, such that each scene describes typical events
in a portion of a script. For example, in restaurant’s script, roles may be waiter,
customer, restaurant itself, and food. Scene may be, ENTER, ORDER, EAT, PAY,
LEAVE. Each scene’s details are represented as a sequence of CD representations.
For example, the ENTER scene in a restaurant may consist of a causal chain as shown
in Fig. 7.18, and the sequence of CDs are:
Scene-name: ENTER
C PTRAN C into restaurant
C ATTEND eyes to tables
C MBUlLD where to sit
7.6 Conceptual Dependencies 211

C PTRANS C to table
C MOVE C to siting position

In above, C is a customer.

7.6.4 Conceptual Dependency Versus Semantic Nets

The conceptual dependency networks and its derivatives are often grouped into the
family of semantic net representations, because the CD is a content theory, whereas
the semantic networks are a structure theory. The distinction between these two
types of theories lies in their emphasis: the semantic nets theory is about how the
knowledge should be organized—there are nodes with arcs connecting them. There
is also some general notion about the structure’s semantics, i.e., how to interpret a
particular semantic network. In any representation, apart from the knowledge orga-
nization, there is also a general notion about inheritance. Both of these points are
about structural information. The semantic networks say nothing about what will be
represented, e.g., what labels should be used for nodes and arcs, and what arcs to use
where? It is only up to the user to decide about these details, and what a semantic
network says is about the (structural) form that the representation will take.
The CD theory was an attempt to enumerate the types of nodes and arcs which
could be used to build representations. Instead of specifying the structure of repre-
sentations, the CD theory specifies the contents. Note that the basic conventions used
for drawing CD graphs also specified structures, but these graphs were not really the
essence of the theory. The essence of the CD was in the primitives, and in the types
(names) of dependencies which could be used to link the primitives together.
The above highlighted distinctions between CDs and semantic nets would become
clear if one thinks of trying to implement the CDs and semantic nets in first-order
predicate logic (FOPL). In CDs, it is not difficult to imagine that the primitives:
ACTs, dependencies, and states would specify a set of predicates to be used when
one writes the predicate calculus statements to represent sentences. There will be a
need to adopt some translation conventions, in order to make all assertions of FOPL
type, but these would be quite straightforward.
But, it does not make a sense to implement a semantic net in FOPL, as the two
representations are in competition with each other: each representation provides
a different syntax for distinguishing between predicates, arguments, and relations.
There would be nothing left to semantic nets, if they are translated into predicate
logic. Putting it another way, the semantic nets would not add anything to FOPL.
This statement is in contrast to CD, which adds the CD primitives to the predicates
used, as and when needed.
212 7 Networks-Based Representation

7.7 Summary

In a network-based representation, the pieces of knowledge are clustered together

into coherent semantic groups. It provides a more natural way of mapping knowledge
between the natural language and these networks. A semantic network has ako (a kind
of), has-parts, color, and has-property are binary relations. Semantic networks have
primitives, and inference in the semantic networks is provided through inheritance.
The semantic networks are based on the associationist theory, which defines the
meaning of an object in terms of a network of associations with other objects.
A natural language understanding program must understand the intentions, beliefs,
hypothetical reasoning, plans, and goals. Conceptual Graphs (CG) portray mental
perception which consists of basic primitive concepts and relationships which exist
between them. In CG, a sentence is represented as a verb node, with various case
links to the node representing other participants in the action.
A frame can be viewed as generalized semantic network. In a frame, there is stress
on instances or classes, rather than nodes, and on slots and their values instead of
links and connections. Inheritance moves the default values in frames from classes to
instances through activation of the appropriate when-constructed procedure. Frames
represent some stereotypical situation, like, a classroom, house, a machine with
various parts, etc, and have slots, which may hold values, or procedures. Procedures in
frames may be sleeping procedures (daemons) and may be awakened when required.
Frames have the language to describe them. The Frames have applications in machine
vision.
A Description Logic models concepts, roles and individuals, and their relation-
ships. The fundamental modeling concept of a DL is the axiom: “a logical statement
relating roles and/or concepts”. It is a family of formal knowledge representation
languages, which is more expressive than propositional logic and has more efficient
decision properties than first-order predicate logic. It is used in formal reasoning on
the concepts of an application domain.
The Conceptual Dependency (CD) framework is a stratified linguistic system
that attempts to provide a computational theory of simulative performance. Every
conceptualization must have a two-way dependency link. The CDs are intended to
be used in reasoning with natural language constructs, independent of any language
or the phrases in the language. The CDs have a set of primitive actions, a set of states
for representation and result of the action, and a set of dependencies, or conceptual
relationships, which could exist between primitives, and states. An English language
sentence can be parsed into CDs. The inference using CDs is unambiguous.
Scripts represent stereotypical sequences of events, such as going to a restaurant.
By using the script, the theory was that the understand had quick access to those
events which always happen in a stereotypical event sequence.
Exercises 213

Exercises

1. Explain the difference between Ontologies and Semantic networks.

2. Describe the logical, structural, semantic, and procedural parts of semantic
networks.
3. There are many words in the English language which can be used as noun and
verb, for example, “book” in “Book my ticket” and “This is my book” have used
the word “book” as verb and noun, respectively. In the following words, what
are their different parts of speech?
milk, house, liquid, airborne, group, set.
Suggest a method in each case, as to how you will reason the true meaning.
4. Suggest a data structure for the implementation of semantic nets such that
retrieval can be as fast as possible.
5. Represent the relationships between quadrangle, parallelogram, rhombus, rect-
angle, and square in the form of a semantic network. Is the semantic network
unique, or are there many different forms it can take?
6. Represent the following statements using semantic networks:
a. “Rajan teaches his students a lot of innovative things.”
b. “Raman tells Rajan’s students a number of useful things.”
c. Mike and Mary’s telephone number is the same.
d. John believes that Mike and Mary’s telephone number is the same.
7. Represent the following knowledge in a semantic network:
Dogs are Mammals Birds have Wings
Mammals are Animals Bats have Wings
Birds are Animals Bats are Mammals
Fish are Animals Dogs chase Cats
Worms are Animals Cats eat Fish
Cats are Mammals Birds eat Worms
Cats have Fur Fish eat Worms
8. Represent the following in partitioned semantic networks:
a. Every player kicked a ball.
b. All players like the referee.
c. Andrew believes that there is a fish with lungs.
9. Represent the following statements using semantic networks:

a. “John tells his students a lot of useful things.”

b. “Andrea tells John’s students an enormous number of useful things.”
Suppose you wanted to build an AI system that was able to work out “who tells
John’s students the greatest number of useful things.” How could you do that?
10. Suppose you learn that “Tom is a cat”. What additional knowledge about Tom
can be derived from your representation? Explain how.
214 7 Networks-Based Representation

11. Suppose Tom is unlike most cats and does not eat fish. How could one deal with
this in the semantic network?
12. Formulate the solutions as to how the semantic networks can be used in the
following cases?
a. Natural language understanding
b. Information retrieval
c. Natural language translation
d. Learning systems
e. Computer vision
f. Speech generation system
13. “The inferencing in semantic networks make use of unification, chaining, modus
ponens, and resolution.” Justify each, taking a suitable example.
14. Explain, using semantic networks, how we can map an object’s perception to
concepts, and identify these concepts. Give examples.
15. How semantic networks help in understanding the meaning of words in natural
language sentences? Explain.
16. Represent the following as a series of frames:
Dogs are Mammals Birds have Wings
Mammals are Animals Bats have Wings
Birds are Animals Bats are Mammals
Fish are Animals Dogs chase Cats
Worms are Animals Cats eat Fish
Cats are Mammals Birds eat Worms
Cats have Fur Fish eat Worms
17. Express the following sentences in Description Logic:
a. All employees are humans.
b. A mother is a female who has a child.
c. A parent is a mother or a father.
d. A grandmother is a mother who has a child who is a parent.
e. Only humans have children that are humans.
18. Translate the logic expressed in Fig. 7.10 into DL .
19. Select one or more answers from the following. Also, justify the answer(s)
selected by you.
a. What type of reasoning is performed using semantic networks?
(A) Deductive (B) Default
(C) Inductive (D) Abductive
(E) Hierarchical
b. In the Description Logic, the domain is always,
(A) Open world (B) Closed world
(C) Depends on the domain used (D) None of above
References 215

References

1. Collins AM, Quillian MR (1969) Retrieval time from semantic memory. J Verbal Learn Verbal
Behav 8(2):240–247
2. Deliyanni A, Kowalski RA (1979) Logic and semantic networks. Commun ACM 22(3):184–
192
3. Faucher C (2001) Easy definition of new facets in the frame-based language Objlog+. Data
Knowl Eng 38:223–263
4. Harmelen FV et al (2008) Handbook of knowledge representation. Elsevier, pp 213–237
5. https://www.inf.unibz.it/~franconi/dl/course/dlhb/dlhb-02.pdf. Accessed 19 Dec 2017
6. http://www.jfsowa.com/pubs/semnet.htm. Accessed 12 Feb 2018
7. Lytinen SL (1992) Conceptual dependency and its descendants. Comput Math Appl 23(2–
5):51–73 Pergamon Press
8. Minsky M (1974) A framework for representing knowledge. MIT-AI Laboratory Memo-306
9. Pike R, Kehler T (1968) The role of frame-based representation in reasoning. Commun ACM
28(9):904–920
10. Quillian MR (1967) Word concepts: a theory and simulation of some basic semantic capabilities.
Behav Sci 12(5):410–443
11. Quillian MR (1968) Semantic information processing. Cambridge, Mass., MIT Press, pp 216–
270
12. Simmons RF (1973) Semantic networks: their computation and use for understanding English
sentences. In: Schank RC, Colby KM (eds) Computer models of thought and language. W.H.
Freeman and Co, San Francisco, CA
13. Simmons RF, Chester D (1977) Inferences in quantified semantic networks. Proceedings of the
fifth international joint conference on artificial intelligence. MIT, pp 267–273
14. Sowa J (1976) Conceptual graphs for a data base interface. IBM J Res Develop 336–355
Chapter 8
State Space Search

Abstract State space search is one of the three fundamental requirements to achieve
AI. This chapter present the basic techniques, called uninformed search, of searching
the goal in problem solution. A search requires the representation of state space in the
forms of a directed graph. The basic methods—depth-first search (DFS), breadth-
first search (BFS), along with their algorithms, analysis of these algorithms, and
along with worked out exercises are presented. The improved techniques—iterative
deepening DFS, and bidirectional search, followed with chapter summary, and then
large number of practice exercises are provide at the chapter end

Keywords State-space · State-space search · Depth-first search · Breadth-first

search · Complexities of search · Bidirectional search · Iterative deepening DFS ·
Search algorithms

8.1 Introduction

Next to knowledge representation, search is another important requirement for Arti-

ficial Intelligence. This chapter relates the problem solving to search of solution,
where search process is a tree search, consisting of generating of new states, and
ultimately leading to a goal state. The presented search is called blind-search or
exhaustive search, and turns out to be an exponential search ultimately. The various
search methods’ analysis and complexities have been derived, and presented.
Search is one of the operational task that characterize AI programs best. Almost
every program depends on search procedure to be performed to carry out its pre-
scribed functions. Problems are generally specified in terms of states, and solution
corresponds to the goal states. Solving a problem then amounts to searching through
the different states, called state-space, until one of the goal state is reached. If goal
state is not found, it concludes that solution does not exist for the problem or the goal
state is not reachable.
The state space consists set of vertices V and set of connections between them
in the form of edges (links) from E. Thus, the search space is a graph G = (V , E).
The state space to be searched here is different than it is found in the information

© Springer Nature India Private Limited 2020 217

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_8
218 8 State Space Search

searching in the conventional trees and graphs, in the sense that there, the space
limits are fixed, i.e., the number of vertices, and edges connecting them are already
known. However, in AI search problems, the alternate moves are generated from each
vertex, like in a chess game, the search tree is to be generated and simultaneously be
searched. Hence, you cannot determine the size of tree to be searched in advance.
Learning Outcomes of this Chapter:
1. Formulate an efficient problem space for a problem expressed in natural language
(e.g., English) in terms of initial and goal states, and operators. [Usage]
2. Describe the problem of combinatorial explosion of search space, search time,
and its consequences. [Familiarity]
3. Select and implement an appropriate uninformed search algorithm for a problem,
and characterize its time and space complexities. [Usage]

8.2 Representation of Search

The search space is generally represented as a directed graph (in fact a tree), with
vertices as states, and edges of the graph as transitions for moves to explore for
the goal state. The initial configuration of the problem description is start state, at
which we apply the specified rules to generate new states (vertices). Thus, it becomes
similar to a tree with start configuration as root node, then there are interior nodes,
and finally child nodes at the lowest level having no further children. To understand
the state space search better, we consider the following example of puzzle game [5].

Example 8.1 8-Puzzle Game.

The 8-puzzle game is shown in Fig. 8.1, with initial configuration, goal configuration,
and transitions (i.e., moves) possible from each state. We refer the configuration of
the game equal to the collective status of tiles on the board. This configuration we
call as state.
Using the allowed moves of a numbered tile (1–8), to tile left, right, up, down,
to the destination blank-tile, we generate new states starting from the start state (S)
as A, B, C, . . . ,. As a state is generated, it checked for the goal state. If yes, we
terminate the process and declare the fact that goal is reached. If not, the generated
states are further expanded by application of generate rules (moves), until the goal is
reached. If we carry on the process of expanding all the states generated, the goal state
shown in the figure with reverse order of tiles from first to last row, will ultimately
be reached, of course after a large number of transitions have been carried out.
The transitions’ diagram like this is called graph, with each configuration as a node,
each move as an edge, and finding the solution, i.e., goal is nothing but searching the
graph through traversing, for the goal state. 
There are many ways in what order we must generate the new child nodes. For
example, generate children at a level k + 1 only after all the nodes at level k have
8.2 Representation of Search 219

1 2 3 S= 8 7 6
Start Goal
state 5 4 3
4 5 6 state 2 1
7 8

1 2 3 1 2 3
6 A B
4 5 4 5
7 8 7 8 6

1 2 3 1 2 3
4 5 6 C 4 6
D
7 8 7 5 8

8 7 6
5 4 3
Goal 2 1
state

Fig. 8.1 The 8-puzzle game

been expanded. This is called BFS (breadth first search). Alternatively, to expand
a node to its full depth without touching its neighbors (siblings), this BFS (breadth
first search), or there can be variants of DFS or BFS.
The Search Strategies are evaluated along the following dimensions:
• Completeness: Does it always find a solution if one exists?
• Optimality: Does it always find a least-cost solution, i.e., in terms of minimum
number of transitions?
• Time complexity: What is to be number of nodes generated? The duplicate nodes
generated due to multiple paths, are also counted.
• Space complexity: What is maximum number of nodes in memory at any time?

8.3 Graph Search Basics

Let there is a directed graph G = (V , E), where V is set of vertices, and E is set
of ordered pairs (u, v), called edges, such that u is head and v is tail of an edge.
In an undirected graph, the edges are unordered pairs of vertices, and there is no
concept of fixed head and tails of edges. Graphs form abstractions for many problems,
e.g., circuits in electrical engineering, molecular structures in chemistry, relations in
social structures, and in modeling for example, in online sales. Due to their enormous
applications, it is important that there should be efficient algorithms for answering
graph-theoretical questions.
220 8 State Space Search

We should be aware of following terminologies, to when thinking of graph algo-

∗
rithms. For the graph G = (V , E), a path p = u − → v is a sequence of connected
edges and vertices from vertex u to v. A path is simple if all the vertices in that path
∗
are distinct, otherwise the path has one or more cycles. A path p = u − → v is called
∗
a closed path, and a closed path p = u − → u is a cycle if all its edges are distinct and
the only vertex that occurs twice in p is u. Two cycles which are cyclic permutations
of each other are considered to be the same cycle.
An undirected version of a directed graph is the graph formed by converting each
edge of the directed graph into an undirected edge and removing the duplicate edges,
if any. An undirected graph is connected if there is a path between every pair of
vertices.
Trees
A rooted tree T is a directed graph, whose undirected version is a connected graph,
with one special vertex (the root), which is head of no edges, but is tail of one or
more edges. All the vertices other than the root are head of exactly one edge. An
edge “(u, v) ∈ T ” is a relation in tree T , and denoted by u → v. A relation of, “there
∗
is a path from u to v, in T ” is denoted by u − → v. If u → v holds, then u is called
∗
“parent” of v, and v is called son of u. If u − → v then u is ancestor v, and v is called
descendant of u. Every vertex is also ancestor and descendant of itself.
If u is a vertex in a tree T , then we call Tu as subtree of T comprising all vertices
of all the descendants of u in T .
For a directed graph G, the tree T is spanning tree of G if T is subgraph of G and
T contains all the vertices of G.

8.4 Complexities of State-Space Search

From the previous section, it is clear that the process of searching can be automated
by applying the rules of moves to cause transitions from one state to next, until you
reach to the goal state. However, in the process a state should not repeat in the path, for
example, like v1 → v2 → v3 → v1 → v4 → v5 → vg . Such repeated vertices form
a loop and we never reach to goal state. Since the previously visited vertex v1 is
repeated, the path may instead be taken as v1 → v4 → v5 → vg .
Let us try to find out number of states to be visited in the worst case, to reach
to goal state, in the search required in Fig. 8.1. Thus, all unique states generated
(like shown in figure as A, B, C, . . . , goal), in the path to goal, in the worst it is
9! = 362880.
If we consider the case of one of the popular board game like Chess, for a
particular strategy of opponent, we definitely have some winning strategy, repre-
sented by a configuration. Our aim is to reach to that configuration from the starting
configuration, in step-by-step way, where each step is a move by us, followed by
move by opponent. Each move transforms the chessboard from one configuration to
8.4 Complexities of State-Space Search 221

another configuration. Suppose we assign the job to a computer to play against us as

opponent. So let us try to compute the number of total moves required in the worst
case to move from start state to winning state, of course, going through all the states
generated so far. Assuming on the average 20 alternate moves for a configuration,
and assuming there are 100 different configurations, the total states which can be
generated and searched are 20100 . This figure is greater than 10130 , a combinatorial
explosion of states. This value is even greater than the number of pico-seconds passed
since Big-bang occurred and also greater than the number of molecules in the known
universe. Winning a game on computer, using an exhaustive search method like this,
amounts to going through these sequence of states, which ultimately should lead to
goal state. From above, we note that the search time is exponential in time.

Example 8.2 Consider the graph shown in Fig. 8.2 for a small size problem, where
it is required to reach to goal state E from the start state A.
For the graph shown in Fig. 8.2a, when it is searched from the start state A, the
states generated are shown in the tree in Fig. 8.2b. Since objective of search is to
reach to the goal, a search process terminates the moment the goal state is reached
through any path. The tree in Fig. 8.2b shows the total possibles states generated in
the worst case, of course many repeated, but in a path no vertex is repeated. 
Thus, We can conclude that problem solution through search process is a tree
search (even if it is a graph originally) with already specified start node and goal
node. Tree search can end-up repeatedly visiting the same nodes, unless it keeps
track of all nodes visited. But, this could take vast amounts of memory, so large that
most computers do not have, even for small size graph of say 50 nodes.
Complexity of Travelling Salesman Problem
The travelling salesman problem’s (TSP) solution requires to find a best path covering
all the nodes in a graph with given start node and the destination node. Consider that
number of nodes are n. Assuming that there is a path from every node to every other
node. Thus, all the possible paths are the all permutations of these n nodes. Having
given the first node, traversing all the nodes and visiting back to the first, can be done
in n! number of ways. This gives worst case complexity of O(n!). For example, for

A
B 1 D
5 4 Start node = A B C
6 Goal node = E
A D E
2 C B
3 E
D
C E E
E
(a) Graph with vertices A-E. (b) Graph search is a tree.

Fig. 8.2 Undirected graph

222 8 State Space Search

the graph shown in Fig. 8.2b, the possible paths set for start A and destination node
E, is {ABCE, ABDE, ACBDE, ACE}.
The possible applications of TSP are distribution networks, VLSI design, and
other optimization problems.

8.5 Uninformed Search

It is also called blind search, because we do not know, which moves ultimately
will lead to the goal faster. Since all the nodes are required to be searched—the
search is called exhaustive search. To search the entire state space, the two impor-
tant approaches are—breadth-first search (BFS) and depth-first search (DFS). What
others approaches exist are variants of DFS and BFS. For example, depth-limited
search, and iterative deepening DFS are variants of DFS, while bidirectional search
may be based on DFS as well BFS.

8.5.1 Breadth-First Search

To carry out the BFS for a graph or to solve a puzzle, like, shown in Fig. 8.1, first,
root node is checked if it is goal, if yes then terminate the process after declaring that
goal is found. If goal is not found, the children nodes are generated for root node
and tested if any one of them is goal. Then further children are generated for each
of them, and so on, until goal is reached or the search is terminated. In the latter, no
new state can be generated in any of the search path. A BFS phenomena is indicated
in Fig. 8.3 where dotted trace is showing order in which nodes are tested for the
goal [2].
The BFS can be implemented using a queue type data structure, named here as
List. The front node of the of the queue is represented by List.Head . The algorithm
for this is shown as, Algorithm 8.1, which checks all the paths at a given length,
before testing the paths of longer length. The inputs to this algorithm are: the graph
G, start node S, and goal node Goal. If goal is reached it returns the success. After
searching all the vertices, which will be indicated by a empty List, if goal is still not
found, the algorithm returns fail and exits.
If the BFS Algorithm 8.1 is applied for generate-and-search, a tree like one shown
in Fig. 8.3 gets constructed, and this entire tree needs to be searched in the worst case
to find the goal state. The BFS order of this tree is S, A, B, C, D, E, F, G, H .
Though, we certainly locate the goal if it exists, the path to goal from start state is
not remembered by the search Algorithm 8.1. This, however can be carried out in the
following way. In the queue data structure used for implementing BFS, each entry is
stored like (x, y), where x is next child node to be explored for goal, and y is its parent
node. The value y = 0 represent the root node. We call this queue as open-list, i.e,
the nodes which have not been fully explored yet. When a node is deleted from front
8.5 Uninformed Search 223

Algorithm 8.1 BFS(Input: G, S, Goal)

1: List = [S]
2: repeat
3: if List.Head = Goal then
4: return success
5: end if
6: generate children set C of List.Head
7: append C to List
8: delete List.Head
9: until List = []
10: return fail

Table 8.1 Trace of state space search (BFS) for Fig. 8.3
Open-list Closed-list
[(S, 0)] []
[(A, S), (B, S), (C, S)] [(S, 0)]
[(B, S), (C, S), (D, A), (E, A)] [(S, 0), (A, S)]
[(C, S), (D, A), (E, A), (F, B), (G, B)] [(S, 0), (A, S), (B, S)]
[(D, A), (E, A), (F, B), (G, B), (G, C), (H , C)] [(S, 0), (A, S), (B, S), (C, S)]
[(E, A), (F, B), (G, B), (G, C), (H , C)] [(S, 0), (A, S), (B, S), (C, S), (D, A)]
[(F, B), (G, B), (G, C), (H , C)] [(S, 0), (A, S), (B, S), (C, S), (D, A),
(E, A)]
[(G, B), (G, C), (H , C)] [(S, 0), (A, S), (B, S), (C, S), (D, A),
(E, A), (F, B)]

Fig. 8.3 Breadth-first search S

A B C

D E F G H

of list, we add this deleted node into a separate list, called closed-list. The entries in
Table 8.1 show the search operations to search the goal node G. Each row in open-list
is constructed by a deletion or append operation on the previous content of the queue
data structure [4].
The search process is terminated when goal node G is encountered as the next
node in open-list. After reaching the goal node we can backtrack to find out the path
from goal node to start as: “(G, B), (B, S)”. Another path is “(G, C), (C, S)”. We
consider the first, or if path lengths are given then the one which is shorter.
Since BFS always explores the shallow nodes before the deeper ones, BFS finds
the shallowest path leading to the goal state. A modified and improved search can
be obtained by expanding only the lowest cost node (n). This is called Uniform cost
search. This modifies BFS such that the cost of a path should remain low but not
224 8 State Space Search

decreasing hence the term uniform. Since, a BFS algorithm is bound to find a goal, if
at all the goal exists, the BFS is complete inference system, as well as optimal (finds
the shortest path, as the nearer nodes are searched before the farther).

8.5.2 Depth-First Search

The depth-first search (also called backtracking search), is a technique, which is

widely used for finding solutions to problems in Artificial Intelligence, and Combi-
natorial Theory. In the following discussions we are trying to analyse it properties.
Consider that G is a graph we wish to explore. At start, all of its vertices are unex-
plored. To explore it, we start from some vertex of G and choose an edge to traverse,
which leads to new vertex. Continuing in this way, at each step we select an edge that
leads from the vertex already visited, and reaches to a vertex that is either already
visited or new. When there is no further edges available in sequence to explore, we
choose an edge yet to be explored from most recently visited vertex, and continue the
process. If there is no most recent edge available, we may try the parent node of this
vertex, and so on. Whenever we run out of edges leading from previous visible ver-
tices, we choose some unexplored vertex, if one exists, begin new exploration from
this point. Eventually, we will traverse all the edges of G, each exactly once [6].
Consider the following choice rule: when selecting an edge to traverse, always
choose an edge emanating from the vertex most recently reached which still has
unexplored edges. A search which uses this rule is called a depth-first search (DFS).
The set of old vertices with possibly unexplored edges may be stored on a stack type
data structure. Thus a depth-first search is very easy to program either iteratively or
recursively, provided we have a suitable computer representation of a graph.
To perform a DFS search, we generate all the next states for the root node, then
pickup the left-most node, generate the children for this, check for goal, and repeat
this, until we reach to goal or the dead end.
When reached to the dead end (the final child node), from then, back track, to the
siblings of this node, apply the DFS, then reach to siblings of its parent, and so on.
Applying the DFS Algorithm 8.2, generates a tree like one shown in Fig. 8.4. The
order of nodes visited in DFS order are: S, A, D, E, B, F, G, C, H . For DFS also, a
table of trace can be constructed as it was done for BFS, to find the path from root
node to goal node.

Fig. 8.4 Depth-first search S

A C
B

D E F G H
8.5 Uninformed Search 225

Algorithm 8.2 DFS(Input: G, S, Goal)

1: List = [S]
2: repeat
3: if List.Head = Goal then
4: return success
5: end if
6: generate children set C of List.Head
7: delete List.Head
8: insert C at begin of List
9: until List = []
10: return fail

8.5.3 Analysis of BFS and DFS

Consider the Figs. 8.3, and 8.4. Assume that the goal node is E. Using the DFS method
it needs only four steps to reach to node E. Where as it requires six steps to reach
to E if BFS search method is used. Thus, for a deeper goal node, DFS is considered
better. If the node to be searched was C, the BFS required four steps, and DFS, which
first search deeper information, then shallow, needs total eight comparisons. Thus,
which approach is best, depends on the position of goal node. Also, if branching
factor b (number of branches per node) is large, the DFS is better suited, and BFS
is worst. Thus, the efficiency of search depends on the structure of tree, the search
method used, and branching factor [2].
Let us assume that in a tree constructed in a search has depth d , and for each
node there are b nodes that gets generated, called branching factor of the tree. For a
BFS tree, the worst-case time spent for any search is the maximum number of nodes
visited to determine the goal. This is the worst-case time-complexity of the search. At
a time how many nodes are in the open-list, determine the space, or space-complexity
of the algorithm.
For a BFS search, total nodes visited for tree of depth d are:

1 + b + b2 + · · · + bd = O(bd ), (8.1)

which is worst-case time complexity, and the maximum number of nodes in the Open-
list will exist at the lowest level of the tree. Thus, space-complexity is O(bd ). Hence,
in the case of BFS, both time and space complexity are O(bd ).
For a DFS search also, in the worst-case, all the nodes are required to be visited.
Hence, time-complexity is same as for BFS, and equal to O(bd ). Since, DFS needs
to store only b nodes per level for a depth of d , the total nodes to be stored are in
memory are b × d , hence space-complexity is O(bd ). Usually the branching factor
b is much smaller than d , results to the space complexity as O(d ).
In the BFS, since all the nodes at a given depth are stored in order to generate
the nodes at the next depth, the maximum number of nodes that must be stored to
search to depth d is bd , hence the space complexity is O(bd ). As with time, the
226 8 State Space Search

Fig. 8.5 Graph with nodes v1

in linear oder, but paths in
exponential order
v2

v3

vn

average-case space complexity is roughly one-half of this, which is also O(bd ). This
space requirement of breadth-first search is its most critical drawback. As a practical
matter, a breadth-first search of most problems, spaces will exhaust the available
memory long before an appreciable amount of time is used.
The Depth-first search avoids the memory limitation of BFS. It works by always
generating a descendant of the most recently expanded node, until some depth cutoff
is reached, and then backtracking to the next most recently expanded node and
generating one of its descendants. Therefore, only the path of nodes from the initial
node to the current node must be stored in order to execute the algorithm. If the depth
cutoff is d , the space required by DFS is only O(d ).
Since the depth-first needs to store the current path at any given depth, it is bound
to search all the paths down to any specified cut-off depth. A new parameter, called
edge branching factor (e), is defined to analyse its time complexity.
e = average number of different operators applicable to a given state (i.e., node).
In case of trees, the edge branching factor and node branching factors are equal.
This is because, the number of branches of a tail node v of an edge (u, v) are same
as that from the edge (u, v). But, for a graphs, the edge branching factor may exceed
the node branching factor. For example, for the graph shown in Fig. 8.5 has an edge
branching factor of two, while its node branching factor is one only. Accordingly,
the worst case number of paths from node v1 to reach to the node vn is 2n−1 , i.e., in
the worst case in which you can traverse the tree of root v1 is O(2n−1 ), which is equal
to O(2n ), an exponential time complexity!
We know that BFS traverses a graph by visiting the nearer nodes before the farther
nodes, hence when Fig. 8.5 is traversed in BFS it will be in linear order of all the nodes
in the graph. However, the DFS, which goes deeper and deeper by sequencing the
edges, hence in general may visit many nearer nodes far later than the farther nodes.
But, considering the Fig. 8.5 for DFS, since it joins the edges up to the maximum
depth, the worst-case paths will be 2n−1 , with time complexity of O(2n ), which is
8.5 Uninformed Search 227

exponential. In general, the time complexity of a depth-first search for depth d and
edge branching factor e is O(ed ). Note that, the space used by depth-first search grows
only as the log of the time required, the algorithm is actually time-bound rather than
space-bound. This is because, the time is ed , while space is log ed = d .
Consequently, for a system in which state repetition is possible (see Fig. 8.5),
every generated state must be stored in a table, and every new state generated must
be looked upon if it is generated again. In principle, this lookup table can be expensive,
but an efficient approach can be used to do it in a time logarithm of the total number
of states or better. A simple way of doing this is discrimination tree, where previously
seen states are stored in the leaves of the tree, whose non-terminal nodes are labeled
with discriminations.
Another drawback of DFS is that it requires an arbitrary cutoff depth. During the
search, if branches are not cutoff, and duplicates are not checked, the algorithm may
run for ever. Generally, the depth at which the first goal may appear, is not known in
advance, hence if cutoff is set too low, the algorithm may not find the goal, in spite of
its existence, and may terminate. If the goal estimate is taken too deep, the algorithm
may spend too much of a time, before it reaches to the goal or terminating without
finding the goal.
It is to be noted that, DFS is not good at finding the goal, particularly if the goal
is on the opposite side of the tree which is being searched, even though the goal is
shallow. So, many a times, a depth-limited DFS search is performed, i.e., searching
the space for predetermined depth only. A method which combines the advantages of
both the BFS and DFS is iterative deepening DFS, discussed in the following part.

8.5.4 Depth-First Iterative Deepening Search

The depth-first iterative deepening DFS (DFID) reaches to shallow goals much faster
than the ordinary DFS. It first sets the tree depth d = 0, performs DFS, hence checks
the root node for goal. Then discards all the nodes generated in the previous search,
starts over and do a depth-first search for d = 1. Next, starts over and do a DFS for
d = 2, and so on. In ordinary DFS, the goal node C (Fig. 8.6) will get searched in
8 comparisons, where as using iterating deepening DFS, search is carried out for
tree depth d = 0, next in d = 1 the node C gets located, requiring total 1 + 4 = 5
comparisons [3].
The standard algorithms of BFS and DFS, both have serious limitations, which
are overcome by the algorithm DFID. The iterative-deepening algorithm, however,
is a general algorithm, and can be applied to uni-directional search, bi-directional
search, and heuristic searches like A∗ , which we will discuss in Chap. 9.
Since, DFID expands all the nodes at given depth, it is guaranteed to find a solution
in shortest-time. The disadvantage of DFID is that it performs wasted computation
prior to reaching to goal depth. In fact, at first glance it seems very inefficient.
However, this wasted computation does not effect the asymptotic growth of the run
228 8 State Space Search

Fig. 8.6 Search tree of S

iterative deepening DFS depth = 0

A B C depth = 1
depth = 2
D E F G K

time for exponential search. The intuitive reason is that almost all the work is done
at the deepest level of the search.
The worst-case complexities for the this method remains the same as ordinary
DFS, however, the average case improves. This is because, in the worst case you
still need all nodes to be compared, taking time equal to O(bd ), and worst case space
requirements is O(d ) only.

8.5.5 Bidirectional Search

If the search is carried out in both the directions, it can be sped-up. The bidirectional
search sacrifices the space for time, i.e., gains in time complexity at the cost of space
complexity. This technique search from the initial state in forward direction, and
from the goal state in backward direction, and both searches are carried out together.
The states generated are stored until the common state is found to both the searches.
Consider that a bidirectional search is carried out with the depth as d and branching
factor b. If each side progresses with same depth, the search required from each, for
d
example, for iterative deepening DFS, has O(b 2 ) for time, and O(b ∗ d2 ) for for
d d
space. When combined from both opposite sides, it becomes 2 ∗ O(b 2 ) = O(b 2 )
for time, and 2 ∗ O(b × d2 ) = O(bd ), which is O(d ), for space.
The depth first iterative deepening (DFID) search can be applied to bidirectional
search as follows: A single iteration consists of a DFS from one direction to a depth k,
and stores the states at depth k. The second DFS searches from the other direction, one
to depth k and other to k + 1, which does not stores the states but simply matches
against the states stored from other direction. Note that the search to depth k + 1
is necessary to find odd-length solutions for the goal. The process is repeated from
k = 0 (solution at depth zero), to k = d /2. If it is assumed that some hashing scheme
is used to perform the matching in constant time for every node, the bidirectional
algorithm will find the optimal solution of total depth d of the graph, in time O(bd /2 )
and space O(bd /2 ).
However, for bidirectional search to be implemented, it is necessary that invertible
functions must be available for generating nodes. Which, in fact, does not exists in
every problem. For example, to search a descendant in a tree of many generations,
one can generate children nodes for each parent, and if a path leads from a forefather
node x to a descendant node y, we can say that y is descendant of x. Similarly, we can
travel in the tree backward, from a descendant y to its parents, and then their parents,
8.5 Uninformed Search 229

until, if it reaches to x through any path, we can declare that x is forefather of y, or

y is descendant of x. Hence, there exists an invertible function. Similar phenomena
exists in the case of board game, when we traverse from, say state x to state y, in case
of chess game. The non-invertible functions are: sin(x), cos(x), and mapping from
composite number to unique primes. For all these you cannot uniquely determine,
the argument from the value of the function.

8.6 Memory Requirements for Search Algorithms

It is important, what amount of memory is used by the graph search algorithms. We

will consider first DFS algorithms for this discussion, followed with BFS, and finally
the frontier search algorithm. The search of a graph can be specified explicitly, by
listing all its nodes and edges in a data structure, which should be large enough to
hold the entire graph. When the search is specified implicitly, the graph is described
by an initial node, and a set of operators that generate all the successor nodes for any
given current node. In graph search, we are usually concerned with very large graphs,
hence they need to be specified implicitly. To generate a successor node means, to
create explicit data structure that represents the node, and to expand a node means
to generate all its children [1, 2].

8.6.1 Depth-First Searches

We consider first the depth-first search algorithm, as it has the minimal space require-
ments. The DFS can be implemented using last-in first out stack data structure. The
nodes are stored and taken off from the top of the stack. Every time a node is taken
from the top of the stack, it is expanded, and the resulting children are are placed on
the top of the stack. When recursive implementation is used, every time a node is
generated, the DFS is called recursively, on each of the child nodes generated. The
memory requirements of DFS, in both the explicit DFS and recursive implementa-
tion, is only linear in the maximum depth search, and it is generally not of much
significance.
The DFS consists number of drawbacks, e.g., in a graph with cycles, a pure DFS
may not terminate. Even in a tree, the first solution found by DFS is not necessarily
optimal. Both of these deficiencies can be solved by depth-first iterative deepening
(DFID) method. This method performs a series of depth first searches, to successively
greater depths, until the goal state is found. At this point, the current node stack is
solution, with top most node as goal, and bottom node to top node sequence as the
path from start node to the goal node. If recursive procedure is used, then the recursive
call stack contains the solution. The space complexity of DFID is linear, as in the
case of DFS.
Another drawback of DFS is such that it cannot be easily rectified. In a graph
having multiple paths to the same state, DFID as well all the DFS can generate far
230 8 State Space Search

more states than the number of actual states. It is due to the inability of DFS and DFID
to detect the duplicate nodes which corresponds to the same states. As an example,
consider a grid graph, where each node has four adjacent nodes (north, east, south,
and west) as its neighbors.1 An efficient DFS would keep track of the operator used
to generate a node from its parent, and will not apply its inverse when generating
its children. This will reduce the branching factor from four to three. Accordingly,
when considering a search in a radius r of this graph, the DFS would generate O(3r )
nodes. The large majority of these nodes are duplicate, because, there are only O(r 2 )
unique states with in a radius r. For example, applying, East followed by North,
generates the same states, as applying the North followed by East. Hence, the DFS
can generate exponentially more nodes than the actual states available in the problem
space [1].

8.6.2 Breadth-First Searches

As a general solution to BFS, the search algorithm stores all the generated nodes and
compares the newly generated nodes with the stored nodes to find out if there are
any duplicates. The BFS maintains two nodes lists, a Closed-list which is a list of
nodes that have been expanded, another list called Open-list that is generated but yet
not expanded. The Open-list is managed as first-in fist-out (FIFO) queue. If a head
node of the Open-list is not the goal state, it is removed from Open-list, expanded,
marked as Closed, and its children generated due to expansion are added to the tail
of the Open-list (i.e., appended).
Apart from the above operations, as each node is generated, it is also added to a
hash table if it is already not in this table. This storing in hash table is needed for fast
checking of the nodes in future as, whether the same node is reached again (through
a different path). For the implementation, every generated node is typically stored in
the hash table. The Open-list consists of pointers originated through the hash table,
and each node of Closed-list is indicated by a flag in the hash table storage.
The main drawback of this method is the memory requirements to store the nodes,
which is exponential.

8.7 Problem Formulation for Search

Given a problem for solution using AI search, we need to formulate its solution
through a state space search. This requires discovering the start and goal states,
which may be explicitly specified in the problem, else we need to discover from the
problem specification. The next step is to find out the possible moves to generate
children at each state. Consider the following example to formulate a search problem.

1 See Fig. 8.10a, p. 235, in exercises at the end of this chapter.

8.7 Problem Formulation for Search 231

Example 8.3 Farmer, goose, grains, fox problem.

A farmer (F) wants to move himself, a fox (X), a goose (G), and some grains (R),
across a river. Unfortunately, his boat is so tiny that he can take only one of his
possessions across on any trip. Worse yet, an unattended fox will eat a goose, and
an unattended goose will eat grains, so the farmer must not leave the fox alone with
the goose nor the goose alone with the grains. What is he to do?
We create a node for each safe position. Then draw the possible next moves by
a link from the one to the other. The end result is a graph with 10 nodes shown in
Fig. 8.7. The states are marked as 1–10. State 1 indicates that all the four objects are
one side of the river bank, and other side (separated by symbol |) is empty. Only one
move is possible from state 1, caused by transporting of goose and farmer by boat
to other bank. This results to state 2. Then, the farmer leaves the goose and returns
back, the new state is 3. From state 3 two alternate moves are possible, to states 4,
5. These states result to states 6, 7, respectively.
The states 6, 7 merges to a single state 8, which has one child state 9, and next
state 10 is goal state, indicating that all objects are to the destination back of river.
What one needs to do is to find a shortest path from the state where all objects are
on one side of the bank, to the goal state where all the object have been transported
to other side of the river bank.
The general approach is to search the paths until either the goal is reached or no
more path is left to be explored. We note that the goal state can be reached by several
paths, but to avoid loops and cycles. We note that each state is invertible also. When,
loops are not considered, there are two paths to goal as: 1–3, 4, 6, 8–10, and 1–3, 5,
7, 8–10. 

Fig. 8.7 Farmer-goose-

grains-fox problem as graph XGRF | - 1
search
XR | GF 2

XRF | G 3

4 X | GRF R | XRF 5

6 XGF | R GRF | X 7

G | XRF 8

GF | XR 9

- | XGRF 10
232 8 State Space Search

8.8 Summary

All the search strategies discussed in this chapter are called blind search. In the
absence of any idea as what direction of search may lead to a solution faster, one
need to carry out the exhaustive search. In all these searches, the time complexity is
exponential.
A graph search for AI turns out to be a tree search. Unlike conventional data
searching in a tree with tree’s physical dimensions are given, in a AI search one need
to generate-and-search simultaneously. The search space is called state space, and
nodes of the tree are called states.
The BFS (breadth-first search) has time and space complexities both as O(bd ), and
in DFS (depth-first search) has these complexities as O(bd ) and O(d ), respectively.
The Depth-first iterative-deepening (DFID) provides better average case in space
complexity, as it can locate the shallow nodes faster. However, the worst case time
and space complexities are same as that in DFS. The bidirectional search reduces
d
the time complexity to half exponential, i.e., O(b 2 ) in both the depth and breadth
searches.
To analyse the time complexity of DFS, a new parameter, called edge branching
factor (e) is important. It is “average number of different operators which are appli-
cable to a given state.” In case of trees, both the edge branching factor and the node
branching factor are equal. However, for the graphs in general, the edge branching
factor may exceed the node branching factor. In a graph with multiple paths to the
same state, any depth-first search, including DFID, can generate far more nodes than
there are states. This is due to the inability of the DFS algorithm to detect most
duplicate nodes that represent the same state. Following table gives comparison in
brief:

Search algorithm Properties

Breadth-first search Uninformed, Non-optimal (Exception:Optimal only if you are counting
total path length), Complete
Depth-first search Uninformed, Non-optimal, Incomplete

Exercises

1. Describe a state space in which iterative deepening search performs worse than
depth-first search.
2. Three missionaries (m1 , m2 , m3 ) and three cannibals (c1 , c2 , c3 ) are on one side
of a river, along with a boat that can hold one or two people. Suggest a solution
to get everyone to the other side, without ever leaving a group of missionaries
on any side of the river outnumbered by the cannibals in that place. Formulate
the solution for this, draw the graph, and demonstrate the goal search through
BFS.
Exercises 233

3. Given a 5-litre jug full of water, and an empty 2-litre jug. The jugs have no
marking or level indicator. The goal is to fill 2-litre jug with exactly one litre of
water. Give the corresponding state diagram. Assume that following are some
example of moves: p5,2 (pour 5 l into 2 l jug), e2 (empty two litre jug), etc.
4. John has two jugs: one holds exactly 3 l and one exactly 4 l, but they have no
subdivision marking. A recipe calls for exactly 5 l. He has a source of water but
no other jug. Draw the state-space graph and determine the sequence of steps in
getting exactly 5 l (into the recipe) that wastes the least amount of water. Suggest
the state-space search approach to be followed.
5. Figure 8.8 shows a search tree with states A to O along with estimated distance
from that state to the goal. The state with distance zero is goal itself.
a. Make use of BFS, and show the steps along with the elements of Queue data
structure as well the distance from goal, for each step.
b. Make use of DFS, and show the steps along with the elements of Stack data
structure as well the distance from goal, for each step.
6. Manually run the Algorithm 8.2 for the graph shown in Fig. 8.4 considering
that start node is S and goal node is H . While running this algorithm, show the
progressive contents of the stack.
7. Given the blocks world indicated in Fig. 8.9, it is required to get to goal state
from the initial start state. Construct the search tree for:
a. BFS
b. DFS
Assume that following rules for moves will be followed by the robot arm for
carrying out this job:

• stack(x, y): stack block x on block y,

• lift(x): lift-up the block x,

A
20

B C
15 17

D E F G
14 12 11 13

H I J K L M N O
10 9 0 5 8 0 7 5

Fig. 8.8 State—space tree

234 8 State Space Search

Robot arm

C B

B A C

Goal State
Start State
Fig. 8.9 Blocks world

• putg(x): put block x on ground,

• unstack(x, y): unstack block x from block y.

8. Find space and time complexity for DFS, for parsing a NL sentence of 50 words.
Suggest the assumptions you would make to reduce the space and time com-
plexity.
9. Give examples of graphs having different edge and node branching factors.
10. For the graphs shown in Fig. 8.10a–d, with start node S and goal node G, find
out the average node branching factor and average edge branching factor.
11. For the following board games, find out the worst case path length to goal, total
number of paths to goal, node branching factor, and edge branching factor.

a. Rubics-cube
b. Tic-tac-toe
c. 8-puzzle
d. Sudoku

12. Consider that a large number of websites are available, with their domain names
in a domain-name-file. Each website has index-page (homepage) and other pages
linked directly or indirectly from the homepage. In addition, there may be links
from website pages to other website pages. Suggest a method and write an
algorithm to crawl (scan) these websites to create an index table as follows:
each row in table has one entry corresponding to each domain name, comprising
domain name and five most frequent keywords: {domain-name, key-word1, …,
key-word5}. Your algorithm must follow a suitable search approach.
a. Give the justification for the type of search you have used.
b. Give the time and and space complexity analysis.
Exercises 235

S S

(a) (b)

G
S S

(c) (d)

G G

Fig. 8.10 A graph with start node S and goal node G

c. Is the search carried out by this algorithm complete?

d. Is the search optimal?
13. a. List four criteria that are generally used to evaluate the search.
b. List two limitations, for each of the following search methods: DFS, BFS,
Bidirectional.
c. Give an example of problem where DFS will work better than BFS.
d. Give an example of problem where BFS will work better than DFS.
14. There are 10 pages of power-point presentation slides, which are inter-linked
through pointers, with each page having not more than 5-links. The “start” page
is given, and “end” page is identified by “Thanks”. Suppose only way to navigate
through these pages is through links, and no idea available about the actual order
of the pages.
a. What is number of transitions in worst case to reach to the end page?
b. What is the number of transitions in best case?
c. What is time complexity (i.e., number of nodes visited) in DFS, BFS, and
iterative deepening DFS?
d. What is space complexity in DFS and BFS?
15. Discuss all the drawbacks of search in case the Closed-list is not used in search
algorithms.
16. Solve the Towers of Hanoi problem using DFS, and find out the space and time
complexity for this for n disks. Also answer the following:
a. What is maximum branching factor for n disks?
b. What is average branching factor for n = 4 disk?
236 8 State Space Search

17. Imagine there are 10 million websites in the entire Internet. Suppose that a search
engine uses certain strategies to index these websites on regular basis, to maintain
and update a very large database of indexes. Answer the following through your
imaginations/logic as well as explore them on web.
a. What strategy should search engine use to navigate through these websites
through the links.
b. What strategy should be used to create the index of the home-page of every
website?
c. What should be criteria to update the indexes?
d. How to avoid the re-indexing of already indexed nodes?
18. Can every AI search be carried out using bidirectional search? What is necessary
condition for this, if any?
19. For a search-tree of depth d and branching-factor also b, carry-out the following
analysis for DFS and BFS:
a. Average case time complexity.
b. Average case space complexity.
c. Best case time complexity.
d. Best case space complexity.
20. Suggest an architecture of hash-table for storage of nodes found in graph search,
and quick searching of the same when it is found again next-time in graph search.
21. What is the memory requirement (in KB or MB) for data-structure for 8-puzzle
for each of the following searches?
a. BFS
b. DFS
Assume that only the open-list is maintained to store the nodes. Each node
correspond a state of the 8-puzzle board, and is equal to 9-bytes (in each one
byte, 4-bits are for square number, and 4-bits for integer number in that, including
the space). Note that, when head nodes are deleted, the memory is freed.
22. Suggest an approach to combine the DFS and BFS in a single algorithm, so that
at certain states you can choose DFS in place of BFS or vice-versa, based on
meeting specific situation. For example, to check for shallow goals BDF can be
used while for deep goals DFS can be tried. What data structures you will use
to implement such type of search, as well the true strategies?

References

1. Korf RE et al (2005) Frontier search. J ACM 52(5):715–748

2. Korf RE (2008) Linear-time disk-based implicit graph search. J ACM 55(6):26:1–26:40. https://
doi.org/10.1145/1455248.1455250
References 237

3. Korf RE (1985) Depth-first iterative-deepening: an optimal admissible tree search. Artif Intell
27:97–109
4. Ludger GF (2009) Artificial intelligence—structures and strategies for complex problem solving,
5th edn. Pearson Education, New Delhi
5. Nilsson NJ (1980) Principles of artificial intelligence, 3rd edn. Narosa Publishing
6. Tarjan R (1972) Depth-first search and linear graph algorithms*. SIAM J Comput 1(2):146–160
Chapter 9
Heuristic Search

Abstract This chapter provides in depth study of heuristic search methods—the

methods for searching the goal (solution) to problems, that are more like human,
and do not follow the exhaustive search approach, making them far more efficient
than the uninformed search methods. The introduction starts with formal defini-
tion of heuristic search, then follows Hill-climbing searches, their algorithm and
analysis, best-first search, its algorithm and analysis, optimization, A-star search,
and approaches to better heuristics. Finally, the search methods—simulated anneal-
ing (based on treatment of metals), Genetic Algorithm (GA)-based search method,
along with their analyses are presented, followed with chapter summary, and at end
an exhaustive list of practice exercises, along with multiple-choice questions are
provided.

Keywords Heuristic search · Hill-climbing search · Best-First Search · A-star

search · Simulated annealing · Genetic algorithm

9.1 Introduction

The subject of combinatorial optimization comprises a set problems that are central
to the domain of computer science. The field of combinatorial optimization aims to
develop efficient algorithms to find out minimum/maximum values of some function
having large number of independent variables. The function is usually called objective
function or cost function, and represents the quantitative measure of the “goodness”
of some complex system. The cost function depends on the total configuration of the
system which comprises many parts.
The best-quoted example of a combinatorial optimization problem is Traveling
Salesman Problem (TSP). It is stated as: given a list of n cities, and distance between
every two cities, it is required to plan a salesman’s route which guarantees to pass
through every city once only, covering all the cities, and finally returns to the starting
point. The order of cities to be visited should be so planned, that the total distance,
to cover all the cities, is minimum. Instead of distance, it can be cost of travel from
city to city, or some other parameter. For the sake of generality, we call this as

© Springer Nature India Private Limited 2020 239

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_9
240 9 Heuristic Search

cost. The problems of this type are common in the areas of scheduling and design.
Two subsidiary problems in these areas are of general interest: 1. Predicting the
estimated cost of a salesman’s optimal route averaged over some arrangement of
cities, having given the arrangement of cities and the distance between pairs of
cities, 2. Estimating/obtaining the upper bounds of computing efforts necessary to
determine the optimal route.
All the exact methods (also called exhaustive) known so far, for determining the
optimal route of the salesman problem, requires the computing efforts, that grows
exponentially with the total number of cities n. Hence, to carryout the solution in
realistic times, the exact solution can be attempted only for a small number of cities
n, may be less than a hundred. The TSP belongs to a class of large set of problems,
called NP-Complete (NP for nondeterministic polynomial in time) problems. So far,
no method of exact solution having computing effort bounded by a polynomial power
of n (say n k , k ≥ 1) has been found for any of these problems.1
However, if such a solution was found for any problem, say A, which is NP-
complete, then it would be possible to map to A, all the remaining NP problems, as
all NP problems are member problems of A (an NP-complete). However, it is yet not
known as what are the features of the individual problems that make it NP-complete,
and that cause this level of difficulty of solving!
The problems of NP-complete class are at common place in many situations
of practical interest, and hence the importance of their solutions. Fortunately, the
solution methods, called, heuristic methods have been developed, which have com-
putational requirements proportional to only a limited powers of n (the size of the
problem). Unfortunately, there is no universal heuristic method which can be applied
to all types of problems. In fact, the heuristics are not general, but problem-specific,
and there is no guarantee that one heuristic procedure for finding near-optimal solu-
tions for one NP-complete problem will be effective for another problem also.
Fundamentally, there exist two basic approaches to heuristics: 1. “divide-and-
conquer” and 2. “iterative improvement”. The first approach is based on the concept
of dividing the original problem into subproblems of manageable sizes, and then the
subproblems are individually solved. At the end, the solutions to the subproblems are
patched back together to get the complete solution. For this method to produce a good
solution, it is necessary that subproblems are naturally disjoint, and the division of
the original into subproblems is proper, in the sense that errors made in patching up
the subproblems do not offset the gains obtained in applying more powerful methods
to the subproblems.
Other approach to heuristics is based on iterative improvement, where one has
to start with a known configuration. And, for this configuration, a standard rear-
rangement operation is applied to all parts of the system in turn, until a rearranged
configuration which improves the cost function is discovered. As a next step, this
rearranged configuration becomes the new configuration of the system, and the pro-
cess is repeated until we reach to a configuration such that no further improvements

1 The NP-Complete problems require exponential power of computing efforts, in terms of n, i.e.,
kn .
9.1 Introduction 241

can be found. The iterative improvement comprises a search space for rearrange-
ment, so that there is a flat region in space, indicating that an optimized solution has
reached—called global maxima.
Instead of settling to global maxima, the search quite often gets stuck-up in a local
maxima. Due to this, it is usual to perform the search several times, starting from
different randomly generated configurations, and save the best result, so as to reach
the global maxima.
This chapter presents the heuristic methods for AI search problems. These meth-
ods are better informed, hence explore the state space in a more right directions. The
analysis and complexities of these methods are also discussed.

Learning Outcomes of this Chapter:

1. Describe the role of heuristics and describe the trade-offs among completeness,
optimality, time complexity, and space complexity. [Familiarity]
2. Select and implement an appropriate informed search algorithm for a problem by
designing the necessary heuristic evaluation function. [Usage]
3. Evaluate whether a heuristic for a given problem is admissible (i.e., can guarantee
optimal solution). [Assessment]
4. Design and implement a genetic algorithm solution to a problem. [Usage]
5. Design and implement a simulated annealing schedule to avoid local minima in
a problem. [Usage]
6. Design and implement A∗ search to solve a problem. [Usage]
7. Compare and contrast genetic algorithms with classic search techniques. [Assess-
ment]
8. Compare and contrast various heuristic searches vis-a-vis applicability to a given
problem. [Assessment]

9.2 Heuristic Approach

The search efficiency can improve tremendously—reducing the search space, if there
is a way to order the nodes to be visited in such that most promising nodes are explored
first. These approaches are called informed methods, in contrast to the uninformed
or blind methods discussed in the previous chapter. These methods depend on some
heuristics determined by the nature of the problem. The heuristics is defined in
the form of a function, say f , which some how represents the mapping to the total
distance between start node and the goal node. For any given node n, the total distance
between start and goal node is f (n), such that

f (n) = g(n) + h(n), (9.1)

where g(n) is distance between start node and the node n, and h(n) is the distance
between node n and the goal node. We note that g(n) can be easily determined
242 9 Heuristic Search

and can be taken as shortest. However, the distance to goal is f (n), which requires
computation of h(n), called heuristics, cannot be so easily determined. In deciding
the next state every time, which is represented by node n, the state is chosen such
that f (n) is minimum.
Considering the case of the traveling salesman problem, which otherwise, is a
combinatorially explosive problem, with exponential time complexity of O(n!) for
n nodes, reduces to only O(n 2 ) if every time the next node selected is the nearest
neighbor, that is, the one having shortest distance from the current node.
Similarly, in the 8-puzzle problem, the next move is chosen the one having mini-
mum disagreement from the goal, i.e., having minimum number of misplaced posi-
tions with respect to the goal.
The heuristic methods reduce the state space to be searched, and supposed to give
the solution, but may fail also.

9.3 Hill-Climbing Methods

The name hill-climbing comes from the fact that to reach the top of a hill, one selects
the steepest path at every node, out of the number of alternatives available. Naturally,
one has to sort the slope values available, pick up the direction of move having highest
angle, then reach to the next point (node) toward the hill top, then repeat the process.
The hill-climbing Algorithm 9.1 is an improved variant of the depth-first search
method. A Hill-climbing method is called greedy local search. Local, because it
considers a node close to the current node, at a time; and greedy because it always
selects the nearest neighbor without knowing its future consequences. The inputs to
this algorithm are G, S, and Goal, which stand for—graph, start(root) node, and the
goal node, respectively.
Consider the graph shown in Fig. 9.1a, where start node is A and goal node is G.
It is required to find out the shortest path from node A to node G, using the method
of hill-climbing. Figure 9.1b shows the search tree for reaching to goal node G from
start node A, with shortest path A, B, D, E, G, and path length 14. It can be easily
worked out that this approach cannot lead to shortest path always.
Though simple, hill-climbing suffers from various problems. These problems are
prevalent when hill-climbing is used to optimize the solution.

• Foothill Problem: This occurs when there are secondary peaks or local maxima.
These are mistaken for the global maxima, as the user is left with false sense of
achieving the goal.
• Plateau Problem: This occurs when there is a flat region separating the peaks.
• Ridge Problem: It is like a knife edge or an edge on top of a hill, both the sides are
valleys. It again gives a false sense of top of the hill, as no slope change appears.
9.3 Hill-Climbing Methods 243

Algorithm 9.1 Hill-Climb(Input: G, S, Goal)

1: Open = [S]
2: Closed = nil
3: if Open = nil then
4: return fail
5: end if
6: repeat
7: if Open.H ead = Goal then
8: return success
9: end if
10: expand Open.H ead and generate children’s set, call it C
11: reject all paths in C having loops
12: delete Open.H ead and insert it into Closed
13: sort C in order of heuristic, with best heuristic node in the front
14: insert C at the front of List
15: until Open = nil
16: Return fail

B 9 A
2 E 2 5
2
4
9 C
A 3 B 9
D 3
7 6 G 2
5 C E
C D
4 6
Start node = A
E G
Goal node = G
6
G
(a) Graph to be searched for goal ’G’. (b) Search-Tree.

Fig. 9.1 Graph with Hill-climbing search

Due to above difficulties, a hill-climbing algorithm is not able to find a goal, even
if the goal exist. Thus, hill-climbing algorithms are not complete.
Consider the following phenomenas:
1. Rotating the brightness knob in control panel of an analog TV does not improve
the quality of picture,
2. While testing a program, running it again and again, with different data sets does
not indicate new discovery of errors,
3. Participating in a sports again and again (without new ideas and training), does
not improve further performance.
In all the above three cases, we strive for optimum performance. In case 1, the
adjustable parameter is TV control, in second types of different input data, and in the
third, adjustable parameter is more energy and preparedness. But it is appears that
optimum has reached (a highest point in performance, from where no improvements
take place, an indication of saturation point, the goal, but not the true goal).
244 9 Heuristic Search

The above scenarios are created due to either of the foothill problem, or plateau,
or ridge.
Local Versus Global Search
The preference for local search where only one state is considered to further expand
at a time, out of the newly explored states, is good for a memory efficiency, in
comparison to expanding all the successor nodes, but appear to be a too extreme
step. Thus, instead, k best nodes out of successors generated are considered for
further expansion. This gives rise to a new method, called local beam search.
However, the local beam search too will lead to concentration to a specific direction
to those successors which generate more potential nodes. Consequently, this search
also ultimately becomes a local search. A better solution is to elect these k nodes
such that they are not the best successors, but randomly selected, out of the next
generation of successor nodes. This new method is called stochastic beam search.
Thus, we grow a population (of nodes) and from that we generate further another
population by random selection, as well as on some criteria of merit. Thus, we reach
closely to the approach used in genetic algorithms-based search.

9.4 Best-First Search

If a problem that has a very large search space and can be solved by iteration (unlike
theorem proving),2 there is usually no alternative to using the iterative methods. In
such problems, there is a serious issue in bounding the effort to make the search
tractable. Due to this, the search should be kept limited in some way, e.g., in terms
of total number of nodes to be expanded, or maximum depth to which it may reach.
Since there is no guarantee that a goal node will be ultimately reached, an evaluation
function is invoked to help us decide the approximate distance to the goal, for any
given node at the periphery of the search. This or a similar function can also be
used for deciding which tip node to sprout next. Hence, the effort for proper design
of valuation functions limits the later stage difficulty in solving the problem. Note
that heuristics-based methods are called informed search methods, in contrast to the
uninformed methods discussed in the previous chapter.
Among all the problem-solving strategies based on search the heuristics, one of
the most popular methods of exploiting heuristic information to cut down search
time is best-first search strategy. This strategy possess general philosophy of using
heuristic information to assess the “merit” latent in every candidate search avenue
exposed during the search, and then continues the exploration along the direction
of highest merit. There are no local maxima “traps” in best-first search like in hill-
climbing methods. The best-first search strategy is usually seen in context of path-
searching problems—a formulation that represents many combinatorial problems
with practical applications, such as routing telephone traffic, layout of printed circuit

2 In problems such as theorem proving, the search must continue until a proof is found.
9.4 Best-First Search 245

board, scheduling, speech recognition, scene analysis, mechanical theorem proving,

and problem-solving.
The heuristic approach typically uses special knowledge about the domain of the
problem being represented by the graph to improve the computational efficiency of
solution to particular graph-searching problem. However, the procedures developed
via the heuristic approach generally have not been able to guarantee that minimum
cost solution paths will always be found.
Given a weighted directional graph G = (V, E, W ) with a distinguished start
node S and a set of goal nodes R, the optimal path problem is to find a least cost
path from S to any member of R where the cost of the path may, in general, be an
arbitrary function of them, weights assigned to the nodes and branches along that
path. A Generalized Best-First Search (GBFS) strategy will pursue this problem by
constructing a tree T of selected paths of G using the elementary operation of node
expansion, that is, generating all successors of a given node, Starting with S, the
GBFS will select for expansion that leaf node of T that features the highest “merit,”
and will maintain in T all paths which have been encountered so far. And, that still
appear as viable candidates for sprouting an optimal solution path. When no such
candidate is available for further expansion, the search terminates. In that case the
best solution path found so far is issued as a solution; if none has been found, a
failure is declared. Due to its nature of search, the best-first search is also called
branch-and-bound method, i.e., branching a search to other directions to which path
cost is minimum, and bounding the cost to that minimum, until a better minimum is
found after next expansion.

9.4.1 GBFS Algorithm

A best-first search algorithm maintains two lists of nodes, an “Open-list”and a

“Closed-list”. The Closed-list contains those nodes that have been expanded, by
generating all their children, and the Open-list contains those nodes that have been
generated, but not yet expanded. At each iteration of the algorithm, an Open node
having smallest total cost from start node, is expanded, moved to the Closed-list and,
its children are added into the Open-list [2].
The best-first search method takes the best node to be explored first, among
the number of nodes in the open-list. When a node is selected as a candidate for
expanding, its all children’s distance is computed from the start node, which serve as
heuristic value. This approach works because there is always a best node available
for expanding, until the goal is reached or the entire search has taken place.
Since, best node is selected every time, it is guaranteed to give the best solution.
The value of heuristic function f (n) for a given node n, does not here include the
distance from current node to the goal node, as required in Eq. (9.1, page no. 241),
however, since the best path is chosen every time, it is likely to provide the optimum
solution for the problem.
246 9 Heuristic Search

Algorithm 9.2 shows the steps for best-first search.

Algorithm 9.2 Best-first Search(Input: S, Goal)

1: Open = [S]
2: Closed = []
3: repeat
4: if Open.H ead = Goal then
5: return success
6: end if
7: generate children’s set C of Open.H ead
8: if n ∈ C already exists in OPEN and new n is reachable by shorter path then
9: remove the old n
10: end if
11: if n ∈ C already exists in Closed and reachable by shorter path then
12: replace n ∈ C by the same node from Closed, along with shorter distance from root
13: end if
14: remove Open.H ead and insert into Closed
15: update distance from root for all C nodes
16: add all C to either side of Open and record their parents
17: sort Open by path length so that least cost path node is at front
18: until Open = nil
19: return fail

Example 9.1 Best-first search.

Fig. 9.2a shows the graph, and Fig. 9.2b shows the search tree using GBFS for reach-
ing to goal node G.
Every node in best-first search shows the node identification along with its distance
from the root node S. The order in which the nodes are explored is shown with dotted
line. Since G is goal, its path from root is S, A, C, G with shortest path length 11.
Note that any other path will be of longer or equal length. 

S,0
4 4 G
C
A A,3 B,4
3 3
5 C,7 B,8 A,9 D,6
S 5
4 4 E
B
2 E,10
D G,11 C,11
D,12

(a) Graph (b) Best-first Search

Fig. 9.2 Best-First (Branch-and-Bound) search

9.4 Best-First Search 247

9.4.2 Analysis of Best-First Search

As we have noted that, the best-first searches tend to put the searching effort into
those subtrees that seem most promising (i.e., they are most likely of providing the
best solution). However, the best-first searches require a great deal of bookkeeping
for keeping track of all competing nodes, contrary to the great efficiencies possible
in depth-first searches.
Depth-first searches, on the other hand, tend to be forced to stop at inappropriate
moments thus giving rise to the horizon effect (number of possible states is immense
and only a small portion can be searched). They also tend to investigate huge trees,
large parts of which have nothing to do with any solution (since every potential arc
of the losing side must be refuted). However, these large trees sometimes turn up
something that the evaluation functions would not have found were they guiding the
search. Sometimes the efficiencies and discovery potential of the depth-first methods
appear to out-weight what best-first methods have to offer. In fact, both methods have
some glaring deficiencies.
Optimizing Best-First Search
In practice, several shortcuts have been devised to simplify the computation of GBFS.
First, if the evaluation function f (n) used for node selection always provides opti-
mistic estimates of the final costs of the candidate paths evaluated, then we can
terminate the search as soon as the first goal node is selected for expansion, without
compromising the optimality of the solution used. This guarantee is called admissi-
bility and is, in fact, the basis of the branch-and-bound method. This we can observe,
for example in a chess game, where goal is property of a configuration and not the
property of path from start node. Hence, once a winning configuration is reached,
there is no need to try it from other paths.
Second, we are often able to purge from tree T , large sets of paths that are rec-
ognized at an early stage to be dominated (i.e., superior) by other paths in T . This
becomes particularly easy if the evaluation function f is order-preserving, that is,
if, for any two paths p1 and p2 , leading from S to n, and for any common extension
p3 of those paths, the following holds (Fig. 9.3) [2]:

f ( p1 ) ≥ f ( p2 ) ⇒ f ( p1 p3 ) ≥ f ( p2 p3 ). (9.2)

The property of Order-preserving is a judgmental version of the principle of

optimality in Dynamic Programming. The principle states that, a path p1 is judged to
be more meritorious than another path p2 , such that both are paths from one source
state S to some future state n, and there is a common extension p3 of p1 and p2 .
In such a scenario, the common extension cannot later reverse the judgment made
earlier. Under such conditions, there is no need to keep multiple copies of nodes in the
tree T . Every time, the expansion process generates a node n, which already resides
in T , only lower path to node n be maintained, and the link from more expensive
parent of n is discarded. This has been illustrated in Fig. 9.3.
248 9 Heuristic Search

Fig. 9.3 Order-preserving in S

GBFS

p1
p2

n
p3

The best-first search allows revisiting the decisions. This is possible when a newly
generated state by expansion of one of the state in Open-list is found in the closed-list
also. The best-first would retain the shorter path to this node, and purge the other to
save space. However, in a variant of best-first, called, greedy best-first search once a
state is visited the decision is final and the state is not visited again, thus eventually
accepting the suboptimal solution. This however, does not require the Close-list, thus
saving the memory space significantly.
Special Cases of Best-First Search
Individual best-first search algorithms differ primarily in the cost function f (n).
If f (n) is the total depth of node n (not the distance from start), best-first search
becomes breadth-first search. Note that breadth-first searches all the closer nodes (to
start) before farther nodes. If f (n) = g(n), where g(n) is the cost of the current path
from the start state to node n, then best-first search becomes Dijkstra’s single-source
shortest-path algorithm.
If f (n) = g(n) + h(n), where h(n) is a heuristic estimate of the cost of reaching
a goal from node n, then best-first search becomes a new algorithm A∗ algorithm.
Breadth-first search can terminate as soon as a goal node is generated, while
Dijkstra’s algorithm and A∗ must wait until a goal node is chosen for expansion to
guarantee optimality. Every node generated by a best-first search is stored in either
the Open- or Closed-lists, for the following reasons:

1. To detect when the same state has previously been generated. This is to prevent
expanding it more than once.
2. To generate the solution path once a goal is reached. This is done by saving with
each node a pointer to its parent node along an optimal path to the node, and then
tracing these pointers back from the goal state to the initial state.
3. To choose only the node, which is at shorter distance, when a newly generated
node already exists in the closed-list.
9.4 Best-First Search 249

The primary drawback of best-first search is its memory requirements. By storing

all nodes generated, best-first search typically exhausts the available memory in very
short time on most machines.
While breadth-first search manages the Open-list as a first-in first-out queue,
generalized best-first searches manage the Open-list as a priority-queue in order to
facilitate efficiently determining the best node to expand next.
All of these algorithms suffer the same memory limitation as breadth-first search,
since they store all nodes generated in their “Open”or “Closed”lists, and will exhaust
the available memory in a very short time.

9.5 Heuristic Determination of Minimum Cost Paths

The objective of heuristic determination of minimum cost path is to find an algorithm

that searches a graph, G = (V, E) to obtain an optimal path from start node S to its
perfect goal node t. In the search process, each time a new node is expanded, two
things are stored with each successor node: 1. The cost of reaching to n through a least
cost path created so far, and 2. A pointer to the predecessor node of n. Ultimately, the
algorithm gets terminated at some goal node t, and no further nodes are expanded. At
this state, we can reconstruct the minimum cost path from S to t, simply by chaining
back the nodes from t to S through the pointers to predecessor nodes.
In order to expand as few nodes as possible for searching an optimal path, the
search algorithm must constantly make an informed decision about what node is to
be expanded next. The expansion of nodes that are not going to be in the optimal
path, will result to wastage of efforts. On the other hand, if the algorithm ignores the
nodes that might be in the optimal path, it will fail to find such a path, in that case
the algorithm is not admissible. Thus, a good algorithm obviously needs some way
to evaluate the available nodes to determine which node to expand next.
Consider that an evaluation function could be calculated for some node n. Let,
f ∗ (n) is estimated minimum distance from start state to goal state, constrained
through node n. Assume that this evaluation function be defined in such a way
that the node with smallest value of f ∗ is expanded next. We will show that for
a suitable choice of the evaluation function f ∗ , the algorithm A∗ is guaranteed to
provide an optimal path to a preferred goal node from the start node S, which is
sufficient condition for the admissibility of the algorithm.

9.5.1 Search Algorithm A∗

By far, the most studied version of best-first-search is the algorithm A∗ , which was
developed for additive cost measures, that is, where the cost of a path is defined as
the sum of the costs of its arcs. The A∗ is in fact a family of algorithms, we will see
it shortly. The algorithm makes use of ordered state-space search and the estimated
250 9 Heuristic Search

heuristic cost to determine the evaluation function f ∗ , a function which provides the
goal state. This process is similar to the best-first search, but now it is unique in the
sense that it defines f ∗ , which will provide a guarantee of optimal path to the goal
state. The A∗ algorithm is in the class of branch-and-bound algorithms, which are
common in use in operations research for finding the solution of a problem, in the
form of a shortest path in a graph [1].
The evaluation function f ∗ (n) estimates the quality of the solution path through
node n, and it is based on values returned from two components: 1. g ∗ (n), and 2.
h ∗ (n). The first component is the minimal cost of a path from a start state to n, and
the second component, called heuristic value, is a lower bound on the minimal cost
of a solution path from state n to goal state.
In graphs, g ∗ can have error only in the direction of overestimating the minimal
cost. In future steps of the algorithm, if a shorter path is found, the value of g ∗ can
be readjusted to lower side. The function h ∗ carries the heuristic information, such
that it has the capability that ensures that value of h ∗ (n) is less than h(n). This later
condition is essential to the optimality of A∗ algorithm. The property, as per which,
h ∗ (n) is always less than h(n), is called admissibility condition.
To match this cost measure, A∗ employs an additive evaluation function f (n) =
g(n) + h(n), where g(n) is the cost of the currently evaluated path from S to n and
h is a heuristic estimate of the cost of the path remaining between n and some goal
node. Since g(n) is order-preserving and h(n) depends only on the description of
the node n, therefore f (n) is also order-preserving, and one is justified in discarding
all but one parent for each node.
The admissible search algorithm for A∗ (A-star) is given as Algorithm 9.3.

Algorithm 9.3 Admissible search A∗ (Input: G, S, Goal)

1: Open = [S]
2: Closed = []
3: compute f ∗ (n) for all n ∈ Open
4: repeat
5: select the open node n whose f ∗ (n) is smallest
6: resolve ties arbitrarily, but always in favor of any node n ∈ Goal
7: if n ∈ Goal then
8: move n to Closed
9: terminate algorithm
10: else
11: move n to Closed
12: apply successor operator to n
13: calculate f ∗ for each successor n i of n
14: move all n i ∈
/ Closed to Open
15: move to Open any n i ∈ Closed and for which f ∗ (n i ) is smaller now than it was when n i
was in Closed
16: end if
17: until Open = nil
18: return fail
9.5 Heuristic Determination of Minimum Cost Paths 251

9.5.2 The Evaluation Function

Let, for any graph G and any set of goal nodes Goal, let us assume that f (n) is
the actual cost of an optimal path that is restricted to go through only the node n,
i.e., from source S to a preferred goal node n. Note that at the begin of the A∗
search algorithm, the constrained node n is nothing but S. Hence, g(n) = g(S) = 0.
Therefore, f (S) = h(S) is the cost of unconstrained optimal path from node S to
preferred goal node, what so ever it is. Actually, for every node n on optimal path,
the condition f (n) = f (S) holds, and for every node n not on an optimal path,
f (n) > f (S) holds. Thus, although f (n) may not be known in advance, it seems
reasonable to use the evaluation function f ∗ (n) as an estimate of f (n). This is
because determination of the true value of f (n) may be main problem of interest.
In the following, we present some properties of search algorithm A∗ where cost
f (n) of an optimal path through node n is estimated using an evaluation function
f ∗ (n). The function f (n) can be written as the sum of two parts, g(n) and f (n):

f (n) = g(n) + h(n). (9.3)

In the above, g(n) is the actual cost of an optimal path from node S to n, and h(n)
is the actual cost of an optimal path from node n to a preferred goal node.
If we had the estimates of g and h, we could easily get the estimate of f (n), as
the simple addition of the two. An obvious choice for g ∗ (n) is, so far smallest cost
path found by the algorithm, from S to n. This indicates that g ∗ (n) ≥ g(n).
Through a simple example we will explain that the above estimate is easy to
compute as the algorithm progresses through its computations [6].

Example 9.2 Consider the subgraph shown in Fig. 9.4, with start node S and three
other nodes n 1 , n 2 , n 3 , with cost on edges as the weights.
Having given this, we trace the algorithm A∗ as it proceeds. With S as start node,
n 1 and n 2 are the successor nodes. The estimates for g ∗ (n 1 ) and g ∗ (n 2 ) are then 4
and 9, respectively. Let, A∗ expands the next node as n 1 , and obtains the successors
n 2 and n 3 . At this stage g ∗ (n 3 ) = 4 + 3 = 7, and g ∗ (n 2 ) is 4 + 4 = 8. The value of
g ∗ (n 1 ) ultimately remains equal to 4, irrespective of the goal node. 

Fig. 9.4 Admissibility test n3

3
n1

4
4
S
9
n2
252 9 Heuristic Search

We have the following arguments for an estimate h ∗ (n) of h(n), which we are not
able to compute for this example, as no criteria for heuristics is specified here. We
usually rely on information from the problem domain for heuristics. For example,
many problems that can be represented in the form of problem of finding mini-
mum cost path through a graph that contains some “physical” information, and this
information is used to form the basis for estimation of h ∗ . When considering the
connection between cities through roads, the value h ∗ (n) might be air distance from
city n to goal city, because this distance is the shortest possible length of any road
connecting city n to the goal city. Thus, it is lower bound on h(n). However, the
above conclusion is based on the assumption that air connectivity between any two
cities follows a straight line rule.
As another example, in an 8-puzzle problem, at node n, the distance h ∗ (n) might
be equal to the number of tiles dislocated with respect to the goal state.
We will discuss later about using information from specific problem domains, to
form estimate of f ∗ . However, first we can prove that if h ∗ is any lower bound of h,
then the algorithm A∗ is admissible.
Example 9.3 A∗ -Search.
Figure 9.5 shows a graph, heuristic function table for h(n) for every node in the
graph, and tree constructed for A∗ search for the graph, for given start state S and
goal state G. To expand the next node, the one having smallest value of function f is
chosen out of the nodes in the frontiers. The function f for a node n is sum of three
values: the g value of the parent of n, the distance from parent of n to the node n,
and heuristic value (estimated distance from n to goal, given in the table) indicated
by h. In the case of a tie, i.e., two states having equal values of f , the one to the left
of the tree is chosen.
If, in addition, h(n) is a lower bound to the cost of any continuation path from
n to goal, then f (n) is an optimistic estimate of all possible solutions containing
the currently evaluated path. Then, terminating A∗ upon the selection of the first
goal node does not compromise its admissibility. Several other properties of A∗ can
be established if admissibility holds, such as the conditions for node expansion,
node reopening, and the fact that the number of nodes expanded decreases with
increasing h.
Based on the criteria set for A∗ (i.e., to always expand the node having smallest
value of f ). The distance to goal ( f ) for each node n is: distance from source S to
parent, plus distance from parent to this node n, plus distance h from this node n to
goal. For current node n = A, these distances are 0, 1, 6, respectively. The order in
which nodes have been expanded for Fig. 9.5a, and shown in search tree in Fig. 9.5c
with start node S and goal node G are: (S, 0), (B, 6), (A, 7), (B, 5), (C, 6), (C, 7)
(with parent A), (C, 7) (with parent B), (G, 8), (G, 9), (G, 12). Finally, we note that
the best path is corresponding to goal (G, 8), and it is: S, A, B, C, G. Note that, in
the A∗ -tree, we followed the sequence (S, 0), (B, 6), with (A, 7) and not (C, 7),
which are equally weighted. We chose the node to the left-side subtree. Had we
chosen, (C, 7) in place of (A, 7), we would have reached to (G, 9) as next node,
which incidentally was not a good choice. 
9.5 Heuristic Determination of Minimum Cost Paths 253

A State h
1 11 S 7
5 A 6
S 2
C G B 2
4 3
C 1
2
B G 0
(a) Graph. (b) Heuristic Table.
f = 0+7
S =7
f=0+1+6 1 4
= 1+6=7 A B f=0+4+2
2 11 =4+2=6
5 2
f=1+2 +2 f=1+5+1
=3+2=5 B =6+1=7 C G C f=4+2+1
f=1+11+0 = 6+1=7
2 3 3
=12+0=12
f=3+2+1 G f=6+3+0
C G =9+0=9
=5+1=6 f=6+3+0
3 =9+0=9

f=5+3+0
=8+0=8 G (c) A*-Search Tree.
Fig. 9.5 Graph and A*-Search tree

9.5.3 Analysis of A∗ Search

The A∗ is actually a family of search algorithms, as mentioned earlier, and many

other search algorithms are special cases of this algorithm.
The estimated value h ∗ (n) for h(n) can be obtained from the problem domain.
For example, in the case of 8-puzzle problem or the 8-queen problem, this value is
inverse of the number of tiles out of place with the goal or the inverse of number of
queens giving checks to other queens. Even better value is, actual number of moves
from n to goal.

Let us consider the specific cases of f ∗ (n) as follows:

• When h = 0, then g = d (the distance to goal in the search tree). This algorithm
is called as A, and it is identical to Breadth-First Search (BFS).
• We claimed that the BFS algorithm is guaranteed to find the minimum path length
to the goal node. If h is lower bound on h ∗ , i.e., h(n) ≤ h ∗ (n), for all n, then the
algorithm will find an optimal path to a goal node. This algorithm is called as A∗ .
254 9 Heuristic Search

9.5.4 Optimality of Algorithm A∗

A search algorithm B is called optimal if there does not exist any other search algo-
rithm performing the searching in less time or space or can do the job by expanding
fewer nodes, having a guarantee of solution quality as that of algorithm A. Hence,
if h(n) is lower bound on h ∗ (n) then the solution of A∗ is admissible. This estimate
also concludes that any open node n may even be arbitrarily close to a preferred goal
node.
In one way, we can define an optimal search algorithm as one that picks the correct
next node at each choice. However, this specification of an algorithm is not of much
use as this much specification is insufficient to design an algorithm. In fact, whether
such an algorithm may ever exist is also an open question in itself [6].

9.6 Comparison of Heuristics Approaches

The heuristic search that finds the shortest path to a goal wherever it exists is called
admissible. We may like to know, in what sense one heuristics is better than other,
is called informedness of the heuristics.
When search is made, it is expected that same node will not be accessible from a
shorter path later on. This property is called monotonicity.
The breadth-first search algorithm is admissible algorithm, because it searches a
path at level n, before searching the paths at level n + 1, hence if the goal exists at
level n it will be certainly found out. However the BFS algorithm is too expensive
as a general purpose algorithm.
The A∗ algorithm does not require g(n) = g ∗ (n). This shows that there may be
subgoals in the path, which are longer than g ∗ (n); this is due to monotonicity [8].

Definition 9.1 (Monotonicity) A heuristic function h is monotonous if for all the

nodes n i , n j , where n j is descendant of n i , such that

h(n i ) − h(n j ) ≤ cost (n i , n j ), (9.4)

where cost (n i , n j ) is actual cost, in number of moves from node n i to n j .

Definition 9.2 (Informedness) For a problem, suppose there are two A∗ heuristic
functions h 1 and h 2 . Then, if h 1 (n) ≤ h 2 (n), for all states n in the search space, the
heuristic h 1 is called more informed that h 2 .

For example, the criteria of number of tiles out of place, in the 8-puzzle is better
informed than the breadth-first or depth-first search methods. Similarly, the heuristics
which calculates the number of transitions to reach the goal is better informed than
the one considering the heuristics based on the number of tiles out of place. In general,
a more informed is an algorithm, there is less expansion of space for searching [6].
9.6 Comparison of Heuristics Approaches 255

Approaches to Better Heuristics

From the above two examples, we note that in some cases of search, the path matters
(TSP), while in other the path does not matter, and only the final configuration matters
(8-puzzle).
A simple algorithm for heuristic search could be considering only single state at
a time rather than many states corresponding to many paths sprouting at the same
time. Such algorithms are called local search in contrast to the global search, which
maintains many active paths at the same time. The local search algorithms consume
much less memory, usually a constant amount.
The Branch-and-bound algorithms are implicitly enumeration algorithms. These
are the principal general methods for finding out optimal solutions for discrete opti-
mization problems. The branch-and-bound algorithms are based on the following
parameters:

(D, E, L , N , P, U ), (9.5)

where

D: Node dominance function,

E: Set of node elimination rules,
L: Node lower bound solution cost function,
N: Next node selection rule,
P: Partitioning or branching rule, and
U: Upper bound solution cost function.

A branch-and-bound algorithm is a two-step algorithm: first step is a splitting

or branching step which returns two or more smaller sets S1 , S2 , . . ., whose union
covers S, where S is set of candidates. Minimum of f (x) over S is min{v1 , v2 , . . .},
where each vi is the minimum of f (x) within Si . The recursive application of this
step defines a tree structure (a search tree) whose nodes are the subsets of S. The
second step, called bounding, computes upper and lower bounds of the minimum
value of f (x), within a given subset of S.
Application of the branch-and-bound technique has grown rapidly. Representative
examples of this include: flow-shop and job-shop sequencing problem, traveling
salesman problem, integer programming problem, and general quadratic assignment
problem. Though, the branch-and-bound algorithms are usually more efficient than
complete enumeration, however, these algorithms have computational requirements
that usually grow exponentially or high degree polynomial of the problem size n. In
these cases, their usefulness is limited to small size problems.
These are other search techniques based on “natural phenomena”. Under this
we are going to discuss two techniques: (1) Simulated annealing, which is based
on changes in the properties of metals and alloys due to heating them to higher
temperature and then slowly decreasing their temperature; and (2) Something based
on the Darwin’s theory of evolution, called genetic algorithms.
256 9 Heuristic Search

9.7 Simulated Annealing

Annealing is process of treatment of metal or alloy by heating to a predetermined

temperature, holding for a certain time, and then cooling to room temperature to
increase ductility and reduce brittleness. The process of annealing is carried out
intermittently during the working of a piece of a metal to restore ductility lost through
repeated hammering or other working. Annealing is also done for relief of internal
stresses. The annealing temperature varies with metals and alloys, and with properties
desired, but must be done within a range, that prevents the growth of crystals. It is an
optimization algorithm, its strength is that it avoids getting caught at local maxima—
the solutions that are better than nearby, but not best.
Simulated Annealing (SA) is a probabilistic search for the global optimization of a
problem for locating a good approximation to the global optimum of a given function,
in a large search space. The name of the process and its inspiration come from
annealing in metallurgical processes, where a function E(S) needs to be minimized.
This function is analogous to the internal energy of the system in that state. The
goal of SA is to bring the system from some arbitrary initial state, to a state having
minimum possible energy [7].
Process
The SA makes use of heuristics to reach to the goal state, such that at each step, the
heuristic considers some neighboring state s  of the current state s, and probabilis-
tically decides of moving the system to move to s  state or staying in s. When the
above sequence of steps are repeated, the probabilities ultimately move the system to
more stable states at lower energy. Typically, the iterations continue until the system
reaches to a state that is good enough for the application, or until a given number of
iterations are exhausted.
In SA we make one change from the normal heuristic search; we attempt to
minimize the function’s value instead of maximizing. So, instead of heuristic function
it is called object function. This is like a valley descending rather than hill-climbing.
Note that in 8-puzzle, for hill-climbing, we compute inverse of number of tiles out
of place with respect to goal to obtain heuristic value. So, if zero tiles are out of
place (i.e., at goal), the heuristic function is infinite. This would correspond to object
function as zero (or minimum).
The physical substances usually move from higher energy configuration to lower
levels, so that the valley descending occurs naturally. But, there is some probability
that a transition to higher energy will occur, given by
ΔE
p = e kT , (9.6)

where ΔE is positive change in energy level (difference, i.e., current cost–new cost,
so ΔE is negative in valley descending), T is temperature in Kelvin absolute tem-
perature scale, and k is Boltzmann’s constant. As per this property, the probability
to a large uphill move will be lower than probability of small move. Also, the prob-
ability that a large uphill move will take place, decreases as the temperature (T )
9.7 Simulated Annealing 257

Fig. 9.6 Probability of

Probability p
uphill move in simulated

Probability p
annealing, as a function of 1 1
temperature and change in
energy

1 (-ve)
Temperaure T ΔE
(a) (b)

decreases. Figure 9.6 shows the effect of increase of temperature and decrease of
ΔE, on probability.
In other words, the uphill moves are more possible when temperature is high, but
as the temperature decreases, relatively small uphill moves are made until finally
process converge to a local minimum configuration.
The rate at which system cools is called annealing schedule. If cooling occurs
too fast, the physical system will form stable regions high energy. That is, local but
not global minima is reached. If a slower schedule is used, a uniform crystalline
structure, which corresponds to a global minimum, will develop.
In search techniques, change in E (ΔE) is equal to change in object function. The
constant k represents the correspondence between the unit of temperature and unit
of energy. Since it is constant, we take probability p in Eq. (9.6) as
ΔE
p = e T . (9.7)

SA mimics annealing process in metallurgy by combination of random search

and hill-climbing. During metallurgical annealing, alloys are cooled at a controlled
rate to allow for the formation of larger crystals. Larger crystals are chemically at a
lower energy state than smaller ones; alloys made of crystals in the lowest energy
state are comparably stronger and more rigid. At a high temperature, the search is a
random walk, and as the temperature lowers the search gradually transits to a local
search. Capturing this idea in an algorithm yields a random process over a space of
configurations where the probability of actually moving to a new configuration is
determined by the difference in energy and the current temperature. Algorithm 9.4
shows the steps for this process.
We note that SA uses
1. Iterative improvement,
2. Local random search,
3. Exploration, and
4. Greedy search.
When the temperature is high, atoms can move anywhere freely, and have equal
probability. When temperature does down, this freedom is reduced.
258 9 Heuristic Search

Algorithm 9.4 Simulated Annealing

1: T = high
2: generate random solution
3: calculate energy (E) of the solution
4: set initial temperature T (sufficiently high)
5: (Gradually decrease the temperature)
6: while T > cut-off temperature do
7: test solution ← solution
8: for n iterations do
9: adjust test solution
10: calculate energy E of test solution
11: ΔE = E 1 − E 2
12: if ΔE < 0.1 then
13: update solution and energy E
ΔE
14: else if e T > random(0 to 1) then
15: update solution and E
16: end if
17: decrease T
18: end for
19: end while
20: end

Formal Approach
The Boltzmann probability function tends to return True at higher temperature and
False at lower temperature; thus in essence the search gradually shifts from random
walk to local hill-climb.
A simulated annealing algorithm is suitable for minimization of an objective
function f , having the mapping, f : S → R, where S is some finite search space,
and R is real number. Typically, the search spaces, designated as Sn , comprise sets
of bit strings {0, 1}n of fixed length or the set of all possible the permutations over
the set {1, 2, ..., n}.
When considering the search space S, it is necessary to define some notion of
neighborhood N , which is a relation N ⊆ S × S. A function

N : S → P(S) (9.8)

refers to the neighborhood of a search point s ∈ S, expressed as

N (s) = {s  ∈ S | (s, s  ) ∈ N }. (9.9)

Simulated annealing is considered efficient if it can locate a global maximum of

f at sufficiently high probability, and at the same time use fewer number of steps.
SA is a widely used heuristic to NP-complete problems that appear in real life
from job-shop scheduling to groundwater remediation design.
Most analysis of search algorithms usually focuses on the worst-case situa-
tion. There are relatively few discussions of the average performance of heuristic
9.7 Simulated Annealing 259

algorithms, because the analysis is usually more difficult and the nature of the appro-
priate average to study is not always clear. However, as the size of optimization
problems increases, the worst-case analysis of a problem will become increasingly
irrelevant, and the average performance of algorithms will dominate the analysis of
practical applications. This large number domain of statistical mechanics, and hence
of simulated annealing.

9.8 Genetic Algorithms

The Genetic Algorithms (GAs) are search procedures based on the process of natural
selection and genetics. These are increasingly used in applications in difficult search
problems, optimization, and machine-learning, across a wide spectrum of human
endeavor. A GA processes a finite population of fixed-length binary strings. In prac-
tice, the strings are: bit codes, k-ary codes, real (floating-point) codes, permutation
(order) codes, etc. Each of these has their place, but here we examine a simple GA
to better understand basic mechanics and principles [3].
A simple GA consists of three operators: selection, crossover, and mutation. Selec-
tion is the survival of the fittest within the GA. Figure 9.7 shows the sequence of
operations on a population Pn and producing next population Pn+1 . To understand
the operation of the genetic algorithm, a population of three individuals is taken.
Each is assigned a fitness value by the function F, called fitness function. On the
basis of these fitnesses, the selection phase assigns the first individual (00111) zero
copies, the second (111000) two, and the third (01010) one copy. After selection,
the genetic operators are applied probabilistically; the first individual has its first bit
mutated from a 1 to a 0, and crossover combines the next two individuals into two
new ones. The resulting population is shown in the box labeled Tn+1 . Algorithm 9.5
shows the steps for search using GA.

Population at Population at
Tn Tn+1
00111 11100 Mutation 01100
11100 11100 11010
01010 Selection 01010 Cross- 01100
......... over
F(00111) = 0.1 applied
F(11100) = 0.9 probabilistically
F(01010) = 0.5

Fig. 9.7 Sequence of operations in GA

260 9 Heuristic Search

Algorithm 9.5 Genetic Algorithm(Input: Initial Population, fitness function, percent

for mutation, selection threshhold)
1: Initialize the population with random candidate solutions
2: Apply fitness function to Evaluate each candidate’s fitness value
3: repeat
4: Select parents based on fitness value
5: Recombine pairs of parents (crossover)
6: Mutate resulting offspring
7: Apply fitness function to Evaluate new candidates’ fitness value
8: until termination condition/goal is reached

There are many ways to achieve effective selection, including ranking, tourna-
ment, and proportionate schemes, but the key notion is to give preference to better
individuals. Consider that individuals are strings of fixed length. In a selection game
of two-party of such individuals, pairs of strings are drawn randomly from the parental
(original) population, and the better/fitting individuals places an identical copy into
the mating pool. When the whole population is selected in this manner, every indi-
vidual will participate in two tournaments; the best individuals in the population will
win both trials. The median individual will typically win one trial and those worst,
do not win at all.
For the selection to function, there must be some way of determining who is fitter
individual. This evaluation can come directly from the formal objective function,
or from the subjective judgment by a human observer. The ordering used is usually
partial ordering.
The population holds representations of possible solutions. It usually has a fixed
size and is a multi-set. Selection operators usually take whole population into account,
i.e., reproductive probabilities are relative to current generation and diversity of a
population refers to the number of different fitnesses.
If we were to do nothing but selection, GAs would not be very interesting because
the trajectory of populations could contain nothing but changing proportions of
strings contained in the original population. In fact, if run repeatedly, selection alone
is a fairly expensive way of—with high probability-filling a population with the best
structure of the initial population [4].

9.8.1 Exploring Different Structures

To do something more sensible, the GA needs to explore different structures. The

main operator used in GAs is crossover, which can be one-point or multi-point
crossover. A simple one-point crossover is performed using these three steps:
1. two individuals structures are chosen from the population using selection operator,
and considered for mating,
9.8 Genetic Algorithms 261

2. a crossover site along the string length is chosen uniformly at random, and,
3. the values following the crossover site are exchanged between the two strings.
Let the two strings be, A = 00000 and B = 11111. If the random choice of a cross
site turns up at 2, the two new strings following the crossover will be A = 00111
and B  = 11000. These resulting strings are placed in the new population pool, and
the process continues pair-by-pair from original population, until the new population
is completely filled with “off-springs” constructed from the bits and pieces of good
(selected) parents [5].

9.8.2 Process of Innovation in Human

Note that, the selection and crossover are simple operators, which do the job of:
generating random number, copying string, and exchange of partial strings. However,
their combined action makes much of the genetic algorithm’s search ability. To
understand this, we need to think the processing required to be done by human (us)
when we innovate. Often we combine the notions that worked well in one context,
with those that worked well in another context, to generate possibly better ideas (new
notions) of how to attack the problem at hand. In similar way, the GAs juxtapose
many different, highly fit substrings (called as notions) through the combined actions
of selection and crossover to form new strings (can be called as ideas).

9.8.3 Mutation Operator

If selection and crossover provide much of the innovative capability of a GA, what is
the role of the mutation operator? In a binary-coded GA, mutation is the occasional
(low-probability) alteration of a bit position, and with other codes a variety of diver-
sity generating operators may be used. By itself, mutation induces a simple random
walk through string space. When used with selection alone, the combination form a
parallel, noise-tolerant hill-climbing algorithm. When used together with selection
and crossover, mutation acts as both insurance policy and as a hill-climber.

9.8.4 GA Applications

The simplest GAs are discrete, nonlinear, stochastic, highly dimensional algorithms
operating on problems of infinite varieties. Due to this, GAs are hard to design and
analyze [5].
262 9 Heuristic Search

The nature of problems GAs can solve are:

• GAs can solve hard problems quickly and reliably,
• GAs are easy to interface to existing simulations and models,
• GAs are extensible, and
• GAs are easy to hybridize.
Because GAs use very little problem-specific information, they are remarkably
easy to connect to extant application code. Many algorithms require high degree of
interconnection between the solver and the objective function. For example, dynamic
programming requires a stagewise decomposition of the problem that not only limit
its applicability, but also can require massive rearrangement of system models and
objective functions. GAs on the other hand have clean interface, requiring no more
than the ability to propose a solution and receive its evaluation.
Although there are many problems for which the genetic algorithm can evolve
a good solution in reasonable time. There are also problems for which they are not
suitable, such as problems in which it is important to find the exact global optimum.
The domains for which one is likely to choose an adaptive method such as the genetic
algorithm are precisely those about which we typically have little analytical knowl-
edge, they are complex, noisy, or dynamic (changing over time). These characteristics
make it virtually impossible to predict with certainty how well a particular algorithm
will perform on a particular problem, especially if the algorithm is nondeterministic,
as is the case with the genetic algorithm. In spite of this difficulty, there are fairly
extensive theories about how and why genetic algorithms work in idealized settings.

Example 9.4 4-Queen Puzzle.

Suppose we choose to solve the problem for N = 4. This means that the board size
is 42 = 16, and the number of queens we can fit inside the board without crossing
each other is 4. A configuration of 4 queens can be represented as shown in Fig. 9.8,
using 4-digit string made of decimal numbers in the range 1–4. Each digit in a string
represents the position of queen in that column. Thus, all queens in the left-to-right
diagonal will be represented by a string 1234.
To solve this problem we take initial populations as [1234, 2342, 4312, 3431]. Let
us recombine by randomly choosing the crossover after digit position 2. We recom-
bine 1, 2 and 3, 4 members of population, producing [1242, 2334, 4331, 3412]. When
this is combined with the original population, we get [1234, 2342, 4312, 3431, 1242,
2334, 4331, 3412]. Next, a random mutation is applied on members 3431 and 2334,
changing the third digit gives 3421, 2324. Thus new population is, [1234, 2342, 4312,
3421, 1242, 2324, 4331, 3412].
The fitness of a string is proportional to the inverse of number of queens giving
check in each string. for example, in the configuration in Fig. 9.8, total number of
checks of Q 1 . . . Q 4 are: 2 + 3 + 2 + 1 = 8. Similarly in configuration 1234, total
number of checks are 12.
This is because each queen is crossing the remaining three. These numbers in
the remaining seven configurations are: 8, 4, 4, 4, 6, 8, 8. Thus fitness functions of
9.8 Genetic Algorithms 263

Fig. 9.8 4-Queens’ board

Q4
configuration
Q1

Q2
Q3

= 2 3 4 1

above population of elements are [ 12 , 8 , 4 , 4 , 4 , 6 , 8 , 8 ]. Thus, if we need to keep

1 1 1 1 1 1 1 1

a population of size 4, of more fitter members their fitness functions are [ 41 , 41 , 41 , 16 ].

These further combine next time, with population size of 4, having population of
[4312, 3421, 1242, 2324]. This sequence can go on until goal is found in one of the
configuration. 

9.9 Summary

The domain of combinatorial optimization consists of a set of problems, which

requires development of efficient techniques for finding the minimum or maximum
of a function having many independent variables. This function is called cost function
or objective function, and represents a quantitative measure of the “goodness” of
some complex system. Because these combinatorial class of problems contain many
situations of practical importance, heuristic methods have been developed, which
require computations proportional to only a small polynomial of n, where n is size of
the problem. These heuristics cannot be generalized, and are unfortunately problem-
specific.
Two basic strategies are common for heuristics: “divide-and-conquer” and “iter-
ative improvement”. The first approach divides/splits the problem into subproblems
of manageable sizes, then solves each subproblem, and finally the sub-solutions are
patched back together, to get the desired solution.
In iterative improvement-based approach, the heuristics starts with the system
in a known configuration, and then, a standard rearrangement operation is applied
to all parts of the system in turn, until a rearranged configuration that improves
the cost function is discovered. The rearranged configuration then becomes the new
configuration of the system, and the process is repeated until no further improvements
are found.
These methods depend on some heuristics determined by the nature of the prob-
lem. The heuristics is defined in the form of a function, say f , which somehow
represents the mapping to the total distance between start node and the goal node.
For any given node n, the total distance between start and goal node is f (n), such
that
264 9 Heuristic Search

f (n) = g(n) + h(n), (9.10)

where g(n) is distance between start node and the node n, and h(n) is the distance
between node n and the goal node.
One of the heuristic methods is hill-climbing. The name comes from the fact that
to reach the top of a hill, one selects the steepest path at every node, out of the number
of alternatives available. A Hill-climbing method is called greedy local search.
Among all the heuristic-based problem-solving strategies, informed best-first
search strategy is one of the most popular methods to exploit heuristic information
to cut down the search time. This method assesses exploration along the direction of
highest merit, using heuristic information.
Several shortcuts help to simplify the computation of best-first search. First, if the
evaluation function f (n) used for node selection always provides optimistic estimates
of the final costs of the candidate paths evaluated, then we can terminate the search
as soon as the first goal node is selected for expansion, without compromising the
optimality of the solution used. This guarantee is called admissibility.
We are often able to purge from search tree T , large sets of paths that are recognized
at an early stage to be dominated by other paths in T . This becomes possible if the
evaluation function f is order-preserving, that is, if for any two paths p1 and p2 ,
leading from start node S to n, and for any common extension p3 of those paths, the
following holds:

f ( p1 ) ≥ f ( p2 ) ⇒ f ( p1 p3 ) ≥ f ( p2 p3 ) (9.11)

The most studied version of best-first-search is the algorithm A∗ , which provides

additive cost measures, that is, where the cost of a path is defined as the sum of
the costs of its arcs. This algorithm uses ordered state-space search and estimated
heuristic to a goal state f ∗ , like the best-first search.3 But, f ∗ is unique in the sense
that it can guarantee an optimal path to goal.
Based on values returned from two components: g ∗ (n) and h ∗ (n), the evaluation
function f ∗ (n) estimates the quality of a solution path through node n. The one
component, i.e., g ∗ (n), is the minimal cost of a path from start node to node n, and
h ∗ (n) is a lower bound on the minimal cost of a solution path from node n to a goal
node.
There are other search techniques based on “natural phenomena”. Under this
there are two techniques: (1) Simulated annealing, which is based on changes in
the properties of metals and alloys due to heating them to higher temperature and
then slowly decreasing their temperature; and (2) Something based on the Darwin’s
theory of evolution, called genetic algorithms. Annealing is process of treatment of
metal or alloy by heating to a predetermined temperature, holding for a certain time,
and then cooling to room temperature to improve ductility and reduce brittleness.
The process annealing is carried out intermittently during the working of a piece of
a metal to restore ductility lost through repeated hammering or other working.

3 f ∗ is also known as evaluation function.

9.9 Summary 265

Table 9.1 Heuristic search methods

S.No. Search algorithm Properties
1. Best-First Search It depends on definition of f (n). If f (n) = h(n), then it is
likely not optimal, and potentially incomplete. But, A∗ is a type
of best-first search, which is complete and optimal. It is due to
its choice of f (n) that combines g(n) and h(n)
2. Hill-Climbing It is Non-optimal, Incomplete like DFS, follows heuristics
3. Beam Search It is like BFS, expand nodes in f (n) order, and Incomplete for
small k (k best nodes from successors are considered for further
expansion). But, Complete, and like BFS for k = infinity. It is
Non-optimal. When k = 1, Beam search is analogous to
Hill-Climbing method without backtracking
4. Branch and Bound It is Optimal, and g(n) is the cost of path from s to node n and
f (n) = g(n) + 0
5. Simulated annealing Escapes local optima, and is complete and optimal given a long
enough cooling schedule

The Genetic Algorithms (GAs) are search procedures based on natural selection
and genetics. A GA processes a finite population of fixed-length binary strings. In
practice, all these are bit codes, kry-codes real (floating-point) codes, permutation
(order) codes, etc.
A simple GA consists of three operators: selection, crossover, and mutation. Selec-
tion is the survival of the fittest within the GA. Each member of population is assigned
a fitness value. On the basis of these fitnesses, the selection phase assigned zero or
more copies of each individual. After selection, the genetic operators of crossover
and mutation are applied probabilistically to the current population to produce new
population.
The special features of GAs are as follows:
• GAs can solve hard problems quickly and reliably,
• GAs are easy to interface to existing simulations and models,
• GAs are extensible, and
• GAs are easy to hybridize.
Table 9.1 gives comparison of various heuristic methods:

Exercises

1. Answer the following short review questions.

a. In what condition the best-first search becomes the breadth-first?
b. What can you infer from the condition: f (n) = g(n)?
c. What can you infer from the condition: f (n) = h(n)?
266 9 Heuristic Search

Fig. 9.9 State-space search Start Goal

2 3 1 2

1 3

d. In what situation the A∗ search become best-first search?

e. What is the primary drawback of best-first search?
f. Which search method(s) use the priority-queue data structure?
2. Consider the 3-puzzle problem, where the board is 2 × 2 and there are three tiles,
numbered 1, 2, and 3, and blank tile. There are four operators, which move the
blank tile up, down, left, and right. The start and goal states are given in Fig. 9.9.
Show, how the path to the goal can be found using
a. Breath-first search.
b. Depth-first search.
c. A∗ search having g(n) equal to number of moves from start state, and h(n)
is number of misplaced tiles.
Assume that there is no possibility to remember states that have been visited
earlier. Also, use the given operators in the given order unless the search method
defines otherwise. Label each visited node with a number indicating the order in
which they are visited. If a search method does not find a solution, explain why
this happened.
3. Explain what algorithms or heuristics are suitable for solving constraint satis-
faction problems under the following situations. Justify your answers.
a. The problem is so tightly constrained that it is highly unlikely that solutions
exist.
b. The domain sizes vary significantly: some variables have very large domains
(over 1,000 values) and some have very small domains (with fewer than 10
values).
c. Eight-Queens Problem: Arrange eight queens on a chess board in such a
manner that none of them can attack any of the others. (Note: A queen will
attack another queen if it is crossing other queen while moving horizontally,
vertically, or diagonally).
d. The set of variables and set of domains are handled by a computer say,
M. Each constraint is handled by a networked computer, say N . Traffic in
the networks is slow. To check a particular constraint, computer M sends
a message to computer N through the network, which in turn will send a
message back to indicate whether the constraint is satisfied or violated.
4. Suggest a heuristic function for the 8-puzzle that sometimes overestimates, and
show how it can lead to a suboptimal solution on a particular case.
Exercises 267

5. Prove that, if the heuristic function h never overestimates by more than a constant
cost c, then algorithm A∗ making use of h returns a solution whose cost exceeds
that of the optimal solution by no more than c.
6. Give the name of the algorithms that results from each of the following special
cases:
a. Local beam search with k = 1.
b. Local beam search with k = ∞.
c. Simulated annealing with T = 0 at all times.
d. Genetic algorithm with population size N = 1.
7. Explain, how will you use best-first search in each of the following cases? Give
the data structure and explain logic.

a. Speech recognition
b. PCB design
c. Routing telephone traffic
d. Routing Internet traffic
e. Scene analysis
f. Mechanical theorem proving

8. What type of data structure is suitable for implementing best-first search, such
that each node in the frontier is directly accessible, and all the vertices behind it
remain in the order they have been visited.
9. Answer the following in one sentence/one word.
a. How will you detect during the search of a graph that a particular node has
been already visited?
b. Is the best-first search optimal?
c. Is the best-first search order-preserving?
d. Is the best-first search admissible?
10. In the Traveling Salesperson Problem (TSP) one is given a fully connected,
weighted, undirected graph and is required to find the Hamiltonian cycle (a
cycle visiting all of the nodes in the graph exactly once) that has the least total
weight.

a. Outline how hill-climbing search could be used to solve TSP.

b. How good results would you expect hill-climbing to attain?
c. Can other local search algorithms be used to solve TSP?

11. Show that if a heuristic is consistent, then it can never overestimate the cost
to reach the goal state. In other words, if a heuristic is monotonic, then it is
admissible.
12. Suggest an admissible heuristic that is not consistent.
13. Can GAs have Local maximas? If it is not, how does the GAs tries to avoid it?
If yes, justify it.
268 9 Heuristic Search

14. Explain different data structures that can be used to implement the open list in
BFS, DFS, Best-first search?
15. Find out the worst-case memory requirements for best-first search.
16. If there is no solution, will A∗ explore the whole graph? Justify.
17. Define and describe the following terms related to heuristics:
Admissibility, monotonicity, informedness.
18. Show that:
a. The A∗ will terminate ultimately.
b. During the execution of the A∗ algorithm, there is always a node in the
open-list, that lies on the path to the goal.
c. If there exits a path to goal, the algorithm A∗ will terminate by finding the
path of to goal.
d. If there is no solution in A∗ , the algorithm will explore the whole graph.
19. Discuss the ways using which h function in f (n) = g(n) + h(n) can be improved
during the search.
20. Why must the A∗ algorithm work properly on a graph search, with graph having
cycles?
21. For the graph shown in Fig. 9.5, find out whether the A∗ search for this graph is
a. Optimal?
b. Oder-preserving?
c. Complete?
d. Sound?
22. Apply the BFS algorithm for robot path planing in the presence of obstacles for
a square matrix of 8 × 8 given in Fig. 9.10. Write an algorithm to generate the
frontier paths. Assume that each move of robot in horizontal (H) and vertical
(V) covers a unit distance, and the robot can take only the H and V moves. The
start and goal nodes are marked as S and G. Shaded tiles indicate obstacles, i.e.,
robot cannot pass through these.
23. Redesign the problem of robot shown in Fig. 9.10 for A∗ search. Assume that
value of h is number of squares equal to V − i + H − j, where V and H are
both 8.

Fig. 9.10 8 × 8 tiles, with j

obstacles in shades
S

G
Exercises 269

24. Solve the 8-puzzle manually for 20-steps, where heuristic is number of tiles
out of place with respect to goal state. Assume that f ∗ (n) = g ∗ (n) + h ∗ (n)
= 0 + h ∗ (n) = h ∗ (n), so that only the heuristics is deciding factor for next node.
Note that algorithm shall be DFS. In case of ties, give preference to those nodes
which are to the left of the search tree.
25. Find an appropriate state-space representation for the following problems. Also,
suggest suitable heuristics for each.
a. Cryptarithmetic problems (e.g., TWO + TWO = FOUR)
b. Towers of Hanoi
c. Farmer, Fox, Goose, and Grain.
26. Suggest appropriate heuristics for each of the following problems:
a. Theorem proving using resolution method
b. Blocks world
27. If P = “heuristic is consistent”, and Q = “heuristic is admissible”. Then show
that P ⇒ Q. Demonstrate by counter example that Q  Q.
28. Consider the magic-puzzle shown in Fig. 9.11. Suggest the formalism for search-
ing the goal state when started from the start state. (Note that in the goal state
all the rows, columns, and diagonals have equal sums equal to 15).
29. Make use of GA to solve 4-puzzle (Fig. 9.12). A move consists, sliding of either
of tiles 1 or 2, or 3 into the blank tile. Such a movement creates blank tile at
a different position, and the process is repeated until goal state is reached. The
solution requires not only reaching to the goal state, but also finds the trace path
to reach the goal. Construct a suitable fitness function to implement search by
GA, the search should consider only those members of the population which
correspond to valid moves.
30. For the graph shown in Fig. 9.13, make use of DFS and certain depth cut-off to
backtrack the search from that cut-off.
31. Use best-first search for Fig. 9.14 to find out if the search from start node A to
goal node G is

Fig. 9.11 Magic-puzzle

1 2 3 6 7 2
4 5 6 1 5 9
7 8 9 8 3 4
Start state Final state

Fig. 9.12 4-puzzle

1 2 1

3 2 3
Start state Goal state
270 9 Heuristic Search

Fig. 9.13 A graph with start 1 4 7 10

node S goal G 3 6 9
S G
2 5 8 11

Fig. 9.14 Graph with start A 2 B

node A and goal node G
8 6
3
H F

1 5 1 4

E 4 G

7 9
5
D C

a. Oder-preserving
b. Admissible
c. Optimal

32. How you will apply the simulated annealing in the following scenarios? For each
case, give the problem formation so as to compute the ΔE, temperature T , the
states s, s  ; the state space S, and object function f : S → R, and perform 5–10
iterations steps manually.
a. 8-puzzle
b. 8-Queen problem
c. Tic-tac-toe problem
33. One fine morning you find that your laptop is not booting. There can be enumer-
able reasons for this. Assume that you are expert in installation and maintenance
of laptops. Represent the search process for trouble-shooting the laptop by con-
structing a search tree.

a. Suggest, what search method you consider as most appropriate for this?
Also, explain the justification of the particular method you have chosen?
b. What heuristics you would like to suggest for making the search efficient?
c. What are the characteristics of this search? Comment for admissibility,
monotonocity, and completeness of this solution.

34. Assume a population of 10 members, and apply GA to find out the solution
by performing five cycles of iterations, each having selection, mutation, and
crossover. Verify that we are far closer to the solution than we were in the begin
after performing these iterations. Represent the members as bit strings {0, 1}n
for some integer n. Also, fix up some criteria for fitness function, as well as the
probability of mutation, and point of crossover.
Exercises 271

a. 8-puzzle
b. square-root of a number
c. Factors of an integer

35. What are the consequences of the following special cases of GA-based search?
a. Only the selection operation is performed in each iteration, based on the
fitness value.
b. Only the crossover operation is performed in each iteration at a random
position.
c. Only the mutation operation is performed in each iteration at a random bit
position.
36. Simulated annealing is guided by a changing “temperature” value that determines
the likelihood of visiting nodes that appear to be worse than the current node.
How does the search behave for very low and very high temperature values, and
why it behaves so?
37. Select the best alternatives in each of the following questions.
i. The mutation operation is good for the following:
(a) noise tolerance (b) hill-climbing
(c) random walk (d) all above
ii. The following operation of GA has maximum contribution to search:
(a) mutation (b) selection
(c) crossover (d) fitness function
iii. What operation of GA is responsible for random walk?
(a) mutation (b) crossover
(c) none above (d) both a and b
iv. GAs are not good for the following purpose:
(a) finding exact global optimum (b) local search
(c) approximate solution (d) global search
v. GAs are good in environments which are :
(a) complex (b) noisy
(c) dynamic (d) all above
vi. GA is always:
(a) P (b) nondeterministic
(c) NP (d) deterministic

References

1. Bagchi A, Mahanti A (1985) Three approaches to heuristic search in networks. J ACM 32:
I:1–27
2. Dechter R, Pearl J (1985) Generalized Best-First Search strategies and the optimality of A∗ . J
ACM 32(3):505–536
3. Forrest S (1993) Genetic algorithms: principles of natural selection applied to computation.
Science 261:872–878
272 9 Heuristic Search

4. Goldberg DE (1989) Genetic algorithms in search, optimization and machine learning. Addison-
Wesley, Reading
5. Goldberg DE (1994) Genetic and evolutionary algorithms come of age. Commun ACM
37(3):113–119
6. Hart PE et al (1968) A formal basis for the heuristic determination of minimum cost paths. IEEE
Trans Syst Sci Cybern 100–107
7. Kirkpatrick S et al (1983) Optimization by simulated annealing. Science 220(4598):671–680
8. Korf RE et al (2005) Frontier search. J ACM 52(5):715–748
Chapter 10
Constraint Satisfaction Problems

Abstract Constraint Satisfaction Problems (CSPs) is a theory about some special

type of problems, where every move of search is subject to fulfillment of certain con-
straints. The chapter introduces with such problems in the beginning, then presents a
formal general model of such problems with analysis, explains the solution approach
with the synthesis of constraints using simple, as well extended theory of synthesis.
Next, it presents the classes of CSP algorithms—generate and test, backtracking,
discusses how efficiency can be increased, propagation of constraints, followed with
cryptarithmetics, chapter summary, and then at the end a list of exercises for prac-
ticing.

Keywords Constraint satisfaction problems (CSP) · Generate and test ·

Backtracking · CSP representation · Constraints · Synthesizing constraints ·
Theory of synthesis · Cryptarithmetics

10.1 Introduction

This chapter presents the theory of some special type of problems, called, Constraint
Satisfaction Problems (CSP), where one need to search the state space, but every move
is subject to the fulfillment of certain constraints, which is different from the selection
of best fitting state in best-first search or in hill-climbing. Many combinatorial search
problems can also be expressed in the form of CSPs, where the aim is to find an
assignment of values to a given set of variables, subject to some specified conditions
(constraints). For example, the SAT (the satisfiability) problem may be viewed as a
CSP where variables are assigned only to the Boolean values, such that it satisfies
the constraints given in the form of clauses.
A general case of a constraint satisfaction problem is known to be NP-hard.
However, a restricted version of the problem that can be obtained as tractable by
imposing restrictions on the form of constraint interconnections [7].
In an algorithm for the solution of CSP, the constraints imposed on the previous
states also propagates to the next states. If there is a violation of constraints, we
need to backtrack and try other branches in the tree constructed. The constraints only

© Springer Nature India Private Limited 2020 273

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_10
274 10 Constraint Satisfaction Problems

specify the relationships without specifying a computational procedure to enforce

that relationship. Ultimately, the algorithm has to find a solution to the specified
problem.
Consider the problem of preparing a time-table for classes in a department, where
there are courses, teachers, classrooms, periods, and days of the week. The teachers
are assigned subjects, rooms are assigned classes, the time-slots in days and weeks
(periods) are assigned the classes, teachers, and subjects, etc. It should be done to
satisfy many things, like, all the teachers should get the courses of their choice (as
far as possible), there is uniform distribution of teaching load among the teachers,
no teacher should be assigned two classes at the same time in a day, neither it should
happen with the room, or even two teachers teaching a class at the same time! This
is an example of a constraint satisfaction Problem.
It is usually common, to encounter many similar problems to be solved, in job
scheduling in project completion, supply chain management, CPU job scheduling,
and other optimization problems.
A CSP is a mathematical problem defined in the form of a set of objects, whose
collective state during the solution must satisfy a number of constraints or restrictions.
A CSP represents entities in a problem in the form of a homogeneous collection of
finite constraints over some variables, which is solved by methods called CSP meth-
ods. These problems are presently the subject of intense research, in both the fields of
Operations Research (OR) and Artificial Intelligence (AI), because the regularity in
their formulation provides a common basis to analyze and solve problems of many
unrelated fields. The CSPs often exhibit high complexity, requires heuristics and
combinatorial search together to solve such problems in a realistic amount to times.
Examples of such problems are Satisfiability Modulo Theories (SMT), Answer Set
Programming (ASP), and Boolean Satisfiability problem (SAT) [2].
Learning Outcomes of this Chapter:
1. Formulate a problem specified in natural language (e.g., English) as a constraint
satisfaction problem and implement it using a backtracking algorithm or stochastic
local search. [Usage]
2. Order of selection variables and their values for CST. [Familiarity]
3. Complexity issues of CST. [Usage]

10.2 CSP Applications

Application of CSPs range from database queries, e.g., find all x, y, z such that x,
y are components, z is assembly, and failure of x causes the failure of y, to scene
analysis where it is required to analyze the scene by segmenting it, for example, sky
region is labeled blue, vegetation regions are green, and cars are totally surrounded
by either grass or sky. Following are some applications of CSPs:
10.2 CSP Applications 275

• Computer Vision (Interpreting objects in 3D scenes): Scene labeling.

• Solid Modeling: constrained-based design, beautification.
• Advanced Planning and Scheduling: activity scheduling, scheduling production,
airline scheduling.
• Assignment problems: stand allocation for Aircrafts, balancing work among dif-
ferent persons, who teaches what class?
• Electrical engineering: fault location, circuit layout computation.
• Network Management and Configuration: planning of cabling of telecommunica-
tion networks, network reconfiguration without service interruptions.
• Molecular Biology: chemical hypothesis reasoning, protein docking, DNA
sequencing.
• Database Systems: Ensure and/or restore data consistency.
• Cryptography.

We consider a classical graph coloring problem. This problem requires to assign

colors to the vertices of a graph in such a way that if any two vertices are joined by
an edge in the graph, they will have a different color. The variables of this constraint
problem are the nodes, constants (values) are colors in a set, and the constraints
are the requirements that neighboring vertices do not have the same color. For our
purposes, we will restrict the problem to have a finite set of permissible colors.

Example 10.1 Map Coloring Problem.

Consider the case of boundary relations between states of Indian territory, where,
Madhya Pradesh (MP) is connected to Uttar Pradesh (UP), Rajasthan (RAJ), Gujrat
(GUJ), Maharastra (MAH), and Chattisgarh (CHG), all surrounded to MP as shown
in Fig. 10.1.
Assuming that we have been given a task of coloring each region either red,
green, or blue in such a way that no neighboring regions have the same color. To
formulate this as a CSP, we define the variables to be the regions: MP, RAJ , GU J ,
MAH , CH G, and U P. The domain of each variable is the set {red , green, blue}.
The constraints require neighboring regions to have distinct colors; for example, the
allowable combinations for RAJ and GU J is a set of pairs:

Fig. 10.1 Cities connected

as directed graph RAJ UP

MP

GUJ CHG
MAH
276 10 Constraint Satisfaction Problems

{(red , green), (red , blue), (green, red ),

(green, blue), (blue, red ), (blue, green)}

A constraint can also be represented as an inequality RAJ = GU J , provided

that the constraint satisfaction algorithm has some way to evaluate such expres-
sions. There is a solution for this problem as: {RAJ = red , GU J = green, MAH =
red , CH G = green, U P = red , MP = blue}. However, there are many other possi-
ble solutions. 

10.3 Representation of CSP

Many problems, especially in AI, can be represented using constraint satisfaction

methods, in a declarative way by identifying variables of interest for the problem,
in a well-defined domain, and assignment of variables is restricted by constraints.
Formally, a constraints’ network R consists of a finite set of variables V , and a set C
of constraints. Thus, a constraint satisfaction problem is defined as a triple,

V , D, C (10.1)

where,
V = {v1 , v2 , . . . , vn }, is a finite set of variables;
D = {D1 , D2 , . . . , Dn } is a finite set (domains) of values (which may be finite or
infinite); and
C = {C1 , C2 , . . . , Cn } is a finite set of constraints.
Associated with each variable v is a finite, discrete domain Dv . A constraint c
on the variable set Vk ⊆ V in its extensional form, is a subset of the Cartesian-
product of the domains of the afflicted variables. The expression var(c) denotes the
tuple of variables on which the constraint is defined. A relation, rel(c) ⊆ Xv Dv , and
v ∈ var(c) are the relational information of the constraint c. The assignment of a
value d ∈ Dv to a variable v is denoted by v := d . A tuple t of a assignments of
variables Vk ⊆ V satisfies a constraint c, if and only if t[var(c)] ∈ rel(c).
Each constraint ci is a pair (Vi , Ri ) (usually represented as a matrix), where Vi ⊆ V
is a set of variables, called the constraint scope, and Ri is a set of (total) functions
from Vi to D, called the constraint relation (an n-ary relation). Thus an evaluation
of variables is evaluation of a function f from the set of variables to the domain of
values, f : V → D. The evaluation function f satisfies a constraint (v1 , . . . , vn ), R
if (f (v1 ), . . . , f (vn )) ∈ R. A solution is an evaluation that satisfies all constraints.
Hence, given an instance of a constraint satisfaction problem, its solution is a
function f from the set of variables (V ) of that instance to the set of values (domain
D), i.e., f : V → D. This mapping is subject to each constraint (Vi , Ri ) ∈ C, the
restriction of f to Vi , denoted f | Vi , is an element of Ri .
10.3 Representation of CSP 277

Example 10.2 Simple CSP examples.

For a set of values in domain D as a set of real numbers R, there is a relation on

some set of variables {u, v, w}, which is a set of total functions from {u, v, w} to R.
Following is a typical relation of this type:

{f : {u, v, w} → R | 5u + 7v + w = 0}. (10.2)

If we fix an ordering on the variables (w, u, v), then the same relation can be
represented as a set of 3-tuples of real numbers:

{(a, b, c) ∈ R3 | a + 8b + 11c = 0}. (10.3)

10.3.1 Constraints in CSP

A constraint can often be regarded as a mathematical equation, and the constraint

network as a system of simultaneous equations. The solutions of simultaneous equa-
tions correspond to the propagation of constraints. However, the constraints are not
always restricted to equations, they can also be expressed as inequalities and non-
numeric relationships. The CSPs are, therefore, excellently suited for qualitative, as
well as quantitative modeling of systems and physical relationships [8].
A constraint relation can be finite or infinite. A finite constraint relation can be
represented by simply giving an explicit and elaborate list of all the elements in
the relation. However, the infinite constraint’s relation cannot be listed exhaustively.
Alternatively, instead of listing the elements’ mapping, the relations in both the
cases can be expressed using suitable specification language, or linear equations, or
formulas.
There can be many ways to provide the restrictions in mapping values to variables.
Following are types of (restrictions) constraints:

1. Unary Constraints: Involves single variable, e.g., MP = green.

2. Binary Constraints: Constraints involve pair of variables, e.g., MP = U P.
3. Higher order Constraints: These contain the {(red, green), (red, blue), (green, red),
(green, blue), (blue, red), (blue, green)} or more variables, e.g., cryptographic
column constraints, and Professors A, B, C cannot be in the same committee
together.
4. Preference or Soft constraints: For example, the red color is better than green,
cost for each variable assignment, etc.
278 10 Constraint Satisfaction Problems

The specifications in the constraint relations indicate the scope for allowed com-
binations of simultaneous values for the different variables. How many variables in
a scope simultaneously satisfy the constraint, is called arity of the constraint. For
example, unary constraint specify the allowed values for a single variable, while the
binary constraint specify the allowed combinations of values for a pair of variables.
In addition, there is an empty constraint (i.e., 0, 0), having scope of variables and
relations both empty.
We represent constraints’ relations as sets of functions, rather than sets of tuples.
In fact, both the representations are equivalent, as by fixing the ordering of variables
in the scope of constraints, we can associate the functions with the corresponding
tuples. But, we shall prefer to use functional representations because it simplifies the
definitions, as the following examples demonstrate.

Example 10.3 CSP for predicate expression.

Let the variables take Boolean values, and the set of possible values for the variables
is {T , F}. Given this the logical formula ‘x1 ∧ x2 ∧ ¬x3 ’can be used to specify the
constraint with scope {x1 , x2 , x3 }, such that the following constraint relation holds.

{f : {x1 , x2 , x3 } → {T , F} | f (x1 ) ∧ f (x2 ) ∧ ¬f (x3 ) ≡ T }. (10.4)

In a similar way, when the possible values is a set of real numbers R, then the
equation ‘x1 + 2x2 = 7’can be used to specify the constraint with scope {x1 , x2 } and
following constraint relation holds.

{f : {x1 , x2 } → R | f (x1 + 2x2 = 7}. (10.5)


For the representation of constraints, such as the relationship between height and
weight of a person, one of the following possibilities can be used:

• Table: as a lookup table.

• Functions: Normal weight = height (in cm) − 100.
• Heuristics: For example, the under weight can be specified as:

actual weight
< 0.8 (10.6)
normal weight

Propagation algorithm may be distinguished according to what can be propagated

through a variable. It can be classified as:

• Fixed Value quantity: For example, C = 5.

• Set of values, for example, C ∈ {3, 4, 6, 7}.
• Symbolic expression, for example, C = 2y.
10.3 Representation of CSP 279

For propagation of fixed value or set, a simple forward chaining is sufficient

as a constraint strategy for rules. The application of symbolic expressions must be
controlled heuristically.

10.3.2 Variables in CSP

A constraint satisfaction problem can be represented as a constraint graph, where

the vertices correspond to the variables and the edges to the binary constraints.
The constraints specify which pairs of values are allowed. If all pairs are allowed,
we say that there is no constraint. If we place the vertices in a linear order, we
obtain an ordered constraint graph. Such an ordering corresponds to the levels of
a backtrack search tree at which the variables are instantiated. The order in which
a backtrack process chooses variables to instantiate is termed the vertical order of
search, corresponding to depth in the backtrack tree.
Discrete Variables
For a finite domain with n variables, and domain size |D|, there are O(|D|n ) total
number of assignments. The example is Boolean Satisfiability problem (SAT), where
there are n variables (v1 , . . . , vn ), and these can be assigned in total 2n ways; for
8-queens, the complexity is O(864 ). For infinite domains, the data are integers,
characters, and strings. Examples of linear constraints are: v1 + v2 + v3 = 1050,
0 ≤ v1 + v2 + 2 ∗ v4 ≤ 5000, v1 − 3 ∗ v5 ≥ 300, etc.
Nonlinear constraint for minimization can be defined as follows: Minimize f (x)
subject to gi (x) ≤ 0 for each i ∈ {1, . . . , m}, and hj (x) = 0 for each j ∈ {1, . . . , p},
and x ∈ X . A maximization problem can be defined in a similar way.
Considering a CSP of job scheduling, where variables are start and end days of job.
They need a constraint language like, startjobi + 5 ≤ endjobj . There are infinitely
many solutions to this problem. If the constraints are linear, then it is solvable, else
there is no general algorithm. Linear constraints are solvable in polynomial time [6].
Continuous Variables
Building an airline schedule or class schedule uses counter variables.

Example 10.4 CSP Problem for an electric Circuit.

In an electrical circuit, R is a constraint, given the voltages V1 and V2 (see Fig. 10.2).

Fig. 10.2 Constraint in an V1 V2

electric Circuit V1 − V2 = RI
280 10 Constraint Satisfaction Problems

In an extension of the above logic, all the circuit components can be represented
as constraints and connected by common variables (voltage and currents) to form a
constraint network. If the voltage and currents of power source are known, the values
are propagated through the constraint network in the same way as a real current, so
that a value can be calculated for each variable. 

In the following we will consider how restricting the allowed constraints affects
the complexity of this decision problem.

Definition 10.1 Decision problem.

A CSP can also be considered as a decision problem as follows: For any set of
constraints CSP() for the solution of the problem represented as a set of formulas
, and with every,

Instance: is a finite set of constraints C ⊆  which are to be satisfied, then we ask,

Question: Does C has a solution?
If there exists some algorithm that solves every instance in this CSP() in poly-
nomial time, then we say that CSP() is ‘tractable’, and refer to C as a tractable set
of constraints. 

10.4 Solving a CSP

It is helpful to visualize a CSP as a constraint graph (network). The nodes of the graph
correspond to variables of the problem and the edges correspond to constraints.
The representation of states in CSP is in accordance with a standard pattern—a
set of variables with assigned values, successor function, and goal test are written in
generic way. In addition, one can develop effective and generic heuristics that require
no additional domain-specific expertise.

It is fairly easy to see that a CSP can be as a standard search problem as follows:

• Initial state: It is an empty assignment {}, where all variables are unassigned.
• Successor function: A value can be assigned to any unassigned variable, provided
that it does not conflict with previously assigned variables, and does not violate
the constraint.
• Goal test: Whether the current assignment is the complete assignment?
• Path cost: A constant cost (e.g., 1) for every step.

All the CSPs have similar steps for the solution. Every solution appears at depth
n with n variables, this indicates that it is a depth-first search approach. The Path is
irrelevant, so it can also use complete-state formulation. For domain size |D| = d ,
the branching factor is,
b = (n − i)d (10.7)
10.4 Solving a CSP 281

at depth i. Hence the total number of leaves are:


n−1
(n − i)d n = n!d n . (10.8)
i=0

10.4.1 Synthesizing the Constraints

A constraint expression is a conjunction of constraints. The constraints can be repre-

sented in the form of a network such that variables are represented as nodes, and each
constraint is a link/arc/edge between two nodes. Further, the constraints are restricted
as unary or binary constraints. A unary constraint is a predicate on single variable,
and a binary constraint, a predicate on two variables. Consider a small constraint
satisfaction problem of coloring a two vertex complete graph1 using a single color,
say red. This graph contains vertices v1 , v2 , and (v1 , v2 ) is the edge joining these
vertices. The coloring problem here can be expressed as a CSP: X1 , X2 are variables
that represent the colors of nodes v1 , v2 , respectively. The binary constraint gives a
restriction that X1 and X2 are not of the same color. Also, a unary constraint on both
the X1 and X2 requires these nodes to be red. Let the initial domain set for X1 and X2
are colors available as {red, green}; having given this, the problem is represented as
a constraint network, shown in figure below [5].
not-same-color
red ⊂ {red green}1 {red green}2 ⊃ red

In the figure, the nodes identified as {red green}1 and {red green}2 contain possible
values for variables X1 and X2 . The loop symbol (⊂) at each node indicates that a
unary constraint is applicable on each node, and the link “not-same-color”, between
nodes corresponds to binary constraint between them. The first level of inconsistency
in constraint networks is node inconsistency. Note that, the potential domain values
in this example are red and green, while the unary constraint for each node is red.
Hence, we can immediately eliminate the green from both the nodes X1 and X2 . On
doing this elimination, we get the following network.
not-same-color
⊂ {red}1 {red}2 ⊃

The second level of inconsistency is arc inconsistency. The arc between X1 and X2
is inconsistent because, for the value of “red” in X1 , it does not satisfy the constraint
“red is not-same-color as red.” To satisfy this constraint, we remove the color “red”
from both the nodes. This cuts down the search space, but unfortunately, this case
reflects the fact that the problem is impossible to solve. Hence, in this case, there is
no global solution, and the network is “unsatisfiable.”

1A complete graph is one in which all possible edges between pairs of vertices are present.
282 10 Constraint Satisfaction Problems

Fig. 10.3 Arc and path {red green}2

inconsistencies

{red green}1 {red green}3

It is quite possible for a network to have no arc inconsistencies, but still, it can be
unsatisfiable. Now, let us consider the problem of coloring a complete three-vertex
graph with two colors, represented in Fig. 10.3.
Let us assume that a set of possible values of color for each variable is {red green}.
The binary constraint between a pair of nodes is the same, “is not the same color
as.” This network is arc-consistent, e.g., given a value “red” for X1 , the X2 may be
assigned “green”, hence “red is not the same as green” is satisfied. But, there is no
way to choose single values a1 , a2 , a3 for X1 , X2 , X3 , respectively, so that all the three
binary constraints are simultaneously satisfied. If we choose red for X1 , the binary
constraint forces us to choose green for X2 . Having chosen green for X2 , the binary
constraint forces us to choose red for X3 , and that forces us to choose green for X1 ,
but for X1 , red color has been already chosen.
Nevertheless, it may be helpful to remove arc inconsistencies from a network.
This involves comparing nodes with their neighbors as we did above. Each node
must be so compared; however, comparisons can cause changes (deletions) in the
network and so the comparisons must be iterated until a stable network is reached.
These iterations can propagate constraints some distance through the network. The
comparisons at each node can theoretically be performed in parallel and this parallel
pass iterated [3].
Thus, removing arc inconsistencies involves several distinct ideas: local con-
straints are globally propagated through iteration of parallel local operations. It
remains to be seen which aspects of this process are most significant to its appli-
cation. The parallel possibilities may prove to be particularly important.
A more powerful notion of inconsistency is called as path inconsistency, where a
network is path inconsistent if there are two nodes X1 and X2 such that ‘a’ satisfies
X1 , ‘b’ satisfies X2 , a, and b together satisfy the binary constraint between them,
yet there is some other path through the network from X1 to X2 , such that there is
no set of values, one for each node along the path, which includes a and b, and can
simultaneously satisfy all constraints along the path. For example, the network in
Fig. 10.3 is path inconsistent: red satisfies X1 , green X3 , red is not the same color as
green: however, there is no value for X2 which will satisfy the constraints between
X1 and X2 , and between X2 and X3 , while X1 is red, X3 is green.
Consider the problem of coloring the complete four-vertex graph with three colors
(Fig. 10.4). Each node contains red, green, and blue, and each arc again represents
the constraint “is not the same color as.” In particular, path consistency does not fully
determine the set of values satisfying the global constraint, which in this inconsistent
case is the empty set.
10.4 Solving a CSP 283

Fig. 10.4 Network with {r g b}2

path inconsistency
{r g b}4
{r g b}1 {r g b}3

In conclusion, arc and path consistency algorithms may reduce the search space,
but do not in general fully synthesize the global constraint. When there are multiple
solutions, additional searches will be required to specify the several acceptable com-
binations of values. Even a unique solution may require further search to determine,
and the consistency algorithms may even fail to reveal that no solutions at all exist.

10.4.2 An Extended Theory for Synthesizing

As the coloring problem suggests, the general problem of synthesizing the global
constraint is NP-complete, and thus unlikely to have an efficient (polynomial time)
solution. On the other hand, it is found that in specific applications it may be possible
to greatly facilitate the search for solutions.
There are two key observations that have motivated this approach.

1. Node, arc, and path consistency in a constraint network for n variables can be
generalized to a concept of k-consistency for any k ≤ n.
2. The given constraints can be represented by nodes, as opposed to links, in a
constraint network; we can add nodes representing k-ary constraints to a constraint
network for all k ≤ n (irrespective of whether a corresponding k-ary constraint
is given or not). These constraints are then propagated in the augmented net to
obtain higher levels of consistency.

Continuing in this manner by adding successively higher levels nodes to the

network and propagating the constraints in the augmented net, one can get the k-ary
consistency for all k, with maximum k equal to a total number of nodes in the network,
i.e., n. It is clear that constraints are not required to be restricted to binary constraints.
Ensuring that there are no lower level inconsistencies, one moves progressively to
achieve higher level consistencies. However, it may lead to a combinatorial explosion
when we try to ensure that there are no inconsistencies when we progress to achieve
higher level consistencies. The final result is a network where n-ary nodes specify
explicitly, the n-ary constraints which we aim to synthesize, and from here, no further
search is required.

Example 10.5 A Preliminary example based on Synthesis Algorithm.

We consider a crude example of the synthesis algorithm in operation, by way of

motivation for the formal description which follows. Assume that following are
284 10 Constraint Satisfaction Problems

{e f }2 {e f }2

{be bf }12 {ed f g}23 {be bf }12 {ed f g}23

{a b}1 {bc bd bg}13 {c d g}3 {b}1 {bc bd bg}13 {d g}3

(a) (b)
{e f }2

{be bf }12 {ed f g}23

{b}1 {bd bg}13 {d g}3

(c)

Fig. 10.5 Applying synthesis algorithm

constraints given on variables X1 , X2 , X3 : a unary constraint C1 specifies that X1

must be either a or b (C1 = {a b}). Similarly, for X2 , the C2 = {e f } and for X3 ,
the C3 = {c d g}. The binary constraint between X1 and X2 states that: X1 = b,
X2 = e, or X1 = b, X2 = f . Hence, binary constraint C12 = {be bf }. In similar way,
C13 = {bc bd bg} and C23 = {ed f g}.
We are interested to find out the choices for X1 , X2 , X3 , if any, that can simulta-
neously satisfy all these constraints. We begin building the constraint network with
three nodes representing the unary constraints on the three variables. Next, we add
nodes representing the binary constraints, and link them to the unary constraints (e.g.
{be bf }12 represents C12 . The combined figure is shown as Fig. 10.5a.
After we add and link node C12 we look at node C1 and find that element a does
not occur in any member of C12 . We delete a from C1 . Similarly, we delete c from
C3 after adding C23 . The constraint network now appears as in Fig. 10.5b.
Now from C3 we look at C13 and find that there is an element bc in C13 which
requires c as a value for X3 , while c is no longer in C3 . We remove bc from C13 , as
shown in Fig. 10.5c. From this we note that the problem has two solutions as follows:

X1 X2 X3
b e d,
b f g.

But, but the consistency of all the variables have not been simultaneously verified
yet, which is covered using the concept of augmented network. 
Augmented Network
So far we have merely achieved a sort of “arc consistency” (though we indicate the
restriction of the pair bc (in C13 ), as well as the elements a and c).
10.4 Solving a CSP 285

Fig. 10.6 Augmented {e f }2

network
{be bf }12 {bed bf g}123 {ed f g}23

{b}1
{d g}3
{bd bg}13

Next, we add a node for the ternary constraint (i.e., k-ary constant k ≤ n, k = 3,
n = 3). Note that, there was no order-three constraint given originally, hence we
could assume initially of no constraint of order three. But, from the given binary and
unary constraints, we could construct the ternary constraint.
Note that, C1 and C23 together allow only the {bed bf g} as set of triples. This is
used as the ternary node and we link it to the binary nodes shown in Fig. 10.6, which
makes an augmented network.
We look at the new node from its neighbors and vice versa, as we did earlier, to
ensure consistency of the sort we obtained earlier between neighboring nodes. C13
is consistent with the new node: bd is part of bed , bg part of bf g. Similarly C12 and
C23 are consistent with the ternary node. If necessary, we could propagate deletions
around until local consistency is achieved on this augmented network. However,
in this case, the network is already stable; no further changes are required. The
ternary node represents the synthesis of the given constraints. There are two ways
to simultaneously satisfy the given constraints: X1 = b, X2 = e, X3 = d ; or X1 = b,
X2 = f , X3 = g.
In a similar way, nodes can be created for higher order, like, quaternary, and k-ry
consistency checks.

Example 10.6 Applying the Synthesizing algorithm to solve the coloring problem.

We apply the Synthesizing algorithm discussed above for this purpose. There are a
total of six unary constraints C1 −C6 , and nine arc constraints. The arc, which is for
binary, indicates Ci “color is not the same as” Cj . To begin with, all the states are
assigned rgb (Fig. 10.7a). After this, to start with MP is constrained to r, therefore, r
is removed from the rest (Fig. 10.7b). In the next and final stage node consistency is
assured, starting from U P and propagating to the rest of the network, simultaneously,
maintaining the arc-consistency also as shown in figure (Fig. 10.7c).
The progressive assignment of colors is also shown in Table 10.1. 

10.5 Solution Approaches to CSPs

A constraint satisfaction problem is a high-level description of a problem, where, a

model of the problem is represented as a set of variables and their domain values. The
problem statement is in the form of constraints that specify the relations between the
286 10 Constraint Satisfaction Problems

{rgb} {rgb} {gb} {gb}

RAJ UP RAJ UP
{rgb} {r}
MP MP
{rgb} {gb}
GUJ {rgb} CHG GUJ {gb} CHG
{rgb} MAH {gb} MAH

(a) (b)
{b} {g}
RAJ UP
{r}
MP

{g}

GUJ {b} CHG

{g} MAH

(c)

Fig. 10.7 Graph through synthesizing algorithm

Table 10.1 Graph coloring as state-space search

Steps↓ Regions→ MP UP RAJ GU J MAH CH G
0. Initial: {} {} {} {} {} {}
1. rgb rgb rgb rgb rgb rgb rgb
2. (MP = r) r gb gb gb gb gb
3. (U P = g) r g b gb gb gb
4. (GU J = g) r g b g b gb
5. (CH G = g) r g b g b g

variables. However, these constraints only specify the relations between variables,
without any computational procedure/method to enforce such relations. The solution
to such CSP is an algorithm which arrives at results, while satisfying the constraints
relations. The finite domains’ constraint satisfaction problems are typically solved
using some type of search. The most common techniques used for solutions to CSPs
comprise three things: 1. some variant of backtracking search, constraint propagation,
and local search [4].
There are several variants for backtracking. The backtracking ensures that the
solution is consistent, and it saves the search time once it finds that a particular
iteration of search is not going to lead to a solution. The backtrack is usually based
10.5 Solution Approaches to CSPs 287

on the value of a single variable, but it can be done on more than one variable. The
learning through constraints infers and saves the new constraints which can later be
used to avoid part of the search. We can foresee the effects of choosing a variable or
value, through look ahead, and can determine in advance whether the sub-problem
is going to be satisfied or not.
The propagation of constraints is useful because it turns a problem into one that
is equivalent but is usually simpler to solve. In addition, it proves satisfiability or
unsatisfiability of the problem. However, the later is not always guaranteed to happen,
but it always happens for some forms of constraint propagation or some kind of
problems, or both.
The process of Constraint propagation in the course of solution of a CSP problem
modifies the CSP, in other words, the constraint propagation enforces a local consis-
tency, which is a set of conditions related to the consistency of a group of variables
or constraints, or both.
The Local search methods are nothing but incomplete satisfiability algorithms.
They may find a solution to a problem, but may fail also even if the problem is sat-
isfiable. The local search works iteratively to improve an assignment over variables.
At each step of the iteration, a small number of variables are changed values, with
the aim of increasing the number of constraints satisfied by this assignment.
A local search algorithm specific to CSPs is min-conflicts algorithm. The local
search appears to work well when changes to variables are of random nature. The
search, when integrated with local search, is called a hybrid search, and found to
give better results.
Yet another approach for CSP solution is Neighborhood substitution. As per this,
whenever a value a for a variable is such that it can be replaced in all constraints by
another value b, then a is eliminated. The neighborhood substitutions are useful to
find a single solution, as well all the possible solutions.

10.6 CSP Algorithms

We will consider three solution methods for constraint satisfaction problems:

Generate-and-Test, Backtracking (possibly Dependency Directed), and Consistency
Driven. Solving a CSP could mean to find:

• One solution, without preference as to which one,

• All solutions,
• An optimal, or at least a good solution.

The nature of methods for solving a CSP are:

1. Combinatorial methods for finite domain of values, D: The Solutions can be

found by systematic search in D, either traversing the space of partial solutions,
or explore the space of complete value assignments.
288 10 Constraint Satisfaction Problems

2. Analytical methods for the infinite domain of values D: Solutions here can be
found by analyzing the constraints as some (generalized) equation system, either
solving the constraints simultaneously or considering the constraints sequentially
one by one.

A systematic search can be carried out in one of the two ways, either using
Generate and Test approach or through Backtracking.

10.6.1 Generate and Test

We generate one by one all possible complete variable assignments and for each, we
test if it satisfies all constraints. The corresponding program structure is very simple,
just nested loops, one per variable. In the innermost loop, we test each constraint.
In most situation, this method is intolerably slow. The steps for generate-and-test is
shown as Algorithm 10.1.

Algorithm 10.1 Generate-and-Test

1: repeat
2: Assign a value to each variable.
3: if it is a solution then
4: return Success
5: else
6: modify the assignment
7: end if
8: until All assignments are done
9: return fail

This algorithm searches all of the domain of values, D. However, improvements

are possible, like, use an informed/smart generator such that the conflicts found by
the tester are minimized. Such algorithms are called stochastic algorithms.
Example for generate and test is assigning colors to nodes in the graph where
we assign values to variables and modify the assignment till we get a solution;
and generating decimal digits for assignment to variables in the cryptoarithmetic
problem 10.9 discussed later.
Merging the generator with the tester results in a new algorithm, called backtrack-
ing algorithm.

10.6.2 Backtracking

In the following, we go into the deeper concepts of backtracking to solve typical

problems, and will try to express the complexity of solutions in mathematical forms.
10.6 CSP Algorithms 289

Using a computer algorithm to answer a question like “How many ways are there
to do so and so?”, or “list all possible ways for ....”, usually requires an exhaus-
tive search out of a set of potential solutions. One technique for organizing these
searches is the backtrack, which works by continually trying to extend partial solu-
tions. At each stage of the search, if an extension of the current partial solution is
not possible or not found, we backtrack to the original partial solution, and try a
different extension of the existing partial solution. This method is, however, used in
a wide range of combinatorial problems, which includes, parsing, game playing, and
optimization [1].
We assume that the solution to a CSP problem consists of a set of vector
(a1 , a2 , . . .) of undetermined length, which satisfies certain constraints on the com-
ponents that makes it a solution. Each ai (i = 1, 2, . . .) is a member of a finite and
linearly ordered set Ai . Thus, an exhaustive search must consider all the elements of
set A1 × A2 × · · · × Ai , for i = 0, 1, 2, . . . as potential solutions. At the begin, we
start with the null vector (say, ) as a partial solution, and the constraints tell us
which of the members of A1 are candidates for a1 . Let, this set be S1 ⊆ A1 , the least
element of S1 is a1 , and the partial solution is (a1 ). In general, various constraints
which describe solutions show that what subset Sk of Ak comprises candidates for the
extension of the partial solution (a1 , a2 , . . . , ak−1 ) to (a1 , a2 , . . . , ak−1 , ak ). Suppose
the partial solution (a1 , a2 , . . . , ak−1 ) allows for no possibilities for ak , then Sk = φ,
hence we backtrack, and make a new choice for ak−1 . If no more new choices exist
for ak−1 , then we backtrack still further and make a new choice for ak−2 , and this
goes on, till a value next to previously selected value comes.
It is helpful to picture the above discussed process in terms of a tree traversal in
Fig. 10.8. The subset of A1 × A2 × · · · × Ai , i = 0, 1, 2, . . . which, when searched,
can be represented as a tree as follows: the node at 0th level /root node of the tree
is null vector. The children of the root are the choices for a1 . In general, the nodes
at the kth level are the choices for ak , given the choices made for a1 , a2 , . . . , ak−1
as indicated by the ancestors of these nodes. In the tree, the dotted lines show the
backtrack traverse. Since the traversal goes as deep as possible in the tree before
backing up to explore other parts of the tree, such a search is called a depth-first
search (DFS).

Fig. 10.8 Backtrack-tree Start

searches as depth-first search

Choices for a1

Choices for a2 , given a1

Choices for a3 , given a1 and a2

290 10 Constraint Satisfaction Problems

Backtracking Algorithm
Backtracking is performed through recursion, where it maintains a partial assignment
of the variables. All variables are initially unassigned. At each step of the algorithm,
a variable is chosen and all possible values are assigned to it in turn. For each
value assigned, consistency of partial assignment is checked to find out whether it
matches with the constraints. For each case of a consistency check, a recursive call is
performed. When all values for a variable have been tried, the algorithm backtracks
and goes back to previous variable to try a different value for this and start again.
In the basic backtracking algorithm, consistency is defined as—satisfaction of all
constraints whose variables are all assigned.
As the backtrack is no more than an “educated” exhaustive search procedure; it
should be stressed that there exist numerous problems that even the most sophisticated
application of backtrack will not solve in a reasonable length of time.
The checking or verifying the assignment of variables is not done in arbitrary
order. The variables are tried in some order, so as to place in the front those variables
that are highly constrained or with smaller ranges. This ordering has an impact on the
efficiency of solution algorithms, as once the more constrained variables are satisfied,
there are less chances of backtracking the solution. In the process we start assigning
values to variables, check constraint satisfaction at the earliest possible time, and
extend an assignment if the constraints involving the currently bound variables are
satisfied.
The Algorithm 10.2 shows the steps for backtracking.

Algorithm 10.2 Recursive-backtrack()

1: assignment={}
2: if assignment is complete then
3: return assignment;
4: end if
5: v = select-unassigned-variable();
6: for each value d in order-of-domain values do
7: if d is consistent with assignment according to constraints then
8: add {v := d } to assignment;
9: result := Recursive-backtrack();
10: if result is NOT failure then
11: return result;
12: else
13: remove {v := d } from assignment;
14: end if
15: end if
16: end for
17: return fail

The variable and value ordering in backtracking are important. Note the line
contained by this backtracking algorithm:
10.6 CSP Algorithms 291

v = select−unassigned −variable();

which selects the next unassigned variable. A static variable never results in a most
efficient search. We pickup that variable which fails quickly (i.e., with fever legal
moves), to help faster exploration of state space.
However, the following are the problems with backtracking:
• Thrashing. This is repeated failure due to the same reason generated by older
variable assignments.
• Redundancy. This is rediscovering the same inconsistencies.
An example, we use a backtracking algorithm to solve the problem of n × n
non-attacking queens, i.e., in how many ways can n queens be placed on an n × n
chessboard so that no two queens are attacking each other? The following example
presents the solution for n = 4 queens problem [10].
Example 10.7 Four/Eight Queens problem.
“Eight queens” is a classic problem of combinatorial analysis, which requires place-
ment of eight “queens” on a 8 × 8 chessboard such that no pair of queen “attack” each
other, i.e., no two queens are on a common row or column or diagonal. This problem
is ideally suited for a solution by backtracking algorithm. Instead of 8 queens, here
we use its simpler version—a scaled down version to place four queens on a 4 × 4
chessboard so that no two queens attack each other.
Instead of beginning by examining of all 16 C4 = 1820 ways of selecting the loca-
tions for four queens on 16 squares of 4 × 4 board, the first observation is that each
row is to contain exactly one queen. Thus, to begin with, a queen is placed on the
first square of the first row (Fig. 10.9a), and then another queen on the first available
square of the second row (Fig. 10.9b). Observing that this prevents placing a queen
anywhere in the third row (Fig. 10.9c), it is necessary to backtrack to the second row
and move the queen there one square to the right (Fig. 10.9d), so that there is room
for another queen in the third row.
We note that now there are no available squares (Fig. 10.9e) in the fourth row, as
every position in fourth row is crossed by some already placed queen. Therefore, it is
necessary to backtrack all the way to the first row (Fig. 10.9f). After this, there is one
available location in the second row (Fig. 10.9g), followed by suitable locations in the
third and fourth rows (Fig. 10.9h), which gives us a solution. Note that, backtracking
reveals that this solution is unique except for its mirror images, which are, of course,
also solutions. The two mirror images can be obtained by rotating the board along
the X axis and Y axis.
A backtrack search techniques cut down the search space, as they need not test all
combinations of possible assignments to variables—a solution backtracks whenever
it comes across an inconsistency of assignment. However, the backtrack often exhibits
costly “thrashing” behavior.
Similar to 4-queen, the eight queens puzzle is the problem of placing eight
chess queens on an 8 × 8 chessboard so that no two queens attack each other (see
Fig. 10.10).
292 10 Constraint Satisfaction Problems

1 1 1 1
. . 2 2 2
. . . . .
3

(a) (b) (c) (d)

1 1 1 1
. . .
2 2 2
3 3
. . . .
4
(e) (f) (g) (h)

Fig. 10.9 Backtrack based solution for four queens Problem

Fig. 10.10 Solutions to

non-attacking queens on the X
8 × 8 chessboard X
X
X
X
X
X
X

If we extend, a solution to n-queens problem requires that no two queens share

the same row, column, or diagonal. The eight queens puzzle is an example of the
more general n-queens problem of placing n queens on an n × n chessboard, where
solutions exist for all natural numbers n with the exception of 2 and 3. The eight
queens puzzle has 92 distinct solutions. 

10.6.3 Efficiency Considerations

Though the backtracking algorithm does not try all the alternatives of assignments,
hence it is a relatively efficient algorithm. But, there is a number of heuristics, using
which we can cut the cost, and make it more efficient. The following are the ways to
improve the backtrack and make it more efficient [1].
Branch merging
The branches of the search path isomorphic to the already searched path should not
be searched and instead be merged with the already searched path. This will reduce
the total search time.
10.6 CSP Algorithms 293

Preclusion
In the generation of solutions through backtracking, the backtracking should occur
as soon as it is discovered that the current partial solution is in fact not leading to a
solution, and will only add to the time required for the solution.
Search rearrangement
The search should be rearranged such that nodes of low degree occur at the begin-
ning of the search, and of high degree occur in the latter part of the search. This
will not contribute to the explosive growth of search tree in the beginning, and high
branching at the ending nodes does not contribute to much complexity. Since preclu-
sion frequently occurs at a fixed depth, in this strategy fewer nodes may need to be
examined. As an example, when faced with the choice of several ways of extending
the partial solution, e.g., which square to tile next or in which column to place the
next queen, we choose the one that offers the fewest alternatives.
Branch and bound
When searching for a solution of minimum cost, once a solution is found all the partial
solutions with greater cost than the minimum are discarded. When this approach is
used, it is possible to get a good solution early in the search. Note that, this technique
is possible only when the costs are additive in nature.

10.7 Propagating of Constraints

The solution to CSP requires propagation of constraints. Let us find out what is
the basic mechanism to propagate these constraints. Consider that we have to locally
propagate a constraint CJ to a neighboring constraint CK . For this, we should remove
all aK ∈ CK which do not satisfy CJ .
A global propagation of a constraint is defined recursively. To globally propagate
a neighboring constraint CJ , first locally propagate CJ to CK , then if anything was
removed from CK due to local propagation, globally propagate CK through all its
neighbors. The rest of the propagation is similar to arc consistency algorithm (see
page no. 284).
Yet another situation is, a constraint is looked at only when a variable is chosen,
by selecting the variable that is yet unassigned. This also will reduce the propagation
overheads as, before this time it is not required, and hence we are not concerned.
In the following, we discuss the various techniques for the propagation of con-
straints.
294 10 Constraint Satisfaction Problems

10.7.1 Forward Checking

We can drastically reduce the search space by early looking at some of the constraints,
even before the search is started. This process is called forwarding checking. When
a variable, say X is assigned a value, the forward checking process checks each of
the variable Y that is unassigned and connected to X by some constraint. And, a
value in Y ’s domain is deleted that is inconsistent with the value chosen for X . To
apply this in Fig. 10.7, for map coloring, we keep check of remaining legal moves
for unassigned variables, as shown in the Table 10.1. When ‘red’is assigned to MP,
the red color is dropped from all the rest, otherwise, it will conflict with any red with
any of the other variables. Hence, in the table, once the MP is assigned R, the R is
deleted from rest of all states (i.e., variables Y ). This has reduced the ranching on
variables all together, by propagation of information from MP to rest of all states.
In the above, we terminate the assignment when any variable has no legal moves.
Try yourself the forward checking for the 4-queen problem yourself.

10.7.2 Degree of Heuristics

Variable with the largest number of constraints first

Select that variable as the next variable for assignment, which is involved in the
largest number of constraints on other unassigned variables. This will leave less
number of assignment to other variables, hence if it does not succeed, it will return
back quickly without wasting must of the space and time.
Least Constraints value heuristics
Given a variable, choosing the least constraining value leaves the maximum flexibility
for subsequent assignments.

Example 10.8 Constraint Propagation.

A constraint graph is shown in Fig. 10.11, for the domain values {1, 2, 3, 4}. Follow-
ing are the constraints:

Fig. 10.11 Constraint graph

v2

v1 v3 v5

v4
10.7 Propagating of Constraints 295

1. There are two unary constraints: Variable v1 cannot take values 3, 4, and variable
v2 cannot take value 4.
2. There are eight binary constraints, stating that variables connected by edges cannot
have the same value.
It is required to find the solution using the following heuristics: minimum value
heuristics (MVH), degree heuristics (DH), and forward checking (FC).

MVH : v1 = 1, the minimum value integer is selected first.

FC + MVH : v2 = 2, next minimum value is selected with forward checking.
FC + MVH + DH : v3 = 3, the v3 is maximum constrained, due to v1 and v2 adja-
cent nodes.
FC + MVH : v4 = 4, forward checking with minimum value heuristics.
FC : v5 = 1

The first variable v1 is assigned the minimum value (i.e., 1) out of the available
set of values at the start. The forward checking and MVH assign next, minimum
available value to v2 , while checking forward, i.e., v3 , v5 , that there is no conflict.
Next using the FC heuristics, MVH and DH, we assign 3 to v3 . Using FC and MVH
we assign 4 to v4 . Finally, the variable v5 is assigned 1, through forward checking.
In the above, the degree heuristics select that variable, which is involved in the
largest number of constraints on other unassigned variables. It is useful for tie-
breaking. In other words, it is selecting the least constraints, where given a variable
choose the least constraining value (leaves the maximum flexibility for subsequent
assignments). 

10.8 Cryptarithmetics

The cryptarithmetic codes are typical of encrypting some information using these
codes, or hiding an encryption key such that it can be recovered on solution of
the given constraints. The cryptarithmatic puzzles are CSPs that lie at the heart
studies of human and computer and computer problem-solving. Usually, the problem
to be solved is to find unique digit assignments to each of the letters such that
the numbers represented by words do the correct addition or subtraction, or even
any given arithmetic. However, this problem turns out to be complex. In decimal
representation, the constraint of a unique digit mapping to a unique letter reduces
the total number of states, from 10n to 10!/(10 − n)!, where n is a number of unique
letters in the problem.
An example of such a problem is:

GERALD
DONALD
-------
ROBERT
296 10 Constraint Satisfaction Problems

This has n = 10, and one solution is:

197485
526485
-------
723970

For solving such a cryptic problem requires an agent to perform a search. The
speed at which these problems can be solved depends on many factors, like, initial
conditions, and the particular sequence of actions chosen by the agent to solve it while
moving through the search space. The chosen sequence depends on the knowledge
about the problem/puzzle to be solved, as well as the knowledge about which steps
to be taken next.
A simple non-cooperative search strategy is—generate-and-test, where an assign-
ment is made to each letter and tested to see if the problem is solved or partly solved.
In the following example, we show the detailed steps of generate and test, and use
backtrack to solve a similar problem [9].
Example 10.9 Cryptographic puzzle.
Given, a sum as below, find out the values of these 8-variables (S, E, N , D,
M , O, R, Y ∈ {0, 1, . . . , 9}) satisfying the validity of this sum, such that each variable
gets a unique value from set {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}.

S E N D
+ M O R E
========
M O N E Y

For the above, the constraints can be formulated as follows:

c1 : D + E = Y + 10 × v1
c2 : v1 + N + R = E + v2 × 10
c3 : v2 + E + O = N + v3 × 10
c4 : v3 + S + M = O + v4 × 10
c5 : M = v4

In above, variables v1 , . . . , v4 ∈ {0, 1} are carries. Also, Since M = 0, so, S = 0.

In the above, c1 . . . c5 are constraints. The value range for variables is as follows:

D, E, N , R, O, Y ∈ {0, 1, 2, 3, 4, 5, 6, 7, 8, 9},
M , S ∈ {1, 2, 3, 4, 5, 6, 7, 8, 9}.

From above, we get straight forward, v4 = M = 1. Thus, 1 can be removed from

the value sets of all the other variables. Hence:
10.8 Cryptarithmetics 297

D, E, N , R, O, Y ∈ {0, 2, 3, 4, 5, 6, 7, 8, 9}.

Assume that v3 = 1.

c4 : v3 + S + M = O + v4 × 10
1 + S + 1 = O + 1 × 10

Hence, S = 8 or S = 9. Since, M = 1 requires O = 1. So, O = 0, and S = 8.

c3 : v2 + E + O = N + v3 × 10
∴ v2 + E = N + v3 × 10 = 10 + N .

The above cannot be satisfied for any constraint value for E, since E < 10. This
concludes that assumption v3 = 1 was wrong. Hence, this and all the derived values
must be reset, and we backtrack to select v3 = 0.

c4 : v3 + S = O + 9
S = O + 9.

Therefore, the only solution is S = 9, and O = 0. C3 is calculated as:

c3 : v2 + E + O = N + v3 × 10
v2 + E + 0 = N + 0 × 10
∴ v2 + E = N .

The assumption v2 = 0 will leads to contradiction E = N . Hence, v2 = 1. There-

fore, 1 + E = N . The c2 reduces to:

c2 : v1 + N + R = E + v2 × 10
v1 + 1 + E + R = E + 1 × 10
v1 + R = 9.

Because, S is already 9, R cannot be 9. Hence, v1 = 1 and R = 8.

Let us try to propagate the set of values. The present sets are:

D, E, N , Y ∈ {2, 3, 4, 5, 6, 7} (10.9)

From the constraint c1 , since v1 = 1, we have,

c1 : D + E = Y + 10 × 1

That is,
D + E = Y + 10 (10.10)
298 10 Constraint Satisfaction Problems

We note that already assigned values to variables are {S = 9, O = 0, M = 1, R =

8}. From Eqs. 10.9 to 10.10, we note that D + E ≥ 12, Y can be assigned value ≥ 2.
Various possibilities are:

• (D, E) = (6, 7). Thus, 1 + E = 1 + 7 = 8 = N . This is not possible, because R

is already 8.
• (D, E) = (7, 6). Thus, N = 1 + E = 7. This fails, due to conflict with D = 7, and
N = 7.
• (D, E) = (7, 5). Thus, N = 1 + E = 6. Y = D + E − 10 = 2.

The unique solution of the problem is given below.

S E N D 9 5 6 7
+ M O R E + 1 0 8 5
========= => =========
M O N E Y 1 0 6 5 2

10.9 Theoretical Aspects of CSPs

The CSPs are also studies in computational complexity theory and in finite model
theory. The CSPs are well-known for their complexities, which are exponential in
nature. To solve a CSP one needs to establish a relation, like mapping nodes in a
graph to certain colors, satisfying some constraints; assigning queens to blocks on a
chessboard, with certain constraints; mapping letters to digits in a cryptarithmetic,
etc. In all these cases we have certain mapping relations, with certain constraints.
From these observations an important question that can be asked is, whether for each
set of relations, the set of all the CSPs that can be represented using only relations
chosen from that set is P or NP-complete? If such a theorem is true, then CSPs
provide one of the largest known subsets of NP. The classes of CSP that are known
to be tractable, are those where hypergraph of constraints have bounded tree-width
and there are restrictions on the set of constraints relations [7].
Some Properties of CSPs are:
• Solution of CSPs is in general NP-complete.
• Before a CSP is solved, identification of restrictions that make the problem
tractable is important.
• Each CSP can be converted into a binary CSP.
• Over-constrained CSP: The CSP contains more constraints than required which
may be inconsistent and / or redundant.
• Under-constrained CSP: The CSP that cannot be solved uniquely.
10.10 Summary 299

10.10 Summary

In Constraint Satisfaction Problems(CSP), one needs to search the state space, but
every move is subject to the fulfillment of certain constraint, which is different from
the selection of the best fitting state in best-first search or hill-climbing. The aim of
their solution is to find an assignment of values to a given set of variables subject to
specified constraints.
The CSP problems are common in job scheduling, supply chain management,
CPU job scheduling, and other optimization problems. A constraint satisfaction prob-
lem is defined as a triple,
V , D, C

where V is a finite set of variables, D is a set (domains) of values, and C is a finite

set of constraints.
A CSP can be represented as a constraint graph, where the vertices correspond
to the variables and the edges to the binary constraints. Types of variables in CSP
can be either discrete variables or continuous variables. The variables can be further
classified as fixed value quantity, set of values, or symbolic expressions. The type
of constraints can be unary constraints, binary constraints, higher order Constraints
or preference/soft constraints. Application of constraints can be carried out using
synthesis algorithms.
CSPs are typically solved using: variants of backtracking, constraint propagation,
and local search.
The efficiency considerations for CSPs are: preclusion, branch merging, search
rearrangement, and branch and bound.
The constraint propagation in a network is done through forward checking.
The selection of the most fitting variables for the first assignment is done through
the criteria of degree heuristics, i.e., either a variable having the largest number of
available constraints is handled first, or the one having the least constraints value
heuristics.

Exercises

1. How many solutions exists for the map coloring of Indian states (see Fig. 10.1)?
2. A problem called, “Minimal Bandwidth Ordering Problem” is described as fol-
lows: Given is a graph G = (V , E), with V as a set of vertices, E as a set of
edges (x, y), and x, y ∈ V . If we order the vertices, the bandwidth of a vertex
x is the maximum distance between x and a vertex y that is adjacent to it. The
bandwidth of ordering is the maximum bandwidth of all the nodes under that
ordering. The minimal bandwidth ordering problem is to find an ordering with
the minimum bandwidth.
300 10 Constraint Satisfaction Problems

a. Formulate the Minimal Bandwidth Ordering Problem as a CSP. Clearly state

the parameters: V , D, C. Ignore the minimization requirement.
b. Given part (a) above, describe the topology of the constraint graph.
c. Given the part (a), what is the size of the search space?
d. Are there any constraint satisfaction techniques effective for solving this
optimization problem? If yes, explain such techniques, and if no, justify it.
3. Sudoku can be viewed as a binary constraint satisfaction problem.
a. What are the variables of this CSP?
b. What are their domains?
c. How would you translate the requirement that no two of the same digit may
occur in the same row, column, or block into binary constraints?
d. Does the requirement that each digit occur at least once in each row, column,
or block have to be directly specified? Why or why not?
4. For what size of n, the n-queen problem has no solution, with n in the range
1-10.
5. Find out, what techniques have been used for propagation of constraints at various
stages in the solution of the problem 10.9.
6. What algorithms or heuristics are relevant to solving some constraint satisfaction
problems in the following situations. Justify your answers.
a. The domain sizes of the problem vary significantly: some variables have very
large domains (over 1, 000 values) and some have very small domains (less
than 10 values).
b. The problem is so tightly constrained that it is quite unlikely that a solution
exists.
7. Solve the following CSP problems:
a. T W O + T W O = FOU R
b. ABC + DEF = GHIJ
c. ALFA + BETA + GAMA = DELTA
8. Give an algorithm for CSP. Justify that this standard algorithm is a depth-first
search (DFS). Can the CSP be implemented as:
a. Breadth-first search,
b. Best-first search.
Justify. Also, modify the existing DFS algorithm to work as BFS, and as best-first
search.
9. Consider the problem of coloring a complete three-vertex graph with three colors,
where the nodes X1 , X2 , X3 are all the set {r g b}; X12 , X13 , and X23 all equal
{rg rb gr gb br bg} and X123 = {rgb rbg brg bgr grb gbr}, the six possible
colorings. Make use of extended synthesizing algorithm to color this graph.
Exercises 301

Fig. 10.12 Tiles to be

colored
X1 X2 X3

X4 X5 X6

X7 X8 X9

10. For the solution of a CSP problem, explain,

a. Why it is good heuristics to first choose a variable that is most constrained?
b. Why it is good heuristics to first choose a value of a variable that is least
constrained?
11. For the solution of a CSP problem, explain,
a. For better heuristics, why it is preferred to first choose a variable that is most
constrained?
b. For better heuristics, why it is preferred to first choose a value of a variable
that is least constrained?
12. The Fig. 10.12 shows 9-tiles in the form of variables (X1 . . . X9 ) for CSP. Write
the steps of an algorithm to color these tiles using R, G, B colors such that no
adjacent tiles have the same color. Also, so the final assignment of colors to these
tiles. (Note: A tile is adjacent to others if it is in the same row or column and
near to the other.)
13. In a department of engineering, there are teachers (T1 . . . T8 ), subjects (courses)
(S1 . . . S6 ), tutorials (t1 . . . t3 ), lecture theaters (LT1 . . . LT3 ), tutorial rooms
(tr1 , tr2 ). The courses and tutorials are to be allocated to teachers and rooms
on days (D1 . . . D5 ), and periods (P1 . . . P6 ) in a day. The constraints are: no two
subjects be given to a teacher on the same day, the teacher should get subject in
the same LT every time, as well as the tutorial class in the same tr. The tutorial ti
shall be engaged by the same teacher who is engaging the course Si . Define the
V , D, C, elements of each set, and show the part of the graph for assignment.
14. Consider the following definition of Comma-Free Codes: They have the property
that, if a1 a2 . . . ak , and b1 b2 . . . bk are in the subset, then none of the “overlaps”
a2 a3 . . . ak b1 , a3 a4 . . . b1 b2 , . . . , ak b1 . . . bk−2 bk−1 are in the subset. Such a sub-
set is called a comma-free dictionary. This is because, in any “message” consist-
ing of consecutive words from this collection, such as a1 a2 . . . ak , b1 b2 . . . bk ,
one may omit the commas or other separations between words, without the pos-
sibility of ambiguity as to word beginning and ending, even if one begins at a
random starting point within the message.
302 10 Constraint Satisfaction Problems

In an investigation of codes with some synchronization properties, we consider

nk “words” which consist of k alphabets from a n-symbol alphabet. In this, we
look for the largest subset of these words which has the comma-free property.
As an example with three and four-letter English words: if BOOK and INK
were in the dictionary, then KIN could not be in the dictionary because of the
ambiguity in booKINk, as it would result to identification of individual words
as boo+KIN+k, which is inconsistent.
Find out other words for n-symbols alphabets, and possible upper limits, if any?

References

1. Bitner JR, Reingold EM (1975) Backtrack Program Tech. Communications of the ACM
18(11):651–655
2. Bordeaux L et al (2006) Propositional satisfiability and constraint programming: a comparative
survey. ACM Comput Surv 38(4):1–54. https://doi.org/10.1145/1177352
3. Cooper MC (1997) Fundamental properties of neighborhood substitution constraint satisfaction
problems. Artif Intell 90:1–24
4. Dorndorf U et al (2000) Constraint propagation techniques for the disjunctive scheduling
problem. Artif Intell 122:189–240
5. Freuder EC (1978) Synthesizing constraint expressions. Commun ACM 21(11):958–966
6. Freuder EC (1985) A sufficient condition for backtrack-bounded search. J ACM 32(4):755–761
7. Jerrum M (2010) Constraint satisfaction problems and computational complexity. Commun
ACM 53(9):98–106
8. Michalowski M (2008) A general approach to using problems instance data for model refine-
ment in constraint satisfaction problems. A Dissertation Presented to the Faculty of graduate
school, USC, Computer Science Department
9. Scott H (1991) Cooperative solution of constraint satisfaction problems. Science 254:1181–
1182
10. Solomon W et al (1965) Backtrack programming. J ACM 12(4):516–524
Chapter 11
Adversarial Search and Game Theory

Abstract Game theory is the formal study of conflict and cooperation, first time
introduced as long back as 1921 by mathematician Emile Borel, then enriched in
1928 by von Neumann and Oskar Morgenstern, and much enriched by Josh Nash,
has enormous applications, including in business, and even in the prediction of
election results, etc. The game playing is also a search process. The chapter presents
the classes of games as combinatorial and games of chance, then further as zero-sum
games and non-zero-sum games, the prisoner’s dilemma, game playing strategies,
the games of perfect information, arbitration scheme in games, minimax search in
game playing, and analysis of specific games like tic-tac-toe. The more efficient
search processes like alpha and beta are presented, as well as the alpha cutoff and
beta cutoff methods to prune the search process are presented and analyzed, followed
with chapter summary, and an exhaustive list of exercises along with a number of
multiple-choice questions provided at the end.

Keywords Game theory · John Nash · Nash equilibria · Zero-sum games ·

Non-zero-sum games · Games as search · Prisoner’s dilemma · Arbitration
scheme · Minimax search · Game theory applications · Tic-tac-toe · Alpha-beta
search · Alpha cutoff · Beta cutoff

11.1 Introduction

The Game theory is the formal study of conflict and cooperation. Game theory’s
concepts apply whenever the actions of several agents are interdependent. These
agents may be individuals, groups, firms, or their combination. The concepts of
game theory provide a language to formulate, structure, analyze, and understand
strategic scenarios. Cooperative game theory concentrate on what agreements the
agents are likely to reach, while noncooperative game theory concentrates on what
strategies are likely to be adopted by the agents. Both of these traditions have adopted
models of bargaining in which each agent’s risk posture is conveyed by comparing
his preferences for risky and riskless alternatives.

© Springer Nature India Private Limited 2020 303

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_11
304 11 Adversarial Search and Game Theory

The formal game theory was introduced in 1921 by mathematician Emile Borel,
which was further advanced by the mathematician John von Neumann in 1928 in
a text by title “Theory of parlor games.” The Game theory became a field of study
in mathematics, subsequent to the publication of the book “Theory of Games and
Economic Behavior” in 1944 by John von Neumann and the economist Oskar Mor-
genstern. John Nash, in the year 1950, demonstrated that finite games have always an
equilibrium state. At this equilibrium state all the players choose actions that are the
best for them given their opponents’ choices. From that time onward, the concept of
equilibrium has remained the focal point of analysis of noncooperative game theory.
In the year 1960, the scope of game theory got widened and it was being in problems
related to war and politics. The whole subject of game theory caught special attention
when Nobel prize was awarded jointly to John Nash, John Harsanyi, and Reinhard
Selten, in the field of Economics, in 1994.
Another branch of game theory, called, Combinatorial Game theory (CGT) stud-
ies strategies and mathematics of two-player games of perfect knowledge such as
chess or go. But generally concentrates on simpler games like nim or solving end-
games or their special cases. An important difference between this subject (i.e.,
CGT) and classical game theory (a branch of economics, after von Neumann and
John Nash) is that in CGT, game players are assumed to move in sequence rather
than simultaneously, so there is no scope in randomization or information hiding
strategies.
The CGT does not study games with importance in computer knowledge called
games of chance, like poker (card games). The combinatorial games include games
like chess, checkers, Go, Arimaa, Hex, and Connect. They also include one player
combinatorial puzzles, and even no player automata.
In CGT, moves are represented as game-tree, and the trees can be searched while
making moves in the game. A game-trees is a special kind of semantic tree where
nodes represent board configurations and branches indicate how one board configu-
ration can be transformed into another configuration by a single move. The decisions
regarding the next move are made by two adversaries, who play the game, hence the
name adversarial search.
Since the number of nodes in a binary or higher order tree generally increases
exponentially with increase in depth of the tree, it is usually unfeasible to do an
exhaustive search in a game-tree. The search-procedures, called as, alpha-beta search
are commonly used to guarantee correct results without exhaustively searching for
such trees. However, such procedures still require complete search of the tree in
special situations. The fact of these game-trees is that some times they are so large that
you cannot fit them even in the hard-disk, hence practically impossible to completely
search such trees.
Improved results can be obtained in the game playing computer programs, by
searching the tree to only a limited depth. This is possible by static evaluation func-
tions to estimate the utility vales at some given nodes. This is followed by computing
the utility values at shallower nodes in the tree assuming that estimated utility values
were correct. This technique is called heuristic game-tree searching, is used by us
in “real-life” decision situations.
11.1 Introduction 305

It is often a universal agreement that when a game-tree is heuristically searched,

increasing the depth of search always improves the quality of decision. This has
been demonstrated with many game-playing computer programs, but such results
are purely empirical.
This chapter presents the idea of game theory, where each of the opponents search
the tree, like DFS and BFS. However, the strategy of each player’s move (in two-
player game) is to defeat others. There is a score value, which one player (called
maximizer) tries to maximize, while the other (called minimizer) tries to minimize.
The search time can be reduced using alpha-beta search method.
The games theory has applications in finance and business systems, economics,
and other areas where one needs to build a model in two more contradicting
situations [7].
Learning Outcomes of this Chapter:
1. Game theory and its applications. [Familiarity]
2. Equilibrium state in games. [Usage]
3. Types of games. [Familiarity]
4. Compare and contrast basic search issues with game playing issues. [Familiarity]
5. Apply minimax search with alpha-beta pruning to prune search space in a two-
player game. [Usage]

11.2 Classification of Games

Games are classified into several categories based on certain significant features, the
most obvious of which is the number of players involved. A game can thus be desig-
nated as one-person, two-person, or n-person (n > 2). A player, i.e., a participant in
a game need not be a single person. If each member of a group has the same feelings
about how the game should progress, and end, the members may be called as a single
player. A player may also be a corporation, or a basketball team, or even a country.
In games of perfect information, such as chess, each player knows everything
about the game at all times. The Poker, on the other hand, is a game of imperfect
information because players do not know about the cards that others are having.
The extent to which the players are opposed or coincidence is another basis of clas-
sifying the games. The zero-sum or to be precise constant-sum games are completely
competitive. The Poker, for example, is zero-sum because the combined wealth of
the players remains constant; if one player wins the other must lose. In non-zero-sum
games, all players can be winners or losers. In a dispute between management versus
labor, if some agreement is not reached, both will be losers. For example, if a labor
strike cannot be avoided.
In cooperative games, the bidders may communicate and make agreement as
binding to both the players, and in noncooperative they may not. The examples
of players in cooperative games are car salesman versus customer, management
versus employees, whereas, bidders for government contracts are supposed to be
noncooperative bidders.
306 11 Adversarial Search and Game Theory

The games can further be classified as finite and infinite games. The games are
finite if the player has to make only a finite number of decisions, and has only a finite
number of alternatives. When either the decisions are infinite or the alternatives are
infinite, the game is called infinite.
The simplest of all is a one-person game, with no opponent. As an example, for
going to the office he/she may choose to go on foot or ride on a public transport
or hire a cab, are the alternate moves. Each one of us as an individual always play
many one-person games daily, like, doing shopping, doing homework given by a
class teacher, managing kitchen, cleaning house, doing meditation; assuming that
no other person or robot is involved in the above tasks. Each of these games (the
tasks), has many steps or moves, and one needs to choose a move among number
of available next move/choices. For example, while shopping, we may have options
to: first collect vegetable items or the grocery items, and before the final move of
making payment of bills, we may choose the move as, pay by cash, or by card.

11.3 Game Playing Strategy

Let there be a facility by which in a board game, every unique board configuration
can be converted into a unique single, overall quality value in the form of an integer
number. Further, consider that positive number indicates the move in favor of one
player, and negative—in the favor of his/her opponent. The degree of favoredness is
the absolute value of that number. Let us call ourself as maximizer player, and the
opponent as minimizer. Also, let us imagine that there is a number (call it static score,
accumulated so far), the maximizer will make a move such that the score increases
to maximum, while the opponent (adversary) makes a move so that by addition of
quality number, the score number minimizes.
The game playing is favorite for AI search, because:
• It is a structured task, often a symbol of “intelligence”,
• clear definition of success and failure,
• does not require large amounts of knowledge (at first glance),
• focus on games of perfect information, and
• multi-player, and chance game.
The difference between the games and search problems is that the games playing
are highly unpredictable, as one does not know what move the opponent is going to
make. And, only based on the opponent’s move the other player has to make a move.1
The other difference is that the time limits for a move are in realistic times, hence
which one is unlikely to find a goal in that time, approximations are necessary for
both the players. The formal system behind these games is called Game Theory [6].

1A player can learn from the opponent’s past moves as to how and what strategy the opponent
followed. But, this requires the learning ability. Hence, the game’s moves shall not be based current
state, but past also. This has not been considered for the present discussion.
11.4 Two-Person Zero-Sum Games 307

11.4 Two-Person Zero-Sum Games

The term “zero-sum” (or equivalently “constant-sum”) means that players have dia-
metrically opposed interests. The term comes from parlor games like poker, where
there is a fixed amount of money around the table. If one player wins some money,
others have to lose an equivalent amount. Two nations trading make a non-zero-sum
game since both are simultaneously in gain. An equilibrium point is a stable outcome
of the game associated with a pair of strategies [2].
Consider the Fig. 11.1, which is played between you and your opposite. The
numerical values in the cells indicate the figures in dollars your opponent will pay
to you. To start the game, assume that you select row B. Next, your opponent will
select a column, say I I . Therefore, your opponent will pay you $4. Next, say, you
select row C, your opponent, suppose selects column I I I . In this case, you pay to
your opponent one dollar, as the value at the junction is −1. The two-player zero-sum
game is a perfect information game.
A strategy in game theory is a complete plan of action that describes what a
player will do under all possible circumstances. There are poor strategies and good
strategies.
Other examples of two-person zero-sum games of perfect information are chess,
checker, and Japanese go. Such games are strictly determined, i.e., rational players
making use of all available information can deduce a strategy that is clearly optimal,
hence the outcome of such games are deterministic. In chess, for example, one of
these three possibilities exists: (1) white has a winning strategy, or (2) black has
a winning strategy, or (3) black and white each have a strategy that is winning or
draw [9].
Two-Person Non-zero-sum Games
Most games that arise in practice are non-zero-sum games, where players have both
common as well as opposed interests. For example, the buyer vs. seller is a non-zero-
sum game. The buyer wants a lower price and seller a higher price, but both want
that deal to be executed. Similar is the case with the game of hostile countries, who
may disagree on many issues but try to avoid the war. Many obvious properties of
zeros-sum games are not available in non-zero-sum games. One of the differences
is the effect of communication on the game, which does not help the opponent in
zero-sum game.

Fig. 11.1 Two-person YOUR OPPONENT

zero-sum game
I II III
A 5 -2 1

YOU B 6 4 2

C 0 8 -1
308 11 Adversarial Search and Game Theory

The game of labor-union vs factory owner is a non-zero-sum game. If the man-

agement is properly informed about the demands of a labor union, probably, the
strike may get withdrawn, benefiting both the laborer and factory owner. The games
where players communicate, make a binding agreement between two parties is called
cooperative games, and where players are not allowed to communicate are noncoop-
erative games. In a non-zero-sum game, unlike in Fig. 11.1, each table entry consists
of a pair of numbers. This is because the combined wealth of the players is not con-
stant. Since it is not possible to deduce one player’s payoff from the payoff of the
other player, both player’s payoff must be specified. The prisoner’s dilemma is the
best example of non-zero two-player noncooperative [2].

11.5 The Prisoner’s Dilemma

Two criminals A and B are arrested and interrogated separately. Each suspect may
either confess (defect the other prisoner) with a hope of a lighter sentence, or may
refuse to talk (cooperate the other prisoner). The police do not have sufficient infor-
mation to convict the suspects unless at least one of them confesses. If they cooperate,
then both will be convicted to minor offense and sentenced to a month in jail in the
absence of supportive evidence. If both are defecting, then both will be sentenced to
jail for six months. If A confesses and B does not, then the A will be set free but the
B will be sentenced to one year in jail, and vice versa when A does not confess and
B confesses.
The police explain these outcomes to both suspects and tell each one that the
other suspect knows the deal as well. Each suspect must choose his action without
knowing what the other will do. This is shown in a table in Fig. 11.2.
If we observe closely the outcome of different choices that are available to the
suspects, it becomes clear that regardless of what one suspect chooses, the other sus-
pect feels better off by choosing to defect, which means he confesses. This concludes
that both the suspects choose to defect, and stay for six months in jail. This choice
is clearly a less desirable outcome than that of only one month in jail, which is the
result if both cooperate to each other by not confessing (i.e., cooperating).
We assume that there is no honor among the criminals, and each one’s sole concern
is to save himself. This game is called prisoner’s dilemma. Since a suspect must

Fig. 11.2 Prisoner’s

dilemma
11.5 The Prisoner’s Dilemma 309

Fig. 11.3 Prisoner’s

dilemma applied for business

make his own decision without knowing that of others, he must consider each of his
partner’s alternatives and anticipate the effect of those on him.
Supposes that A confesses, B must go to jail for one year or six months. Alter-
natively, if A does not confess, then either B is set free or given one month’s jail. In
either case, for B it is better to confess!
However, if both cooperate to each other, both will not confess, and both go for
miner term in jail of 1 month. If both confess, they go to jail for 6 months each.
Further, we conclude that both may be cooperative or uncooperative. In coopera-
tive case, each does better than they were uncooperative. Also, for any fixed strategy
of the opponent, other player does better by playing uncooperatively [2].

Example 11.1 Prisoner’s dilemma applied for business problem.

Consider the Fig. 11.3, which shows that there are two business firms A and B, trading
in different regions, and each spending either $1,000 or $2,000 for publicity. The total
of publicity and sale figure is $12,000 (12k) always. If both the firms spend $1k for
publicity, their sale is $5k each, if both spend $2k, the sale of each is $4k. Similarly,
if A spends $1k and B spends $2k for publicity, A has sale of $3k, while B’s sale is
$6k. Similarly in the reverse case. We assume that in $2k publicity expenditure for
each together, the extra market efforts is wasted, hence the sale comes down.
If any party is spending $2k each year, it will induce the other firm to spend also
$2k next time. A more optimistic strategy is to signal your intent to the opponent of
$1k expenditure, so he also spends $1k, and this cooperation results to $5k sale for
each. But, two competing firms know that they would not do the business forever,
hence they loose the cooperation, and each spends $2k, and the profit comes down
to $4k each. 
Similar conditions to the prisoner’s dilemma may appear in water consumption by
the citizens. Imagine that there is water shortage and citizens are urged to cut down
the water consumption. Each citizen responds to these requests by considering his
own interest, so no citizen will conserve water. This is with the prevailing concept:
any saving by an individual has a negligible effect on the city’s water supply. If
everyone acts as per his own interest, the results will be catastrophic for all.
Imagine that two rival nations are preparing their military budget. Each nation
wants to maintain a military advantage over the other by building a powerful army,
and each spends accordingly. They ultimately end up having the same relative power
and good deal of power.
310 11 Adversarial Search and Game Theory

11.6 Two-Player Game Strategies

There are two key assumptions in playing the games:

1. Every player in the market acts as per his/her self interest. The players pursue
well-defined exogenous objectives, i.e., they are rational. The players understand
and try to maximize their payoff functions.
2. While selecting a strategy/plan of action, a player estimates the potential response/
reactions of other players. The player also takes into account his/her knowledge
or expectation of other players’ behavior, that is, the player reasons strategically.

Dominant Strategies
A dominant strategy is one that provides roughly the highest payoff for a particular
player, irrespective of what strategies are employed by other players [9].
A formal definition of dominant strategy can be given as follows: Let there be a
game played between n players, and let pi (s) is the pay off to player i as a function
of all other players’ strategies s, where,

s = (s1 , s2 , . . . , sn ). (11.1)

For the sake of simplicity, let us define s−i as a combined strategy of all n players,
except player i as,

s−i = (s1 , s2 , . . . , si−1 , si+1 , . . . , sn ). (11.2)

The strategy si is called dominant for player i if for any strategy s−i and any
alternative strategy si , provided that the following pay off relation, holds for player
i with respect to all players s−i .

pi (si , s−i ) ≥ pi (si , s−i ). (11.3)

Definition 11.1 (Dominant strategy) For a player i, a strategy si is called a dominant,

if no matter what the other players choose, playing si maximizes it’s pay off.

Definition 11.2 (Strictly dominated strategy) For a player i, a strategy si is called

strictly dominated, if there exists another strategy si such that no matter what the
other players choose, playing si gives player i a higher payoff than playing any
alternate strategy si . That is,

pi (si , s−i ) > pi (si , s−i ). (11.4)

In the game of prisoner’s dilemma, “defect” is a dominant strategy and “cooperate”

is a strictly dominated strategy for both players. To understand the idea of a dominant
strategy, let us try to understand the following example.
11.6 Two-Player Game Strategies 311

B B
s1 s2 s3 s1 s2 s3
s1 2, . . . 1, . . . 4, . . . s1 ...,2 ...,4 ...,3
A A
s2 1, . . . 0, . . . 3, . . . s2 . . . , 1 ...,2 ...,1

(a) (b)

Fig. 11.4 Dominant strategies

Example 11.2 Dominant Strategies.

Consider the strategies s1 , s2 and s1 , s2 , s3 for players A, B, respectively, for a two-
player game shown in Fig. 11.4a, b.
We note from (a) that strategy s1 always provide a higher payoff to player A, than
strategy s2 , it is irrespective of what strategy is chosen by player B. Hence, s1 is
dominant strategy for player A. If we focus to player B (Fig. 11.4b), s2 is a dominant
strategy for it, no matter what the player A chooses. Thus we predict that the outcome
of the game will have each of the players choosing their dominant strategy, so the
outcome is (s1 , s2 ). 
It is not necessary that a dominant strategy should exist in a game. However, if it exists,
it greatly simplifies the choices the players should make, helps in the determination
of the outcome of the game. A rational player would never choose to play with
a strictly dominating strategy; hence these strategies can be eliminated from any
consideration. In case of a prisoner’s dilemma, when strictly dominating strategy
“cooperate” is eliminated from the possible choices of both players, there remains
only (defect, defect) as the possible outcome of this game. However, if a strictly
dominant strategy exists, it will remain the only choice for a player. This scheme is
called modified prisoner’s dilemma game.
Given a set of strategies, say, s−i chosen by the other players, a strategy si max-
imizing player i’s pay off is called the best response of player i. Formally, in an
n-player game, the best response function Ri of player i is expressed as,

Ri (s i ) = argmax pi (si , s−i ). (11.5)

In the above equation, pi (si , s−i ) is the payoff function of player i for the strategy
profile (si , s−i ).
By simply eliminating the strictly dominating strategies in a modified prisoner’s
dilemma, it is always not possible to predict the final outcome of the game that,
however, may simplify the solution. Note that, in such a case we may find only the
(defect, defect) strategy. This is possible because many a time in a game, there are
no dominated strategies! Hence, there is a need for a stronger solution concept—a
theory that constructs a notion of equilibrium state to which a complex chain of
thinking finally converges. In that case, the strategies of all players will be mutually
312 11 Adversarial Search and Game Theory

consistent, i.e., each player will choose his/her best response against the choices of
other players. For such a theory to work, the equilibrium state must exist. We will
discuss this in the Nash arbitration section.

11.7 Games of Perfect Information

An important question in game theory is whether all participants in a game are equally
updated to the progressive information about the game, and even in the presence of
any external factors that might affect the pay offs or outcomes of the game. In the
case of “prisoner’s dilemma,” the common-sense knowledge rules form the game
rules, and these rules are known to everyone. Such games are the games of complete
information.
On the other hand, in a game of incomplete information, at least one player is
uncertain about other player’s pay off function. For example, a sealed bid quotation
of a software by many bidders for government purchase, is a game of incomplete
information, because a bidder does not know how much other bidders have quoted.
In other words, in a game of incomplete information, the bidders’ pay off functions
are not common knowledge. In game theory these situations are called static games.
In contrast to static games, dynamic games have multiple stages, such as in the
game of chess or bargaining while buying. The static games are also called simulta-
neous move games, due to the way they are played. However, this simultaneity is not
reflected in the actual sequence of the moves in the game, but it is in respect of the
information available to the players,2 while they make their choices. For example, in
a case of sealed bid tender, the process of receiving bids may run for hours or days,
and the participants are free to submit their bids any day during that week (in online
it is all the hours of the day!), subject to some deadline time and date. Such bids
are still called simultaneous move games because the bidders are not affected due to
others’ actions while they are choosing their own actions.
In the above discussions, the word strategy may be taken as a complete action plan
for all possible ways the game can proceed. For example, a player could delegate the
strategy to its representative, or write a computer code so that the representative or
computer would know exactly what to play in all possible situations, every time the
player is supposed to make a move [11].

11.8 Games of Imperfect Information

The simplest two-person zero-sum games of imperfect information has saddle points.
Such games have predetermined outcomes provided it is rational play. The predeter-
mined outcome is called the value of the game. Consider the table given in Fig. 11.5.

2 This information is helpful to players to choose alternative moves.

11.8 Games of Imperfect Information 313

The two campaigning political parties, A and B must each decide how to handle a
disputed issue in a village. They can support it, oppose it, or evade it. Each party must
make a decision without knowing what its rivals will do. Each pair of decision (x, y),
where x, y either of three decisions by party A and B, determine the percentage of
votes that each party receives in this village. Each party wants to maximize its own
percentage of the vote. The numbers in the matrix in Fig. 11.5 indicate the percentage
share of votes for party A, the balance (100 min this value) is that of party B. Suppose
party A oppose the issue and B supports it, then A gets 80% and B gets 20%.
The A’s decision seems difficult at first because it depends upon B’s strategy. The
player A does best to support if B evades; A does best to oppose if B supports; and A
does best to evade if B evades it. We note that party A must consider the B’s strategy
before deciding its own strategy. Whatever A decides, B gets the largest share by
opposing it. Interestingly, once this is realized by A, he prefers to settle for 30% !
This 30%:70% votes for A:B is called game’s saddle point [11].
A better way of finding the saddle point is to determine the maximin and minimax
values. Using this method, first A determines the minimum percentage of votes it can
obtain for each of its strategies, then selects the maximum of these three. This can be
shown by representation in the form of a tree (see Fig. 11.6), where B’s decisions are
nodes of a sub-tree, and its root is A’s decision. The minimum percentage A will get
if it supports, oppose, and evade are: 20, 25, 30, respectively. The largest of these,

Fig. 11.5 Imperfect information game

Fig. 11.6 Selection of maximin and minimax

314 11 Adversarial Search and Game Theory

i.e. 30, is maximin (maximum of the minimums) value. Similarly, each strategy the
B chooses, it determines the maximum percentage of votes A can win, i.e. minimum
B can win (similar to what we did for A). In this case, if B supports, oppose, and
evade, the maximum A gets is: 80, 30, 90, respectively. Now, the B obtains its highest
percentage by minimizing A’s maximum percentage of the vote. Here, the smallest
of A’s maximum is 30. This 30 is B’s minimax (minimum of maximums) value.
Because both the maximin and minimax are 30, hence 30 is a saddle point. This
selection process is shown in Fig. 11.6.

11.9 Nash Arbitration Scheme

The player’s in a bargaining game often want to make the most favorable agreement
that they can, and make all the attempts to avoid any risk of losing the agreement. For
example, a seller does not want to lose the customer, but at the same time keeping the
maximum profit as far as possible. Also, the customer would like to buy it at the best
price. A similar situations occur when one of the two fighting nations tries to arrive
at peace. In the above situations, the problem is to reach arbitration that somehow
reflects the strengths of the players, so that you get the effects of negotiation without
risk. The Nash arbitration suggests the following procedure [10]:

Assume that two parties are in the process of negotiating a contract. The parties
may be two nations, seller-customer, management vs. labor, etc. Also, assume
that failure to agree, i.e., no trade, no sale, or labor strike to go on, would have
utility of zero to both parties. Nash then selects a single arbitrated outcome
from all agreements that are within the power of the players, such that the
outcomes of the product of the players’ utilities is maximized.

The four properties for such outcomes are the following:

1. The arbitrated outcome should be independent of the utility function, and should
depend only on the preferences of the players.
2. The arbitrated outcome should be Pareto optimal, i.e., there should not be any
other outcome in which both the players simultaneously do better.
3. The arbitrated outcome should be independent of irrelevant alternatives.
4. In symmetric games, the arbitrated outcome has the same utility for both players.
That is, in a situation if player A has utility x and B has utility y, then there exists
also an outcome y for A and x for B player.

Definition 11.3 (Pareto domination) An outcome ‘A’ of a game is considered as

Pareto dominating some other outcome ‘B’ in a game, if at least one player in the
game is better off in outcome ‘A’ and no other player is worse off.
11.9 Nash Arbitration Scheme 315

The disadvantage of Nash is that before the Nash arbitration is applied, the utility
function of both the players must be known. In practice, the utility functions, are
often misrepresented by the player to turn them to their advantage. Further, the Nash
scheme is neither predictable nor enforceable. It is rather a priori agreement obtained
by abstracting away many relevant factors, such as bargaining strengths of players
A and B, cultural norms, etc. It is often found that Nash arbitration often happens
unfair, the poor become poorer and the rich as richer.
The idea of a Nash equilibrium (NE) in a game played by two or more players is
based on the fundamental principle of a system at rest.

Definition 11.4 (Nash equilibrium) A Nash equilibrium is a set of strategies for all
n players in a game, such that every player is playing a best response to every other
player’s best response. Formally, an NE is a combined strategy s ∗ , as a consequence
of all players strategies s1 , . . . , sn , of n-players, expressed as,

s ∗ = (s1∗ ; s2∗ ; . . . ; si∗ ; . . . ; sn∗ ). (11.6)

In other words, the NE is a pay off of strategies (si∗ , s−i

∗
) such that each player’s
strategy (si ) is an optimal response to the other players’ strategy (s−i ), expressed as:
∗
pi (si∗ , s−i
∗
) ≥ pi (si , s−i
∗
). (11.7)

Example 11.3 Nash Equilibrium.

Consider that a rich man and a poor man get one million dollars if they agree on how
they share it between them. If they fail to agree, they get nothing. In this case, the
Nash equilibrium will give more money to the richer person [9].
When more money is received than a person already possess—people play it safe.
Most, unless they are very rich, prefer to sure of thousand dollars against a chance
of a million dollars But a large insurance company would even prefer the chance of
one million losings and in fact,gladly accepts much less attractive risks every day.
This indifference is reflected in the poor man’s utility function more strongly than
in the rich man’s utility function. A utility function that correctly reflects the poor
man’s situation would be square-root function: $100 would be 10 tiles, $16 as four
tiles, and $1 as one tile only. Thus, the poor man would be indifferent on a chance
of $10,000 and sure of $2000. 
The general approach for solving imperfect-information games is shown in the
Fig. 11.7. To begin with, the original game is abstracted to generate a similar but
smaller game. The size of the game is reduced to the extent that the equilibrium-
finding algorithm can be easily applied on this. In the next step, the abstract game is
solved for equilibrium or near-equilibrium state.
A Nash equilibrium is based on the notion of rational play, which is a profile of
strategies of all the players, with one strategy per player, such that no player can
increase or decrease the payoff by switching to a different strategy. The strategy for
a player means the characteristic of the play carried out when his/her turn comes.
316 11 Adversarial Search and Game Theory

Fig. 11.7 A general approach to solve imperfect-information games

For each information set (i.e., collection of game states), it is the probability with
which the player selects each of his available actions. Due to the private nature of
the information of other players, an information set cannot be distinguished by any
player as whose turn is it. At the end of the game, the strategies from the abstract
game are mapped back to the original game [13].
There are two kinds of abstract games: (1) Information abstraction, where it is
assumed that a player does not know some information—that he knows, due to the
reason that information sets are bundled, and, (2) Lossless abstraction algorithm,
which results to an abstract game, and each equilibrium from this abstract game is
also equilibrium in the original game.

11.10 n-Person Games

Theoretically, n-person game where players do not communicate, is in no way dif-

ferent from two-person games. As per the von Neumann and Morgenstern theory,
in such games coalitions can be formed by joining together various subgroup of
players, where each of them has a value associated with them—a minimum amount
the coalition can attain. Usually, such subgroups may exist in a legislative body or
in merging business groups. These n-person games are described in characteristic
function form—first listing the players, then various coalitions, and then the value of
these coalitions. The value of coalition function is additive, and in any given game
there are many solutions.
Equilibrium Points in n-person Games
We can formally define the concept of an n-person game follows: each player in
the game has associated with it a finite set of strategies, called pure strategies, and
there is a well-defined set of payments to n-players. The set of pure strategies is a
n-tuple of strategies, one for each player. The mixed strategies are nothing but the
probability distributions over the pure strategies, for which the payoff functions are
the expectations of the players.
11.10 n-Person Games 317

For n-players, a point in product space is representation of any n-tuple of strategies,

which is obtained by multiplying n-strategy spaces of the players. One such n-tuple
(say, t1 ) counters another n-tuple (say, t2 ), if the strategy of each player in t1 results to
highest possible expectation for its players against the n − 1 strategies of the players
in t2 . An equilibrium point occurs when an n-tuple is countering itself.
A correspondence of each of the n-tuple ti (i = 1 . . . n), with its set of all coun-
tering n-tuples, say, ti j , provides a one-to-many mapping of the product space into
itself. By definition of countering it is clear that a set of countering points t j of a
point ti is convex. Assuming that the payoff function is continuous, the mapping
graph will be a closed one. This closedness is formally defined as follows.

Definition 11.5 (Closed-mapping) If p1 , p2 , …, pi , . . . and q1 , q2 , …, qi , . . . are

sequences of points in the product space, such that qi → q and pi → p, and qi
counters pi , then q counters p. Such a mapping is called closed mapping. 

In the above discussions, because the graph is closed and the image of each point
under the mapping is convex, the mapping has a fixed point in its image. Therefore,
there is an equilibrium point.
In the case of two-person zero-sum game, the “main theorem” and the “existence
of an equilibrium point”, are two equivalent statements. In such case, any two equi-
librium points ultimately lead to the same expectation for the players. However, its
occurrence is not always necessary [5].

11.11 Representation of Two-Player Games

Chess has certain attractive features that many, more complex tasks do not have. The
available options (moves) and goal (checkmate) are sharply defined. The discrete
model of chess is called a game-tree, and it is the general mathematical model on
which the theory of two-player zero-sum games of perfect information is based.
Since all the legal moves are known to both the players at all the times, therefore this
is a game of perfect information. This also is a game of zero-sum, because when one
player looses, the other gains it.
There is a root node at the top of a chess tree, that represents the initial setup
(configuration) on the board, say c0 . Corresponding to each opening move, there is
an arc leading to another node that represents a new board configuration (say, c0i ) after
this move is made. If there are n number of such maximum configurations to which
a legal move can be taken from the root node, then the nodes corresponding to these
configurations are c01 . . . c0n . This process of construction of news configurations or
new nodes goes on as we progress in playing the chess game. Thus, constructing a
game-tree—a recursive structure of the tree, with the root node and sub-trees under
that, with arcs pointing to the potential next states, such that each of the game-trees
would be a smaller game-tree. The number of arcs/edges leaving a node is referred
to as branching factor, and distance of any node from the root, in terms of a number
of in-between nodes, is called depth. If b and d are the average branching factor and
318 11 Adversarial Search and Game Theory

depth of the tree, respectively, then the tree contains at least bd nodes. When the
current state of the game is a leaf node, the game terminates.
Each leaf node has a value associated with it, that corresponds to the payoff of
that particular outcome. A game can have any payoff associated with each outcome,
which helps in deciding to choose the move of higher payoffs. However, the standard
parlor games have result like win, loss, or some times draw.
In the case of two-player games, the players take alternate turns to make a next
move, which chooses a state from among the children states of the current state. If
the two-player game is a zero-sum game, one player attempts to choose a move of
maximum value, and other players would always choose a move of minimum value.
An approach that tells a player what to choose next, is called the strategy of the
game. The chosen strategy (called control strategy) controls the flow of the game.
In principle, the decision to choose which node should be the next node is a simple
one—any state that is one move from the leaf node can be assigned a value which is
best of its children—this can be either maximum or minimum value of the children
under the node.
This maximum or minimum depends on which player out of the two is playing.
The states that are two moves away from the leaf then take on the value of their
best children. This goes on until each child node of the current node is assigned a
value. The best move is chosen based on this strategy. In the two-player zero-sum
game, the player that maximizes is called maximizer, and the one minimizing is
called minimizer. Accordingly, the method of assigning values and choosing moves
is called the minimax algorithm, and it defines the optimal move to be made from
each state in the game.
Complexities
The network of configurations and move (links) come up as a search tree, called
game-tree. The start node is at level 0, and other nodes levels from the top are at
levels 1, 2, . . . , d for a game-tree of depth d. If b is the maximum branching factor,
the worst-case time and space complexities for breadth-first search (BFS) are both
O(bd ). For depth-first search (DFS), time and space complexities are O(bd ), O(d),
respectively.

11.12 Minimax Search

The Minimax search is based on minimax theorem, which is the fundamental theorem
of game theory, stated in 1928 by mathematician John von Neumann.

Theorem 11.1 Minimax Theorem.

The minimax theorem says that every two-person zero-sum game of a finite number
of states has a solution in mixed strategies. If A and B are the players in such game,
there is a value V and mixed strategies for these players, such that if player A adopts
11.12 Minimax Search 319

=Maximizing level, =Minimizing level

P Level 0 Backed-up value P 3 P

2 3 2
3 P2
P1 P2 Level 1 P1 P2 P1

Level 2

P2,2 P1,1 P2,2 P1,1 P1,2 P2,1 P2,2

P1,1 P1,2 P2,1 P1,2 P2,1
static
evaluations 3 8 2 9 3 8 2 9 3 8 2 9

(a) (b) (c)

Fig. 11.8 MINIMAX search

its mixed strategy, the outcome will be at least as favorable to A as V . Whereas, if

B adopts this mixed strategy, the outcome will be no more favorable to A than V .
Thus A and B have the power to enforce the outcome of V . 

Definition 11.6 Utility Theory.

In some of the previous examples, we discussed that players want to maximize

their profit. But in real-life situations, the game players have some other goals. The
examples are: buying lottery tickets, playing gambling in casinos, and buying so and
so insurance policies. The game theory states, how to attain these goals. Hence, it
is important to define a utility function that would reflect an individual’s preferences
while playing/making moves. Basically, the utility function assigns to each of the
player’s alternatives (i.e., moves), a number, that convey the relative attractiveness
of the alternative. Maximizing the player’s utility automatically determines his most
preferred action. This is called Utility theory. 
Consider the Fig. 11.8, where the two-players (maximizer and minimizer) play
alternately. The game has static values at the lowest level, i.e., at the bottom of the
tree. The game-tree is shown with three levels 0 (for root), 1, and 2. The maximizer
chooses move at level 0, minimizer at level 1, and at the next level, i.e., 2, the static
values are available. The maximizer might hope to get to the situation yielding score
9, he knows that minimizer might choose a move deflecting the play towards the score
2. In general, the maximizer (here at level 0) must take the choices available to the
minimizer at the next level, into cognizance. And, similarly, the minimizer must also
take into cognizance the choices available to maximizer at the next level down [1].
Eventually, the limits of exploration are reached and the static evaluator provides
the direct basis for selecting among the alternatives. The minimizer may choose the
values 3 and 2, at the level just up from the static evaluations (Fig. 11.8b). Knowing
these scores the maximizer at level one up makes the best choice between 2 and 3,
at level zero (see Fig. 11.8c).
The procedure by which the scoring information passes up the game-tree is called
MINIMAX algorithm because the score at any level i is either minimum or the
maximum of what is available at the level i + 1. The recursive form of the MINIMAX
320 11 Adversarial Search and Game Theory

algorithm is shown as Algorithm 11.1. The idea of minimaxing is to translate the board
into a static number. However, the process may be expensive due to the generation
of a large number of paths.

Algorithm 11.1 MINIMAX Algorithm

1: if limit of search has reached then
2: compute the static value of the current position relative to the appropriate player;
3: Report the result;
4: else
5: if level is minimizing level then
6: call MINIMAX Algorithm on the children of current position;
7: Report minimum of the result;
8: else
9: (level is maximizing level.)
10: call MINIMAX Algorithm on the children of current position;
11: Report maximum of the result;
12: end if
13: end if

The minimax algorithm has the following properties [14]:

1. The algorithm (i.e., inference) is complete if tree is finite.
2. The worst-case Time complexity is O(bd ), where b is branching factor, and d is
total distance.
3. The worst-case Space complexity is O(d) equal to depth-first exploration.
Example 11.4 Grundy’s game.
Consider a typical player game, called Grundy’s game, where there is a stack of seven
coins of equal size, indicated by initial configuration as (7). The starting player MAX
makes a move so that the stack is broken into two stacks of unequal size. Figure 11.9
shows the alternate moves as: (6, 1), (5, 2), (4, 3). The other player, MIN, has the next
move, which breaks the sub-stacks created further into unequal parts. This process
goes on for alternate players. No Further move is allowed from a node if unequal
partitioning is no more possible from that node. The player who first cannot play the
next move is game the loser.
We note that for the node (2, 2, 1, 1, 1) MIN is not able to make the next move,
hence MIN loses. The player that makes the first move is generally a MAX player.
Thus after this move, the node positions occupied are called MAX nodes, and from
these nodes, the move is due to MIN player. This leads to the transition to MIN-MAX
nodes.
At every configuration, there are choices available either for MAX or MIN player
for the next move. For example, MAX will try to foresee the possible next moves
available to MIN after he has made a move, and accordingly chooses this next move
so the MIN cannot have a good move, and he is more close to a winning strategy.
The MIN also keeps a similar strategy. Finally, the game ends at a node where no
further move is possible by any of the players. 
11.13 Tic-tac-toe Game Analysis 321

Start (7)
MAX’s move

(6, 1) (5, 2) (4, 3) MAX

(5, 1, 1) (4, 2, 1) (3, 2, 2) (3, 3, 1) MIN

(4, 1, 1, 1) (3, 2, 1, 1) (2, 2, 2, 1) MAX

(3, 1, 1, 1, 1) (2, 2, 1, 1, 1) MIN

(2, 1, 1, 1, 1, 1) MAX

Fig. 11.9 MINIMAX game-tree

11.13 Tic-tac-toe Game Analysis

The tic-tac-toe is a board game, with 3 × 3 = 9 positions, and played between two-
players through alternate moves. The game has a total 765 different positions, and
26,830 possible games. The initial configuration is the empty board. The one player,
called MAX makes the first move with X , and the other player, called MIN player,
makes a O move. The player which is first in placing all X or O array as row or
column or diagonal wins the game [8].
Figure 11.10 shows some initial configuration and of some moves of this game.
The game of tic-tac-toe generates all the possible moves at level i before generating
the moves of level i + 1, hence, it is breadth-first search (BFS). We note that while

Level 0 Start MAX

MAX’s move

Level 1 X
X MIN
X
MIN’s move
O
X O X O MAX
X

O X O X MIN
X X

Fig. 11.10 Tic-tac-toe with some initial moves

322 11 Adversarial Search and Game Theory

Fig. 11.11 A tic-tac-toe 1 2 3 4

game
X O X O X O X
O
X X

5 6 7 8
O X O X O X O X
O O X O O X O
X X O X X O X X O X X

in start configuration (level 0) having all empty positions, the player MAX can make
moves to level 1. The player MAX will choose moves in such a way that it leads
to a configuration ultimately such that all the X ’s are continuous and are in a line,
or column or diagonal. At the same time, the MAX will choose such a move that it
becomes difficult for MIN later to make a move to a configuration with all O’s in a
line, or column or diagonal. Thus, each of MAX and MIN makes the moves so that
it is winning for them and blocking for the other player.
The Fig. 11.11 shows total of 8 moves of this game, for a particular sequence of
moves by alternate players X and O, numbered as 1–8 Finally, the player O wins
the game.
To decide which configuration is superior to move into, we associate a weight
function that will be maximized by the MAX player, and the MIN player will try to
minimize it. If a configuration is indicated by c, then this weight function is defined
as:
f (c) = m − n, (11.8)

where, m = number of complete rows, columns, and diagonals that are still open
(fully) for the next move of MAX, and
n = number of complete rows, columns, and diagonals that are still open for the
next move of MIN.
If c is a winning configuration for MAX, then f (c) = ∞, and if c is winning
configuration for MIN then f (c) = −∞. If a configuration is like in Fig. 11.12, then
f (c) = 4 − 6 = −2. The m are solid arrows (Fig. 11.12a), and n are shown as dotted
arrows (Fig. 11.12b).

Fig. 11.12 Computing the

weight function X X
O O

(a) (b)
11.13 Tic-tac-toe Game Analysis 323

O O
X X X X
O O

Fig. 11.13 Identical configurations in tic-tac-toe

If we make use of symmetries in generating successor positions, then all the game
states shown in Fig. 11.13 are identical.
For efficiency reasons, early in the game, the branching factor is kept small by
symmetries, and late in the game, it is kept small by a number of open positions
available. The first move is always important.

Example 11.5 Computation of static values in tic-tac-toe game.

The Fig. 11.14 demonstrates the computation of this weight function at each of the
node. To win the game, the MAX at level 0, should make move to such a node at level
1 that produces a maximum value of function f (c) at level 1. It is this maximum
value that is backed up to level 0. This value is as a result of the maximum value
of the corresponding nodes that are under the MAX node at level say 1, (in general,
for MAX at level i it backup the maximum of values those at level i + 1). The MIN
node at level i + 1 should make a move that produces minimum value from those at
level i + 2. This process goes on until the maximizer sees the static values at the last
level (bottom) of the tree. Similarly, the minimizer will also, choose the next move
by attempting to see the static values at the last node.

Fig. 11.14 Tic-tac-toe with backup of static values

324 11 Adversarial Search and Game Theory

The static values at the bottom nodes in this example are shown as 1, 3, 4, 0, −1.
The values at the upper levels, i.e, parents of static nodes, and their parents are called
backup values. The backup values are based on the “look-ahead” of the game-tree,
and this depends on the lower levels nodes, in fact even those that are near the end
of the game-tree.
Some of the nodes may represent a win for MIN, hence they may have value
−∞. When evaluations are backed-up, the MAX’s best move is that which avoids
its immediate defeat. In the case of MAX, the win corresponds to a node value of
+∞. 

11.14 Alpha-Beta Search

Alpha-Beta is a game-tree search, which is an improvement over the minimax pro-

cedure, but equivalent to minimax in some respect—both the procedures will always
choose the same depth successor at best, and they will assign the same value to it. The
Alpha-beta method turns out to be several orders magnitude faster than minimax—it
saves the time by not searching certain branches of three, which are not fit for any
contribution for winning strategy of the game. This is because under certain condi-
tions, the values of certain branches do not affect the value(s) which is ultimately
backed-up3 to higher levels of the tree. Therefore, these branches are not required to
be evaluated. When the alpha-beta procedure detects these conditions, it stops work
on one branch and skips to another. This event is called an alpha or beta cutoff [12].
At first, it appears that static evaluator must be used on each leaf node at the bottom
of the search tree. But, it is not required. To see how alpha-beta search works, consider
the example shown in Fig. 11.15. Alpha-beta starts just like a minimax procedure
by evaluating all the successors of P1 . The minimum of these static values is then
backed-up to P1 , since P1 is a MIN position. The backed-up value of P1 is alpha and
has a value 5 in this example.
Alpha is the lower limit for the backed-up value of the top position, P. Since, P
is the max position, we backup the value of the largest value successor of P. Since
we have evaluated only one successor at this time, we do not know that it will be 5
or larger. The value of alpha may change as the other successors are evaluated, but
it can only increase, not decrease.
Having evaluated P1 , the procedure begin work on P2 . An Alpha cutoff takes place
at P2,1 = 3, as it is less than alpha. Since P2 is a Min position, V2,1 is an upper limit
for V2 (value of node P2 ). Since V2 is less than alpha, P2 is definitely eliminated as
a candidate for the largest valued successor of P. There is no point in evaluating the
other successors of P2 so the procedure begins work on P3 next.
The above cutoffs save the machine a good deal of time. The alpha cutoff at P2,1
means that the machine need not bother to evaluate P2,2 and P2,3 . A second alpha

3 Selection
of minimum of the payoffs or maximum of the payoffs (depending on which player has
a move) of children of any node vk and moving it to node vk , is called backing-up of values.
11.14 Alpha-Beta Search 325

P MAX Position
Depth 0

Backed-up
value P1 P2 P3 MIN Position
5
Depth 1

P1,1 P1,2 P1,3 P2,1 P2,2 P2,3 P3,1 P3,2 Bottom level
static P3,3
positions
evaluations 5 10 12 3 0 0 7 1 5 1 Depth 2
Positions not X X X
evaluated due
alpha-cutoff

Fig. 11.15 Alpha-beta search

Fig. 11.16 Deep alpha cutoffs

cutoff occurs at P3,2 , which eliminates P3,3 . Thus, in the Fig. 11.15 alpha-beta search
program would evaluate only six of the bottom level successors, while a minimax
program would evaluate all nine.
Although the example is given only for a tree of three levels, it is clear that the
procedure will work just the same below any Max position Px , at any depth in a large
tree, provided only that there are at least two levels below Px . If there were more
levels below P1,1 in the example, then we could use the backup value of P1,1 instead
of the static value. If there were more levels above P, then we could backup of P.
Moreover, it is possible to pass a value of alpha down from the top of a large tree.
Thus, an alpha established at depth 1 could be used to produce cutoffs at depths 2,
4, and 6. These deep cutoffs are illustrated in Fig. 11.16.
326 11 Adversarial Search and Game Theory

Alpha is defined by the values of the successors of a Max positions (i.e., odd
depths, here it is P1 , P2 , P3 nodes, in Fig. 11.15), while alpha cutoff occur among
the successors of a Min position (even depths, here it is P1,1 . . . P3,3 , in Fig. 11.15).
It is possible to define another variable, Beta, which is established at even depths
and generates cutoffs at odd depths. The action of the beta cutoff is exactly inverse
of alpha cutoffs.
The effect of alpha-beta cutoff is to make the tree space grow slower with depth.
Thus, the advantage over simple MINIMAX. It is about twice as good as at Maximum
depth Dmax = 3 and about thirty times as good at Dmax = 6. This is typical. It is
good to have depth-dependent program. Such a measure is called as depth relation
(DR),
log N
DR = (11.9)
log N M M

where, N is a number of nodes at the bottom of the tree, and N M M is the number of
nodes at the bottom of the tree in minimax search. D R = [0, 1], is effective depth in
comparison to minimax procedure.
Alpha-beta search is equivalent to minimax in the sense the two procedures will
choose the same depth successor as best and will always give the same value for the
successor.

11.14.1 Complexities Analysis of Alpha-Beta

The benefit of alpha-beta pruning lies in the fact that branches of search tree can be
eliminated so that search tree can be limited to more promising sub-trees and deeper
search can be performed in lesser time. Thus it belongs to the branch-and-bound
class of algorithms.
If move ordering for alpha-beta is optimal (best moves are always selected
√ first),
d
then the total number of leaf nodes to be evaluated are O(b 2 ) = O( bd ). The
explanation of above is that all the first players’ moves must be studied to find the
best one, but for each, only the best second player’s move is needed to refute all but
the first (and best) player move-alphabet ensures that no second player moves need
d
to be considered. Thus, the complexity (O(b 2 ) is possible if nodes are ordered. If
3d
they are at random then it is O(b 4 ).
Thus, it can be demonstrated that the number of static evaluations needed to dis-
cover the best move in an optimally arranged tree is O(bd/2 ). Let the static evaluator
function be represented by a variable c, which needs to be maximized and minimized
at different levels. We note in Fig. 11.15 that evaluation of c is not required at level
1 (shown with values 5 and ≤3 for minimizing level). This is because we do not
need to evaluate for ≤3, as the value 5 is carried forward to the root. Since static
evaluations are required only at the alternate levels, the depth of search reduces to
half. Therefore the time time complexity is O(bd/2 ).
11.14 Alpha-Beta Search 327

11.14.2 Improving the Efficiency of Alpha-Beta

The number of cutoffs generated by alpha-beta procedure depends on the order in

which successors are evaluated. In Fig. 11.15, if the machine had evaluated P2 , P3
then P1 , alpha would have been 0, then 1, then 5, and there would have been no alpha
cutoff.
This fact suggests the possibility of improving on the alpha-beta procedure by
ordering successors of a position in order to generate a large number of alpha-beta
cutoffs.
The ordering can be achieved by ordering the nodes of a search tree at their static
values. The largest value successor of a max position is put first and reverse order is
used for the successors of min position. This procedure is based on the assumption
that their static value of a position is positively correlated with the deeper, backup
value at that position. Since ordering obtained may not be correct at other levels,
dynamic ordering may be adopted.
Consider the symmetric configurations in Fig. 11.17. We note that the cost func-
tion in all these is equal. Thus, if these configurations are merged in the alpha-beta
search, or even in the minimax search, the number of nodes in the game-tree can
be substantially reduced. Similarly, many other symmetric configurations exist, and
those, similar, can be merged. All this helps in boosting the efficiency of the alpha-
beta search.
The common techniques for assignments of tip values to the game-trees are based
on probabilistic assumptions, simulations, and closed form results. Though the alpha-
beta search is considered as optimal minimax algorithm under sequential traversal of
nodes of the game-tree, it can be far efficient if game-trees are traversed in parallel.
The problem, if any with sequential alpha-beta search occurs when the optimal
path moves towards the right of the tree. The later is because the alpha-beta search
actually explores the game-tree from left to right. Since a parallel alpha-beta search
can explore multiple paths simultaneously in all the regions, it obtains a better global
perspective of the tree and can cut off the search in the regions where sequential alpha-
beta cannot do. A parallel search algorithm can perform a parallel tree-traversal using
state-space search.

O O
X X X X Symmetric configurations
O O

O O
X X Symmetric configurations
O O

Fig. 11.17 Symmetric configurations

328 11 Adversarial Search and Game Theory

11.15 Sponsored Search

Due to the prevalence of more and more business, commerce, and trading activities
on the WWW, the sponsored search is becoming very important. The sponsored
search is based on input from many resources, like Web-search engines, users, and
content providers. Here, contents providers are companies or individuals owning the
Website hosting the contents.
The keywords selected by the contents provider are matched against the users’
query by the web search engines, and the corresponding link is displayed in high-
lighted form. The content providers pay to the search engine company in many cases,
whenever a user enters one of these terms for search, and clicks on the sponsored
link displayed. On occasions, the user is required to go one step further and perform
some specified action on a web-page on the content provider’s website. The content
providers can identify those phrases from the user queries that are most likely to be
related to the required Web sites. Often the sponsored search platforms provide the
capability to the content providers to tailor the presented material on the sponsored
link that conforms to user queries.
In fact, it is not necessary that these simple approaches would always work.
Consider a situation, where many different content providers assign the work of
responding to user queries to a search, all of them want the search engine to respond
to the same term or phrase of the user query. In such a scenario, an electronic auction
system will rank the sponsored links, in such a way that the highest bidder will be of
the topmost rank, the next lower bidder will get the next link, and so on. That means,
the one who pays more will get the links of his/her product and services listed at top,
and only they will get the customers, because the customers usually do not open all
the links.
The web search engines also adopt other ranking factors apart from the price
paid by the bidder. For example, in the past, some sponsored link used to get much
higher clicks, is likely to be displayed among the top, or at the top. This is because
a link that gets more clicks, is likely to be more relevant. And, when the link is
relevant, the user is potentially a good customer for the content provider. A sponsored
link with maximum clicks will generally produce a major part of the profit for the
search engine. Hence, both the content providers and the search engines have higher
monetary incentives for providing the contents to the user which are relevant.
The key-phrase selection can be conceptually viewed as dynamic form of Web
site meta-tagging focused on the user. The contents providers can change one or
more of following:
phrases to be searched,
bid price for various phrases,
degree of matching of terms,
time based restrictions, like displaying at 10 AM, evening, at early morning, etc,
geographical restrictions, and,
amount they bid for a given period.
From the above discussions, we conclude that the content providers become active
participants in the sponsored search process [3, 4].
11.16 Playing Chess with Computer 329

11.16 Playing Chess with Computer

One of the well known, landmark example of the game is the chess game, played
between grandmaster Garry Kasparov and IBM Deep Blue Computer in 1997. Fol-
lowing are some characteristics of this competition:
• The system used for this purpose comprised of 30 number of IBM RS6000
machines, running a program for search, and a large number of custom VLSI
chess processors that performed the job of chess move generation, and ordering
of these moves. This hardware searched for the last few levels of the game-tree,
and evaluated the leaf nodes.
• Deepblue searched more than 100 million nodes per second on average, with peak
speed as high as 330 million nodes per second.
• This software generated 30 million positions per move reaching a depth of 14 quite
often.
• Heart of the search algorithm was a standard iterative deepening alpha-beta search.
• In some cases, the search reached a depth of 40 plies.
• The evaluation function for a proper move comprised over 8000 features, many of
them described a highly specific pattern of pieces.
• A large, end-game database, which comprised solved positions, were also used.
• The chess program also used an opening book with about 4000 positions, and a
database of 700,000 grand-master games for consulting for consensus for next
moves.
In the above, the IBM’s Deep Blue software beat the World Chess Champion in
a series of six matches.
In the more recent chess playing computer games, varieties of heuristics for prun-
ing are commonly used to reduce the effective branching factor from 35 down to less
than 3, hence they are more powerful with even far less processing capabilities of
processors.

11.17 Summary

The Game theory is the formal study of conflict and cooperation. Another branch
of game theory, called, combinatorial Game theory (CGT) studies strategies and
mathematics of two-player games of perfect knowledge such as Chess or Go. In
games of perfect information, such as Chess, each player knows everything about
the game at all times. Poker is a game of imperfect information, as other players do not
know your cards. In zero-sum games, if one player loses, the other will gain; Poker,
Chess, and Parlor are games in this category. Games of bargaining, trade, agreements,
are non-zero-sum games, where all the players gain on successful completion of the
game, and loose in the event of failure.
330 11 Adversarial Search and Game Theory

Strategy for two-player game is: every unique board configuration is converted
into a unique single integer number. While one player tries to maximize this number,
the opponent does the reverse.
Prisoner’s dilemma is a game of conflict and cooperation; and complete informa-
tion, such that each player is aware of strategies of the other player.
A dominant strategy provides roughly the highest payoff for a player, irrespective
of what strategies are employed by the other players.
For the games of the bargain, Nash arbitration scheme is used to arrive at an
agreement, called Nash equilibrium (state), which is a set of strategies for all n players
such that each player is playing the best response to every other player’s best response.
The equilibrium state follows the principle of a system at rest.
The scoring in two-player game is done using the MINIMAX algorithm, where
the score at any level i is either minimum or the maximum of what is available at the
level i + 1.
Games like, tic-tac-toe, and chess requires the search, which can be conducted
using the minimax search. However, this approach has exponential complexity for
searching the game-tree.
An improved search method, called alpha-beta search approach, saves the search
time by pruning the game-tree through alpha and beta cutoffs.

Exercises

1. Demonstrate the min-max search for the tic-tac-toe puzzle (for 10 moves).
2. Demonstrate the alpha-beta search for the tic-tac-toe puzzle.
3. Apply alpha-beta search (from left to right) to the game-tree in Fig. 11.18. Show
the backed-up value of each node. Mark with an X any branches that are not
searched. Identify these as alpha/beta cutoffs. Mark the best move with an arrow
from the root node.
a. Identify these as alpha/beta cutoffs, mark the best moves with an arrow from
the root node.
b. Compute the time complexity of search with worst-case branching factor of
3 and height of the tree as h.
4. A game nim is played as follows: there are two players who remove one, two, or
three coins, alternately from a stack of five coins. Represent this game playing
as a search tree. Suggest any suitable strategy for winning this game?
5. Given the tree in Fig. 11.19, explore this tree using the alpha-beta procedure.
Indicate all parts of this tree that are cutoff. Also, indicate the winning path(s),
and strike out all the values that are not required to be computed.

a. Identify these as alpha/beta cutoffs, mark the best moves with an arrow from
root node.
b. compute the time complexity of search with worst-case branching factor of
3 and height of tree as h.
Exercises 331

Fig. 11.18 Game-tree

Fig. 11.19 Alpha-beta Maximizer

search

Minimizer

Maximizer

6 4 9 6 4 0 3 3

6. Consider the min-max tree shown in the Fig. 11.20, whose leaves are labeled
with natural numbers; n and m are variables.
a. Assign values to n and m such that, to compute the value at the root node, no
alpha-beta cutoff is possible. Compute the value of the root node.
b. Assign values to n and m such that, to compute the value at the root node, an
alpha-beta cutoff is possible. Indicate the cut and compute the value of the
root node.
7. Following are either two-player or many player games, and standard minimax/
alpha-beta search techniques are to be applied for game playing. Explain, how
well these techniques apply to these games? Give the formal approach for each,
where possible.
a. Basketball
b. Badminton
c. Football
d. Soccer
e. Tennis
f. Kabbadi
8. Consider two firms A and B, that give service in the same market. The firms
incur constant average costs of $2 per unit, and are free to choose a high price
of $10 or a lower price of $5 per unit for marketing. When both firms set a high
price, the total demand is 5,000 units which are split evenly between the two
332 11 Adversarial Search and Game Theory

Fig. 11.20 Min-max tree MAX

3 4 n m

Fig. 11.21 A two-player B

game
s1 s2 s3
s1 2, 3 6, 1 2, 3
A
s2 2, 3 3, 4 5, 1
s3 4, 1 5, 4 4, 2

Table 11.1 Two-player game move choices

s1 s2 s3
s1 (1, 1) (3, 4) (2, 1)
s2 (3, 3) (0, 4) (0, 9)
s3 (2, 4) (2, 5) (8, 1)

firms. When both set a low price, the total demand is 10,000 units, which is again
split evenly. If one firm sets a low price and the second a high price, the low
priced firm sells 8,000 units, and the high priced firm only 1,000 units.
Analyze the pricing decisions of the firms A and B as a noncooperative game.

a. Construct the payoff matrix, where the elements of each cell of the matrix
are the two firms’ profits.
b. Derive the equilibrium set of strategies.
c. Is it the game of prisoners’ dilemma? Justify.

9. Find out the Nash equilibrium for the game shown in Fig. 11.21, which is played
between two-players A and B, having strategies s1 , s2 , s3 and s1 , s2 , s3 , respec-
tively.
10. Consider the Table 11.1 for two-player game.

a. Find out the maximum moves by the game to win/loose.

b. Which moves are dominated?
c. Identify all the best responses.
d. Is there any Nash equilibrium?
11. Draw a payoff matrix for the following game of scheduling a party. Also, find
maximin moves, domination, best responses, and Nash equilibrium (if it exists).
Exercises 333

Table 11.2 Two-player perfect information game

Cooperate Defect
Cooperate (4, 2) (2, 6)
Defect (6, 0) (0, 4)

Let the friends A and B do not speak to each other, but have many common
friends. Both of them want to invite these friends to the party, either on Saturday
or Sunday. However, both of them prefer Sunday over Saturday. If both decide
the party for the same day, it will be considered as a disaster with a (negative)
payoff of $500 for both. If they plan the party on different days, the one proposing
Sunday gets a payoff of $250, and the other of $200.
12. For two-players game shown in Table 11.2, find out all Nash equilibria states,
possible domination, maximin moves, and best responses.
13. Find out the saddle point for the game shown in Fig. 11.5.
14. What can be the true considerations, for example, to reach to Nash equilibrium
faster in a buyer vs seller game?
15. Are we in a position to determine in advance that arbitration between two parties
will result in an equilibrium state? Justify for yes/no.
16. Justify the following statement: “It is often found that Nash arbitration is often
unfair, in which rich becomes richer and poor becomes poorer.”
17. For any minimax game of even height game-tree, show that final results of the
game shall be the same, irrespective of who plays first.
18. Apply the minimax algorithm for the problem of prisoner’s dilemma (Fig. 11.2),
and demonstrate the backup of values from static levels.
19. Why the game of tic-tac-toe is BFS search? Justify.
20. Write an algorithm, to compute the weight of any general configuration of tic-tac-
toe given any number of X s and Os on arbitrary position. Note that if 0 ≤ |X | ≤ 5
then 0 ≤ |O| ≤ 4 and vice versa.
21. In a game search tree with the root node at level 0, suppose the alpha cutoff
occurs at level i. What is the minimum required depth of the tree? Justify your
answer.
22. For a game-tree of depth d, with branching factor b uniform at all the sub-tree,
let alpha cutoff occurs at every alternate depth due to the second node from left in
each sub-tree. What is an expression for the time-complexity of the search using
only alpha cutoff? Assume that the time-complexity of the minimax algorithm
is O(bd ).
23. Assume that in a game-tree, the alpha and beta cutoff occurs at every alternate
level. Find out the time complexity of this search.
24. Show that if static nodes are ordered in order of their static values, then the time
d
complexity of alpha-beta search is O(b 2 ).
25. Show that if static node weights are in random order, then the time complexity
3d
of alpha-beta search is O(b 4 ).
334 11 Adversarial Search and Game Theory

26. In the Fig. 11.15, find out the alpha-beta cutoffs in each case, when order of
sub-trees are P3 P2 P1 , P2 P3 P1 , and P2 P1 P3 . What you conclude by this change
in order?
27. Are the winning results of minimax and alpha-beta search tree identical for a
given search tree? Justify.
28. Show that cost functions of symmetric configurations in tic-tac-toe are identical.
29. Two firms Alpha and Beta serve the same market. They have constant average
costs of $2 per unit. The firms can choose either a high price ($10) or a low price
($5) per unit for their product. When both firms set a high price, the total demand
is 10,000 units, which is split evenly between the two firms. When both set a low
price, the total demand is 18,000 units, which is again split evenly. If one firm
sets a low price and the other a high price, the low priced firm sells 15,000 units,
while the high priced firm sells only 2,000 units. Analyze the pricing decisions
of the two firms as a noncooperative game.
a. In the scenarios mentioned above, form representation, construct the payoff
matrix, where the elements of each cell of the matrix are the two firms’ profits.
b. Derive the equilibrium set of strategies.
c. Explain why this problem an example of the prisoners’ dilemma game.
30. Write a parallel search algorithm for an alpha-beta search to reduce the search
space.
31. Select one or more than one, or True/False, or nil answer in the following
multiple-choice questions:
a. Prisoner’s dilemma is following type(s) of game:
(A) Noncooperative game (B) Non-zero-sum game
(C) Imperfect information game (D) all the above
b. Poker is following type(s) of game:
(A) Zero-sum (B) Perfect Information
(C) Non-zero sum (C) None of above
c. Which of the following are static games?
(A) Prisoner’s dilemma (B) Bid of winning a contract
(C) Chess (D) Poker
d. Which of the following are called simultaneous move games?
(A) Prisoner’s dilemma (B) Bid of winning a contract
(C) Chess (D) Poker
e. The tic-tac-toe is following type of search:
(A) Iterative deepening (B) DFS
(C) Best-first search (D) BFS
f. Alpha is defined by the values of successors of:
(A) MAX position (B) MIN position
(C) Both MAX and MIN (D) None of above
g. Alpha cutoff occurs among the successors of:
(A) MIN position (B) MAX position
(C) Depends of the structure of the tree (D) None of above
Exercises 335

h. Beta is defined by the values of successors of:

(A) MAX position (B) MIN position
(C) Can be any position (D) None of above
i. Beta cutoff occurs among the successors of:
(A) MIN position (B) MAX position
(C) Depends of the structure of the tree (D) None of above
j. The alpha-beta search belong to the following class of algorithm:
(A) Hill-climbing (B) Branch-and-bound
(C) Divide-and-conquer (D) Iterative approximation
k. Alpha is defined at odd level of the tree (T/F)?
l. Beta is defined at even level of the tree (T/F)?

References

1. Abramson B (1989) Control strategies for two-player games. ACM Comput Surv 21(2):137–
161
2. Davis MD (1983) Game theory—a nontechnical introduction. Dover, New York
3. Jansen BJ, Spink A (2007) Sponsored search: is money a motivator for providing relevant
results? Computer 8:52–57
4. Jansen BJ et al (2009) The components and impact of sponsored search. Computer 5:98–101
5. Nash JF Jr (1950) Equilibrium points in n-person games. Proc of the Nat Aca of Sciences
36(1):48–49
6. Nau DS (1983) Decision quality as a function of search depth on game trees. J ACM 30(4):687–
708
7. Neumann JV, Morgenstern O (2007) Theory of games and economic behavior (Commemor
edn), Princeton University Press
8. Nilsson NJ (1980) Principles of artificial intelligence, 3rd edn. Narosa Publishing, India
9. Prisner E (2014) Game theory through examples, Electronic edn. Mathematical Association of
America. ISBN 978-1-61444-115-1
10. Roth AE (1983) Towards a theory of bargaining: an experimental study in economics. Science
220:687–691
11. Schaeffer J, Herik HJ (2002) Games, computers, and artificial intelligence. Artif Intell 134:1–7
12. Slagle JR, Dixon JK (1969) Experiments with some programs that search game trees. J ACM
16(2):189–207
13. Sandholm T (2015) Solving imperfect-information games. Science 347(6218)
14. Stockman GC (1979) A minimax algorithm better than alpha-beta? Artif Intell 12:179–196
Chapter 12
Reasoning in Uncertain Environments

Abstract Real-life propositions are neither fully true nor false, also there is always
some uncertainty associated with them. Therefore, the reasoning using real-life
knowledge should also be in accord. This chapter is aimed to fulfill the above objec-
tives. The chapter presents the prerequisites—foundations of probability theory, con-
ditional probability, Bayes theorem, and Bayesian networks which are graphical rep-
resentation of conditional probability, propagation of beliefs through these networks,
and the limitations of Bayes theorem. Application of Bayesian probability has been
demonstrated for specific problem solutions. Another theory—the Dempster–Shafer
theory of evidence—which provides better results as the evidences increase is pre-
sented, and has been applied on worked examples. Reasoning using fuzzy sets is yet
another approach for reasoning in uncertain environments—a theory where mem-
bership of sets is partial. Inferencing using fuzzy relations and fuzzy rules has been
demonstrated, followed by chapter summary, and a set of exercises at the end.

Keywords Reasoning in uncertainty · Probability theory · Conditional

probability · Bayes theorem · Belief networks · Belief propagation ·
Dempster–Shafer theory · Fuzzy sets · Fuzzy relations · Fuzzy inference

12.1 Introduction

The real-life propositions are neither fully true nor false, and there is uncertainty asso-
ciated with them. Therefore, the reasoning drawn from these are also probabilistic.
This requires probabilistic reasoning for decision-making. This chapter presents two
approaches for probabilistic reasoning—the Bayes theorem and Dempster–Shafer
theory. The first is implemented as Bayesian belief networks, where nodes (events)
are connected using directed graph with edges of the graph representing cause–effect
relationships; the beliefs propagate in a network as per the rules of Bayes conditional
probability. The Dempster–Shafer (D-S) theory is based on evidential reasoning such
that evidences are conjuncted, and as more and more evidences take part in the rea-
soning, the ignorance interval decreases. The classical probability theory is a special
case of D-S theory of Evidential Reasoning.

© Springer Nature India Private Limited 2020 337

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_12
338 12 Reasoning in Uncertain Environments

In the case of Classical Logic, a proposition is always taken either as true or

false. However, in a real-life scenario one cannot always say a proposition to be
either 100% true or 100% false, but known to be true with a certain probability. For
example, one can say that the proposition: “Moon rover will function alright,” may
have a probability of being true as 30%, which is calculated based on the success
rate of previous probes sent to the Moon. However, the proposition “Sun rover will
function alright” has a probability of being true as 0%, due to extreme temperature
at the Sun’s surface. Because such uncertainties are common in expert systems when
they are deployed in real-life applications, the knowledge representation as well as
the reasoning system of the expert system must be extended to work like in a real-life
situation.
Fundamentally, there are two basic approaches to reasoning in uncertain domains:
probabilistic reasoning and non-monotonic reasoning. The basis of probabilistic
reasoning is to attach some probability value to each proposition to express the
uncertainty of the event. These uncertainties may be derived from existing statistical
information available as the database of a corpus, or these may be estimated by the
experts.
Facilities for handling uncertainty have long been an integral part of knowledge
based system. In the early days of rule-based programming, the predominant methods
used variants on probability calculus to combine certainty factors associated with
applicable rules. Although it was recognized that certainty factors did not conform
to the well-established theory of probability, these methods were nevertheless favored
because the probabilistic techniques available at the time required either specifying
an intractable number of parameters or assumed an unrealistic set of independence
of relationships.
The most important area, for understanding the uncertainties is medical diagnos-
tics or troubleshooting of any system; in the first it is required to identify the patterns
(diagnoses) on the basis of their properties (symptoms), while in the second it is com-
mon to recognize the faults on the basis of system behavior. However, there is no
one-to-one mapping of symptoms and diagnoses, and the uncertainties of mapping
symptoms with diagnoses prevail due to the following reasons:

• ascertaining of the symptoms,

• proper evaluation of symptoms, and
• insufficient criteria about reckoning the scheme.

Generally, the end user of an expert system identifies and ascertains about the
symptoms and their uncertainties, while the uncertainties in the evaluation of symp-
toms are carried out by the experts. Because different people judge these uncertain-
ties, they are likely to judge it differently, and there is no scope of normalization.
Consequently, it is likely to be a large uncertainty. Often the reckoning methods
are also simple, which are some “ad hoc representations”, on account of lack of
theoretical foundations.
Many different methods for representing and reasoning with uncertain knowledge
have been developed during the last three decades, including the certainty factor
12.1 Introduction 339

calculus, Dempster–Shafer theory, possibilistic logic, fuzzy logic, and Bayesian net-
works (also called belief networks and causal probabilistic networks).

Learning Outcomes of this Chapter:

1. Identify examples of knowledge representations for reasoning under uncertainty.
[Familiarity]
2. Make a probabilistic inference in a real-world problem using Bayes theorem to
determine the probability of a hypothesis given evidence. [Usage]
3. Apply Bayes rule to determine the probability of a hypothesis given evidence.
[Usage]
4. Describe the complexities of temporal probabilistic reasoning. [Familiarity]
5. State the complexity of exact inference. Identify methods for approximate infer-
ence. [Familiarity]
6. Design and implement an HMM as one example of a temporal probabilistic
system. [Usage]
7. Explain how conditional independence assertions allow for greater efficiency of
probabilistic systems. [Assessment]
8. Make a probabilistic inference in a real-world problem using Dempster–Shafer’s
theorem to determine the probability of a hypothesis given evidence. [Usage]

12.2 Foundations of Probability Theory

In the following, we list the commonly used terminology of the probability theory.
Considering an uncertain even E, the probability of its occurrence is a measure
of the degree of likelihood of its occurrence.
A sample space S is the name given to the set of all possible events.
A probability measure is a function P(E i ) that maps every event outcomes,
E 1 , E 2 , . . . , E n from event space S, to real numbers in the range [0, 1].
These outcomes satisfy the following axioms (basic rules) of probability:
1. for any event E ⊆ S, there is 0 ≤ P(E) ≤ 1,
2. an outcome from space S that is certain is expressed as P(S) = 1, and
3. when events E i , E j are mutually exclusive, i.e., for any events E i , E j , there
is E i ∩ E j = φ, for all i = j. In that case, P(E 1 ∪ E 2 ∪ . . . ∪ E n ) = P(E 1 ) +
P(E 2 ) + · · · + P(E n ) = 1.
Using these three axioms, and the rules of set theory, the basic laws of proba-
bility can be derived. However, the axioms alone are not sufficient to compute the
probability of an outcome, because it requires an understanding of the corresponding
distribution of events. The distribution of events must be established using one of
the following approaches:
1. Collection of experimental data to perform statistical estimates about the under-
lying distributions,
340 12 Reasoning in Uncertain Environments

2. Characterization of processes using theoretical argument,

3. Understanding about the basic processes are helpful to assign subjective proba-
bility.

Some examples of computation of probability are given below. The upper case
variable names represent the sets.

count o f (All A)
P(A) = (12.1)
count o f (all events)

count o f (All A and B together)

P(A ∩ B) = (12.2)
count o f (all B’s)

P(A ∩ B) = P(A | B)P(B)

= P(B | A)P(A). (12.3)

Therefore,
P(B | A)P(A)
P(A|B) = . (12.4)
P(B)

Most of the knowledge we deal with is uncertain, hence the conclusions based
on available evidences and past experiences are incomplete. Many a time it is only
possible to obtain partial knowledge about the outcomes.

12.3 Conditional Probability and Bayes Theorem

The Bayesian networks provide a powerful framework for modeling uncertain inter-
actions between variables in any domain of concern. The interactions between the
variables are represented in two ways: 1. Qualitative way, using directed acyclic
graph, and 2. Quantitative way, which makes use of conditional probability distri-
bution for every variable in the network. This probability distribution-based system
allows to express the relationship between the variables, as functional, relational, or
logical.
Along with the above, the dynamic probability theory provides a mechanism
for revising the probabilities of events in a coherent way, as the evidences become
available. We represent the conditional probability, i.e., the probability of occurrence
of an event A, given that the event B has already occurred, as P(A|B). Note that, if the
events A and B are totally independent, then P(A|B) and P(A) will have the same
result, otherwise, due to its dependency, the occurrence of A will be effected, and
accordingly the probability P(A|B). Here, occurrence of event A is called hypothesis
and occurrence of event B is called as evidence. If we are counting the number of
sample points, we are interested in the fraction of events B for which A is also true,
12.3 Conditional Probability and Bayes Theorem 341

i.e., (B ∩ A) is what fraction of B? The meaning of fraction is like the A and B are
fractions of the universe set Ω. In (B ∩ A), both A and B are partly included, which
is also represented as (A, B). From this it becomes clear that

P(A, B)
P(A|B) = . (12.5)
P(B)

The above equation is often written as

P(A, B) = P(A|B)P(B). (12.6)

Equation (12.6) is also called “product rule” or simple form of Bayes theorem. It
is important to note that this form of the rule is not often stated a definition, but a
theorem, which can be derived from simpler assumptions. The term P(A|B) in the
equation is called posterior probability.
The Bayes theorem is used to tell us how to obtain a posterior probability of a
hypothesis A once we have observed some evidence B, given the prior probability
P(A) of event A, and probability P(B|A)—the likelihood of observing B—were A
to be given. The theorem is stated as

P(B|A)P(A)
P(A|B) = . (12.7)
P(B)

The formula (12.7), though simple, has abundant practical importance in the area
such as diagnosis, may it be patient or a fault diagnosis in software or an electric
circuit, or any other similar situation. Note that it is often easier to conclude the
probability of observing a symptom given a disease, than the reverse—that is, of a
disease, given a symptom. Yet, in practice it is the latter (i.e., find out the disease or
fault given the symptom), which is required to be computed. See what the doctors
do, it is the disease (hypothesis) they try to find out given the symptom (evidence).
There are certain conditions under which the Bayes theorem is valid. Let us
assume that there is only one symptom (S) and corresponding only one diagnosis
(D), and having given this we will try to extend the formula of the Bayes theorem. The
conditional probability for this is expressed as P(S|D), which says about relative
frequency of occurrence of the symptom S when diagnosis D is true. The term
P(S|D) can be expressed by

|S ∩ D|
P(S|D) = , (12.8)
|D|

where

|S| is frequency of occurrence of symptom S,

|D| is frequency of the occurrence of diagnosis D, and
|S ∩ D| is frequency of simultaneous occurrence of both S and D.
342 12 Reasoning in Uncertain Environments

The probability P(D|S), called posterior probability or conditional probability

of diagnosis D, assuming a symptom S, is given by Bayes theorem as

P(S|D)P(D)
P(D|S) = . (12.9)
P(S)

The term P(S|D) is called likelihood of symptom S when diagnosis is D, and

P(D) is called prior probability.

The above can be proved as follows:

|S ∩ D|
P(D|S) =
|S|
|D| |S ∩ D|
= ∗
|D| |S|
|S ∩ D|
= |D| ∗
|D| ∗ |S|
|S ∩ D| 1
= |D| ∗ ∗
|D| |S|
P(S|D) ∗ P(D)
= ; using Eq. (12.8)
P(S)

The denominator terms P(S) is called the normalizing factor.

Example 12.1 Suppose that the following statistics hold for some disease:
P(tuber culosis) = 0.012,
P(cough) = 0.10, (for over 3 weeks)
P(cough | tuber culosis) = 0.90, i.e., if a patient has tuberculosis, the coughing
is found in 90% of the cases.
Given this, compute the probability of having tuberculosis, given that coughing
exists in those persons, for over 3 weeks.

P(cough|tuber culosis) × P(tuber culosis)

P(tuber culosis|cough) =
P(cough)
0.9 × 0.012
=
0.10
= 0.108.

Hence, the relative probability that a patient who had a tuberculosis was found
to be having also the cough is, 8.33 (= 0.9/0.108) times the probability of finding
tuberculosis in the coughing patients. 
12.3 Conditional Probability and Bayes Theorem 343

Relative Probability
The Bayes theorem allows the determination of the most probable diagnosis, assum-
ing the presence of symptoms S1 . . . Sm . The probability is determined using: 1.
a priori probabilities P(Di ) that belong to a set of diagnoses, and 2. likelihood
P(S j /Di ). The latter is the relative frequency of the occurrence of a symptom S j
given that the diagnosis is Di . In the Bayes probability, an absolute probability is
not important, but a relative probability Pr is determined. Pr is the probability of a
diagnosis compared to the other diagnoses, which is computed as

P(Di ) P(S1 |Di ) . . . P(Sm |Di )

Pr (Di |S1 ∧ S2 ∧ . . . ∧ Sm ) = . (12.10)
Σ nj=1 P(D j ) P(S1 |D j ) . . . P(Sm |D j )

Note that in the above definition, the normalization factor of P(S1 ∧ . . . ∧ Sm ),

which is a common denominator, in the numerator and denominator terms, has been
dropped.
When we are considering many symptoms, as in the above case, it is necessary that
the correlation of every relevant combination of symptoms to disease be determined.
However, if this happens, it would lead to a combinatorial explosion of diagnoses,
e.g., with 100 all possible symptoms, there are 2100 different constellations of diag-
noses. This figure is more that 1030 !
Therefore, while using the Bayes theorem it is always assumed that the symptoms
are independent of each other, i.e., for two symptoms, Si and S j

P(Si ∧ S j |D) = P(Si |D)P(S j |D). (12.11)

The exception to the above is allowed when symptoms are caused directly by the
same diagnosis.
From Eqs. (12.10) to (12.11), it follows that

P(S1 ∧ . . . ∧ Sm |D)
P(D|S1 ∧ . . . ∧ Sm ) =
P(S1 ∧ . . . ∧ Sm )P(D)
P(S1 |D) . . . P(Sm |D)
= .
P(S1 ∧ . . . ∧ Sm )P(D)

The Bayes theorem in the above form is correct, subject to the fulfillment of the
following conditions:
1. The symptoms may depend only on the diagnoses, and must be independent of
each other. This becomes a critical point, when more and more symptoms are
ascertained.
2. The set of diagnoses is complete.
3. Single fault assumption or mutual exclusion of diagnoses—presence of any one
diagnosis automatically excludes the presence of other diagnoses. This criteria is
justified only in a relatively small set of diagnoses.
4. The requirement of correct and complete statistics of prior probabilities (P(D))
of the diagnoses, and conditional symptom-diagnosis probabilities(P(S|D)).
344 12 Reasoning in Uncertain Environments

12.4 Bayesian Networks

The Bayesian network is a graphical representation of the Bayes theorem—a set of

variables represented as cause–effect relations, where variables are nodes and edges
represent the relevance or influence relation between variables. Absence of an edge
between two variables indicates independence of the variables, i.e., nothing can be
inferred about the state of one variable, having given the state of the other variable.
A variable may assume the values from a collection such that this collection is a
mutually exclusive as well as collectively exhaustive set. A variable is allowed to
be discrete—having a finite countable number of states—or it may be continuous—
infinitely many possible states.
The lowercase letters in Bayesian networks represent single variables, while
uppercase letters represent sets of variables. To assign state k to variable x, we
write as x = k. When the state of every variable is comprised in set X , then this
set (an observation) is an instance of X . Set of all the instances is represented by
a universal set U , which is the joint space of all the instances. A joint probability
distribution over set U is the probability distribution over joint space U . Considering
X, and Y as sets, the expression P(X |Y ) denotes the joint probability distribution
over the set X , one for each conditional corresponding to every instance in the joint
space of Y .
The Bayesian networks are good for structuring probabilistic information about a
situation in a systematic way. This information is provided in localized and coherent
way. These networks also provide a collection of algorithms, which help in deriving
many information as implications of the conditionals, that can lead to important
conclusions, and decisions about the given situation. These may be for example, in
the area of genetics—mapping genes onto a chromosome, in message transmission
and reception—finding the most likely message that was sent across a noisy channel,
in system reliability—computing the overall reliability of a system, in online sales,
e.g., to identify the most likely users who would respond to a given advertisement,
in image processing, e.g., restoring a noisy image, and so on.
In more technical terms, a Bayesian network is a representation in a compact form
of probability distribution using graphical method. The traditional methods were
usually too large to handle probability and statistics computations, such as using
tables and equations. Typically, a Bayesian network of 100s variables can be easily
constructed and used for reasoning successfully about probability distribution [4].

12.4.1 Constructing a Bayesian Network

Though the definition of a Bayesian network is based on conditional independence,

these networks are usually constructed using the notions of cause and effect. In simple
words, for a given set of variables, we construct a Bayesian network by connecting
every cause to its immediate effect using arrows in that direction. In almost all the
cases, this results in Bayesian networks, whose conditional independence are implicit
and accurate.
12.4 Bayesian Networks 345

Fig. 12.1 Cause–effect c1

Bayesian network

c2
e

cn

Figure 12.1 shows a Bayesian network with many causes (c1 , c2 , . . . , cn ) and a
single effect e. Due to this, specification of its local distribution can be quite onerous
in the general case. As per the definition for a Bayesian network, it is necessary to
compute the probability distribution of the node conditional on every instance of its
parents. For example, if a node x has n number of parents, each is binary in nature
(either 0 or 1 only), then we need to specify total 2n probability distributions for the
node x. We can reduce the complexity of this computation in such cases by adding
more structures in the model. For example, we can convert our model into a n-way
interaction model by associating with each cause node an inhibitory mechanism that
prevents the cause from producing the effect. In such a model, the effect will be
absent if and only if all the inhibiting mechanisms associated with the present cause
node are active. And, if any of the inhibiting mechanism for this node is inactive, the
cause node will produce the effect.

12.4.2 Bayesian Network for Chain of Variables

A problem domain is nothing but a set of variables. Thus, a Bayesian network for the
domain {x1 , . . . , xn } can represent a joint probability distribution over these variables.
This distribution consists of a set of local conditional probability distribution, that
when combined with a set of assertions of conditional independence, gives a joint
global distribution.
The composition is based on the chain rule of probability (see Fig. 12.2), which
dictates that

n
P(x1 , . . . , xn ) = P(xi |x1 , . . . , xi−1 ). (12.12)
i=1

For each given variable xi , let us assume that Si ⊆ {x1 , . . . , xi−1 } is a set of
variables which causes the variable xi , and {x1 , . . . , xi−1 } to be conditionally inde-
pendent. That is

Fig. 12.2 Chain of domain x1 x2 x3 xn−1 xn

variables
346 12 Reasoning in Uncertain Environments

P(xi |x1 , . . . , xi−1 ) = P(xi |Si ). (12.13)

The approach to this idea is that the distribution of xi can be represented as condi-
tional on a parent set of Si which is substantially smaller than {x1 , . . . , xi−1 }. Given
these two sets, a Bayesian network can be represented as a directed acyclic graph
such that each variables x1 , . . . , xn corresponds to a node in that graph and the parents
of the node corresponding to the variables in Si . Note that since the graph coincides
with the conditioning set Si , the assertions of the conditional dependence are directly
encoded by the Bayesian network structure. This is expressed in Eq. (12.13).
A conditional probability distribution P(xi |Si ) is associated with each node
xi , such that there is one distribution for each instance of Si . By combining the
Eqs. (12.12) and (12.13), we obtain a Bayesian network for {x1 , . . . , xn } which
uniquely determines a joint probability distribution for these variables. That is,

n
P(x1 , . . . , xn ) = P(xi |Si ). (12.14)
i=1

Because the joint probability distribution for any network can be determined
using Bayesian network for a given domain, this network can be used in principle to
compute any probability of interest. Consider that we are given a simple Bayesian
network, with structure w → x → y → z, and we are interested to find out the
probability “of occurrence of event w given that the event z has already occurred,”
which can be expressed as P(w|z). From the Bayes rule, and making use of the above
structure w . . . z, the equation can be expressed as follows.

P(w, x, y, z)
P(w, z) x,y
P(w|z) = =  . (12.15)
P(z) P(w, x, y, z)
w,x,y

In the above, P(w, x, y, z) is called the joint probability distribution computed for the
Bayesian network w → x → y → z. On observation of numerator and denominator
terms of Eq. (12.15), we can understand that this approach is not feasible, due to the
reason that it requires summing over exponential number of terms. However, we can
exploit the conditional independence encoded in a Bayesian network to arrive at a
more efficient computation. Using this feature, the network structure of Eq. (12.15)
can be transformed to

P(w, x, y, z)
x,y
P(w|z) = 
P(w, x, y, z)
w,x,y
 
P(w) P(x|w) P(y|x)P(z|y)
x y
=   . (12.16)
P(w) P(w)P(x|w) P(y|x)P(z|y)
w x y
12.4 Bayesian Networks 347

Equation (12.16) indicates that using conditional independence (i.e., indepen-

dence of conditional variable), we can often reduce the dimensions of an equation
by rewriting the “sums over multiple variables as a product of sums over a single
variable,” or at least to a lesser number of variables.
A probabilistic inference is a general problem of computing probabilities of inter-
est from a joint probability distribution, which may be possibly implicit. It may be
noted that all the exact algorithms for probabilistic inference using Bayesian net-
works exploit conditional independence in the manner described above.
For drawing probabilistic inferences, it is possible to exploit the conditional inde-
pendence in a Bayesian network, however, the exact inference in an arbitrary Bayesian
network is NP-hard.1 But, actually, for many applications, the Bayesian networks are
of small size, or they can be simplified sufficiently, such that the complexity issues
are not important. For applications where usual inference are impracticable, there
are equivalent applications existing which are particular network tailored or suited
for specific queries.

12.4.3 Independence of Variables

The notions of independence and conditional independence are fundamentals notions

of probability theory. It is the combination of quantitative information of numerical
parameters and the qualitative information, that makes the probability theory so
expressive. Consider x and y as two variables that are independent of each other.
The corresponding probabilistic expression of occurrence of events corresponding
to these variables is
P(x, y) = P(x)P(y). (12.17)

Given three variables x, y, z, the probability distribution of these can be decom-

posed into a joint probability distribution that comprises terms, each of two variables,
as shown in Eq. 12.18 and Fig. 12.3.

P(x, y, z) = P(x, y|z)P(z)

= P(x|z)P(y|z)P(z). (12.18)

For the situation shown in Fig. 12.3, the variables x and y are called marginally
independent, but they are conditional dependent, for a given variable z. It is possible
to convince ourselves using the following example. Let the first variable have the
assignment x = “rain”, and the second is y = “sprinkler on”. When both are True,
they will cause the lawn to become wet.
So far we have not made any observation about the lawn, and occurrence of x
(rain) and y (sprinkler on) are independent. However, once it is observed that the lawn

1 NP-Hard:Non-deterministic Polynomial-hard, i.e., whose polynomial nature of complexity is

unknown, and these problems are considered as the hardest problems in Computer Science.
348 12 Reasoning in Uncertain Environments

Rain Sprinkler on
x y

z Wet lawn

Fig. 12.3 Variables x and y are “Conditionally Dependent Given” z

x z y

Fig. 12.4 Variables x and y are conditionally independent given z

is wet, and it is confirmed that it raining, it automatically influences the probability

of the sprinkler, that it is “on”. This probability distribution is therefore

P(x, y, z) = P(z|x, y)P(x)P(y). (12.19)

The next example shows the cause–effect relationships shown in Fig. 12.4. In this
case, the probability distribution is given by

P(x, y, z) = P(y|z)P(z|x)P(x). (12.20)

12.4.4 Propagation in Bayesian Belief Networks

The aim of constructing a Bayesian network is—for a given observation (evidence),

answer a query about the probability distribution over the values of query variables.
A Bayesian network having full specification contains all the information needed to
answer all the queries of probability distribution over these variables. The propagation
of evidence in these networks helps in drawing conclusions. Usually, we make use
of the term “propagation” only, instead of propagation of evidence.
This kind of representation is called Bayesian belief network. The inference in this
network amounts to the propagation of probabilities of a given and related information
through the network to one or more conclusion nodes. If we consider representing
the knowledge related to a set of variables, say, x1 , x2 , . . . , xn , using their point
probability distribution, P(x1 , x2 , . . . , xn ), it will require a total of 2n entries to store
the entire distribution explicitly. Further, for the determination of probability say xi
will require summing all the remaining xi , which will make it prohibitively expensive
in terms of computation.
However, once it is represented using causal relationships, the joint probability
can be computed much faster. Figure 12.5 shows the cause and effect relationships
12.4 Bayesian Networks 349

Fig. 12.5 Bayesian belief x1

network

x2 x3

x4 x6

x5

between variables x1 . . . x5 , and the joint probability distribution can be given as

P(x5 |x4 ∧ x3 )P(x4 |x2 ∧ x3 )P(x2 |x1 )P(x3 |x1 )P(x6 |x3 )P(x1 ). (12.21)

Example 12.2 Some examples of Bayesian belief networks.

Following are causal dependencies of variables, their corresponding belief networks,
and probability distributions:
(i) Let W = “Worn piston rings”, that causes O = “excessive oil consumption”,
which in turn causes L = “low oil level in fuel tank”. Figure 12.6 shows the
cause–effect relationships, and the joint probability distribution for this is given
by
P(W, O, L) = P(W ) P(O|W ) P(L|O). (12.22)

(ii) As another example, let W = “Worn piston rings” cause both B = “blue exhaust”,
as well as, L = “low oil level” (see Fig. 12.7). The joint probability distribution
is given by
P(W, B, L) = P(W ) P(B|W ) P(L|W ). (12.23)

(iii) Consider another example, with variables, L = “Low oil level”, which can
be caused either by C = “excessive consumption” or by E = “oil leak” (see
Fig. 12.8). The joint probability distribution is given by

P(C, E, L) = P(C) P(E) P(L|C, E). (12.24)


Based on the above discussions, we have the following steps for the construction
of a Bayesian network:

W O L

Fig. 12.6 Bayesian network-1

350 12 Reasoning in Uncertain Environments

Fig. 12.7 Bayesian

W
network-2

B L

Fig. 12.8 Bayesian

C E
network-3

Fig. 12.9 Bayesian

Fire(F) Tempering(T)
network-4

Smoke(S) Alarm(A)

Leaving(L) Report(R)

1. Construct belief (event) nodes corresponding to the events,

2. For each node in belief network, make all links corresponding to the cause–effect
relationships,
3. Compute probability at each nodes based on the probabilities of the premises
nodes, and then,
4. Compute the joint probability distribution of the network.

Example 12.3 A Bayesian network for fire security system.

Figure 12.9 shows a small network of six binary variables given in Tables 12.1 and
12.2. The use of such network may be to compute answers to probabilistic queries, for
example, we may like to know the probability of fire, given that people are reported
to be leaving from the room, or answers for queries like what is the probability of
smoke given that the alarm is off.
A Bayesian network has two components: a structure in the form of a directed
acyclic graph, and a set of conditional probability tables. The acyclic graph’s nodes

Table 12.1 Two-variable table

Fire Smoke P(S|F)
False True 0.01
True True 0.90
12.4 Bayesian Networks 351

Table 12.2 Four-variable table

Fire(F) Tempering(T ) Alarm(A) P(A|F, T )
True True True 0.50
False True True 0.85
True False True 0.99
False False True 0.0001

correspond to variables of interest (see Fig. 12.9), and the graphs edges have a
formal interpretation in the form of direct causal influences. A Bayesian network
must include the conditional probability tables (CPTs) for each variable quantifying
the relationship between a variable and its parents in the network.
Consider that there are variables A = “Alarm”, F = “tempering”, and F =
“Fire”, each of which may be True or False. Figure 12.9 shows the acyclic graph,
and Table 12.2 shows the conditional probability distribution of A, given its par-
ents as “Fire” and “Tempering”. As per this CPT, the probability of occurrence
of the event A = T r ue, given the evidence F = T r ue and T = False (row 3),
is P(A = T r ue|F = T r ue, T = False) = 0.99. This probability is called the net-
work parameter.
The guaranteed consistency and completeness are the main features of Bayesian
networks. The latter is possible because there is one and only one probability distri-
bution, which satisfies the network constraints.
A Bayesian network with n binary variables will induce a unique probability
distribution for over 2n instantiations; such instantiations for the network shown in
Fig. 12.9 are 64. This distribution provides sufficient information to predict prob-
ability for each and every event we can express using variables S, F, T, A, L , R,
appearing in this network. An event may be a combination of these variables with
True/False values. An example of an event may be to found out as “probability of an
Alarm and Tempering, given no smoke, and a report has indicated people leaving the
building.” 
An important advantage of Bayesian networks is the availability of efficient algo-
rithms for computing such probability distributions. In the absence of these networks,
it would require explicit generation of required probability distributions. Generating
such distributions explicitly is infeasible for a large number of variables. Interestingly,
in areas such as genetics, reliability analysis, and information theory, the already
existing algorithms are subsumed by the more general algorithms for Bayesian net-
works.

12.4.5 Causality and Independence

We will try to discover the central insight behind Bayesian networks, due to which it
becomes possible to represent large probability distributions in a compact way. Con-
sider Fig. 12.9 and the associated Conditional Probability Tables (CPTs), Tables 12.1
352 12 Reasoning in Uncertain Environments

and 12.2. Each probability appearing in a CPT specifies a constraint which must be
satisfied by the distribution induced by the network. Consider that a distribution
must assign the probability of 0.1 to the event of “having smoke without fire”, i.e.,
P(S = T r ue|F = False), where S and F are variables, respectively for “smoke”
and “fire”. These constraints, however, are insufficient to help in concluding a unique
probability distribution. So, what additional information we need? This answer lies
in the structure of a Bayesian network; as we know, that additional constraints are
specified by the Bayesian network in the form of probabilistic conditional indepen-
dence. As per that, once the parent of every variable is known, the variable in the
structure is assumed to be independent of its non-descendent parents. Figure 12.9
shows that the variable L that stands for “leaving the premises” is taken as indepen-
dent of its non-descendant parents T , F, and S, once its parent A becomes known.
Put differently, as soon as the variable A becomes known, the probability distribu-
tion of variable L will not change on the availability of new information about the
variables T , F, and S.
As a different case in Fig. 12.9, we assume that variable A has no dependence
on variable S (a non-descendant parent of A). This assumption becomes true as
soon as the parents F and T of variable A becomes known. These independence
constraints, due to non-descendant parents are called Markovian assumptions of a
Bayesian network [2].
From the above discussion, do we mean that whenever a Bayesian network is
constructed, it is necessary to verify the conditional non-dependencies? This, in fact
depends on the construction method adopted. There are three main approaches to the
construction of Bayesian networks:

1. subjective construction,
2. a construction based on synthesis from other specifications, and
3. construction by learning from data.

The first method in the above approaches is somewhat less systematic, as one rarely
thinks about the conditional independence while constructing a Bayesian network.
Instead, one thinks about causality, i.e., adding an edge E i → E j , from event E i
to event E J , whenever E i is recognized as a direct cause of E j . This results in a
causal structure where Markovian assumption is read as: “each variable becomes
independent of its non-dependents, once direct causes are known.”
A distribution of probabilities induced due to a Bayesian network also satisfies
additional independencies beyond the Markovian. All these independent variables
can be identified using a graphical test, called d-separation.2 As per this test, any two
variables E i and E j shall be considered to be independent if every path between E i
and E j is blocked by a third variable E k . For instance, consider the path α as shown
in Fig. 12.9.
α : S ← F → A ← T. (12.25)

2 d-separation: A method to determine which variables are independent in a Bayes net [3].
12.4 Bayesian Networks 353

Now, consider that alarm A is triggered. The path in Eq. (12.25) can be used to
show a dependency between variables A and T as follows: we know that observing S
(smoke as evidence) increases the likelihood that fire F has taken place. This is due to
the direct cause–effect relation. Also, the increased likelihood that F has taken place
explains away the tempering T as cause of the alarm, i.e., there are lesser chances
that the alarm is due to tempering. Hence, a path in Eq. (12.25) can be used as a
dependence between S and T . Therefore, the variables S and T are not independent
due to the presence of unblocked path between them.
However, instead of variable A, if F is a given variable, this path cannot be used
to show a dependency between S and T , and in that case the path will be blocked
by F.

12.4.6 Hidden Markov Models

The Hidden Markov Models (HMMs) are useful for modeling dynamic systems with
some states not observable, and when it is required to make inferences about changing
states, given the sequence of outputs they generate. The potential applications of
HMMs are those that require temporal pattern recognition, like speech processing,
handwriting recognition, recognizing gestures, and in bioinformatics [2].
Figure 12.10a shows an HMM—a model of a system with three states a, b, c, and
with three outputs x, y, z, and possible transitions between the states. There is prob-
ability associated with each transition, e.g., there is a transition from state b to c with
20% probability. Each state can produce certain output, with some probability, e.g.,
state b can produce output z with probability 10%. An HMM of Fig. 12.10a has been
represented by a Bayesian network as shown in Fig. 12.10b. There are two variables,
St (for state at time t), and Ot (for system output at time t), with t = 1, 2, . . . , n.
The variable St has three values a, b, c, and Ot also has three values x, y, z. Using
d-separation on this network, one can derive the characteristic properties of HMMs,
i.e., once the system state at time t is known, its states and outputs are independent
at time > t, and also independent at time < t. Figure 12.10b is a simple dynamic
Bayesian network.

.20

a b c S1 S2 S3 Sn

.10

O1 O2 O3 On
x y z
(a) HMM (b) Dynamic Basian network.

Fig. 12.10 Hidden Markov model

354 12 Reasoning in Uncertain Environments

12.4.7 Construction Process of Bayesian Networks

One approach for the construction of Bayesian networks is mostly subjective—

reflecting on the available knowledge in the form of say, perceptions about causal
influences. This knowledge is captured into a Bayesian network, e.g., the network
shown in Fig. 12.9 we discussed earlier.
Automatic Synthesizing of networks
The second method for Bayesian networks construction synthesizes these networks
automatically using some other type of formal knowledge. Many jobs, related to
reliability and diagnosis, require system analysis. A Bayesian network can be auto-
matically synthesized from the formal system design of such systems. We consider
a job of reliability, where a reliability block diagram is used for reliability analysis,
with system components being connected to effect their reliability and dependencies,
as shown in Fig. 12.11a. The various components shown are power supply to power
the Fan-1 and Fan-2, which collectively cool the processor-1 and 2, and these proces-
sors are interfaced to the a hard disk drive. The processor-1 requires the availability
of either Fan-1 or 2, and each fan requires the availability of power supply. We are
interested to compute the overall reliability of the system, i.e., probability of its avail-
ability, given the reliability of each of the components in the system. Figure 12.11b
shows the systematic conversion of each block of the reliability block diagram into
a Bayesian network fragment, where Subsystem-1, Subsystem-2, and Block-B are
assumed to be available [2].
Figure 12.12 shows the corresponding Bayesian network constructed using reli-
ability blocks. Using the reliability of individual components we can construct Con-

Fan1 Processor 1
Hard
Power disk
supply drive
Fan2 Processor 2

(a) Reliability block diagram.

Subsystem-1 Block B
Subsystem-1

Block B OR AND

Subsystem-2 Subsystem-2

(b) Bayesian network fragments

Fig. 12.11 Syntheizing a Bayesian network

12.4 Bayesian Networks 355

Fig. 12.12 A Bayesian P1

F1
network for reliability
analysis A3
A1
O2 S
E O1

A4
A2
D
F2 P2

ditional Probability Tables (CPTs) of this system. Consider the variables as follows:
E is for power supply, F1 and F2 for fans, P1 , P2 for processors, and D for hard disk.
Let these variables represent the availability of the corresponding component. Let us
assume that variable S represents the availability of the whole system. The variables
Ai and O j represent the logical AN D and O R, respectively.
Network Constructing by learning from data
A third approach for the construction of Bayesian networks works on the principle
of learning from data, like patients medical records, airline ticket records, or buying
patterns of customers, etc. Such data sets can help the network to learn parameters
given the structure of the network, or it can learn both the network and its parameters
when the data set is complete. However, learning only the parameter is an easier task.
Since the learning itself is an inductive task, the learning process is guided using
the induction principle. Two main principles of inductive-based learning are 1. max-
imum likelihood function, and 2. learning using the Bayesian approach. The first
approach is suitable to those Bayesian networks that maximize the probability of
observing the given data set, while the second approach uses some prior information
that encodes preferences on a Bayesian network, along with the likelihood principle.
Let us assume that we are interested in learning the network parameters. Learning
using the Bayesian approach allows to put a prior distribution on each network
parameter’s possible values. The data set and the prior distribution together induce
new distribution on the values of that parameter, called posterior distribution. This
posterior distribution is then used to pick a value for that parameter, e.g. distribution
mean. As an alternative, we can also use different parameter values while computing
answers to queries. This can be done by averaging over to all possible parameter
values according to their posterior probabilities.
Given a Bayesian network as in Fig. 12.12, we can find out the overall reliability
of this system. Also, for example, given that the system is unavailable, we can find
the most likely configuration of the two fans, or the processors. Further, given a
Bayesian network, we can answer the questions such as What single component can
be replaced to increase the system reliability by, say, 10%? These are the example
questions which can be answered in these domains using the principles of Bayesian
probability distributions.
356 12 Reasoning in Uncertain Environments

12.5 Dempster–Shafer Theory of Evidence

The Dempster–Shafer Theory (DST) of evidence can be used for modeling several
single pieces of evidences within a single hypothesis relations,3 or a single piece of
evidence in relations that are multi-hypotheses type. The DST is useful for the assess-
ment of uncertainty of a system where actually only one hypothesis is true. Another
approach to DST, called the reliability-oriented approach, contains the system with
all hypotheses—pieces of evidence and data sources. The hypothesis is a collection
of all possible states (e.g., faults) of the system, which is under consideration [5].
In DST, it is a precondition that all hypotheses are singletons (i.e., elements) of
some frame of discernment—a finite universal set Ω, with 2|Ω| subsets. However,
we will write it simply as 2Ω subsets. A subset of Ω is a single hypothesis or a
conjunction of hypotheses. The subsets 2Ω are unique and not all of them disjoint,
however, it is mandatory that all the hypotheses are disjoint and mutually exclusive,
in addition to being unique.
The pieces of evidences above are symptoms or events. An example of a symp-
tom or evidence is the failure that has occurred or may occur in the system. An
evidence is always related to a single hypothesis or to a set of hypotheses (multi-
hypothesis). Though theoretically it can be debated, it is not allowed in DST that
many different evidences may lead to conclude the same hypothesis or the same
collection of hypotheses. A relation between a piece of evidence and a hypothesis
is qualitative, which correspond to a chain of cause–effect relations—an evidence
implies an existence of hypothesis or hypotheses, like in the Bayes rule.
The DST represents a subjective viewpoint in the form of a computation for an
unknown objective fact. The data sources used in DST are persons, organizations,
or other entities that provide the information. Using a data source, the mapping in
Eq. (12.26) assigns an evidential weight to a diagnosis set A ⊆ Ω. Since A ∈ 2Ω ,
the set A contains a single hypothesis or a set of hypotheses. The probability of any
diagnosis A is a function whose value is between 0 and 1, and is expressed by

m : 2Ω → [0, 1]. (12.26)

Note that the difference between DST with probability theory is that, the DST map-
ping clearly distinguishes between the evidence measures and probabilities.
Definition 12.1 (Focal Element) Each diagnosis set A ⊆ Ω, for which m(A) > 0,
is called a focal element.
Definition 12.2 (Basic Probability Assignment) The function m is called a Basic
Probability Assignment (BPA) that fulfills the condition:

m(A) = 1. (12.27)
A⊆Ω

3H → E is a hypothesis relation, where H is hypothesis and E is evidence.

12.5 Dempster–Shafer Theory of Evidence 357

Meaning of the above statement is that, for the evidences presented by each data
source which are equal in weight, it is necessary that all statements of a single data
source are normalized. In other words, no data source is more important than the
others. We assume that if a data source is null, then its BPA should also be null, i.e.,
m(φ) = 0.
A belief measure is given by the function, which also ranges between 0 and 1,

bel : 2Ω → [0, 1], (12.28)

such that 
bel(A) = m(B). (12.29)
B⊆A; B=φ

Counterpart of belief (bel) is called plausibility measure pl, which again is a

mapping pl : 2Ω → [0, 1], and is defined as

pl(A) = m(B). (12.30)
B∩A=φ

It is important to note that pl(A) above is not be understood as a complement of

bel(A). For a focal element m(A) > 0, and A ⊆ Ω, it always holds that bel(A) ≤
pl(A). The difference between plausibility and belief is evidential interval range,
called uncertainty interval. Thus,

uncertant y interval = pl(A) − bel(A). (12.31)

As we add more and more evidences in the system, the uncertainty interval
reduces, as will be seen in examples in the following. This is a natural and logi-
cal property of DST.

12.5.1 Dempster–Shafer Rule of Combination

The DST allows combination of evidences to predict the hypothesis with a stronger
belief. It is called Dempster’s rule of combination, which aggregates two independent
bodies of evidence, into one body of evidence, provided that they are defined within
the same frame of discernment. The Dempster’s rule of combination combines two
evidences and computes new basic probability assignment as

m 1 (B) ∗ m 2 (C)
A=B∩C
m(A) = m 1 ⊗ m 2 (A) =  . (12.32)
1− m 1 (B) ∗ m 2 (C)
B∩C=φ
358 12 Reasoning in Uncertain Environments

In the above, A = φ and numerator part in the equation represent the accumulated
evidences for the sets B and C, that supports the hypothesis A (since A = B ∩ C),
and the denominator is called the normalization factor [5].

12.5.2 Dempster–Shafer Versus Bayes Theory

The Dempster–Shafer Theory (DST) (also called theory of belief functions) is the
generalization of the Bayes theory of conditional probability. To carry out the com-
putation of probability, the Bayes theory requires actual probabilities for each ques-
tion of interest. But in DST, the degree of belief of one question is based on the
probabilities of the related questions. The “degree of belief” may or may not have
real mathematical properties of probability, but, how much they actually differ from
probabilities will depend on how closely the two questions are related.
The DST is based on two ideas: 1. obtaining degrees of belief for one question from
subjective probabilities of related question, and 2. use Dempster’s rule for combining
such degrees of belief when they are based on independent items of evidence.
The DST differs from the application of Bayes theorem in the following respects:
1. Unlike the Bayes theorem, the DST allows the representation of diagnoses in
the form of classes and hierarchies, as shown in Fig. 12.13.
2. For a probability of X % for a diagnosis, the difference (100 − X )% is not taken
as against the diagnosis, but interpreted as an uncertainty interval. That is, the
probability of diagnosis lies between X % and 100%, and as more and more
evidence is added the diagnosis may tend toward 100%.
3. Probability against a diagnosis is considered as the probability of the complement
of the diagnosis. For example, if an event set is {A, B, C, D}, the probability
against the event “A” is probability of {B, C, D}, which will have further distri-
bution.

In spite of many advantages of DST, evaluation of probabilities in DST is more

complex than in the Bayesian probabilities. This is because, probabilities for a set of
diagnoses are related to one another in DST, hence probability of a set of diagnoses
needs to be calculated from the distribution of probabilities over all sets of diagnoses.

Fig. 12.13 Hierarchy of

evidences
12.5 Dempster–Shafer Theory of Evidence 359

Example 12.4 Computing Probability distribution of a biased coin.

Consider a biased coin, where the probability of arriving at the head is P(H ) = 0.5.
Since it is biased we do not know as to what the rest 50% is attributed to. In the DS
theory, this is called the ignorance level. Since we do not know where the rest of the
probability goes, we assign it to the universe U = {H, T }. Which means that it can
attribute to {T } and {H, T }.
Now let us assume that there is an evidence, that probability for arriving at the tail
is 46%. Due to this new evidence, the balance probability assignment is 1 − 0.45 −
0.46 = 0.09, and this can be assigned to the set {H, T }, that cannot be attributed to
{H } or {T }, unless there is some further evidence available. Hence, 0.09 probability
assigned stands for the coin standing vertical when it is tossed. 

Example 12.5 Using DST for finding probabilities for certain diseases.

Let us assume that in a certain isolated island the universe set of disease is Ω =
{malaria, typhoid, common cold}. A team of Doctors visit this island and perform a
laboratory test (T1 ) on a patient, which shows that it is the case of 40% for Malaria.
Hence, as per the DS theory, the balance 60% goes to the ignorance level for the
entire universe {malaria, typhoid, common cold}.
Fearing lack of confidence, the Doctors ask the patient for a second laboratory
test, for some other parameters; the evidence now shows a belief case of 30% for
Malaria. Hence, the rest 30% goes to the entire universe {malaria, typhoid, common
cold}. Thus ignorance is 30%.
The Doctor then asks for a third laboratory test, to further add to the evidences, and
the finding shows that the belief for Typhoid is 20%. This makes the total belief of
90%, and the ignorance level reduces to 10%, which again is assigned to the universe
set {malaria, typhoid, common cold}. 
The above example shows that as more and more evidence is added, we get more
and more strong belief about the distribution of probability on various diagnoses,
and the ignorance level reduces. This, in fact, is what it should be, and the natural
case of reasoning in a probability case.
Let us assume that the kind or errors which can occur in software testing is the
set: {data validation errors, computation errors, transmission errors, output error}. In
this case also we use DST, similar to the cases discussed above, and can compute
the joint probability distribution of errors.

Example 12.6 Find out the distribution of probabilities for some diseases shown
by a hierarchically structured diagnoses as shown in Fig. 12.13. Consider that the
following are the probabilities of certain events: against A = 40%, and evidence for
Y = 70%.

Let event E 1 correspond to against A = 40%, and event E 2 : correspond to Y =

70%. The evidence against A means evidence for the complement of A, which is the
set of evidences {B, C, D}. From Fig. 12.13, the evidence for Y is evidence for set
{A, B}.
360 12 Reasoning in Uncertain Environments

Table 12.3 Combination of evidences

E1 ↓ E2 → {A, B} = 0.7 {A, B, C, D} = 0.3
{B, C, D} = 0.6 {B} = 0.42 {B, C, D} = 0.18
{A, B, C, D} = 0.4 {A, B} = 0.28 {A, B, C, D} = 0.12

In DST, when the evidences are combined, the probabilities get multiplied and the
resultant set is the intersection of original sets. Table 12.3 shows the new distribution.
The probability of an event or a set of events is expressed not as an absolute value
but an uncertainty interval [a, b], where a and b are lower and upper probability
limits, respectively. The value a is computed as the sum of the probabilities of the
set and its subsets, while b is a value 100 − c, where c is the sum of the probabilities
of complements of the set and complements of the subsets corresponding to a.
The probability range [a, b] for diagnoses {A} and {B} are computed as follows:

For diagnosis {A}:

[a, b] = [{A}, 1 − ({B, C, D} + {B, C} + {B, D} + {C, D} + {B} + {C} + {D} + {})]
= [0, 1 − 0.6] = [0, 0.4].

Hence, the probability of diagnosis {A} lies between 0 and 40%.

We know that in DST, probabilities are multiplied and intersection is performed.
Since probability for {A, B} is 0.7 and for {B, C, D} is 0.6, the probability for {B}
({A, B} ∩ {B, C, D} = {B}) is 0.42. Thus, the probability range for diagnosis B is

[a, b] = [{B}, 1 − ({A, C, D} + {A, C} + {A, D} + {C, D} + {A} + {C} + {D} + {})]
= [0.42, 1 − 0.0] = [0.42, 1].

Hence, the probability for {B} lies between 42 and 100%.

If a further evidence E 3 of 20% for D is added, the distribution of probabilities
is changed as shown in Table 12.4.
The empty bracket {} shows that there is no diagnosis attributed to this fraction
of probability. Since it is assumed that the set of diagnoses are complete, we can

Table 12.4 Combination of evidences after a new evidence E 3 is added

E 1 &E 2 ↓ E 3 → {D} = 0.2 {A, B, C, D} = 0.8
{B} = 0.42 {} = 0.084 {B} = 0.336
{A, B} = 0.28 {} = 0.056 {A, B} = 0.224
{B, C, D} = 0.18 {D} = 0.036 {B, C, D} = 0.144
{A, B, C, D} = 0.12 {D} = 0.024 {A, B, C, D} = 0.096
12.5 Dempster–Shafer Theory of Evidence 361

eliminate the empty set and the remaining sets of the probabilities can be taken as
100%. This can be done by dividing all probabilities of non-empty sets by a factor
equal to (1 − s), where s is the sum of the probabilities of empty sets. For this
example, s = 0.084 + 0.056 = 0.14 and 1 − s = 1 − 0.14 = 0.86. Now, the sum
of probabilities without an empty set is 1.0. Next, we obtain the following results
using Dempster’s rule of combination.

0.336
{B} = = 39.0%
0.86
0.06
{D} = = 7.0%
0.86
0.224
{A, B} = = 26.0%
0.86
0.144
{B, C, D} = = 16.8%
0.86
0.096
{A, B, C, D} = = 11.2%.
0.86
The probabilities for {A} and {B} are computed as follows:
{A} = [{A}, 1 − ({B, C, D} + . . . + {B} + {D})] = [0, 0.372], i.e., 0–37.2%.
{B} = [{B}, 1 − ({A, C, D} + . . . + {D})] = [0.39, 0.93], i.e., 39–93.0%. 
From the above exercise, we note that the uncertainty interval (probability of
{A, B, C, D}) has decreased after addition of new evidence E 3 . In addition, the
probabilities of A and B have turned out to be more precise, i.e., narrow.
However, we note that the computation required in DST is combinatorial, since
with n diagnoses a total of 2n number of sets need to be computed [7].

12.6 Fuzzy Sets, Fuzzy Logic, and Fuzzy Inferences

In classical set theory, an element x of the universe U either absolutely belongs to

a set A ⊆ U or does not belong to it at all. This membership relation between the
element x and set A is called crisp, i.e., either Yes/No (or On/Off). A fuzzy set is a
generalization of a classical set by allowing the degree of membership, which is a
real number [0, 1]. In extreme cases the degree is 0, i.e., the element does not belong
to the set or 1, or the element fully belongs to the set.
To understand the idea of a fuzzy set, let us assume that people in an organization
are the universe and consider a set of “young” people in this universe. The youngness
is definitely not a step function from 1 to 0, as one attains a certain age, say 30 years.
In fact, it would be natural if we associate a degree of youngness for each element of
362 12 Reasoning in Uncertain Environments

age, for example, ({Anand/1, Babita/0.8, Choudhar y/0.3}). That means, perhaps
Anand is 23 years old, Babita is 27 years old, and Choudhary is in his 50s. The
membership function of a set maps each element of the universe to some degree of
association with the set.
The fuzzy set concept is somewhat similar to the concept of the set in the classical
set theory. When we say, “Sun is a star”, it means the Sun belongs to the set of stars,
and when we say, “Moon is a satellite”, it means the Moon belongs to the set of
satellites of the Earth. At the same time, both these statements are true, hence, they
both map to true or 1. The statement “the Sun is a black hole” is false as it does
not belong to the set of blacks holes yet. Hence, the statement maps to false or 0.
Thus, mapping of a statement in classical logic is either to 1 and 0, having crisp
values. Since, with each statement in the classical set theory having logical value 0
or 1, the set of these statements is called a crisp set, and also, these sets have a logic
associated, (T /F) due to the membership of their elements. Thus, sets and logic are
two sides of the same coin, they go together. Various properties of fuzzy sets, like
relations, logic, and inferences, have counterparts in the classical sets [1].
To propose a formal representation, a fuzzy set A is represented as A = { a(u)u
|u∈
U }, where u is an element, and a(u) the membership function, called characteristic
function, and represents the degree of belongingness; U is universe set. The notation
u ∈ U means “every u ∈ U ”.
The Fuzzy operations have counterparts operations in classical set theory as union,
intersection, complement, binary relation, and composition of relations.
The common definitions of fuzzy set operations are as follows.
 u 
A∪B = |u∈U (12.33)
max(a(u), b(u))
 u 
A∩B = |u∈U (12.34)
min(a(u), b(u))

and complement of A is expressed as

 u 
A = |u∈U . (12.35)
1 − a(u)

It is possible to derive fuzzy versions of familiar properties of ordinary sets, such

as commutative laws, associative laws, and De Morgan’s laws, based on the above
definitions [6].
The definition of fuzzy logic has semantic issues also; as per that fuzzy logic
it has two different senses: 1. It is a logical system, which is aimed to formalize
the approximate reasoning, 2. It is rooted in a multivalued logic, but its objective is
quite different from traditional multivalued logical system. It accounts for a concept
of linguistic variables (large, small, big, old, cloudy, etc.), canonical form, fuzzy
if-then rule, fuzzy quantifiers, etc. In a broad sense, fuzzy logic is governed by the
fuzzy set theory.
12.6 Fuzzy Sets, Fuzzy Logic, and Fuzzy Inferences 363

µ µ
1 1

0 0
20 40 25
Age Age
Linguistic variable ‘young’ Numerical variable ‘young’

Fig. 12.14 Linguistic and numerical variable “young”

The fuzzy arithmetic, fuzzy mathematical programming, fuzzy topology, fuzzy

graph theory, and fuzzy data analysis are other branches of fuzzy set theory. It is
indeed quite likely that most theories will be fuzzified in this way. The impetus
for the transition from a crisp set theory to fuzzy theory results from the fact that
both—the applicability to real-world problems and generality of the theory—are
substantially enhanced when the concepts of a set are replaced by a fuzzy set.
The concept of a linguistic variable plays a central role in the applications of fuzzy
logic. Consider the linguistic variable, say Age whose values are young, youth, and
old, with “young” defined by a membership function μ as shown in Fig. 12.14. It is
interesting to note that a numerical value such as 25 years is definitely simpler than a
function like young. The value young represents a choice out of three values—young,
youth, old—but 25 represents a choice, which is out of 100 values. Consequently,
the linguistic variable young may be viewed as a method of data compression.
We achieve the same effect in quantization (conversion of an integer into a binary
number), where the values are in intervals. In contrast to quantization, the values
are overlapping in fuzzy sets. The advantage of granulation over quantizations are 1.
granulation is more general, 2. it mimics the way humans interpret linguistic values
(i.e., not as intervals), and 3. a transition from one linguistic value to the next higher
and lower is gradual rather than abrupt. This results in a continuity and robustness
of the system.

12.6.1 Fuzzy Composition Relation

Let us assume that A and B are fuzzy sets of universes U and V , respectively. The
Cartesian product U × V is defined just like for ordinary sets. We define the Cartesian
product A × B as follows,
 (u, v) 
| u ∈ U, v ∈ V . (12.36)
min(a(u), b(v))
364 12 Reasoning in Uncertain Environments

Every fuzzy binary relation (or a mapping relation or fuzzy relation), from fuzzy
set U to fuzzy set V , is a fuzzy subset of relation R defined below. Here m(u, v) is
called a membership function, having the range [0, 1].
 (u, v) 
R :U ×V = | u ∈ U, v ∈ V . (12.37)
m(u, v)

A fuzzy relation R from set A to B can be defined as

 (u, v) 
R AB = | m(u, v) ≤ a(u), m(u, v) ≤ b(u), u ∈ U, v ∈ V . (12.38)
m(u, v)

Consider that a similar relation R BC exists from fuzzy set B to C. As in classical

sets, a composition relation of relations R AB and R BC , i.e., R AB ◦ R BC is obtained
as
  (a, c) 
R AB ◦ R BC = max . (12.39)
min(m R (a, b), m S (b, c))

The fuzzy sets are extensions and more general forms of ordinary sets; accordingly,
the fuzzy logic is an extension of the ordinary logic. Just as there are correspondences
between ordinary sets and ordinary logic, there exists correspondence between fuzzy
set theory and fuzzy logic. For example, the set operations of union (OR), intersection
(AND), and complement (NOT) exists in classical set theory (classical logic) as well as
in fuzzy system. The degree of belongingness of an element in a fuzzy set corresponds
to the truth value of the proposition in fuzzy logic.
A fuzzy implication is viewed as describing a relation between two fuzzy sets.
Using fuzzy logic we can represent a fuzzy implication such as A → B, i.e., if “A
then B”, where A and B are fuzzy sets. For example, if A is fuzzy set young, and B
is fuzzy set small, then A → B may mean, if young then small. A simple definition
is A → B = A × B, where A × B is the Cartesian product of fuzzy sets A and B.

12.6.1.1 Inferencing in Fuzzy Logic

Let R be a fuzzy relation from set U to V (i.e., R : U → V ), X be a fuzzy subset of

U , and Y be a fuzzy subset of V . Given these, we can define a composition rule of
fuzzy inference as
Y =X◦R (12.40)

where Y is said to be induced by X and R. A fuzzy inference is based on fuzzy

implication and the compositional rule of inference. A fuzzy inference is defined as
follows:

Given:
Implication: “If A then B”
Premise: “X is true”,
12.6 Fuzzy Sets, Fuzzy Logic, and Fuzzy Inferences 365

Derive:
Conclusion Y .
To derive a fuzzy inference, we perform the following steps:

1. Given, “if A then B”, compute the fuzzy implication as a fuzzy relation, R =
A × B,
2. Induce Y by Y = X ◦ R.

12.6.2 Fuzzy Rules and Fuzzy Graphs

We have noted that linguistic variable in fuzzy logic behaves as a source of data
compression, as a linguistic variable may stand for a large number of values. The
fuzzy logic, in addition, provides fuzzy if-then rule or simply the fuzzy rule, and a
fuzzy-graph. The fuzzy rule and fuzzy graph bear the same relation to numerical
value dependencies that the linguistic variable has with numerical values [8].
Figure 12.15 shows a fuzzy graph f ∗ of a function dependence f : X → Y ,
where X ∈ U , Y ∈ V , are linguistic variables. Let the set of values in linguistic
variables X and Y be Ai , Bi , respectively. This graph is an approximate compressed
representation of f in the form of f ∗ as

f ∗ = A1 × Bi + A2 × B2 + . . . + An × Bn , (12.41)

which is equal to

n
f∗ = Ai × Bi . (12.42)
i=1

Ai , Bi (for i = 1 . . . n) are called contiguous fuzzy subsets of sets U and V ,

respectively. Also, Ai × Bi is the Cartesian product of Ai , Bi , and “+” in Eq. (12.41)
represents the operation of disjunction—a union operation. When expressed more
explicitly in the form of membership functions, we have

Fig. 12.15 A function and f∗

corresponding fuzzy graph Y f
Bi

Ai × Bi
(Fuzzy point)

Ai X
366 12 Reasoning in Uncertain Environments

Table 12.5 Fuzzy relation f∗ A B

f∗
A1 B1
A2 B2
. .
An Bn

μ f ∗ (u, v) = Vi (μ Ai ∧ μ Bi ). (12.43)

where μ is the membership function. The operation of ∧ is min, ∨ is max, u ∈ U ,

and v ∈ V .
A fuzzy graph can also be represented as a fuzzy relation f ∗ , as shown in
Table 12.5.
Also, a fuzzy graph can be represented as a collection of if-then rules as follows:

f ∗ : i f X is A1 then Y is B1
i f X is A2 then Y is B2
...
i f X is An then Y is Bn . (12.44)

In other words, “(X, Y ) is Ai × Bi ”. Given a fuzzy if-then rule set as follows,

f ∗ : i f X is small then Y is large

i f X is medium then Y is medium
...
i f X is large then Y is small. (12.45)

This rule can be represented equivalently as a fuzzy graph (see Fig. 12.16) using the
fuzzy expression for f ∗ ,

f ∗ = small × large + medium × medium + · · · + large × small. (12.46)

An important concept in a fuzzy graph is any type of function (or relation) that can
be approximated by a fuzzy graph. For example, a normal distribution for probability

Fig. 12.16 Fuzzy graph for Y

a fuzzy relation (Eq. 12.46) f

f∗

0
X
12.6 Fuzzy Sets, Fuzzy Logic, and Fuzzy Inferences 367

Fig. 12.17 Fuzzy graph for P

a probability distribution

0
X

is represented as a fuzzy graph, as shown in Fig. 12.17. This representation is useful

in decision-making and fault diagnosis. Here, the probability P is represented by

P = small × medium + medium × large + large × small. (12.47)

The constraint for probability is that it sums to 1, i.e., Σi Pi = 1.

12.6.3 Fuzzy Graph Operations

The interesting point in fuzzy calculus is that it aims to develop computational pro-
cedures for basic operations on fuzzy graphs. These operations are generalizations
of basic operations on crisp sets and functions, and can be defined as follows: The
required computations can be simplified if the operation of “∗” is taken as mono-
tonically increasing. That is, if a, b, a  , and b are real numbers, then the following
holds:

a  ≥ a, b ≥ b ⇒ a  ∗ b ≥ a ∗ b
a  ≤ a, b ≤ b ⇒ a  ∗ b ≤ a ∗ b. (12.48)

From these operations, it can be easily deduced that “∗” is a distributive operator
over “∨ (max)” and “∧ (min)”. Accordingly,

a ∗ (b ∨ c) = a ∗ b ∨ a ∗ c
a ∗ (b ∧ c) = a ∗ b ∧ a ∗ c. (12.49)

From the above, we conclude that if


f∗ = Ai × Bi (12.50)
i

is a fuzzy graph and C is a fuzzy set, then

  
C∗ Ai × Bi = C ∗ (Ai × Bi ). (12.51)
i i
368 12 Reasoning in Uncertain Environments

Assignment Intersection
Y f Y f

f (u) g
0
X X

Composition Combination
f
Y Y f (u)
g◦f g(u)
g(f (u)) f (u) ∗ g(u)

X X

g∗
f∗
V Interection of fuzzy-
graphs g ∗ and f ∗

Fig. 12.18 Basic operations on functions and relations of fuzzy sets

Consider finding the intersection of fuzzy graphs (Fig. 12.18) f ∗ and g ∗ , where

f∗ = Ai × Bi (12.52)
i

and 
g∗ = Cj × Dj. (12.53)
j

From f ∗ and g ∗ , we obtain the fuzzy intersection as


f ∗ ∩ g∗ = (Ai × Bi ) ∩ (C j × D j ) (12.54)
i, j

and, the view of distributivity of “∩” ultimately reduces to


f ∗ ∩ g∗ = (Ai ∩ C j ) × (Bi ∩ D j ). (12.55)
i, j

These operations on fuzzy functions and fuzzy relations are shown in Fig. 12.18.
12.6 Fuzzy Sets, Fuzzy Logic, and Fuzzy Inferences 369

12.6.4 Fuzzy Hybrid Systems

Different forms of hybrid systems are designed as combinations of Fuzzy Logic,

Neural Networks, and Genetic Algorithms. The fundamental concepts of such hybrid
systems is to complement each other’s weaknesses, that creates a new approach to
solving problems. For example, in fuzzy logic systems, there is no learning capability,
there is no memory, and there is no capability of pattern recognition. Fuzzy systems
combined with neural networks will have all these three capabilities.
Fuzzy systems are extensively used for control, like fuzzy controllers, specially
fuzzy PID (proportional, integral and derivative) controllers.
In spite of their wider applicability, the fuzzy systems have the following lim-
itations: The stability is a major issue in their control. There is no formal theory
guaranteeing their stability. Fuzzy systems have no memory and lack the learning
capabilities. Determining good membership functions and tuning fuzzy systems are
not always easy. In addition, verification and validation of fuzzy systems require
extensive testing. In light of these limitations, the fuzzy systems are being used more
and more, and they become better suited when combined with the neural networks
and genetic algorithms.

12.7 Summary

Probability and Bayes theorem

In classical logic, a proposition is always taken as either true or false, while in real
life, truth value of a proposition may be known only with a certain probability. This
requires probabilistic reasoning. Two basic approaches for reasoning are probabilistic
reasoning and non-monotonic reasoning.
Many different methods exist for representing and reasoning with uncertain
knowledge, these include Bayesian networks, Dempster–Shafer theory, possibilistic
logic, and fuzzy Logic.
Bayesian networks offer a powerful framework for the modeling of uncertainties.
These networks are represented in two different ways: (1) Qualitative approach, by
means of directed acyclic graphs, and (2) Quantitative approach, by specifying a
conditional probability distribution for every variable present in the network. The
conditional probability is represented by

P(A, B)
P(A|B) = (12.56)
P(B)

where comma denotes the conjunction of events.

The Bayes theorem is used to find out conditional probability. The theorem is
stated as
P(B|A)P(A)
P(A|B) = . (12.57)
P(B)
370 12 Reasoning in Uncertain Environments

In a Bayesian network, a node corresponds to a variable, which assumes values

from a collection of mutually exclusive and collective exhaustive states, and the edges
represent relevance or influences between variables. As per the definition of Bayesian
networks, it is required to access the probability distribution of the node conditional
on every instance of its parents. Thus, if a node has n-binary-valued parents, there
are 2n probability distributions for that node. A probabilistic inference is nothing
but computing probabilities of interest from a (possibly implicit) joint probability
distribution. All exact algorithms for probabilistic inference in Bayesian networks
exploit the criteria of conditional independence.
Solving the problem of finding the exact inference in an arbitrary Bayesian net-
work is NP-hard.
The inferencing in a network amounts to propagation of probabilities of a given
and related information through the network to one ore more conclusion nodes.
The subjective approach for the construction of Bayesian networks reflects on the
available knowledge (typically, perceptions about causal influences). Other method
for Bayesian networks is based on automatically synthesizing these networks for the
purpose of reliability and diagnosis. This approach is used to synthesize a Bayesian
network automatically with the help of formal knowledge of system design.
The third method for constructing Bayesian networks is based on learning from
data, such as medical records, customers’ purchases data, or data records of customers
applied for finances. These data sets are used to learn the network parameters, given
their structures, or can learn both the structure and network parameters.
Dempster–Shafer Theory
The Dempster–Shafer Theory (DST) of evidence is used to model several single
pieces of evidence within a single hypothesis relations, or a single piece of evidence
within multi-hypotheses relations.
The combination rule of DST aggregates two independent bodies of evidence
into one body of evidence, provided that they are defined within the same frame of
discernment.
The DST is based on two ideas: 1. to obtain the degree of belief for one question
using the subjective probabilities of a related question, and 2. making use of Demp-
ster’s rule to combine degrees of two belief which are based on independent items
of evidence.
In DST, the difference (100 − X )% with respect to a probability of X % for a
diagnosis is not valued against the diagnosis, but interpreted as an uncertainty interval,
i.e., the probability of a diagnosis is not X %, but lies between X and 100%.
Evaluation of probabilities in DST is more complicated than in Bayes theorem,
because the probabilities for a set of diagnoses must be related to one another, and the
probability of a particular set of diagnoses must be calculated from the distribution
of probabilities over all sets of diagnoses.
Fuzzy logic, sets, and reasoning
A fuzzy set is a generalization of ordinary set by allowing the degree of membership a
real number [0, 1], in contrast to the membership of 0 and 1 in ordinary set theory. The
12.7 Summary 371

fuzzy operations that have counterparts in classical set theory are union, intersection,
complement, binary relation, and composition of relations.
Other branches of fuzzy set theory are fuzzy mathematical programming, fuzzy
arithmetic, fuzzy topology, fuzzy graph theory, and fuzzy data analysis.
In fuzzy logic systems, there is no learning capability, no memory, and no capabil-
ity of pattern recognition. Fuzzy systems combined with neural networks will have
all these three capabilities.
Different forms of hybrid systems are constructed using a combination of fuzzy
logic and other areas. These are Fuzzy-neuro systems, Fuzzy-GA (fuzzy and genetic
algorithms), and Fuzzy-neuro-GA. They help to complement each other’s weak-
nesses, and create new systems that are more robust to solving problems.

Exercises

1. We assume a domain of 5-card poker hands out of a deck of 52 cards. Answer

the following under the assumption that it is a fair deal.
a. How many 5-card hands can be there (i.e., number of atomic events in joint
probability distribution).
b. What is the probability of an atomic event?
c. What is the probability that a hand will comprise four cards of the same
rank?
2. Either prove it is true or give a counterexample in each of the following state-
ments.

a. If P(a | b, c) = P(a), then show that P(b | c) = P(b).

b. If P(a | b, c) = P(b | a, c), then show that P(a | c) = P(b | c).
c. If P(a | b) = P(a), then show that P(a | b, c) = P(a | c).

3. Consider that for a coin, the probability that when tossed the head appears up is
x and for tails it is 1 − x. Answer the following:

a. Given that the value of x is unknown, are the outcomes of successive flips
of this coin independent of each other? Justify your answer in the case of
head and tail.
b. Given that we know the value of x, are the outcomes of successive flips of
the coin independent of each other? Justify.

4. Show that following statements of conditional independence,

P(X | Z )P(Y |Z ) = P(X, Y | Z )

372 12 Reasoning in Uncertain Environments

are also equivalent to each of the following statements,

P(Y | Z ) = P(Y | X, Z ),

and
P(X | Z ) = P(X | Y, Z ).

5. Out of two new nuclear power stations, one of them will give an alarm when
the temperature gauge sensing the core temperature exceeds a given threshold.
Let the variables be Boolean types: A = alarm sounds, F A =alarm be faulty, and
F G = gauge be faulty. The multivalued nodes are G = reading of gauge, and T
= actual core temperature.
a. Given that the gauge is more likely to fail when the core temperature goes
too high, draw a Bayesian network for this domain.
b. Assume that there are just two possible temperatures: actual and measured,
normal and high. Let the probability that the gauge gives the correct reading
be x when it is working, and y when it is faulty. Find out the conditional
probability associated with G.
c. Assume that the alarm and gauge are correctly working, and the alarm
sounds. Find out an expression for the probability in terms of the various
conditional probabilities in the network, that the temperature of the core is
too high.
d. Let the alarm work correctly unless it is faulty. In case of being faulty, it
never sounds. Give the conditional probability table associated with A.
6. Compute the graphical representation of the fuzzy membership in the graph 12.19
for following set operations:
a. Small ∩ T all
b. (Small ∪ Medium)–T all

Fig. 12.19 Membership

functions: Small, Medium, Small
Tall

Medium

Tall
Exercises 373

7. Given the fuzzy sets A = { 0.5

a
, 0.9
b e
, 1 }, B = { 0.7
b
, 0.9
c
, 0.1
d
}, Compute A ∪ B, A ∩
 
B, A , B .
8. Let X = {a, b, c, d, e}, A = { 0.5
a
, 0.3
c
, 1e }. Compute:

a. Ā
b. Ā ∩ A
c. Ā ∪ A

References

1. Chowdhary KR (2015) Fundamentals of discrete mathematical structures, 3rd edn. EEE, PHI
India
2. Darwiche A (2010) Bayesian networks. Commun ACM 53(12):80–90
3. http://web.mit.edu/jmn/www/6.034/d-separation.pdf. Cited 19 Dec 2017
4. Heckerman D, Wellman MP (1995) Bayesian networks. Commun ACM 38(3):27–30
5. Kay RU (2007) Fundamentals of the dempster-shafer theory and its applications to system safety
and reliability modelling. J Pol Saf Reliab Asso 2:283–295
6. Munakata T, Jani Y (1994) Fuzzy systems: an overview. Commun ACM 37(3):69–77
7. Puppe F (1993) Systematic introduction to expert systems. Springer
8. Zadeh LA (1994) Fuzzy logic, neural networks, soft computing. Commun ACM 37(3):77–84
Chapter 13
Machine Learning

Abstract To adapt to the environment it necessary that intelligent machines must

have the capability to learn. This chapter presents the basic concepts and techniques
of learning found in humans, as well as their implementation aspects for machines.
The chapter presents the challenges of building learning capabilities in machines,
types of machine learning, and the relative efforts needed to build these learning
capabilities. The philosophy of the discipline of machine learning is presented. The
basic model of learning is discussed, followed by the classes of learning—supervised
and unsupervised—then various techniques of inductive learning—argument based
learning, online concept learning, propositional and relational learning, and learn-
ing through decision trees—are presented in sufficient details. Other techniques
like discovery-based learning, reinforced learning, learning and reasoning through
analogy, explanation-based learning are presented, with some worked examples.
Finally, the potential applications of machine learning, the basic research problems in
machine learning, followed by chapter summary, and a set of exercises are appended.

Keywords Machine learning · Machine learning classes · Machine learning

model · Supervised learning · Unsupervised learning · Inductive learning ·
Decision-tree-based learning · Argument-based learning · Propositional learning ·
Discovery-based learning · Learning through analogy · Reinforced learning ·
Explanation-based learning · Machine learning applications · Research problems

13.1 Introduction

Learning is one of the most important activities of human beings and living beings in
general, which help us in adapting to the environment. Humans learn from nature as
well as from special learning environments through different techniques that vary in
complexity. The learning requires transformations of ideas and information structures
in the human mind.
This chapter presents the basic concepts and examples of learning techniques
found in humans, and their implementations aspects for learning in machines. A
common view holds that learning involves making changes in the learning system

© Springer Nature India Private Limited 2020 375

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_13
376 13 Machine Learning

that will improve it in some way. These changes are adaptive, because it is found
that the system does some task or tasks more effectively the next time when they are
drawn from the same population of tasks.
Most of the knowledge of the world we have around us, and about the explicit
set of tasks, is not formalized, and even not available in text form. The latter is a
necessary requirement for it to be understood by any computer program. This is the
reason it is not easy to write a program for a computer to do many tasks that we
humans do so easily, such as understanding spoken sentences, images, languages, or
driving a car. Any attempt to achieve this, i.e., organizing sets of facts in elaborate
data structures to enable computers to understand it, has achieved very little success.
The concept of learning is based on the principles of training the computing
machines, and enabling them to teach themselves. Tenet of these machines/programs
is related to what we consider as good decision. For humans and animals, evolutionary
principles dictate that any decision made by them should lead to such behaviors that
optimize the chances of survival and reproduction. In the societies of human beings,
a good decision is one that might include social interactions that bring status, or a
sense of well-being. For a machine such as a self-driving car, such criteria cannot
exist! So the quality of decision-making for a self-driving car is measured based on
the criteria as to how close the autonomous vehicle imitates the behavior of trained
human drivers.
It is important to note that the knowledge required to make a good decision in a
particular situation is not necessarily clear to the extent to code it into a computer
program. Consider, for example, that a mouse has knowledge about its surroundings,
and has an innate sense of where to sniff and how to find food or mates, and at the
same time save itself from predators. It is not easy for any programmer to specify
step-by-step procedures or a set of instructions to produce such behaviors for some
artificial mouse. However, this all remains encoded in the brain of a mouse in the
form of knowledge.
Before we think about creating computers (programs) that can train themselves,
it is important to answer the fundamental question as to how we humans acquire
knowledge. Partly, the knowledge in humans may be innate, but most of it is
acquired through experience and observation. What we know intuitively has been
often learned from examples and practice—the process which cannot be turned into
a clear sequence of steps as a computer program. Since 1950 people have looked for
ways and means and tried to refine general principles that allow animals or humans,
to acquire knowledge through experience. These are the subjects of discussion in this
chapter. Machine learning aims to create theories and procedures—learning algo-
rithms—that allow machines to learn. One such is learning from examples presented
to them, called learning by examples [2].

Learning Outcomes of This Chapter:

1. List the differences among the three main styles of learning: supervised, rein-
forcement, and unsupervised. [Familiarity]
13.1 Introduction 377

2. Identify examples of classification tasks, including the available input features

and output to be predicted. [Familiarity]
3. Explain the difference between inductive and deductive learning. [Familiarity]
4. Implement simple algorithms for supervised learning, reinforcement learning,
and unsupervised learning. [Usage]
5. Determine which of the three learning styles is appropriate to a particular problem
domain. [Usage]
6. Evaluate the performance of a simple learning system on a real-world data set.
[Assessment]
7. Characterize the state of the art in learning theory, including its achievements and
its shortcomings. [Familiarity]

13.2 Types of Machine Learning

The field of machine learning is concerned with the development of computational

theories for learning processes and building learning machines. Because the ability to
learn is a fundamental to any intelligent behavior, the concerns and goals of machine
learning are central to the discipline of Artificial Intelligence. In the learning process,
the learner transforms the information provided by a teacher (or environment) into
some new form, and these are stored in that form for use in the future. The nature of
this knowledge and transformation are the deciding factors for the type of learning
strategy used. Following are the fundamental strategies of learning.

Rote Learning
Learning by Instruction
Learning by Deduction
Learning by Analogy
Learning by Induction (or Similarity)
Reinforcement Learning
Discovery-based Learning

These strategies are in increasing order of complexity of transformation required

in the initial information, which is stored as knowledge base. The complexity orders
of these methods are such that, an increasing difficulty level on the part of the student
(learner) results in a correspondingly decreasing complexity on the part of the teacher
(environment), and vice versa. Following is the brief introduction about each of these
methods.
Rote Learning
The rote learning is the simplest among all the learning, and requires the least amount
of inferencing from the knowledge. To accomplish the inferencing later, the original
knowledge is copied in the same form in which it will be used later. Hence, the
information from the teacher is more or less directly accepted and memorized by the
378 13 Machine Learning

learner. This kind of learning is used when we learn the tables in the early school
classes. A major concern in this learning is how to index the stored knowledge for
future retrieval.
Learning by Instruction
The learning which is next higher in efforts after rote learning is learning by instruc-
tion. This approach requires the knowledge to be transformed into an operational
form before it can be integrated into the knowledge base. This learning takes place
among the students in a class when the class teacher presents a number of facts
directly in a properly organized form to the students.
In learning by instruction (or learning by being told), the basic transformations
performed by a learner are selection and reformation (mainly at a syntactic level) of
information provided by the teacher.
Note that, depending on how the teachers presents the knowledge in the class,
different types of learning take place, and learning by instructions is not the only
learning induced by a teacher into his/her students.
Learning by Deduction
Deductive learning is carried out through a sequence of deductive inference steps
using known facts. Using this, new facts and knowledge are logically derived. For
example, if we have the knowledge of a rule, say “a father’s father is a grand father,”
and we have the case that a is the father of b and b is the father of c, then we can
say that a is the grand father of c. Here we have deduced the result, hence the name
deductive learning.
Learning by Analogy
Analogical learning provides learning of new concepts through the use of similar
known concepts or solutions. This is used, for example, when we solve some problems
in our college examination, based on the solution of somewhat similar solutions we
have already done at home. Also, for example, learning to drive a bike when we
already know how to drive a bicycle is an example of analogical learning.
Learning by Induction
We make use of an inductive learning when after seeing a number of instances or
examples of a concept we formulate a general concept.
In any human learning activity, mostly the strategies out of those discussed above
are involved. It is important to understand the difference between these strategies
on the part of designing a learning system. Though most current artificial systems
focus on any single learning strategy, one may expect that machine learning research
will give increasing attention to a multi-strategy approach, the one which is close
to the human learning system. We as a learner use almost always a combination of
strategies in our daily life. For example, remembering a telephone number, we use
rote learning; seeing someone wearing a new dress every day, we learn that so-and-so
is rich through inductive learning. Learning that those students who complete their
13.2 Types of Machine Learning 379

homework in time get more Cumulative Percentile Index (CPI) is also an example
of inductive learning.
Another example, when we learn Java language having learned C++ is analogical
learning. When we are asked to solve a new puzzle, it is discovery-based learning.
Whenever, a teacher appreciates the student’s efforts by saying “very good”, for
asking an interesting question, or answering a question, it is reinforcement-based
learning for the student.
When you have concluded based on some clues that a teacher is not in a mood of
teaching, most likely you have used deductive learning (or reasoning).

13.3 Discipline of Machine Learning

Any field of scientific study is best defined by the central question it asks. That way,
the field of Machine Learning seeks the answer to the following questions:
1. How to build a computer system that can automatically improve its performance with
experience, and 2. What are the fundamental laws that govern the learning processes?

The faculty of Learning, like the faculty of Intelligence, involves such a vast
range of processes that it is difficult to define it precisely. A dictionary definition of
learning includes phrases such as “to gain knowledge, or understanding of, or skill in,
by study, instruction, or experience.” Another dictionary definition says, learning is
“modification of a behavioral tendency by experience.” Zoologists and psychologists
also study the learning processes in animals and humans.
This chapter focuses on learning in machines. There are several similarities
between learning in animals and in machines. Many techniques in machine learning
are derived from the theories of animal and human learning and are presented in
the form of computational/cognitive models by psychologists. It is quite likely that
the concepts and techniques being explored by researchers in machine learning may
illuminate certain aspects of biological learning.
Regarding the machines, it is roughly said that a machine learns whenever it
changes in its structure or program or data, based its input or in response to external
information. This change takes place in such a manner that its expected future per-
formance improves. A change like addition of a record in a database, though makes a
change in data structure, does not fall in the discipline of learning. As another exam-
ple, the performance of a speech-recognition machine improves after hearing several
samples of a person’s speech—we feel convinced that the machine has learned. This
process is obviously a learning process.
Machine learning usually refers to the changes in systems that perform tasks such
as recognition, diagnosis, planning, robot control, prediction, and their combinations.
These tasks are usually associated with artificial intelligence (AI). The “changes”
mentioned above is in one of the two types: 1. enhancements in already performed
task by a system, or 2. creating special ability in the system.
380 13 Machine Learning

One obvious question related to machine learning is “Why should machines be

built to learn, and why not to build them to perform the desired task in the first
place?” There are several reasons why machine learning is important. The one we
have already mentioned is that understanding the learning process will help us in
building theories which will help us understand how the animals and humans learn.
It is like when we learned optics, we were able to better understand human vision
system. However, there are the following important reasons why we should study
machine learning, and why there should be machines that learn:

(a) There are many tasks that cannot be defined well in advance, except by examples.
Like many times we are able to specify input/output relations as pairs of data but
a precise relationship between inputs and desired outputs cannot be specified.
For example, the weight-to-height ratio of a male as input and health as output.
We would like that based on these input–output sample examples, the machine
be able to adjust its internal structure, so that for new inputs it produces correct
outputs, and thus suitably constrain its input/output function to approximate the
relationship implicit in the given examples.

(b) It is possible that among large sets of data, like—credit cards, medical records,
weather data, astronomical data, etc., there are hidden important relationships
and correlations. The machine learning methods can often be used to extract
these relationships and patterns through the process of data mining, a field of
machine learning.

(c) The machines produced by the human designers often do not work as desired in
certain environments. In fact, when the machine was designed, certain charac-
teristics of the working environment might not be completely known at design
time, hence could not be incorporated in the design. This often happens in space
explorations—like missions to Moon, Mars, Saturn, and even to Sun. Machine
learning methods can improve/modify the performance of the system on the site
in the existing machine design.

(d) The amount of knowledge available and required for certain tasks may be too
large for explicit encoding by humans. But machines can be designed that learn
this knowledge gradually with time, and might be able to capture it continuously
on the fly as it needs.

(e) There is no need to redesign the machines if there are machines that can adapt to a
changing environment. Such machines, once designed, can learn with situations,
and functions for a long time—for example in long planetary missions.

With the continuous redesign of AI, the changes are taking place in the world
knowledge as well as vocabulary.
The questions in the domain of AI cover a broad range of learning tasks, like:
How to design an autonomous mobile robot that learns to navigate using its own
13.3 Discipline of Machine Learning 381

experience? How to mine the data of historical medical records to learn that a given
patient X will respond best to what treatments? How to build search engines that
automatically customize to the user’s interests and other similar questions.
The following is a more precise definition of machine learning.

Definition 13.1 (Machine Learning) A machine learns with respect to a particular

task T , performance metric P, and type of experience E, if it reliably improves its
performance P at task T , following the experience E.

Depending on how we specify T , P, and E, the learning task might also be called
by names such as data mining, autonomous discovery, and database updating, pro-
gramming by example.
The field of machine learning focuses on how to make computers that program
themselves through their experience, with provision of some initial structures. This is
in contrast to the field of computer science that so far focused primarily on manually
programming computers.
Other fields that are closely related to Machine Learning are Psychology and
Neuroscience, which are concerned with the study of human and animal learning,
and are collectively called Cognitive Science. The questions—How do animals learn,
and how can computers be made to learn—most probably have highly intertwined
answers. The insights into machine learning far outweigh due to contributions from
the fields of statistics and computer science than due to studies in human learning
behavior. The main reason behind this is the poor state of understanding about human
learning processes. However, the relation between machine learning and human
learning behavior is continuously becoming stronger, mainly due to the new machine
learning algorithms, such as temporal difference learning. These algorithms are being
suggested as the explanations for neural signals observed in learning in the animals.
The primarily goal of machine learning is to design computers that can adapt and
learn from their experience without intervention from programmers. Following is
an example: the present software engineering, data-intensive software construction
and testing, is aimed to construct programs that are correct and robust. It is data-
intensive software construction, i.e., given a set of input data to be processed, code
functions that produce output (data), such that the code should correctly function
even when input data changes, and even on the face of incorrect data, should reject
them rather than producing wrong results. A machine can be designed to learn from
the pattern of data and continuously improve its performance until it reaches the
desired performance. The error (difference) between the expected performance and
the actual performance can be fed back to act as a tuning to the software, that is,
to make it learn. This latter is the automation of human-intensive task of software
engineering that we call testing.
Machine learning methods for software engineering are appropriate whenever
hand engineering of software is difficult and yet data is available to be analyzed by
learning algorithms. Following are the situations where machine learning algorithms
are expected to outperform humans:
382 13 Machine Learning

(i) human expertise is lacking compared to the machine intelligence in genome

programs—the genome programs are basically evolution process, which can be
used in applications, like drug design;
(ii) the characteristics of the task changes frequently, like in airline seat sales strate-
gies, where the price is decided by supply-versus-demand ratio;
(iii) humans have limited capability of introspection, like required in computer vision,
speech recognition, and motor control;
(iv) a learning program can be customized for individual users, e.g., information
filtering, user interfaces, language, etc.

However, there is yet one limitation of machine learning—the field of machine

learning is inherently stochastic—hence its application may not be appropriate for
tasks where completely error-free performance can be guaranteed, e.g., in life-support
control in intensive care.

13.4 Learning Model

The process of learning requires that new knowledge structures be created due to
some form of input stimuli. Then, this new knowledge is assimilated in the existing
knowledge and is tested for its utility and efficiency. That is, it will be used in per-
forming some task. Figure 13.1 shows a general learning model. The environment is
part of the overall system for learning. The environment may be taken as a teacher or
the one provided by nature which produces random stimuli, or a room with obstacles
where a robot needs to navigate from one wall to another, or may be a classroom
with a teacher and students who are taught using a blackboard and chalk. The com-
munication language between a teacher and a learner may be the same in which a
knowledge base was created or may be different.
The input to the learner component may be physical stimuli or some kind
of description or is symbolic. The information by the learner component cre-
ates/modifies the knowledge structures in the knowledge base, which is used by
the performance component to carry out some tasks, like solving some problems,
game playing, etc. The response produced by the performance component is evalu-

Fig. 13.1 A learning model

13.4 Learning Model 383

ated by the performance evaluator component, and a feedback generated by this is

sent to the learner to decide if the performance is acceptable, based on which the
learner updates its knowledge base. This cycle is repeated until the performance of
the system is reached to the desired level.
In the sections, we present specific learning techniques in more detail. We present
first the fundamental categories of learning; a learning may be basically classified as
belonging to supervised or unsupervised learning.

13.5 Classes of Learning

When there is essentially a teacher present for learning to take place, it is supervised
learning, and a learning achievable without a teacher is called unsupervised learning.
Teacher, here, means not necessarily a physical teacher, but even nature is a teacher.
For example, we put fingers in hot water and we get it burnt. So here, feedback is
from hot water, which teaches us that if you again put the fingers in hot water any
time in future, it would be burnt in similar way.

13.5.1 Supervised Learning

Supervised learning methods require an oracle to be made, such that it forecasts the
classes (categories) to which the given examples would belong and then classifies
the examples in those categories. These methods assume

(a) Existence of some teacher (environment),

(b) A fitness function to measure the fitness of an example for a class, and
(c) Some external method of classifying the training instances.

A supervised learning method uses machine learning techniques to induce a clas-

sifier from sense-annotated data sets. As an example of supervised learning for text
classification, a classifier picks one word at a time, perform its word sense dis-
ambiguation (WSD), and then performs a classification task in order to assign the
appropriate sense to each instance of the word.
A classifier typically learns with the help of a training set containing examples
in which any given target word has been manually tagged with the sense form the
sense inventory of some reference dictionary.
In general, a learning method is called supervised learning if the learning process
requires some intervention from the user. Some approaches in supervised learning
require the user to provide training examples. This is done as a pre-process to appro-
priately tag the examples occurring in the training set. Alternatively, it can be done
as an online process where examples are dynamically tagged as needed during the
learning process.
Propositional learning is a category of supervised learning in which examples of
concepts are represented in terms of either zero order logic or attribute-value logic.
384 13 Machine Learning

This representation has equivalent expressiveness in a strict mathematical sense.

The rules are often learned from positive examples. The examples of concepts are
represented as sets of attributes in the form of slots in knowledge frames, whose values
are heads of syntactic phrases occurring within the training documents. The rules
learned through these approaches are used in performing the tasks, like Information
Extraction (IE) from parsed free text, text classification, question-answering, etc. The
learning process identifies the syntactic phrases that contain the heads to fill these
slots. Often, there is a partial match between these phrases and the exact slot filler,
which is sufficient when the aim is to extract only an approximate of the relevant
information. If an approximate match fails, there is necessity of post-processing.
Other class of supervised learning is relational learning, which is based on repre-
senting the examples of concepts in terms of First-Order Predicate Logic (FOPL)—as
attributes and relations between textual elements. More details of propositional and
relational learning are covered in sections that follow.

13.5.2 Unsupervised Learning

This learning eliminates the need of a teacher, and the learner is solely responsible
to form his own concepts and evaluate these for learning. In fact, the unsupervised
methods have to discover the concept classes to which the given examples belong,
i.e., mapping the examples to concept classes.
For example, scientists are usually not blessed to have a teacher to help them pursue
the research, instead they propose hypothesis to explain the observations made by
them, and evaluate their hypotheses based on criteria like generality, simplicity, and
elegance, and test these hypotheses through experiments designed by themselves.
Another example of unsupervised learning is unsupervised WSD, which is based
on unlabeled corpora, and do not exploit any manually sense-tagged corpus to provide
a sense choice for a word in the context.
The discovery-based learning is also a category of unsupervised learning.

13.6 Inductive Learning

The ability of human beings to master meeting the complex demands is mainly
based on the human ability to exploit previous experiences, like if we have previous
experience of walking through the Thar desert unaided for five miles from 10 years
ago, we will not hesitate to repeat the next adventure now even for still more distance.
Based on our previous experiences, we are able to predict the characteristics or
relations of man-made or natural objects, we can reason about possible outcomes of
actions, and we can apply the previous successful routines and strategies to new tasks
and problems, as in the case of journey through a desert example. It is understood that
the basic process to expand the knowledge is—first construct a hypothesis such that
13.6 Inductive Learning 385

we can transfer the knowledge from previous experience/patterns to new situations.

This approach of learning is called inductive inference [6].
Inductive inference is a mechanism of generalization over the observed regularities
in the examples. Every learning algorithm must make some a priori assumptions to
allow for the generalization, called inductive bias, which also provides a rational basis
to allow for the transfer of learned hypothesis to the new situation. For this transfer,
the language bias or restrictions characterizes the language in which the induced
hypothesis is represented. A bias for search or preference characterizes the method
of selecting the proper hypothesis. Every machine learning algorithm, be it symbolic,
or statistical, or neural, makes such a priori assumptions. This basic learning approach
is also valid for humans, whose learning is taken as a base mechanism, that takes
place at all levels, from concept acquisition to learning high-level schemata.
The acquired concepts and strategic rules need to be communicated, so that they
are available as declarative structures of knowledge. Hence, inductive learning pro-
gramming should be symbolic programming.
In the abstract, we can view the output of inductive learning as a set of Production
rules of the form,

In situation do action

where action may be something overt, or an internal action, or even an inference. We

use the concept sometimes for the right-hand side (i.e., in place of action), and the
word concept definition or pattern for the left-hand side (i.e., in place of situation), to
formalize the inductive learning from one or more instances in which actioni was an
appropriate action response to situationi . Based on these instances, we infer that the
general version actiongen is the appropriate type of action in response to the general
situation type situationgen .

situation1 action1
situation2 action2
… …
situationgen actiongen .

The above example shows that inductive learning is the generalization from a set
of examples; and with little pondering we note that this is one of the most fundamental
learning types among humans. Learning of concepts is a typical inductive learning
approach—given the examples of some concepts, such as “cat”—it will allow us to
correctly recognize future instances of this concept; similarly, learning the concept
that something is a “good stock investment”, it will allow us to correctly perform
instances of this type in the future, i.e., if an investment has been done intelligently,
and has been found rewarding, there are good chances that, in future also we may
adopt a similar strategy. Such instances are very common in life, hence, the most
common application of inductive learning is in concept learning.
386 13 Machine Learning

Definition 13.2 (Concept Learning) Concept learning is, given the number of pos-
itive and negative examples of some concepts with some background knowledge, to
find a general description or hypothesis of the concepts describing all the positive
examples only, and ignoring all the negative examples. 

Definition 13.3 (Deductive System) A logical system whose conclusions logically

follow from a set of input facts, and the system is sound, then the system is called a
deductive system. 

To contrast the deduction with an induction in logic-based systems, consider that

we have the training set with the following formulas:

Ball(Obj1 ), Round (Obj1 ), Ball(Obj2 ), Round (Obj2 ),

Ball(Obj3 ), Round (Obj3 ). (13.1)

Using these formulas if a learning system draws the conclusion (∀x)[Ball(x) →

Round (x)] then that system is an inductive learner. Note that this conclusion does
not logically follow from the facts. Hence, conclusion is useful when there are no
rules of the form [Ball(x), Ball(x) → Round (x)].
Unlike deduction, which is based on general axioms, induction is based on specific
facts (examples). Induction has two goals: 1. formulate acceptable general assertions,
that explain the given facts, and 2. to predict unseen facts. Thus, an inductive inference
is aimed to provide a complete and correct description about a given phenomenon
using specific/partial observations about it. Such inductively obtained description is
true for at least already seen examples, but there is no guarantee of its correctness
about the new examples.
Inductive inference methods can be used in two types of applications: 1. as an
interactive tool for acquiring knowledge using examples, or 2. the method is used
as part of some learning system. In the first case, a user has strong control through
the examples provided by him/her, which decides the type of knowledge going to
be acquired. Whereas in the second application, the inductive procedures (methods)
are activated when some system component wants to learn using positive/negative
examples. These examples constitute the feedback from which the system can achieve
the completion of the current task.
In any relevant domain, the background knowledge is a deciding factor for assump-
tions and constraints imposed on examples and descriptions generated. The back-
ground knowledge is represented using either declarative format or procedural
format. Concept learning requires searching through a large space of hypotheses
(descriptions), each as a sequence of instructions for executing a specific task, with
a goal to finding a hypothesis that best explains the examples. The instruction in the
hypothesis are implicitly defined using the hypothesis representation language.
Inductive Programming
Inductive learning (also called concept learning) can be further classified according
to the following perspectives, depending on whether it is based on
13.6 Inductive Learning 387

• Supervised learning,
• Unsupervised learning,
• Concept hierarchies, or
• Single/multi-label learning.

Inductive programming is concerned with learning computer programs, designed

to work with incomplete specifications—samples of desired input/output behavior,
along with some constraints. These programs are called induced programs, and are
represented in the declarative form using functional programming or logic program-
ming languages. To be used in machine learning, inductive programming creates
program hypotheses in the form of generalized recursive programs. In contrast to the
classification-based learning, inductive program hypotheses are supposed to cover
all the given examples correctly. This is due to the fact that for a program it is expected
that the desired input/output relation holds for legal inputs.
The inductive programming is used with two general approaches: 1. generate-and-
test, and 2. analytical approach. The first enumerates syntactically correct programs
and tests each against the given examples, with search guided through some search
strategy. It is obvious that a cognitive system does not learn rules by the generate-
and-test approach.

13.6.1 Argument-Based Learning

A challenge with machine learning is to deal with large spaces of possible hypotheses,
and the search associated with that space. This problem is better addressed using
expert domain knowledge to constrain the search. Further, the problem is somewhat
simplified, as machine learning allows to use to some extent the general domain
knowledge, which holds true for the entire domain. For constraining the search, it
allows the use of “local” expert’s knowledge relating to specific situations, which is
valid for chosen learning examples. For this, a domain expert provides some learning
examples, these together with other examples are input to the learning algorithms.
These examples are explained in the form of arguments: for and against, and called as
argumented examples; hence the learning using such examples is called Argument-
Based Learning (ABL) [5].

Definition 13.4 (Learning from examples) Having given some examples (argu-
ments), find a theory that is consistent with them. 

To explain the idea of argument-based learning, we consider a simple learning

problem called “Learning about approval of credit limit.” Each example consists of
information, as shown in Table 13.1, which comprises, apart from other things, the
decision of the manager (Yes/ No), indicting whether the credit limit is approved or
not.
388 13 Machine Learning

Table 13.1 Examples for credit limit approval

Customer’s name Repayments’ Rich Height Credit approved
regularity
Mrs. Red No Yes Tall Yes
Mr. Green No No Short No
Miss Blue Yes No Medium No

An algorithm after learning from these examples will induce the following rule:

IF Height = Tall THEN Credit approved = Yes. (13.2)

Now, we want to see how this rule will modify when it faces the arguments.

Definition 13.5 (Arguments-based learning) Given the examples and supporting

arguments, find a theory that explains the examples using arguments. 

To understand what it means to “explain the examples using given arguments,”

see the examples in Table 13.1. Let us assume that an expert gave an argument: “Mrs.
Red was allowed credit because she is rich.” Next, we refer to the rule (13.2), which
states that all tall people were approved credit. Note that this rule correctly classifies
Mrs. Red (as tall), but does not include in the classification the argument (of rich)
for Mrs. Red. Neither the rule mentions Mrs. Red’s property, namely she is rich.
Hence, we say that this rule (13.2) does not “AB-cover” Mrs. Red. Therefore, an
ABL algorithm induces other rule,

IF Rich = yes THEN Credit approved = Yes (13.3)

which explains Mrs. Red’s example, using the given argument: credit was approved
because Rich = Yes; now, this rule AB-covers Mrs. Red. Based on this discussion we
note that the use of arguments in learning contributes to the following advantages:

1. They impose constraints over the space of possible hypotheses, thus reducing
both time and space complexity of each search, enabling faster and more efficient
induction of theories;
2. An induced theory is more meaningful for an expert system since it is consistent
with given arguments as well as examples.

In fact, there may be many possible hypotheses from the perspective of a machine
learning method that may explain the given examples sufficiently well. But these
hypotheses should be simple enough to be understandable to expert (systems). Using
arguments should lead to hypotheses that explain given examples in similar terms to
that used by an expert, and should correspond to full justifications.
The argument-based examples is a resource for domain expert, which is added
as partial domain knowledge in a learning system in advance. This prior knowledge
13.6 Inductive Learning 389

in the form of arguments as individual examples is simple to explain, concrete, and

specific, in contrast to the general theories, which do not possess these properties.
Since an argument is for a specific example, it is not required to be true for the
entire domain, and suffices that it is true in the context of the given argumented
example. As a test, consider a possible expert’s argument to reject credit approval
to Mr. Green: “Mr. Green did not get credit approval because he does not repay
regularly” (see Table 13.1). This is a reasonable argument, however it does not hold
for the entire domain, as Mrs. Red was approved credit, although she did not pay
regularly. This generalization of the context of an argumented example is carried
out by an argument-based learning algorithm only to the extent that it preserves the
consistency with other examples.

13.6.2 Mutual Online Concept Learning

A mutual online concept learning system is an agent-based system (see more details
in Chap. 16) with communication channel for exchange of information among the
agents (see Fig. 13.2). It is assumed that the communication channel is an ideal
one without any noise. A one-agent framework includes the basic elements like
concept, instance producing mechanism, instances, instance interpreting mechanism,
and concept adaption mechanism (or online concept learning), with communication
channel shared among agents.

Adapt Adapt

Concept Concept
Produce

Interpret

Interpret
Produce

Agenti Agentj

Agentk

Instances Instances

Communication
Channel

Fig. 13.2 Mutual online concept learning framework

390 13 Machine Learning

In simple terms, a concept is a mapping from inputs to outputs. When taken as a

function, it is a deterministic function, i.e., if the input is repeated, the output will
follow the same sequence. When we think of it as a mapping with some probability
distribution associated, then the output is also probabilistic in nature. As a deter-
ministic function f , it can be formalized as f : X → Y . Considering it a Boolean
concept, the domain is X = {0, 1}n , and its range is Y = {0, 1}. For implementing
a concept, we can choose from a variety of representations, such as propositional
logic, first-order logic, or neural networks. When a concept is viewed as a function,
an instance (x) is just a specific case in the function’s domain X , i.e., x ∈ X [9].
When it is required for an agent to express a concept to the outside world, it
uses an inverse function to generate an instance using a concept, shown as produce
operation in Fig. 13.2. There be many instances that can be used to express the same
concept if the mapping from input to output is not one-to-one.
An agent may receive an instance from another agent through a communication
channel, and may adapt its concept like probability, network connection weights, etc.
This adaptation may become helpful from the point of view that the receiving agent
may perform its task better next time due to updated concept or knowledge, in terms
of getting more benefits/payoff. Let us assume that, for an agent i, an associated
concept is Cit at time t, and for agent j it is Cjt . Let the instance instance(Cjt ) be
generated by j at time t, and this instance be received by agent i. At the next time
t + 1, the agent i updates its concept to Cit+1 by adapting from its previous concept
(at time t) together with the instance of concept of agent j at time t, i.e.,

Cit+1 = adapt(Cit , instance(Cjt )), j = i. (13.4)

The representation in the learning processes, as well as the knowledge represen-

tation, is an important consideration for system performance. For real values-based
instances, the space in concept learning can be of n-dimensions: Rn , and {0, 1}n for
Boolean values. The concepts learned from instances are based on linear threshold
values, such that there is a hyperplane in Rn that separates the points on which the
function is 1 from the points on which it is 0.
Figure 13.2 indicates the logic of concept based on a broad sense, but does not
specify the time order of agent interactions between the agents. This interaction
can be performed in any of the two modes: serial or concurrent. In the first case,
at each time step only one agent is generating instances or updating concepts, and
other agents should be idle. In the second (concurrent) case, two or more agents may
simultaneously generate the instances and update the concepts. There is no difference
between these agents while working in the concurrent mode, however when in the
serial mode, the two agents are not strictly equivalent since there must be one agent
that starts the mutual process first, and so makes an initial impact on the direction of
formation of a concept.
13.6 Inductive Learning 391

13.6.3 Single-Agent Online Concept Learning

For the sake of simplicity, we consider a single-agent for online concept learning
from a teacher/environment, that is static. In contrast to a single-agent system, the
environment is continuously changing (not fixed) in a multi-agent online concept
learning system, because the agents act on the environment shared by themselves.
The learning task to be performed is as follows: obtain the target concept or
function f ∗ : {0, 1}n → {1, −1}, that maps each instance1 to the correct label or
class in response to the learning instances and feedback from the teacher [9].
A sequence of trials is needed for the online learning task to take place. In a trial,
the following order is followed by the events:
1. an instance is received by a learner from a teacher;
2. the instance is labeled as 1 or −1 by the teacher;
3. the teacher tells the learner whether the label is correct or not;
4. this feedback is used by the learner to update its concept.
In the above steps, each new trial begins when the previous trial ends.
A learning algorithm called the Perceptron algorithm takes real-valued inputs in
the form of a vector, and outputs as 1 if the result of computation is greater than a
threshold (θ ), otherwise −1 is the output [3].
To be precise, if an input instance is −
→
x = (x1 , . . . , xn ), the output computed by
−
→
the perceptron is f ( x ):
 
−
→ +1 if ni=1 wi xi > θ
f( x )= (13.5)
−1 otherwise.

In the above, −
→ w = (w1 , . . . , wn ) ∈ Rn is called the current weight vector of the
perceptron. For the sake of convenience we take threshold θ as 0, and instead add
a constant x0 = 1 with a weight w0 . For the sake of compactness the perceptron
function in Eq. (13.5) is written as

f (−
→
x ) = sgn(−
→
w .−
→
x ).

Let −
→
w .−
→
x = y. Now, we can rewrite Eq. (13.5) as

+1 if y > 0
sgn(y) =
−1 otherwise.

It is to be noted that each weight vector − →w defines a perceptron function or

concept, hence updating a weight vector is equivalent to updating a concept.
Learning of a concept requires choosing values for the weight vector − →w =
(w1 , . . . , wn ). At the start of the algorithm, the initial weights are chosen as

1 Instance and example are the same here, which are input to the function.
392 13 Machine Learning

−
→w = (0, . . . , 0). On receiving an input instance −
→
x = (x1 , . . . , xn ), the learning algo-
−
→ −
→
rithm predicts the label of x to be f ( x ). For the sake of compactness we represent
f (−
→x ) by y.
If it is found that the predicted label is correct, then there are no changes in the
weight vector. But if the predicted label is wrong, the weight vector is updated using
the perceptron learning rule given below. Here, λ is the learning rate.
−
→
w =−
→
w + λy.−
→
x. (13.6)

Multiple Agent Concept Learning

It is an extension of the single-agent concept learning. In the multi-agent system,
every agent behaves both as a teacher and a learner, with a dynamic environment.
This is due to the reason that the other agents being part of the learner environment
also change their concepts as they learn from each other. Since there is no teacher or
environment existing initially, there is no fixed concept to be learned. Accordingly,
at the begin, the concept to be learned is dynamically formed due to the interaction
among the agents.

13.6.4 Propositional and Relational Learning

The learning methods making use of “propositional calculus” are called attribute-
value/propositional learners. They use objects with a fixed set of attributes, selected
from a predefined set of values. The learning based on first-order relational
descriptions—the relational learning—induces descriptions of relations and uses
the objects described in terms of their components and relations among the compo-
nents. As an example, if an object X has attributes a, b, c then this can be represented
as a tree with X as the root and a, b, c as leaves, showing X having relation with
each leaf. This relation is the propositional relation. The background knowledge of
relations is formed using the language of examples and concept descriptions, which
is typically a first-order logic. Particularly, the learners that induce hypotheses in the
form of logic programs, i.e., Horn Clauses, are called inductive logic programming
systems.
The inductive learning has application in automatic construction of knowledge
base in expert systems, which are an alternative to classic knowledge acquisition
systems. The inductive techniques are also used for refinement of existing knowledge
base, since they facilitate the detection of inconsistencies, redundancies, lack of
knowledge, and also are helpful in the simplification of the rules provided by the
domain expert. Due to their capability to detect patterns present in the input examples,
the inductive methods are also used in the domains of biology, psychology, medicine,
and genetics.
13.6 Inductive Learning 393

13.6.5 Learning Through Decision Trees

This is a combination of induction-based learning as well as supervised learning.

In decision trees, every record is associated with a unique leaf node of the decision
tree. The criteria for choosing a unique leaf node is to start from the root node,
and repeatedly choose a child node based on the splitting criteria, which evaluates a
condition on the input record at the current node.
The decision tree is a predictive model for learning, which represents the clas-
sification rules using a tree structure that recursively partitions the training data set
into two subsets. Each internal node of the tree (all, except the leaf nodes) performs
a test on the feature value, which results in two outcomes, one of those is selected,
and ultimately reaches the terminal node where the prediction is performed.
An algorithm for decision-tree construction comprises two states: 1. construction
of decision tree, and 2. pruning the tree. In the first step, the decision tree grows
top-down in the greedy way. The data is examined by selecting the split condition
at each node, starting with the root node. The data is then partitioned and procedure
applied recursively. In state 2, the tree already constructed is pruned to control its
size. Sophisticated pruning methods select the tree in such a way that it minimizes
the prediction errors.
The decision trees tests one feature at each internal node, with one branch for
each feature value, and the class predictions is carried out at the leaves.
Algorithm 13.1 is an algorithm for decision-tree learning (relational learning).
It recursively calls the function DT (T , Ecurr ) representing knowledge that has been
learned using decision trees. The variable T represents the decision tree constructed
so far, and Ecurr is the set of input examples. If all input examples in Ecurr are in
the same class C (i.e., representing the same concept), the decision tree will contain
only one leaf corresponding to C. For example, the proposition E1 = “A smiling
person is a friend,” can be represented as a single leaf tree as shown in Fig. 13.3,
with C = friendly. If the examples in the input belong to n number C1 , . . . , Cn , an
attribute is chosen by the algorithm and based on this attribute, E is divided into sets
E1 , . . . , En , which are disjoint. Each partition Ei consists of the examples with the
same value in the attribute selected. Each set Ei is applied to the algorithm until the
sets which have been obtained are left with elements belonging to the unique class.
Finally, terminal nodes of the tree are the resultant disjoint classes C1 , . . . , Cn into
which the examples are classified.

Fig. 13.3 A decision tree E1 = is-smiling

with single proposition

yes

friendly
394 13 Machine Learning

Algorithm 13.1 Algorithm for Decision tree

1: Initialize: Ecurr = E, T = nil
2: function DT (T , Ecurr )
3: if all examples in Ecurr ∈ Class Cs then
4: generate a leaf labeled Cs
5: else
6: ; {generate new node}
7: select the most informative attribute A with values {v1 , . . . , vn }
8: Split Ecurr into subsets E1 , . . . , En according to values v1 , . . . , vn
9: for i = 1 to n do
10: DT (Ti , Ei )
11: end for
12: end if
13: Output: Decision-tree T with the root A and subtrees T1 , . . . , Tn
14: End

Fig. 13.4 A decision tree for Is-smiling

root domain yes no

Holding Unfriendly
d b
or all
oo
sw
flag
n

Friendly Friendly
Friendly

Any unseen example is classified as follows: start from the root node, attributes
are tested at internal nodes, repeat the process and proceed to a terminal node. It
is assumed that in an example that belong to a different class, at least one of the
attributes has a different value.
In a decision tree, each internal node is labeled by an attribute and links. The
attribute is selected from a set of possible values. In the example of Fig. 13.4, one of
the attributes is “is smiling” and the set of its values is {yes, no}. Another attribute
is “holding” with the set of values as {sword , flag, balloon}. The decision-tree con-
struction has two important objectives: one is selection of the most informative
attribute having the closest sense to the context in the example, and the second is
to minimize the number of tests. Note that the number of tests is height (or depth)
of the tree—paths length from the root to terminal node, and it directly defines the
difficulty level of the problem [4].

Example 13.1 Decision tree for financing a client.

Consider financing an individual by a financing company. The finance may be based

on the good credit history or other factors.
Figure 13.5 shows the details, which causes learning. The learning helps in clas-
sifying the individual loan applications 1 . . . 16 into various categories shown in this
figure. 
13.6 Inductive Learning 395

Fig. 13.5 Learning through

k
decision tree 15
0−
Credit history?
High Risk Go
od

OK
Ba
Credit history?

d
1, 6, 5 Low moderate low
n go
ow od risk risk risk
kn
un bad 11, 13 14 15, 16
Debit?
High Moderate
gh Lo Risk
Hi w Risk
10, 12 8, 9
High Risk Moderate
Risk
2, 4
3, 7

Fig. 13.6 Disambiguation- Bank account?

tree for bank
no yes

bank of? bank/FINANCE

no yes

-
bank of COUNTRY?
no yes

bank/RIVER bank/FINANCE

Example 13.2 Decision tree for Word Sense Disambiguation (WSD).

Consider an example of a decision tree for Word Sense Disambiguation (WSD), i.e.,
to find out the correct sense of a given word in a phrase. We are interested in the word
“bank” in a sentence: “we sat on the bank of Ganges.” The process of disambiguation
is demonstrated through a decision tree shown in Fig. 13.6.
The figure shows that the tree is traversed such that for every internal node it
asks a question, having the answer. The answer to each branching is provided from
the hidden knowledge in the sentence, as well as through commonsense knowledge.
We note that it follows the path of no-yes-no from the root node until the terminal
node, and the choice of sense bank/RIVER is made. In the tree, the node “bank of
COUNTRY” stands for, for example, “bank of India”, “Bank of America”, etc. The
terminal nodes for which no choice can be made based on specific feature values are
indicated with empty values by “−”. 
Construction of the decision tree requires extensive access to the training database.
Sometimes the training database is so huge that it does not fit the memory, in that
case some efficient data access method is needed to achieve the scalability.
During the construction of a decision tree it requires construction of statistics in
memory for each node, and is carried out in a single scan. The split at any node is
based on some database partition that satisfies the split criteria.
396 13 Machine Learning

The statistics for splits is updated in the memory at each node in a path starting
from root node up to terminal node. The partitioning at each node should satisfy the
splitting criteria corresponding to that node, and sufficient statistical data in the form
of corpus is necessary for the construction of any learning decision tree.

13.7 Discovery-Based Learning

The discovery-based learning requires the maximum learning efforts among all types
of learning. One of the earliest discovery-based programs is Automated Mathe-
matician (AM), which derives a number of concepts in mathematics. It is based
on the concepts of set theory. The AM discovers the natural numbers by modify-
ing a concept called “bag”, which is a generalization of the concept of a set in
set theory. The concept “bag”permits multiple occurrences of set elements, e.g.,
{a, a, a, b, c, c, d , d }. Hence, the natural number 4 is represented as {1, 1, 1, 1}. An
addition of the natural numbers is shown as the union of sets: {1, 1, 1, 1} ∪ {1, 1} = 6.
The operation of multiplication is the series of additions: 3 ∗ 4 = {1, 1, 1} ∪ {1, 1, 1}
∪{1, 1, 1}∪ {1, 1, 1} = 12. The operation of subtraction is the difference of the sets:
4 − 3 = {1, 1, 1, 1} − {1, 1, 1} = {1}. The division is obtained by repeated the set
differences. A prime number is an integer having only two dividers, the unity and
the number itself.
One approach for discovery-based learning is learning by concept hierarchies, as
explained in the following.
Learning by Discovering the Concept Hierarchies
One of the most general approaches to solving a complex problem is to decompose
the problem into smaller, less complex, and manageable subproblems. This principle
is the foundation for structured induction in machine learning, where a concept
hierarchy is defined and rules are learnt for each of the (sub)concepts. In this process,
the concept hierarchy can be manually constructed or it can be automated, instead
of learning a single complex classification rule from examples.
Consider that given five Boolean attributes x1 , . . . , x5 , it is required to learn
Boolean function y, expressed as

y = (x4 ∧ x5 ) ⊕ (x3 ∧ (x1 ⊕ x2 )). (13.7)

For five attributes there is space of 32 points. Consider that out of these, there
are total 24 randomly selected points (i.e., 75% of the total) are to be selected as
examples for learning. These are called as attribute-value vectors, and are useful for
hiding the actual structure of the original expression. In Eq. 13.7 it is possible to
justify that the best possible structure of sub-concepts will be as given in Fig. 13.7,
with definition of functions as f1 = ⊕, f2 = ∧, f3 = ∧, and f4 = ⊕.
An approach for learning by concept hierarchies is based on function decompo-
sition. Consider decomposing a function y = f (X ) into y = g(A, h(B)). The orig-
13.7 Discovery-Based Learning 397

Fig. 13.7 Hierarchy of y

concepts as a tree

f4

f2 f3

x4 x5 x3 f1

x1 x2

Fig. 13.8 a Original y y

decision tree, b Decision tree
after decomposing
f g

x1 x2 xk xm x1 x2 c

xk xm

(a) (b)

inal function is shown in Fig. 13.8a and b shows the decomposed function. In the
original function, X = {x1 , x2 , ..., xk , ..., xm }, and after decomposition, A = {x1 , x2 },
B = {xk , ..., xm } are sets of input attributes, and y is variable, which represent the
class. Also, input attributes X = A ∪ B. The f , g, h are functions, which are partially
specified as examples, i.e., by sets of attribute-value vectors with assigned classes.
The functions g and h are determined in the decomposition process, such that their
joint complexity is lower than the complexity of F. The attribute c is called inter-
mediate concept and defined as c = h(B), which is discovered in the decomposition
process.
The function decomposition can be applied recursively on the functions g and
h, which should result in a hierarchy of concepts, as shown in Fig. 13.8b. For each
concept in the hierarchy, like h(B), there is a corresponding function (h here), that
decides the dependency of the concept, e.g., concept c, on its immediate descendants
in the hierarchy [11].
398 13 Machine Learning

13.8 Reinforcement Learning

Our learning is through feedback of our actions in the real world. Human beings
always learn by interactions with the environment/teacher, which are cause and effect
relations. The environment or the world around us is the teacher, but its lessons are
often difficult to detect or grasp by the mind or analyze, hence hard to learn. The
best example is learning by a dog or a young child, where good actions are rewarded
and bad actions are discouraged. It is not defined as a more general example, but the
actions with the world.
The programs that improve their performance at some task by rewards and pun-
ishments from the environment are examples of Reinforcement Learning (RL). For
many automatic learning procedures for real-world tasks, like job-shop scheduling
and elevator scheduling, RL has been found to be very successful. The total dis-
counted reward the learner receives is optimized in the reinforcement learning, i.e.,
a reward that is received after one time step is considered equivalent to a fraction of
the same reward received immediately before [1, 8].
There are four components of RL:

1. a policy,
2. a reward function,
3. a value mapping, and
4. a model of environment.
Learning agent’s choices and methods of action are defined under a given policy.
The policy is represented using a set of production rules. Sometimes the policy could
also be probabilistic in nature. The reward function is a relationship between the state
and goal, which maps each state into a reward measure, and indicates the need of
that action to achieve the goal.
A reinforcement-based learning differs from supervised learning in the sense that
a reinforcement system may not have a teacher to respond to each action, instead the
learner itself creates a policy for interpreting the feedback. For example, playing in a
particular style leads to winning a chess game for a player is reinforcement learning.
The RL is a commonly used framework for problems requiring a number of agents,
but their presence does not contribute to much of complexity as RL needs only limited
feedback for learning. With the objective to maximize the expected reward, RL
algorithms attempt to learn policies that help in taking optimal sequential decisions.
However, this learning may be intractably slow as the methods scale up to more
real-world problems with large state spaces.
A commonly used approach to solving RL problems makes use of a function,
called value function approximation, using which an agent tries to learn a value for
each state. This value is a long-term expected reward from entering a particular state,
and such values are used for deriving a policy.
The state space for real-world problems contains infinitely a large number of
possible states features. Hence, the designer of any such task must pick up only
the most relevant features. Consider the task: Travel to Mumbai for work, and the
13.8 Reinforcement Learning 399

weather report for New Delhi is not likely to be relevant. From these we should
construct a feature set. In most real-world situations, the feature set for a problem
is broken down into a number of subsets, such that each subset can learn a specific
concept of the domain. In this case, some concepts/feature-set subsets may be more
important than the others.

Example 13.3 Formulate a task of navigation to be carried out by an agent, with a

goal to investigate the best plan to go from point A to point B, and may choose a path
and transport method of walking, driving, taxi, bus, train, or air.

The feature set of the agent’s state space may include

– Positions of A and B,
– Raining (yes/no),
– Type of shoes of agent,
– Agent is with umbrella (Yes/no),
– Current time, and
– Day of week.

Using positions as features, the agent can learn the concept of position and basic
path planning. The features raining, shoes, and umbrella are useful in learning as
to how the weather governs the policy, and the features of time, and day in a week
may be useful in learning to handle traffic, for example. A conventional approach to
solving this problem through RL is to learn in a 6D space (positions, raining, shoes,
umbrella, time, weekday) when all the features are taken into account. 
It is important to note that, the order in which features have been selected plays
a significant role, in terms of difficulty level of the problem, and robustness of its
solution.

13.8.1 Some Functions in Reinforcement Learning

In general, in a RL system, an agent recognizes itself in some state p ∈ S, then takes

some action a ∈ A, and then recognizes itself in a new state q. The new state q in
which the agent enters from its previous state p is decided by the agent’s transition
function:
T (S × A) → S. (13.8)

In addition to the state change, the agent receives a reward r for arriving to state
q, based on the reward function:

R(S × A) → R. (13.9)
400 13 Machine Learning

A function called value function V π (p) is based on the average sum-of-rewards

received when an agent starts in state p, and enters state q following a policy π . In
this condition, we express the relation between the value function and optimal policy
V ∗ (p) as
∀π, p : V ∗ (p) ≥ V π (p). (13.10)

The RL is found successful in many domains, which includes many real-world

domains, where it is used to optimize discounted total rewards. However, the most
natural criteria in many other domains are based on optimizing the average reward
per time step.

13.8.2 Supervised Versus Reinforcement Learning

The learning tasks are generally divided into two classes:

• One-shot decisions are tasks like classification and prediction, and
• Sequential decision tasks are like control, optimization, and planning.
The one-shot tasks are decision-based tasks that are usually formulated as super-
vised learning tasks. The learning algorithm in these tasks is provided with a set
of input–output pairs, which act as relations between input and output. The input
is the description of information used for making the decision, while the output is
the description of the correct decision. For example, in handwriting recognition, the
input is an image of the handwritten text of a digit/letter and the output is the identity
of the digit/letter. Once such an algorithm is trained, it can recognize continuous
handwritten text comprising sequence of letters and digit combinations, with good
accuracy depending on how well the learning algorithm has been trained.
The sequential decision tasks are often formulated as reinforcement learning tasks.
Learning algorithm in such tasks is part of an agent that interacts with the external
environment. At each step, the agent observes the current state of the environment,
then selects and executes some action, and on execution of this action the environment
changes its state. Due to the state–action combination, the environment provides some
feedback in the form of immediate reward to the agent, and the process goes on.
The reinforcement learning process maximizes the long-term rewards received by
the agent. For example, when this learning is used in a setting in manufacturing, the
current state is equal to the list of orders to be fulfilled along with the current status
of manufacturing facility. The actions may change the settings of various production
machines to maximize the overall profit. The latter can be achieved, for example by
reducing energy consumption by machines as well as the idle times of machines.
The immediate rewards include, saving in cost of each action, and revenue received
when the order is completed.
13.9 Learning and Reasoning by Analogy 401

13.9 Learning and Reasoning by Analogy

For analogical reasoning, an example in the form of a known solution is sufficient

for learning a new solution. We use analogy when we learn about electric currents
(e.g., Kirchhoff’s current law, I3 = I1 + I2 (see Fig. 13.9) by thinking about water
pipes flow rates Q3 = Q1 + Q2 . While learning the subjects, like medicine, law, and
economics, we also make use of analogy. Other examples are learning to play a card
game of bridge from the knowledge of hearts, or programming a new algorithm using
previously learned programming examples and concepts.
Some specific competence is necessary in using analogies to do certain kinds of
learning and reasoning. The learning takes place when a constraint description is
generated in one domain using analogy, having given a description with constraints
in another domain. For example, we can better learn the Kirchhoff’s current law
using the knowledge about water flows in pipes, as illustrated in Fig. 13.9. Learning
takes place when an analogy is used to answer a question about one situation, having
given another situation that has preceded the current situation [7].
Analogy plays an important role in our reasoning process. We frequently explain
or justify one phenomenon with another, where a previous experience often serves
as framework or pattern for a new analogous experiences. A familiar experience is
often in dealing with new experiences. Since so many of our acts are near repetitions
of our previous acts, analogical-based learning is predominant in our living style. An
implemented system for analogical learning has ingredients [10], discussed below.
Representation of Extensible Relations
The situations for analogical learning are represented based on relations between
pairs of parts. In addition, more descriptions can be attached to the relations when
needed.
Importance-Dominated Matching
Best possible match is found based on what is important in the situations, and then the
similarity is measured between two situations. The importance is computed based
on various kinds of constraints, and cause is a common importance-determining
constraint.

Fig. 13.9 Analogical Q1 Q2 I1 I2

problem-solving

Q3 I3
From known flow rates: Infer Kirchoffs current Law:
Q3 = Q1 + Q2 I3 =??
402 13 Machine Learning

Analogy-Driven Constraint Learning

A constraint such as the one based on Kirchhoff’s law is learned as a result of mapping
parts of a situation belonging to a well-understood domain into the parts of another
situation in a poorly understood domain.
Analogy-Driven Reasoning
Many times we are interested to find out if a particular relation holds. The causes
found in a remembered situation can supply suggestive precedents.
Analogical reasoning shall best work subject to the following restrictions:

• Symbolic sufficiency. A situation to a previous instance can be represented using

certain classes, properties, actions, etc. However, these collections should be sim-
ple, so that matching can be performed efficiently.
• Description-based similarity. A situation can be said to be similar to a previous
one, if it can be mapped.
• Constraint-based similarity. Two situations are said to be similar if they are gov-
erned by the same constraint.

A program written for implementing reasoning by analogy should function based

on identifiable principles, and must do the reasoning and learning by analogy, as in
the following situations:

1. A class teacher explains to a student that a voltage drop across a resistance is

calculated similar to the way we find the pressure drop across a section of a pipe
having known the water flow rate and friction in it. With this analogy, a student
can find the voltage drop across a resistance without the knowledge of Ohm’s
law.
2. A class teacher tells about two different resistances r1 , r2 , their voltage drops
v1 , v1 , and currents i1 , i2 , respectively, and the student will find out the Ohm’s
law.
3. A class teacher suggests to generalize the principle using the flow of water in a
pipe and the flow of current in a resistance, and the student concludes that there is
a linear relation between the forces (potential/pressure) and flows (current/water
flow rate).

From the above examples of analogical reasoning, we understand that a practice

with some specific situations in one domain enables the invention of a law. Alterna-
tively, once several examples of one law are known, the comparison results in the
generalization of the law, like in the second example above. Analogies are likenesses
or similarities between things that are otherwise different. The participants or things
in analogies are unlimited: they may be concepts, objects, problems, solutions, plans,
situations, episodes, and so on.
Analogies play a dominant role in human reasoning and learning processes—
previously remembered experiences are transformed and extended to fit into new
situations; old experiences may provide the explanations or scenarios that tend to
13.9 Learning and Reasoning by Analogy 403

match with the new situations in some aspects and, therefore, may offer the promise
of suggesting a solution to the new situation.
The difference between inductive learning and analogical learning is that, induc-
tion or similarity-based learning is based on the observation of a number of training
examples, but analogical and explanation-based learning requires only one exam-
ple in the form of a known solution or a past experience, as a sufficient criteria for
learning to take place.
Analogical Reasoning
Consider the statement: “I have a Pomeranian dog, who is friendly with the children.
Whenever I see another Pomeranian dog, I assume that he too will be friendly with
the children.” This type of reasoning we usually perform in our daily life. But it also
appears that it does not guarantee the truth. The reason being that is the existence of
one or few shared characteristics does not mean that all the remaining characteristic
are identical.
For analogical reasoning to work, it is necessary that some relevant and known
similarities between two objects or situations be give a sufficient base to a reason
to believe that there are further similarities between them, and that should help to
answer an open question. However, it is not always that way. At the most, analogical
thinking can give the base to think of some probabilities, and these may be the base for
probabilistic judgment. The kind of thinking that takes place in analogical reasoning
has a simple structure:
1. Object A possesses a set of characteristics, say X ,
2. Object B shares the characteristics X ,
3. The object A has characteristic Y also,
4. Because the object A and B share characteristic X , based on this we conclude what
is not known yet, that is, B shares Y as well, which is not necessarily correct.

In Fig. 13.10a, the analogical reasoning will work, while in (b) it may not.

Definition 13.6 (Reasoning by Analogy) It is the process of inferring that a con-

clusion property Y holds of a particular object or situation B (called target), given
that B shares a property or set of properties X with another object/situation A (called
source), which has the set of properties X .

Fig. 13.10 Similarities A B A B

computing for analogical
reasoning X Y Y X

(a) (b)
404 13 Machine Learning

A(X ) ∧ B(X )
A(Y )
−−−−−−
∴ B(Y ). (13.11)

The set of common properties X is the similarity between A and B, and the
conclusion property Y is projected from A onto B. The process may be summarized
by the following statements.

Justification. The analogical reasoning method defined in Eq. (13.11) can be justified
by a two-step process: 1. from the first premise A(X ) ∧ B(X ), conclude a general-
ization ∀ X [A(X ) → B(X )], and 2. instantiate the property X and generalize it to
Y , then apply the inference rule of modus ponens to obtain the conclusion B(Y ).
Due to this process, only the first step is non-deductive. Hence, it appears as if the
problem of justifying the analogy has been reduced to the problem of justifying a
single-instance inductive generalization. 
A criteria for the evaluation of strength of an (enumerative) induction states that,
as the number of instances confirming the generalizations increases, an inference
increases in plausibility. If this is the only criteria used for analogical inference, then
all the conclusions projected by any analogy without counterexamples should also be
plausible. But, this fails. Consider the example, if the inspection of an Indian peacock
reveals that its tail is more colorful than its legs, a projection of this conclusion onto an
unseen peacock is plausible. But, projecting that a hanging feather on the peacock’s
tail will be observed on a second peacock is not plausible. Anyone who has closely
looked at the tail of even a single peacock will have no counterexamples to these
conclusions. Also, both conclusion properties can be projected, so the difference in
agreeableness is accounted for some other criterion. Hence, the problem of analogy
is different from (enumerative) induction because the analogy requires a stronger
criterion for plausibility.
The analogical learning is a nontrivial form of inference, i.e., the conclusions
based on the analogies are not necessarily logical entailments of the previous (source)
knowledge. Analogical is a form of plausible reasoning for the current situation.
The steps in Eq. (13.11) require many other things, like the criteria for similarity
measure, some form of indexing, efficient recall mechanism, a flexible mapping
between base and object domain and, various levels of abstraction.
Following are the steps to perform the analogical reasoning:

1. Analogue Recognition. A new situation or problem is encountered and recognized

as being similar to the previously known situation.
2. Criteria for competence. It is concerned with the knowledge required for any
particular set of algorithms to exhibit the specified behavior.
3. Representation of situation by extensible relations. Extending the mapped experi-
ence, i.e., the newly mapped analogous are modified and extended to fit the target
situation.
13.9 Learning and Reasoning by Analogy 405

4. Determining Analogy and Matching. Here, two patterns are selected for their
similarities and mapped from the base to the target domain.
5. Similarity measures. Various similarity measures are defined, and matchings are
evaluated as per those.
The Access and Recall mechanisms are useful features, as per which the similarity
feature of a new problem serves as an index to the previously solved problem.
This will help us in recalling the previous problem when required.
6. Abstraction. The matchings are abstracted away to generalize it.

The newly formulated solution is validated for its applicability through some
trial process, like theorem provers or simulation. If the validation is supported, a
generalized solution is formed, which accounts for old and new, both, which is equal
to a new learning.
When a case is inferred analogous to another on the basis of a unifying principle
such that it is accepted without having been tested against other possibilities, such
inference due to the analogical reasoning will be faulty. In addition, when some
similarities between two cases are considered decisive on the basis of insufficient
investigation of relevant differences, such analogical reasoning will also go wrong.

13.10 A Framework of Symbol-Based Learning

Given the data, one way to characterize the learning problem is based on the
goal/target of the learner algorithm, which may be a concept, or description of a
class of objects. A learning algorithm may also acquire plans, heuristics for problem
solution, or other form of procedural knowledge in any structure discussed so far,
including the predicate logic.
The set of operations performed by the algorithm may include generalizations
rule, heuristics rule, or a plan that satisfies the goal. The typical operations are
generalizations or specializations of symbolic expressions, adjusting the weight of
a neural network, etc. In concept learning, a learner may generalize a definition as
follows:

color(bird , black) → is(bird , crow).

∃x color(X , black) → is(X , crow).

There is a potential concept space, which is searched by the learner algorithm and
to find the concept. The complexity of this concept space is the main complexity of
the problem solution. The learning program/algorithm must be committed to order
and the direction of the search, as well as it should make use of the available training
data and heuristics, to carry out the search efficiently. The learner may make use of
the heuristics, and, for example, learn the concept ball, by using the first example as
a candidate concept:
size(object1 , small) ∧ color(object1 , red ) ∧ shape(object1 , round )
406 13 Machine Learning

and generalize using the second example:

size(object2 , small) ∧ color(object2 , red ) ∧ shape(object2 , round ).

Here we have used heuristics and not induction for learning.

13.11 Explanation-Based Learning

One type of deductive learning is Explanation-based Learning (EBL). Following

types of information are required to implement the explanation-based learning:

1. A formal statement of the goal concept to be learned;

2. Domain theory: It relates to the concept and the training example;
3. A training example: At least one positive training example of the concept is
needed. It is an instance of the target, what the training program sees in the
world;
4. A Domain History: It is a set of facts and rules used for explaining, how a training
example is an instance of the good concept; and
5. Operational criteria: These are some means of describing the form, which a
concept definition may assume.

The EBL makes use of an explicitly represented domain theory to construct an

explanation of a training example, which is usually a proof that logically follows the
knowledge base. An EBL begins with following prerequisites:

– A target Concept
– Learning by Analogy
– Learning by Instruction
– Learning by Induction
– Learning by Deduction
– Reinforcement Learning
– Discovery-based Learning, which determines the effective definition of this con-
cept, called local concept, which requires a high-level description.

Consider using EBL to learn that a given specific object “this cup” is a cup. The
target concept is a rule in predicate form, which is to be used to infer whether a given
object is in fact a cup. We express this as

premise(X ) → cup(X ). (13.12)

In this implication, the premise is a conjunctive expression with variable X , and pred-
icate expression part(X , Y ) indicates that X  s part is Y . Let the domain knowledge
include
canbelifted (X ) ∧ holdsliquid (X ) → cup(X )
13.11 Explanation-Based Learning 407

light(X ) ∧ part(X , handle) → canbelifted (X )

part(X , Y ) ∧ concave(X ) ∧ pointsup(X ) → up(X )
small(X ) → light(X )

where predicate holdsliquid (X ) means that object X holds liquid, canbelifted (X )

means X can be lifted, and pointsup(X ) means object X points upwards.
A training example should be an instance of the concept for the goal concept, as
represented by the following facts of predicate logic.

cup(thiscup).
small(thiscup).
put(thiscup, handle).
own(bob, thiscup).
part(thiscup, bottom).
part(thiscup, bowl).
color(thiscup, red ).
pointsup(bowl).
concave(bowl).

The operational criteria should be such that the target concept can be defined
in terms of observable structured properties of object, like it points up and its
parts:bottom, handle, bowl, etc. The domain rules are provided in such a way that
these enable the learner (algorithm) to infer whether a description is operational or
not. Making use of this approach, a theorem prover can construct an explanation as
to why the given example is in fact an instance of the training concept—a proof that
confirms that the target concept logically follows from the example. This is shown
in Fig. 13.11, as a generalized proof tree.
The EBL has the following advantages:

1. The training examples often contain many irrelevant information (or noise), such
as “make of the cup”. However, the domain theory allows the learner to select
only the relevant aspects of training example.

Fig. 13.11 Explanation- cup(mycup)

based learning to identify the
object cup

canbelif ted(mycup) holdsliquid(mycup)

part(mycup,
handle) pointsup(bowl)
light(mycup)
part(mycup, bowl)
concave(bowl)
small(mycup)
408 13 Machine Learning

2. Usually, an example may allow for many possible generalizations, and most of
them may not be useful. However, since the EBL inference logically follows
the example, the generalization of EBL are logically consistent with the domain
theory.

13.12 Machine Learning Applications

The domain of machine learning applications is continuously expanding. In the fol-

lowing we present some of the major application areas of machine learning:
Machine Perception
These approaches are used for learning to perform segmentation, feature extraction,
and classification. All these provide higher performance and greater ease of imple-
mentation than traditional systems built from individual hand-built components.
Large-Scale Information Management and Data Analysis
In the present time, the data are being gathered in businesses and scientific applica-
tions far faster than ever before. This volume of data cannot be disseminated by any
manual method. The machine learning methods are helpful in finding patterns and
relationships in these data sets.
Information Filtering and Information Retrieval
Finding out useful information from an abundant size of collections is impossible
using any manual methods. Models of the information needs of users are constructed
based on machine-learning methods for various available information sources. Using
these models it is possible to process the huge volume of information, both offline and
online, to filter the undesired information and deliver to users only the information
they need. However, for this the users should first express their need in natural
language or symbolic form.
Control and Optimization
The adaptive control methods are more of futuristic systems, which are aimed to pro-
vide large-scale optimization, e.g., over the entire enterprises for planning, schedul-
ing, manufacturing, inventory management, and logistics. The machine learning
methods in controls and optimizations may also make it possible for new kinds
of applications, such as smart buildings and smart vehicles.
Speech Recognition
All the presently available speech recognition systems in the market, make use of
machine learning in some form, to train the system to recognize speech of unknown
users as well of specific users. Before shipping, such systems (programs) are trained
in speaker-independent mode, and after it is delivered to the individual users, it is
13.12 Machine Learning Applications 409

trained in speaker-dependent mode to accurately recognize his/her voice. Chapter 20

presents the speech-recognition techniques in more detail.
Computer Vision
The computer vision (or machine vision) applications range from face-recognition
systems to sophisticated systems that automatically classify microscope images of
cells. Many of the computer systems are more accurate than hand-crafted programs.
Some of the important applications of computer vision are handwriting recognition,
fingerprint recognition, and face recognition, which are available at a commercial
scale.
Model Building
Many areas in science and engineering require construction of complex models, to
support monitoring, diagnosis, repair, prediction, and extrapolation. Such models are
a combination of programming and adaptation. Programming is used for the con-
struction of the model, while adaptation is aimed to modify the model and calibrate
it for data. The designers of such models need better tools for construction, calibra-
tion, and managing the models. These models have applications in medical science,
strategic systems, space research, etc.
Machine Learning Techniques to Make Computers Easier to Use
Computers are still not easier to use due to ignorance of the user. While using a
computer, each user has certain goals (about tasks, resources, etc.) and different
preferences (that include, styles, habits, and abilities). Computer systems have very
small features for these yet. Some examples are spell checker, command repetition,
exception handling in high-level languages, in operating system, and in databases.
However, these things are progressively appearing in plenty in smart phones, which
include suggesting the next word while sending email or message, suggesting navi-
gation path for drivers through map applications, storing the schedules and appoint-
ments and reminding through alarms, progressive learning of languages, etc. There is
trend now of smart-watches which are claiming to provide solution for many things
related to health and health monitoring.

13.13 Basic Research Problems in Machines Learning

There are wide-ranging machine-learning algorithms, from general—“off-the-shelf”

methods for a variety of problems to custom methods suited for specific tasks. For
example, general methods are learning through decision trees, rule sets, probabilistic
networks, and feedforward neural networks, which have been applied to many dif-
ferent problems. On the other side, speech recognition algorithms based on hidden
Markov models (HMMs), and methods for inferring using binding-site geometry for
drug design are application-specific.
410 13 Machine Learning

Many fundamental research problems in machine learning are concerned about

the entire range from general to specific spectrum.
1. It is not known yet, as to what all the possible general methods are, as so far
only the four techniques exist in the general category: algorithms for decision
trees, rules, Bayes nets, and neural nets. It is not clear as to what other convenient
representational formalism should be explored.
2. It is not known as to what extent these four general methods are optimal, or if
there is scope for improvement in these algorithms.
3. What are truly the theoretical and engineering principles, that can be used as guide-
lines for the development of application-specific machine learning algorithms?
The development of such algorithms is currently expensive, time-consuming, and
mostly ad hoc in nature.
4. What are the efficient and convenient methods that can be used to acquire, rep-
resent, and incorporate the application-specific knowledge into general learning
methods?

In addition, there are several other challenging questions of fundamental nature

in the sequential decision-making tasks.

1. What are the situations in which it is preferable to represent and learn a policy
directly, instead of learning a value function first? The difference between learning
a policy versus value function is a policy specifies what action should be chosen
in each state, but a value function provides accumulated long-term reward for
each state. The best action can be selected based on the criteria: the one that will
lead to the state with the highest value as per the value function. Some methods,
like genetic algorithms, genetic programming, and parameterized policies, search
directly in policy space. However, the other methods, like temporal difference
learning and real-time dynamic programming, construct value functions.
2. What should be the properties of an ideal value function approximator?
3. What are the best ways to resolve the trade-off between exploitation (i.e., execu-
tion of the current policy) and exploration (i.e., search for better policies)?
4. What are the ways to solve large and partially observable Markov decision prob-
lems?

13.14 Summary

In every learning situation, a learner transforms the information received from an

environment/teacher into some new form, and stores in that format for future use. The
type of learning strategy used decides the type and magnitude of this transformation.
Following are the basic learning strategies:

1. Rote Learning.
2. Learning by Instruction.
13.14 Summary 411

3. Learning by Deduction.
4. Learning by Analogy.
5. Learning by Induction.
6. Reinforcement Learning.
7. Discovery-based Learning.
These strategies are in increasing order of complexity of learning as well as the
magnitude of transformation required.
A general learning model comprises: learner unit, knowledge base, performance
evaluation, and the teacher, which may or may not include a physical teacher.
The learning techniques are fundamentally classified as supervised learning and
unsupervised learning. The major difference between these is that the supervised
learning requires the presence of a teacher, while the unsupervised technique does
not require the teacher component.
Inductive Learning is a generalization carried out using a set of examples. It
is one of the most commonly used learning techniques used by humans. Learning
through concepts is a typical inductive learning process: given the examples of some
concepts, such as “cat”, we attempt to conclude a definition such that it will be helpful
to correctly recognize future instances of the concept “cat”. Note that, inductive
learning makes use of specific facts/examples rather than general axioms as in the
deductive processes.
While searching for possible hypotheses in the large space, the space can be
constrained using the domain knowledge of the experts. Learning from examples is
given the examples find a theory that is consistent with the examples. The explanation
examples are called argumented examples, and these are in the form of arguments:
for and against.
Based on the idea, whether it is a single concept or multiple concepts that are
to be learned, we call it single-concept or multiple-concept learning. A concept is a
function (a mapping), such that the learning task is supposed to discover this function,
called target function f ∗ : {0, 1}n → {0, 1}. The function maps each instance to the
correct label (also called class). In the learning process, the system responds to
learning instances and also to feedback received from the teacher.
The attribute-value learners or propositional learners are the methods that use
a formalism of propositional calculus, where objects are defined with a fixed set
of attributes. The learning methods based on first-order relational descriptions are
called relational learners; they induce descriptions of relations, and make use of
objects described in terms of their components and relations among components.
Automated Mathematician (AM) is one of the earliest discovery-based learning
programs, deriving a number of concepts in mathematics. It is based on the concepts
of set theory. AM discovers the natural numbers by modifying concept of “bag”,
which is a generalization of set. An approach to solving a problems, which is among
the most general, is to decompose the problem into smaller, less complex, and better
manageable subproblems. This principle is the foundation for learning by concept
hierarchies. The rules for each concepts and sub-concepts are learned as a discovery
learning.
412 13 Machine Learning

Reinforcement Learning (RL) is an example of programs, which improve their

performance for some task based on the criteria of rewards and punishments from
the environment. Most approaches to reinforcement learning optimize the total dis-
counted reward received by the learner.
For analogical reasoning, an example in the form of a known solution is sufficient
for learning a new solution. It is the process of inferring that a conclusion property
Y holds a particular object or situation B (called target): from the fact that B shares a
property or set of properties X with another object/situation A (called source), which
has the set of properties X . The set of common properties X is the similarity between
A and B, and the conclusion property Y is projected from A onto B.
Goal of most learning algorithms is a concept, or a general description of a class
of objects, represented in predicate logic. There is a potential concept space, which
the learner searches to find the concept. A learner may generalize a definition as
follows:

color(bird , black) → is(bird , crow)

∃x color(X , black) → is(X , crow).

The above is symbol-based learning.

One type of deductive learning is Explanation-based Learning (EBL). The fol-
lowing information is required to implement it: a goal concept, domain theory, a
training example, a domain history, and operational criteria.
Potential Applications of machine learning are machine perception, information
management and large-scale data analysis, information filtering and retrieval, control
and optimization, software engineering, speech recognition, computer vision, model
building, etc.

Exercises

1. At the end of the day, try to recollect any five important events you have witnessed
this day. Try to associate the type of learning you have used in these events.
2. Analyze and propose the type of structures you consider are appropriate in the
following learning processes:
a. Rote learning
b. Inductive learning
c. Supervised learning
d. Unsupervised learning
e. Learning by example
f. Analogical learning

Answer the above questions, in reference to some programming language, e.g.,

C. And, describe the estimated algorithmic steps. For example, for rote learning
Exercises 413

you may take the example of the number tables, and indexing you need, so that
one can speak the table of say “5”, and also one can answer 5 × 6 = 30.
3. What is the difference between the learning by induction versus learning by
examples.
4. Explain the major difference between analogical learning and inductive learning
in respect of the approach used for learning.
5. Suggest a learning method for each of the following, explaining why the sug-
gested method is appropriate, and provide logical steps to learn using it.

a. To learn how to drive a car after having observed a trained driver, while you
are riding along with the same.
b. To learn how to drive a car after having known the driving of a bullock-cart.
c. To learn how to drive a car after having learned the driving of a tractor.
d. To learn how to fly an aircraft after having very closely observed the birds
flying, like eagle, crane, and hawk.
e. Learning to keep your wallet protected after having lost it due to theft.

6. The helicopter takes off straight, instead of having a run before taking off. It
has similarity with peacock in the birds. Explain a mathematical model, which
supports the learning by analogy of a copter with a peacock.

References

1. Ahmadi M et al (2007) IFSA: incremental feature-set augmentation for reinforcement learn-

ing tasks. In: The sixth international joint conference on autonomous agents and multi-agent
systems, pp 1128–1135
2. Bengio Y (2016) Machines who learn. Sci Am 6:38–43
3. Domingos P (2012) A few useful things to know about machine learning. Commun ACM
55(10):78–87
4. Mantaras RL, Armengol E (1998) Machine learning from examples: inductive and Lazy meth-
ods. Data Knowl Eng 25(1998):99–123
5. Mozina M et al (2007) Argument based machine learning. Artif Intell 171:922–937
6. Schmid U, Kitzelmann E (2011) Inductive rule learning on the knowledge level. Cogn Syst
Res 12:237–248
7. Sunstein CR (1993) On analogical reasoning. Harv Law Rev 106:741–791
8. Tadepalli P, OK D (1998) Model-based average reward reinforcement learning. Artif Intell
100:177–224
9. Wang J, Gasser L (2002) Mutual online concept learning for multiple agents. In: AAMAS’02,
July 15–19, Bologna, Italy
10. Winston PH (1980) Learning and reasoning by analogy. Commun ACM 23(12):689–703
11. Zupan B et al (1999) Learning by discovering concept hierarchies. Artif Intell 109:211–242
Chapter 14
Statistical Learning Theory

Abstract A machine learning system, in general, learns from the environment, but
statistical machine learning programs (systems) learn from the data. This chapter
presents techniques for statistical machine learning using Support Vector Machines
(SVM) to recognize the patterns and classify them, predicting structured objects using
SVM, k-nearest neighbor method for classification, and Naive Bayes classifiers. The
artificial neural networks are presented with brief introduction to error-correction
rules, Boltzmann learning, Hebbian rule, competitive learning rule, and deep learn-
ing. The instance-based learning is treated in details with its algorithm and learning
task. The chapter concludes with a summary, and a set of practice exercises.

Keywords Statistical machine learning · Support Vector Machines (SVM) ·

K-nearest method · Naive Bayes classifier · Artificial Neural Networks (ANN) ·
Boltzmann learning · Hebbian rule · Deep learning · Instance-Based Learning
(IBL)

14.1 Introduction

Statistical machine learning systems help the programs automatically learn from the
data. This is an attractive option as an alternative to manually coding every rule
in a program. Machine learning is used in computer science and beyond, e.g., in
spam filters in email, web searching, placement advertisement, stock trading, credit
scoring, drug design, fraud detection, and in many more applications. The statistical
machine learning can be said to be a kind of a data mining.
The field of machine learning developed in a new direction during 1979–80,
with innovations like decision trees and rule learning. These methods were applied
in expert systems. In the late 1980s, there were renewals of research interests in
machine learning architectures that used perceptron as the basic building block.
This was particularly because the limitations which were highlighted by Minksy and
Papert in the early days of AI were overcome by multilayer networks that used
simple computing elements. The latter used perceptrons like nodes called neural
networks. These networks were trained using biologically inspired backpropagation

© Springer Nature India Private Limited 2020 415

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_14
416 14 Statistical Learning Theory

algorithms, which performed gradient descent on error-based cost functions. This

new approach raised the hopes not only of creating machines that were capable to
learn, but also of understanding the basic mechanisms used by biological learning
systems.
The machine learning algorithms can determine how to perform important tasks
through generalization from examples. A machine learning-based solution, at the
end, turns out to be a feasible and cost-effective solution when compared with the
manual programming approach. In statistical-based machine learning approaches, as
more data becomes available, it becomes possible to tackle more ambitious problems.
In the previous chapter we concentrated on broad classification of machine learn-
ing techniques, and basic principles of each of them. In the present chapter we will
discuss new and emerging techniques, which are mainly statistical based. For the
purpose of illustration, we focus on the most mature and widely used classification
for these techniques.
Learning Outcomes of This Chapter:
1. Apply the simple statistical learning algorithm such as Naive Bayesian Classifier
to a classification task and measure the classifier’s accuracy. [Usage]
2. Apply the simple statistical learning algorithm such as SVM to a classification
task and measure the classifier’s accuracy. [Usage]
3. Evaluate the performance of a simple learning system on a real-world data set.
[Assessment]
4. Compare and contrast each of the following techniques, providing examples of,
in what condition each strategy is superior: decision trees, neural networks, and
belief networks, deep learning. [Assessment]

14.2 Classification

One of the sub-fields of predictive modeling is supervised pattern classification. It

is a task of training a model based on labeled training data, such that the model can
later be used to assign predefined class labels to new objects. Figure 14.1 shows this
process.

Definition 14.1 (Classifier) We refer to a classifier as a system that typically has

inputs vector of discrete and/or continuous feature values and outputs a single discrete
value, the class. 

In that sense, for email filtering we use a classifier that classifies the email mes-
sages into “spam” and “no spam” classes. The input to this classifier may be a Boolean
vector,

x = (x1 , . . . , xj , . . . , xd ), (14.1)
14.2 Classification 417

Training data set

Machine Learning
Algorithm

Prediction
New Data set Classifier (Infernce)

Fig. 14.1 A classifier’s block diagram

where x1 , . . . , xd corresponds to words from some standard dictionary, such that the
element xj = 1 (True) if the jth word of that dictionary is present in the email under
consideration. If the jth word from the dictionary is absent in email, then xj = 0. The
aim is naturally to classify the emails based on these words. A learner (program)
inputs a training set of examples, comprising (xi , yi ), where xi = (xi,1 , . . . , xi,d ),
as an observed input, yi as the supposed to be corresponding output, and also the
classifier, which is supposed to be the class of this email. The test, that the learner
program (i.e., classifier) has successfully learned is, whether this classifier produces
the correct output yt for future examples xt . That is, whether the spam filter has
correctly classified the first-time seen email messages as spam or not spam, or some
other classes as dictated by the classifier [6].
Though there are thousands of learning algorithms today, the learning can be
expressed as comprising three parts: representation, evaluation, and optimization.
Representation
The classifier must be represented in some formal language. Choosing a representa-
tion amounts to choosing a set of classifiers with the capability to learn. The repre-
senting set is called the hypothesis space of the learner, i.e., it covers those features
of the inputs to which we give importance. The classifier must be in the hypothesis
space, otherwise it cannot learn.
Evaluation
The evaluation function, called object function or scoring function of a learning
algorithm, should be able to distinguish between good classifiers from bad classifiers.
Optimization
This function is useful for searching a method for a classifier that helps in scoring
the highest. Proper choice of optimization technique is key to the efficiency of the
learner.
The learners can be divided into two major types: 1. whose representation com-
prises fixed size, e.g., linear classifiers, 2. those whose representation can grow with
data, e.g., decision trees.
418 14 Statistical Learning Theory

14.3 Support Vector Machines

Due to Support Vector Machines (SVMs), a new approach is introduced to machine

learning; the method is based more on statistical foundations and less on biological
plausibility. SVMs are trained by minimizing a convex cost function, which is a
measure of the margin of correct classification. This margin, in the case of the per-
ceptrons, is used for measuring the mistake complexity. In SVMs, the optimization
of this margin is justified by a rigorous statistical analysis. The SVMs use binary
classification, i.e., input is always split into two classes. This approach is proved use-
ful in a wide range of applications, which vary from text classifications, e.g., “Is this
article related to my search query?” to bioinformatics, e.g., “Do these micro-array
profiles indicate cancerous cells?”
The SVMs were originally used for problems in binary classification, where it was
required to distinguish an object belonging to one of the two categories. For a long
time, people attempted to solve such problems by simply reducing them to binary
classification problems. However, these reductions failed to exploit about the struc-
ture of the predicted objects. But SVMs are capable of exploiting the structure also,
e.g., the conventional SVMs can be applied for parts-of-speech tagging applications,
which requires resolving the sense of each word in a sentence using the context of
the word, i.e., surrounding text.
The support vector machines are falling in the category of supervised learning
models based on associative learning algorithms that analyze the data and recognize
patterns, hence they have applications in classification and regression analysis. After
having trained by some training examples, each belonging to one of the two cate-
gories, a SVM training algorithm builds a model that assigns new examples into one
of the two categories. Thus, an SVM is a non-probabilistic linear classifier. They are
powerful approaches to predictive modeling with success in a number of applica-
tions, which include handwritten digit and alphabet recognition, face detection, text
categorization, etc.
An SVM is a model where examples are nothing but representation of points
in space, that are mapped into two classes, such that the examples of the separate
categories are divided by a clear gap that should be as wide as possible. Once the
learning has taken place, any amount of new examples are then mapped into that
same space and predicted to belong to a category depending on which side of the
gap they fall on.
The SVMs fit in the context of classification where the attribute of the objects
whose value is to be predicted, called dependent attribute, has two possible values:
0 and 1. The classification is performed on a surface (in three or more dimensions)
in the space of predictor attributes, that separate the points with dependent attribute
= 0, from those with dependent attribute = 1 (or it may be −1 and +1 respectively).
An optimal separating surface is computed by maximizing the margin of separation
as shown in Fig. 14.2, where, data point with 0 attribute are labeled by circle (◦s)
and those with attribute 1 are labeled as squares (s). The figure shows a separable
problem in a 2D space. The margin of separation is the distance between the boundary
14.3 Support Vector Machines 419

Fig. 14.2 Classification

through SVM

w x=y+1

wx=y
w x=y−1

Fig. 14.3 A case of different 1 3

margin in SVM 2

of points with dependent attribute = 0 and the boundary of those with dependent
attribute = 1. The variables x, y and w are vectors.
The margin (which is a hyperplane) is the measure of “safety” in separating the
two sets of points, the larger is better. Figure 14.3 shows three separator lines between
the set of data values of two categories, with separator line 3 producing the maximum
margin. Computing the optimal separating surface, in a standard SVM formulation
requires solving an optimization problem, which is quadratic in nature [2, 3].

14.3.1 Learning Pattern Recognition from Examples

The SVMs are based on Analogical Learning, which we discussed in the previous
chapter. The Support Vector Machine (or Support Vector Network (SVN)) maps the
input vectors into some high-dimensional feature space that we call Z, through some
nonlinear mapping, which is chosen a priori. A linear decision surface is constructed
in this space, with special properties that ensure high generalization capability of the
network. This approach gives rise to two problems: the first is conceptual in nature,
while the other is technical. The conceptual problem is regarding how to find out a
420 14 Statistical Learning Theory

separating hyperplane that will generalize well in the presence of very high dimension
of the feature space. The technical problem is about how to computationally treat
such a high dimensionality space problem?
The conceptual part of this problem can be solved for the case of optimal hyper-
planes for separable classes. An optimal hyperplane is a linear decision function with
maximal margin between the vectors of two classes (see Fig. 14.2). To construct such
a hyperplane, only a small amount of data is required, called support vectors, which
determine the margin. The optimal hyperplanes should separate the training data
without errors. The optimal hyperplane algorithm is described as follows.
The set of labeled training patterns,

(y1 , x1 ), . . . , (y , x ), yi ∈ {−1, 1}, (14.2)

are linearly separable if there exists a vector w and a scalar b (for bias) such that the
following inequalities are valid for all elements of the training set given in Eq. (14.2).

w.xi + b ≥ 1 if yi = 1,
w.xi + b ≤ 1 if yi = −1.
(14.3)

The inequalities of Eq. (14.3) can be rewritten as

yi (w.xi + b) ≥ 1, i = 1, . . . , . (14.4)

The optimal hyperplane, which is the unique one separating the training data with
a maximal margin, can be expressed by

w0 .x + b0 = 0. (14.5)

As an application of character recognition, Fig. 14.4 shows a classification of an

unknown pattern (which appears like the decimal digit 7) using a support vector
machine. A pattern in the input space is compared with support vectors xk , result-
ing in values u1 , ui , uj , us . The resulting values are nonlinearly transformed using
Lagrangian multipliers. A linear function (F) of these transformed values deter-
mines the output of the classifier [5].
For the hyperplane expression (14.4), we use a standard optimization technique
through which we construct the Lagrangian function,

 
1
L(w, b, Λ) = w.w − αi [yi (xi .w + b) − 1], (14.6)
2 i=1

where ΛT = (α1 , . . . , α ) is the vector of nonnegative Lagrange multipliers corre-

sponding to the constraints.
14.3 Support Vector Machines 421

Fig. 14.4 Classification of Output of the

an unknown pattern classifier
F

α1 αi αj αs

c
u1 ui uj us uk = K(xk , x)

xk

14.3.2 Maximum Margin Training Algorithm

As discussed above, in SVMs it is attempted to find out the support vectors with
maximum margin. An algorithm for this purpose finds a decision function for n-
dimension pattern vectors x, which belong to either of the two classes, A or B.
Input to the training algorithm is a set of k examples x1 , ..., xk with labels y1 , ..., yk ,
respectively.
(x1 , y1 ), (x2 , y2 ), . . . , (xk , yk ) (14.7)

yi = +1 if xi ∈ class A
where (14.8)
yi = −1 if xi ∈ class B.

Making use of these training examples during the learning phase, the SVM algo-
rithm finds the decision functions D(x), where x is an unknown pattern. Once the
training phase is over, the class of unknown pattern is predicted using the following
rule:

x ∈ A, if D(x) > 0
x ∈ B, otherwise.

It is mandatory that the decision functions D(x) are linear in their parameters,
however they are not restricted to linear dependencies of x. These functions can be
expressed either in the direct space, or in dual space. The notation used for direct
space is identical to the one used in perceptions, i.e.,
422 14 Statistical Learning Theory


k
D(x) = wϕ(xi ) + b. (14.9)
i=1

In the above equation, ϕ is a predefined functions of x, and w and b are the

adjustable parameters of the decision function. In the dual space, the decision func-
tions are of the form,
k
D(x) = αi K(xi , x) + b, (14.10)
i=1

where coefficients αi are the parameters to be adjusted, and xi are the training patterns.
The function K is a predefined kernel, for example, radial bias function or a potential
function.

14.4 Predicting Structured Objects Using SVM

A potential drawback of SVMs and other statistical learning methods is that they
treat the category structure as “flat”and do not consider the relationships between
categories, which are commonly used for expressing the concepts as hierarchies or
as taxonomies. The taxonomies or taxonomical structures offer clear advantages in
supporting tasks like browsing, searching, or visualization. It is to realize that all the
real-word concepts have complex hierarchical structures. For example, in browsing,
the list of answers to a query can be taken as the children of the root (keyword).
Further, each answer link would correspond to a web page, which may have many
links, thus making a tree structure.
To better appreciate the above, consider the problem of natural language parsing
as shown in Fig. 14.5, where the parser receives the input in the form of a sentence of
natural language, and the output is a parser-tree as a decomposition of the sentence
into its constituents. While support vector machines are used in NLP applications,
such as Word Sense Disambiguation (WSD), parsing is not suited as a problem to be
solved through SVMs for classification. This is because, in parsing, the output is not
a classification of yes/no but a labeled tree, representing the structure of a sentence.
So the question is, how can we take an SVM and learn predicting a parse-tree?
The question above arises not only for predicting the structured trees but also for
a variety of other structures, like DNA sequence, or sequence ordering for image
segments, etc.
Let us assume that there is multi-class SVM for document classification, where
each document belongs to exactly one category. Let {(xi , yi )}ni=1 be a set of n labeled
training documents. Here, xi ∈ d denotes a standard vector representation for the
i-th training document, and there are total d number of such documents. Each label
is yi , where yi ∈ Y ≡ {1, . . . , k}, and k is the total number of categories.
14.4 Predicting Structured Objects Using SVM 423

NP VP

Det N V NP

Det N

Fig. 14.5 Parse-tree for “The dog chased the thief”

National elections
[Musk’s] campaigning in
the [National elections] position
has motivated [many new young this group
voters.] [His] [position] on many new young voters
[bitcoin money] was well
Musk’s
received by [this group]. His bitcoin money

Fig. 14.6 Resolving the equivalent noun-phrase co-references

Let us assume that there is weight vector wy for every class 1 ≤ y ≤ k. We will
refer to the stacked vector of all weights by w = (w1 , . . . , wk ). The structured output
prediction is the term used for this kind of prediction, where our objective is to learn
a function h : X → Y , that maps the inputs x ∈ X to complex and structured
output y ∈ Y . In the structured output prediction, tasks range from predicting an
equivalence relation, say, in noun-phrase co-reference resolution, to predicting a
correct and well-formed machine translation.

Example 14.1 Resolving the equivalent noun-phrase co-references.

Consider a situation that a man called “Musk” contests in the elections and hence
campaigns to seeks votes as a presidential candidate, and suggests election mani-
festo of promoting the bitcoin money if he wins the elections. Figure 14.6 shows an
example of resolving noun-phrase co-reference, where there is an equivalence rela-
tion between a noun phrase (e.g., “Musk”) and its co-reference “His”, and similarly
between “this group”and “many young voters”.
The problems of this type exists elsewhere also, for example, in image segmen-
tation, for determining an equivalence relation, say y, over a matrix of pixels x, and
the problem of web search to predict a document ranking y for a given query x. The
structural SVMs (SSVMs) are suitable for use, as well as to address these all, and a
large range of prediction tasks with structured output. 
424 14 Statistical Learning Theory

14.5 Working of Structural SVMs

How to map the input to some structured output? Basically, a task of prediction
of a structure is similar to a task of multi-class learning. Each possible parse-tree
y ∈ Y may correspond to one class, and classifying a new example x is nothing but
predicting its correct class out of many possible classes, as shown in Fig. 14.7.
However, the problem is that there is a very large number of classes Y . In case of
parsing, the number of possible parse-trees are exponential as a function of the length
of the sentence. This situation is similar for most other prediction problems, which
output some structure. Hence, there is need for finding a compact representation of
output spaces, which are large in size. In addition, a single prediction for an example
is computationally challenging, this requires the enumeration of all the outputs. The
third challenge is, how to distinguish between two wrong parse-trees, where one may
be closer to the correct one. This, we call as prediction error [9, 11].
We can derive a structural SVM from a multi-class SVM, such that for each class y,
the multi-class SVMs use a weight vector wy , and each input x has a score for each
class y via the function,
f (x, y) ≡ wy .Φ(x), (14.11)

where Φ is a function that extracts the vector Φ(x) of binary or numeric features, from
x. Hence, every feature has additive-weighted influence in the modeled compatibility
between inputs x and classes y. For the classification of x, a prediction rule h(x)
chooses simply the highest-scoring class as the predicted output. This is expressed by

h(x) ≡ argmaxy∈Y f (x, y). (14.12)

Fig. 14.7 Resolving to S

correct structure through
structural SVM NP VP
Class 1
Det N V NP

Det N
”The dog chased the thief.”

S
Class 2 NP VP

Det N V PP

IN N

S
Class k
NP VP

NP CC NP V

Det N N
14.5 Working of Structural SVMs 425

This will result in a correct prediction of output y for input x, provided that the weights
w = (w1 , . . . , wk ) are chosen such that the inequalities f (x, ȳ) < f (x, y) hold true
for all incorrect outputs ȳ
= y. For a given training sample (x1 , y1 ), . . . , (xn , yn ), this
leads directly to a margin (a hard-margin) formulation of the learning problem by
requiring a fixed margin (= 1) separation of all training examples, while using the
norm of w as a regularizer:

1
minw w2 , such that f (xi , yi ) − f (xi , ȳ) ≥ 1, (∀i, ȳ
= yi ). (14.13)
2
For a k-class problem, the optimization problem has a total n(k − 1) inequalities,
that are linear in w. This is because, one can expand

f (xi , yi ) − f (xi , ȳ) = (wyi − wȳ ).Φ(xi ), (14.14)

which is a convex quadratic program.

The drawback of Eq. (14.13) for structured output is there is a generalization across
the inputs x, but output is without generalization due to a separate weight vector wy
for each class y. It is therefore not advisable to reduce the output prediction to multi-
class classification, as the number of possible outputs are likely to become very
large. This problem is solved using a new function Ψ (x, y) in place of Φ(x, y) to
extract features from input–output pairs. This new function is called as joint feature
map. This will yield compatibility functions due to contributions from the combined
properties of both inputs and outputs. Since the compatibility functions is defined via
f (x, y) ≡ w.Ψ (x, y), the number of parameters simply will be equal to the number
of features extracted via Ψ , and that may not depend on the number of classes |Y |.

14.6 k-Nearest Neighbor Method

The nearest neighbor method is a statistical learning method, also called sometimes
as memory-based method. The method is used for clustering of objects based on
some similarity in their attributes. The k-nearest neighbor (k-NN) or simply nearest
neighbor method is an analogical type of learning. Given a training set θ of N number
of labeled patterns, each with n attributes, a nearest neighbor algorithm decides that
some new pattern x belongs to the same category to which its closest neighbors in θ
belong. In other words, a k-nearest neighbor algorithm assigns a new pattern, x, to
that category xi to which the majority of its k-closest neighbors belong.

Example 14.2 Nearest Neighbor.

Consider there are five cars manufactured by company A, and for every car ten
important attributes have been identified, like engine size, color, wheel size, fuel
used, etc. Then, training set size is N = 5, and attribute size of each training set is
n = 10. Let a new model is introduced by company B, and the nearest neighbors’
426 14 Statistical Learning Theory

size is taken as k = 4. In this situation, a car from the company A having four
attributes common to the new car is, say model-Zeta. Therefore, Zeta being the nearest
neighbor to the new model from company B, the new model is put in the class of
model-Zeta. 
If k is relatively large, there are less chances of the decision going wrong due
to noisy training pattern close to x. However, the large values of k also reduce
the sharpness or acuteness of this method. The k-NN method can be thought of as
estimating the values of the probabilities of belongingness to class, given an input
pattern x.
The k-NN approach has been shown to be a useful non-parameteric technique
for regression and classification. In both cases, the input comprises the k-closest
neighbors of the vector. However, finding the k-NNs for a test sample, among N
design samples, is a time-consuming process, particularly for large N . The reordering
of these samples requires N (N − 1)/2 pairwise distance computations, which is a
time-consuming process for large values of N .
When used for object classification, the k-NN algorithm has many practical appli-
cations, e.g., in the areas of artificial intelligence, pattern recognition, statistics, cog-
nitive psychology, vision analysis, and medicine, to name a few. The decision rule
in k-NN provides a simple non-parameteric procedure for the assignment of a class
label to the input pattern based on the class labels represented by the k-closest (for
example, in terms of Euclidean distance) neighbors of the vector.

14.6.1 k-NN Search Algorithm

Let us assume that we are given N design samples, {x1 , . . . , xN }, each of which is n-
dimensional sample. With this, it is required to compute the k-nearest neighbors (k <
n) in a test sample y, as measured by an appropriate distance function d (y, xi ) [7].
There is no need of preprocessing of the labeled sample set in the nearest neighbor
classifier. The nearest neighbor classification rule assigns an input sample vector
y of unknown classification, to a class which is THE nearest neighbor to it (see
Algorithm 14.1). The idea of the nearest neighbor can be extended to k-nearest
neighbors with the vectors y being assigned to the class that is represented by a
majority among the k-nearest neighbors. In the algorithm, dk stands for distance
from y to xk , for k-nearest neighbor.
When more than one neighbor is taken into account, it is likely that there may
be a tie among the qualifying classes, which corresponds to maximum number of
neighbors in the group of k-nearest neighbors. There is a simple way to handle this
problem by restricting the possible values of k. For example, if there is two-class
problem, and k is restricted to odd values only, no tie can occur.
Occurrence of a tie can be handled using the following approach. An unclassified
(i.e., sample) vector is assigned to the class, of those labels that are tied, and for
that class the sum of distances from the sample to each neighbor in the class is a
minimum. There are chances that there may still remain a tie.
14.6 k-Nearest Neighbor Method 427

Algorithm 14.1 K-Nearest Neighbor Algorithm

1: Let W = {x1 , x2 , . . . , xn } //set of n labeled samples.
2: Input y //whose class is to be determined
3: Set k to any value in 1...n
4: Let S = φ //initial set of k-nearest neighbors (for given k)
5: for i = 1 to n do
6: compute distance di (y, xi ) from y to xi
7: if (di ≤ dk ) then
8: S = S ∪ {xi }
9: else
10: if xi is more close to y than some previous neighbor then
11: Delete farthest in the S
12: S = S ∪ {xi }
13: end if
14: end if
15: end for
16: Determine class in S that is in majority
17: if (there is tie) then
18: Compute the sum of distances of neighbors in each class with the one having tie
19: if (there is no tie) then
20: Classify y into minimum found class
21: end if
22: else
23: Classify y into the class of majority
24: end if
25: End

For numerical attributes, the distance metric used in the nearest neighbor is
Euclidean distance. Considering two patterns, x1 = x11 , x12 , . . . , x1m , and x2 =
x21 , x22 , . . . , x2m , distance between two using Euclidean is given by


 m
d (x1 , x2 ) =  (x1j − x2j )2 . (14.15)
j=1

To keep the spread of attribute values along each dimension approximately same,
the above value of the distance is usually modified by scaling of the features. In
that case the distance is represented by the following equation, where aj2 is the scale
factor for the dimension j.


 m 2
d (x1 , x2 ) =  aj (x1j − x2j )2 . (14.16)
j=1

Example 14.3 Finding the nearest neighbor with k patterns.

Figure 14.8 shows a problem of k-nearest neighbor classification, where k = 6. There

are two patterns of each categories 1, 2, and 3, with numbers 1, 2, 3 marked after
each pattern symbol, respectively. A collection of dots makes a pattern. For example,
428 14 Statistical Learning Theory

Fig. 14.8 k-nearest Training patterns

neighbor problem
y pattern
to be classified
2 3 3
2
3 3
2 2 3
3
2 3
2 3
1 2 3
1 2
2
1
1 1

1 1

in the two patterns of category 1, there are three and four dots, each followed by 1s.
The pattern y, represented by a box, is to be classified. In the circle enclosing,
since the majority are in the category of 1, the new pattern y should be classified in
category 1. 

14.7 Naive Bayes Classifiers

The Bayesian networks are powerful tools for decision-making and reasoning under
uncertainty. These networks are specified by two components:

– Graphical component. It expresses uncertainty of causal relations, and consists

of Directed Acyclic Graph (DAG) to represent causal relations and conditional
probabilities of each node, given the parent of each. The vertices represent events
and edges which are the relations between them.
– Numerical component. It quantifies different links in the DAG through conditional
probability distribution of each node in the contexts of its parents.

Basically, the Bayes theorem permits optimal prediction of a class of new

instances, given a vector x = {x1 , ..., xn }, of attribute’s values. Since there is always
insufficient training data to obtain accurate prediction of full joint probability dis-
tribution, the straight forward application of Bayes theorem is impracticable for
machine learning. Therefore, there is need for making assumptions of independence
to make the inference feasible.
The Bayesian classifier, called Naive Bayes, i.e., “straw man”, approach takes
this to the extreme when it is assumed that attributes are statistically independent,
given the value of the class attribute. However, the above assumption never holds
true, but still the Naive Bayes performs exceptionally well in most classification
problems. Apart from this, the Naive Bayes approach is computationally efficient,
as the training is linear in both the number of instances and attributes, and simple to
implement [4].
14.7 Naive Bayes Classifiers 429

Fig. 14.9 Naive Bayes yi

network structure

x1 x2 xn

A Naive Bayes is composed of a DAG with one root node (called parent), which is
an unobserved node representing class of each object to which the testing set belongs,
and several leaf nodes, corresponding to observed nodes, with strong assumption of
independence among these leaf nodes in the contexts of their common parent. The
leaf nodes represent different attribute (features) specifying this object. Note that
the Naive Bayes are composed of two levels only as shown in Fig. 14.9, whereas a
Bayesian network may consist of many levels.
The decision of a Bayesian classifier is represented as a matrix of P(xj |yk ), which
specifies the probability of occurrence of each feature value (x1 , . . . , xj , . . . , xn ) given
each class yk . To classify a new example having features among x1 . . . xn , we make
use of Bayes theorem as

 P( xj |yi )P(yi )
argmaxyi P(yi | xj ) = argmaxyi  (14.17)
k P( xj |yk )P(yk )

that computes P(yi | xj ), which is the probability of the example in class yi given the
features xj . The subexpression xj denotes a conjunct of attribute values all occurring
in an example. The summation is performed over N (1 ≤ k ≤ N ) classes, and the
above probability is calculated for each class yi , and then the class of the highest
probability is selected. The probability P(yk ) is estimated from the distribution of
the training examples among classes. If independence
 of all attributes is assumed
under a given context (i.e., class), then P( xj |yk ) can be calculated using


P( xj |yk ) = P(xj |yk ). (14.18)

j

The values P(xj |yk ) are calculated from the probability matrix. Here, xj is the evidence
on attribute nodes, which can be dispatched into pitches of n evidence of features,
x1 , . . . , xj , . . . , xn . Because Naive Bayes is working on the assumption that these
attributes are independent of each other (given the parent node), their combined
probability is obtained by substituting Eq. (14.18) into (14.17), which results in,

 j P(xj |yi )P(yi )
argmaxyi P(yi | xj ) = argmaxyi 
k [ j P(xj |yk )P(yk )]

= argmaxyi P(xj |yi )P(yi ). (14.19)

j
430 14 Statistical Learning Theory


Note that, there is no need to explicitly compute the denominator k [ j P(xj |
yk )P(yk )], since it is common among all the computation being a normalization
constant. In some practice, log is computed instead of a product, because probabilities
involved can be very small.
The Bayesian learning method, as a classifier, builds the matrix P(xj |yk ) from
training examples, by examining the frequency values in each class. It is possible to
compute this matrix incrementally by incorporating one instance at a time. Alterna-
tively, it can be constructed (i.e., without incremental approach), using all the data
at a time.
Due to its simple structure, the Naive Bayes has many advantages. It is efficient,
as the inference (i.e., classification) is achieved in a linear time. However, the Bayes
network with general structure has complexity of NP-complete. Naive Bayes con-
struction is incremental, in the sense that it can be easily updated, e.g., addition of
new cases. The major problem with Naive Bayes is the assumption of strong indepen-
dence relation, i.e., assumption of independence of features in the context of session
class is not always true, and leads to negative influence on the inferred results.
Naive Bayes is most commonly used in text classification, where words are fea-
tures, and presence/absence of a word can be used to determine the topic of a docu-
ment.
Given N number of training examples, each with n number of attributes, com-
plexity of generating probability matrix for Bayes classifier is O(n.N ). Hence,
the classifier is substantially faster as the runtime is independent of the decision
“rule”generated. Apart from this, the basic operation is performed only once.

14.8 Artificial Neural Networks

The science of machine learning is mostly experimental as there is a no universal

learning algorithm existing yet. That is clear from the fact that, given a number of
tasks, none can make a computer to learn every task well. A knowledge-acquisition
algorithm is always required to be tested on learning tasks and data, that are specific
to a given situation, and it is irrespective of whether it is recognizing a sunrise or
it is doing a language translation. There is no method to prove that the given algo-
rithm will be consistently better for all the situations. However, the human behavior
apparently contradicts. We are fairly good at general learning abilities due to which
we are able to master a number of tasks, like playing chess and playing cards. These
arguments suggest and might serve as inspirations for building machines with some
form of general intelligence. Therefore, the use of Artificial Neural Networks (ANN),
which is a brain model, appears to be a logical justification for building intelligence
systems [8].
The basic unit of the brain for performing the computation is a cell, called neuron;
each one of them sends a signal to other neurons through very small gaps between the
cells, called synaptic clefts. The property of any neuron of sending a signal through
14.8 Artificial Neural Networks 431

this gap, and the amplitude of the signal together, is called as synaptic strength. As
a neuron learns enough, its synaptic strength increases, and in that situation, if it
is stimulated by an electrical impulse, there are better chances that it would send
messages to its neighboring neurons.
The current neural-based learning algorithms need close involvement of humans,
for producing better results. Majority of these algorithms are based on supervised
learning, where each training example is carried out using human-crafted labels,
about what is being learned. Consider an example of a picture of sunrise associated
with the caption: “Sunrise”. In that instance, the goal of the learning algorithm is to
take the photograph as an input, and produce output, the name of the object in the
image, i.e., “sunrise”. We know that the mathematical process of transforming an
input to the output is a function. The synaptic strength, which is a numerical value,
produces this function, which corresponds to the solution to the learning through
ANN.
It is interesting to note that it can be achieved through rote learning also, but it
would not be useful. In fact, when we want to teach to the algorithm “what the sunrise
is”, then to have the algorithm recognize any sunrise, even the one for which we have
not trained! This is the ultimate goal of machine learning algorithm.
The Artificial Neural Networks (ANN) possess the following important
properties:
Learning ability,
Massive parallelism,
Adaptability,
Fault tolerance,
Distributed representation and computation,
Generalization ability, and
Low energy consumption.
which make them candidate for many applications. Although the details of the pro-
posals vary, the most common models for learning and computation take the neuron
as the basic processing unit. Each processing unit is characterized by the following:

– an activity level to represent polarization state of a neuron,

– an output value to represent firing rate of the neuron,
– a set of input connections,
– synapses on the cell and its dendrite,
– a bias value to represent an internal resting level of a neuron, and
– a set of output connections to represent a neuron’s axonal projections.

Each of these aspects of the unit are represented mathematically by real numbers.
Hence, each connection of a neuron has an associated weight, called synaptic
strength, which influences the effect of the incoming input on the activity of the unit.
This weight is either positive, called excitatory or, negative, called inhibitory.
The basic model of artificial neuron with binary threshold is shown in Fig. 14.10.
The mathematical neuron computes the weight as the sum of its n number of input
432 14 Statistical Learning Theory

Fig. 14.10 Basic model of x1 w1

an artificial neuron
x2 w2 y
h
wn u
xn

Fig. 14.11 Single-layer

feedforward and feedback
networks

Feedforward Feedback
network
network

signals x1 , . . . , xn , and the generated output is 1 if the sum is above some threshold
value u, otherwise the output is zero. This can be represented by


n

y=θ wi xi − u (14.20)
i=1

where θ (.) is called a unit step function at 0, and wi is the synapse weight associated
with the ith input. For the sake of simplicity, we consider the threshold u as another
weight, w0 = −u attached to the neuron with a constant input x0 = 1. A properly
chosen weight allows a synchronous arrangement of such neurons to perform uni-
versal computations. There is a crude analogy of this neuron model to biological
neurons as follows: the wires and interconnections model the axons and dendrites,
respectively, in the biological neuron; connection weights in this model correspond
to synapses in biological neuron; and threshold function approximates the activity in
soma. However, this model is an overly simplified one of a true biological neuron.
The ANNs can be considered as weighted directed graphs, where artificial neurons
act as nodes, and directed edges with weights are connections between neurons, and
between outputs and inputs of neuron.
Based on the connection pattern an ANN can be classified as
1. Feedforward networks: In these, the direct graphs have no loops, and
2. Recurrent feedback networks: There are loops because of feedback connections.
Figure 14.11 shows the single-layer feedforward and feedback neural networks.
The most common family of feedforward networks is the multilayer perceptron. The
neurons in these networks are organized into layers with unidirectional connections
between them. Usually, the feedforward networks are of a static type, as they produce
only one set of output values instead of a sequence of values for a given input. These
networks are without memory, as their output is independent of the previous states
of the network.
14.8 Artificial Neural Networks 433

The feedback networks are called recurrent networks, and are dynamic systems.
Neurons’ output patterns are computed when a new input pattern is presented to these
networks. Due to feedback paths, the input to each neuron is then modified, which
causes the network to enter a new state.
The problem of learning in neural networks is simply the problem of finding
a set of connection strengths which allows the network to carry out the desired
computation [10].
An ANN usually learns of the connection weights from the available training pat-
terns, the performance of which gets improved over time through iterative updating
of the weights. The ability of ANNs to automatically learn from examples makes
them attractive and exciting. From a given collection of representative examples, the
ANNs learn underlying rules as a form of input–output relationship, instead of fol-
lowing the rules specified by human experts. This simple factor is a major advantage
of neural networks over traditional experts systems.
To understand this learning process, first there is need of a model of learning
environment in which these neural networks operate. That means, we must know
about what information is available to the network. This model is called the learning
paradigm. The second requirement is to know about how these weights are updated.
This means to know, as to which rule controls the updating process of the weight.
In fact, a learning algorithm refers to some procedure which uses learning rules to
adjust the weights of inputs.
All the three paradigms of learning exist in the neural networks, i.e., supervised
learning, unsupervised learning, and hybrid learning. Supervised learning requires
a teacher, where the network is provided with correct answers, i.e., output, for every
input pattern. The weights are so determined that they allow the network to produce
answers, which are as close as possible to known correct answers.
The unsupervised learning-based neural networks explores the underlying struc-
ture in the data, or correlations between patterns in the data, and organizes patterns
into categories based on the correlations.
A hybrid learning ANN combines both the supervised and unsupervised
approaches.
The learning rules in the neural networks are of four basic types. These are error
correction, Boltzmann, Hebbian, and Competitive learning. Their basic principles
are presented in the following.

14.8.1 Error-Correction Rules

A learning based on error-correction rules makes use of simple concepts. During the
training, the input x1 , . . . , xn is applied to the network, and the flows through the
network generates a set of values in the units of output y. As the next step, the actual
output y is compared with the desired target d . If the output and the target match,
no change is made to weights. If there is no match, change is made to weights of
some of the connections. The problem is to find out as which connections in the
434 14 Statistical Learning Theory

network were at fault that caused the error and resulting in a mismatch. Obviously,
it is supervised learning paradigm. The principle used for error-correction learning
rules is based on the error signal (d − y), to modify the connection weights so as to
gradually reduce the error magnitude.
The perception-based learning works on this principle of error correction. A
perceptron comprises a single neuron with adjustable weights, wi , i = 1, . . . , n, and
a threshold value u, as shown in Fig. 14.10 (page no. 432). Given an input vector x =
(x1 , . . . , xn )t , where t is the iteration number, net input to the neuron is expressed as


n
v= wi xi − u. (14.21)
i=1

If v > 0 the output y of the perceptron is +1, otherwise it is 0. In a classification

problem (of two classes, say A and B), the perceptron assigns an input pattern to
class A if y = 1, and to class B if y = 0.

Algorithm 14.2 Perceptron learning Algorithm

1: Initialize weights w1 , ..., wn , and threshold u to some small random numbers,
2: Apply a small pattern input vector (x1 , . . . , xn )t and evaluate the output of the neuron, as per
equation (14.21),
3: Update each weight wi (i = 1, ..., n) according to: wi (t + 1) = wi (t) + η(d − y)xi .

Algorithm 14.2 is backpropagation learning algorithm, based on error-correction

principle, where d is the desired output, t is the iteration number, and η (0.0 < η <
1.0) is the gain, i.e., size of the step.

14.8.2 Boltzmann Learning

The learning in ANN based on Boltzmann machines has many properties: they are
symmetric and recurrent (i.e., feedback) networks. It consists of binary units (+1 for
“on”, and −1 for “off”). The symmetric property means, weight on the connection
from unit number i to unit j is identical to weight from unit j to unit i, formally,
wij = wji . The neurons are also of two types: 1. Visible neurons are a subset of the
entire set of neurons, and they interact with the environs, and 2. Hidden neurons are
the remaining neurons, which do not interact. Each neuron is a stochastic unit, which
generates and outputs (or it is a state) according to the Boltzmann distribution of
statistical machines.
A Boltzmann machine also operates in one of two modes: 1. Clamped mode, where
visible neurons are clamped onto a specific state determined by the environment; and
2. Free-running mode, in which both the visible and hidden neurons are allowed to
operate freely.
14.8 Artificial Neural Networks 435

The Boltzmann learning algorithm works to adjust the connection weights in such
a way that the desired probability distribution is satisfied by the states of the visible
units. According to the rule of Boltzmann learning, the change in the connection
weight wij is given by

Δwij = η(ρ ij − ρij ), (14.22)

where η is learning rate, ρ ij and ρij are correlations between the states of units i and j
when the network operates in the clamped mode and free-running mode, respectively.
The values of ρ ij and ρij are estimates using Monte Carlo experiments [8].

14.8.3 Hebbian Rule

The Hebbian learning rule specifies the magnitude of the weight by which the con-
nection between two units is increased/decreased in proportion to the product of
their activation. It builds on the Hebbs’s learning rule, which states that the connec-
tions between two neurons might be strengthened if the corresponding neurons fire
simultaneously. This rule works well as long as the input patterns are uncorrelated,
however, this condition places serious limitations on the Hebbian learning rule. A
Hebbian rule for ANN is described as

wij (t + 1) = wij (t) + ηyj (t) xi (t), (14.23)

where xi , yj are the output values of neurons i and j, respectively. These are connected
by the weight wij , and η represents the learning rate. An important property of this
approach is that learning is carried out locally, that is, the change in synapse weight
depends only on the activities of two neurons connected by it. This simplifies the
implementation of the circuit.
A more powerful learning rule is the delta rule, that utilizes the discrepancy
between the desired and actual output of each output unit to change the weights
feeding into it.

14.8.4 Competitive Learning Rules

The competitive-learning units compete among themselves for activation, which is

in contrast to the Hebbian learning, where multiple output units can be fired together.
Hence, only one output unit is active at any given time. The biological neurons follow
this type of learning.
The competitive learning often categorizes or clusters the input data, where similar
patterns are grouped by the network and represented by a single unit. The grouping
is carried out automatically, which is actually based on the correlations.
436 14 Statistical Learning Theory

The simplest possible competitive learning network has a single layer of output
units, as shown in Fig. 14.11, where each output unit j connects to all the input units
xi s, through the weights wij , i = 1, . . . , n. Each output unit also connects to all other
units via inhibitory weights, but has a self feedback with an excitatory weight. Due
to the competition, only the unit i with the largest net input becomes the winner. This
can be expressed by
∀i wi∗ x ≥ wi x. (14.24)

In this learning, only the weights of the winner unit gets updated.

14.8.5 Deep Learning

Beginning from 2005, deep learning—a neural net-based approach, which is driven
for its inspiration from brain science—began to come into its own, and has now
become a singular force propelling AI research forward.
The deep learning is concerned with simulation of ANNs that “learn” gradually,
in the areas of image processing, speech recognition, and understanding, and even to
make the decisions of their own. The basic technique relies on ANNs, which do not
precisely mimic as to how the actual neurons are working. Instead of this they are
based on the general principles of mathematics that allow them to learn from examples
to recognize people or objects in a photograph, and translate the spoken language
from one to another. The technologies based on deep learning have transformed the
AI research, and have produced far accurate results in speech recognition, computer
vision, natural language processing, and robotics.
To be successful in generalizing after observing a number of examples, deep
learning network needs more than just the examples. For example, it depends on
hypotheses about the data and assumptions about what can be a possible solution
for a particular problem. A typical hypothesis that can be built into a system might
conclude that if data input for a particular function in two situations are almost similar,
the output should not change drastically. For example, on altering a few pixels in an
image of a dog will not transform it into a picture of cat.
One type of a neural network that consists of hypotheses about images is called
convolutional neural network. These networks when used in deep learning have
many layers of neurons, which are organized in such a way that the output is less
sensitive to the deviation from the original object, due to changes in the input image.
For example, we will note the changes in a face when viewed from different angles,
however we will still recognize it correctly. A well-trained deep learning network
will also do a similar job.
The design of convolutional networks take their inspirations from multilayered
structure of visual cortex—part of the brain that receives input from eyes. Too many
layers of virtual neurons in a convolutional neural network are what that make the
network “deep”, and hence it is better able to learn about the world about it.
14.9 Instance-Based Learning 437

14.9 Instance-Based Learning

Instance-based learning (IBL) approaches uses supervised learning techniques. There

are several variants of IBL, e.g., exemplar-based learning, case-based reasoning, and
memory-based learning. Though all these methods emphasize somewhat different
aspects, all these approaches are founded on the basic concept of an instance or a
case, as a basis of knowledge representation and reasoning. The meaning of case
here is, observation or example, or incident, which is a single experience, e.g., a
pattern, along with its solution, is a problem of pattern recognition.
In general, a problem along with its solution is a case-based reasoning. To highlight
the main properties of IBL, it is important to understand its difference with model-
based learning. As a typical case, IBL methods learn by simply storing some of the
observed examples. The processing of these inputs are differed until a prediction or
some other query is actually requested. Later, predictions are derived by combining
information from stored examples, in some way. Once the query is answered, the
prediction and intermediate results are discarded.
In contrast, the model-based or inductive-based learning derive predictions in an
indirect way as follows: as a first step, observed data is used in order to induce a
model, say, a decision tree or a regression function. As a second step, the predictions
are obtained using this model, which can also serve as the function of explaining.
Generally, the model-based algorithms, also called eager algorithms, carry higher
complexity during the training phase than the instance-based algorithms. The latter
are also called lazy algorithms, where learning is basically storing of selected algo-
rithms. The lazy methods also need more storage requirements, in a linear order
of the size of the input, and higher computational cost compared to deriving of a
prediction.
The IBL algorithms make use of specific instances, instead of pre-compiled
abstractions during the prediction. They also describe the probabilistic concepts,
because they use similarity functions to yield graded matches between instances.
The IBL algorithms are derived from the nearest neighbor (NN) pattern classifiers,
which also do not save and use only selected instances to generate the classification
predictions. Thus, they are also called as edited NN algorithms. They also maintain
the prefect consistency with the initial training set.

14.9.1 Learning Task

The instance-based learning makes use of the supervised approach, or learning from
examples. The input is a sequence of instances, where each instance is a set of n
attribute–value pair, thus creating an n-dimensional instance space. Only one of these
attributes corresponds to a category attribute, and the other attributes are predictor
attributes. A category is a set of all instances in the instance space which have the
same value for their category attribute. For the sake of simplicity, it is assumed that
categories are disjoint.
438 14 Statistical Learning Theory

The main output of IBL algorithms is concept or concept description. The algo-
rithm is a function that maps instances to categories, i.e., for a given instance from
a instance space, it produces the classification—a predicted value for instance’s cat-
egory attribute.
An instance-based concept description comprises a set of stored instances and,
some information about their past performances, e.g., the number of correct and
incorrect classification predictions. The set of instances can change once each training
instance has been processed. The concept descriptions are decided based on how the
selected similarities and classification functions of the IBL algorithm make use of the
current set of saved instances. In all the IBL algorithms, at least two of the following
three components constitute these functions [1].
Function of Similarity
This function computes the similarity between two instances: one is a training
instance, and the other is in the concept description. The similarities are represented
in numeric values.
Classification Function
The input to this function are 1. similarity function’s result, and 2. classification
performance records of instances in the concept description, which produces output
of the classification.
Concept Description Updater
An updater keeps records of classification performance and resolves as to which
instance to include in the concept description. Inputs to the concept description
updater are training instance, similarity results, classification results, and current
concept description. The output of the updater is the modified concept description.
The first two functions in the above decide how the saved instances in the concept
description can be used to predict the category attributes. Thus, concept description
in IBL not only comprises a set of instances, but also these two functions.
There is an important difference between IBL algorithms and most other super-
vised learning methods: the IBL algorithms do not construct explicit abstractions,
like decision trees or rules. Most of the learning algorithms produce generalizations
using the instances, and use simple matching procedures to classify the instances
when presented in future. This eliminates the need of storing rigid generalizations
of concept descriptions for IBL algorithms.

14.9.2 IBL Algorithm

Algorithm 14.3 is the simple instance-based learning algorithm. The similarity func-
tion, sim(x, y), used in the algorithm is expressed by
14.9 Instance-Based Learning 439

 n

sim(x, y) = − f (xi , yi ). (14.25)
i=1

To compute the similarity, the instances have n attributes. For numerical attributes,
the relation used is
f (xi , yi ) = (xi , yi )2 . (14.26)

For attributes with Boolean and symbolic values, f (xi , yi ) = (xi

= yi ), i.e., f is false
when xi
= yi . The missing attribute values are taken as having maximum difference
from the value present. If both are missing, then f (xi , yi ) = 1. IBL Algorithm 14.3
is very similar to the nearest neighbor algorithm (see page no. 425) we have studied
earlier. The only difference in the IBL algorithm is that it normalizes its attributes’
ranges, processes instances incrementally, and has a simple policy that allows missing
values of attributes [1].
It is important to note that this algorithm’s concept description changes over
time. In the k-NN algorithm, the classification function simply assigns classifications
according to the nearest neighbor policy. In the IBL algorithm, we can find out
instances in the instance space that will be classified by each of the stored instances.
The term CD in the algorithm stands for concept description, and TS is the training
set.

Algorithm 14.3 IBL algorithm

1: Initialize: CD = φ
2: for every x ∈ TS do
3: for every y ∈ CD do
4: SIM [y] = sim(x, y)
5: end for
6: ymax = ∃ y ∈ CD with maximal SIM [y]
7: if class(x) = class(ymax ) then
8: classification = True
9: else
10: Classification = False
11: end if
12: CD = CD ∪ {x}
13: end for
14: End

14.10 Summary

Machine learning systems help the programs automatically learn from the data; it can
be said to be a kind of a data mining. Machine learning is used in computer science
and beyond, e.g., in search engines, spam filters, advertisements, credit scoring, fraud
detection, stock trading, drug design, and in many other applications.
440 14 Statistical Learning Theory

A learning algorithm can be expressed as comprising three parts:

1. Representation,
2. Evaluation, and
3. Optimization.
The popular approaches of statistical machine learning are support vector machine,
k-nearest neighbor algorithm, Naive Bayes, and instance-based learning.
The support vector machines (SVMs) are supervised learning models which use
associative learning algorithms, which can analyze the data and recognize patterns.
Thus they have applications in classification and regression analysis.
An SVM is a model where examples are nothing but a representation of points
in space, which are mapped to two classes, so that the examples of the separate
categories are divided by a clear gap that is as wide as possible. The latter is to
help in making clear distinctions between the categories. New examples are then
mapped into that same space and predicted about their category based on which
side of the gap they fall on. The support vector machine (or support vector network
(SVN)) maps the input vectors into some high-dimensional feature space through
some nonlinear mapping chosen a priori. A limitation of SVMs and other statistical
learning methods is that the category structure is treated as “flat”and that they do
not consider any relationships between categories, which are commonly expressed
in concept hierarchies or taxonomies. The structured SVMs are capable of learning
the taxonomical architectures.
The problem with taxonomical structures is that the number of classes are very
large. For example, in parsing, the number of possible parse-trees is the exponential
factor of the length of the sentence, and this scenario is similar for a majority of other
problems that are designed to predict the output, which is structured in nature. Thus,
there is need for exploring a more compact representation for these large output
spaces.
The k- nearest neighbor method is a statistical method, where, given a training set
θ of n labeled patterns, a nearest neighbor algorithm decides that some new pattern, x,
belongs to the same category, as do the closest neighbors in θ . The k-NN algorithms
are used for the classification of objects, in many practical applications, e.g., in the
areas of artificial intelligence, pattern recognition, statistics, cognitive psychology,
vision analysis, and in medicines. Under many circumstances, the k-NN algorithm
is used to perform the classification.
Bayes networks are powerful tools for decision-making and reasoning under
uncertainty. These networks are specified using two components: 1. a graphical com-
ponent, composed of directed acyclic graph (DAG) to represent causal relations, and
2. numerical component, consisting in a quantification of different links in the DAG
by conditional probability distribution. The Bayesian classifier (Naive Bayes), i.e.,
“straw man” assumes that the attributes in the examples are statistically independent
of each other, given the value of the class attribute, which makes it computationally
efficient.
14.10 Summary 441

The most common models for learning and computation take the neuron as the
basic processing unit. Each such processing unit has following characteristics:
1. Activity level. It is the state of polarization of a neuron.
2. Output value. It depends on the firing rate of the neuron.
3. Input connections. It is the collection of synapses on the cell and their dendrite.
4. Bias value. It is an internal resting level of the neuron.
5. Output connections. These are neuron’s axonal projections.
The Artificial Neural Networks (ANNs) are weighted directed graphs, where
nodes are treated as artificial neurons and directed edges (with weights) are connec-
tions between neurons, and also they act as connections between input and outputs.
The problem of learning in neural networks is the problem of finding a set of con-
nection strengths which can allow the network in future to carry out the desired
computation. There exist all three paradigms for learning in a neural network: super-
vised, unsupervised, and hybrid.
Instance-based learning (IBL) approaches is supervised machine learning tech-
nique. Several variants of instance-based approaches have been devised, e.g., memory-
based learning, exemplar-based learning, and case-based reasoning. The informa-
tion provided by the stored examples is used in some way, to perform the predictions
in the IBL. Once the query has been answered, the prediction itself and other inter-
mediate results are discarded.
As opposed to IBL, the model-based or inductive learning methods derive pre-
dictions in an indirect way: in step one, the observed data is used in order to induce a
model, say a decision tree or a regression function. Then in the second step, the pre-
dictions are obtained on the basis of this model, which can also serve other purposes
like explaining or justifying the inferences.

Exercises

1. “The task of text categorization is to assign a given document to one of the

categories out of a fixed set of categories. This is done on the basis text contents.
The Naive Bayes model is often used for this purpose, where a query variable is the
document category and the “effect” variables are presence/ absence of each word
in the language. It is assumed that words occur independently in the documents,
and their frequencies determine the document category.“ For this statement,
a. Explain how such a models can be constructed, given a set of “training data”
in the form of documents that have been already assigned to categories.
b. Explain how to categorize a new document.
c. Is the independence assumption reasonable? Justify your answer.
2. What linear or nonlinear function is used by an SVM for performing classifica-
tion? How is an input vector xi (instance) assigned to the positive or negative
classes.
442 14 Statistical Learning Theory

Fig. 14.12 Training data for (1, 1)

SVMs (0, 1) (0, 1)

(0, 0) (0, 0)
(1, 0) (1, 0)
(a) (b)

Table 14.1 Data set for machine learning

Department Class Age Salary K$s
Sales Programmar 36 . . . 40 51 . . . 55
Sales Assistant 31 . . . 35 31 . . . 35
Sales Assistant 36 . . . 40 36 . . . 60
Production Assistant 26 . . . 30 51 . . . 55
Production Programmar 36 . . . 40 71 . . . 75
Production Assistant 31 . . . 35 51 . . . 55
Marketing Programmar 41 . . . 45 51 . . . 55
Marketing Assistant 36 . . . 40 46 . . . 50

3. Consider the SVM for the training data given in 2 , in Fig. 14.12a, b; find out the
separating hyperplanes in both the cases.
4. Why is the Naive Bayes classification called Naive? What are the main goals
behind this classification?
5. Consider the data given in Table 14.1,
and use these to train a Naive Bayes classifier with designation attribute as the
class label and all the remaining attributes regarded as input. Once you have your
Naive Bayesian classifier, test the following unseen instances to find out the class:
a. Marketing, 36 . . . 41, 51K . . . 55K
b. Sale, 36 . . . 41, 71K . . . 75K

References

1. Aha DW et al (1991) Instance-based learning algorithms. Mach Learn 6:37–66

2. Boser EB et al (1992) A training algorithm for optimal margin classifiers. In: Proceedings of
the fifth annual workshop on computational theory, COLT’ 92. ACM, New York, pp 144–152
3. Bradley P (2002) Scaling mining algorithms. Commun ACM 45(8):38–43
4. Clark P, Niblett T (1989) The CN2 induction algorithm. Mach Learn 3:261–283
5. Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20:273–297
6. Domingos P (2012) A few useful things to know about machine learning. Commun ACM
55(10):78–87
References 443

7. Fukunaga K, Narendra PM (1975) A branch and bound algorithm for computing K-nearest
neighbors. IEEE Trans Comput 750–753
8. Jain AK et al (1996) Artificial neural networks: a tutorial. Computer 3:31–46
9. Joachims T (2009) Predicting structured objects with support vector machines. Commun ACM
52(11):97–104
10. Rummelhart DE et al (1994) The basic ideas in neural networks. Commun ACM 37(3):86–92
11. Shawe-Taylor J (2009) Machine learning for complex predictions. Commun ACM 52(11):96–96
Chapter 15
Automated Planning

Abstract Automated planning deals with the tasks of finding out ordered sets of
actions that allow a system to transform an initial state to a state satisfying goal spec-
ification. The set of actions is a plan, and it belongs to PSPACE-complete. Automatic
planing or scheduling generates a set of actions automatically. This chapter presents
the nature of automated planning, the classical planning problem, agent types that
execute the problem, and worked examples. It also covers, the concepts and imple-
mentation aspects of forward planning, partial-order planning, planning languages,
a case study of general planning language—STRIPS, and search strategies. Plan-
ning with propositional logic, planning graphs, and hierarchical network planning
are demonstrated. The multiagent planning techniques are presented for goal and
task refinement, decentralized planning, and on how to do coordination after it is
planned. This is followed by the chapter summary, and a set of exercises.

Keywords Automated planning · Scheduling · Forward planning · Partial-order

planning · Planning languages · STRIPS · Multiagent planning

15.1 Introduction

Automated planning is concerned with the problem of finding a set of actions to

be carried out in an ordered way to complete a job, such that they allow a system
to transform an initial state into a state that satisfies the goal specification. In a
deterministic system, these sets of actions are called plans, while in nondeterministic
systems they are called policies. Finding a plan, or even deciding its existence,
has shown to be PSPACE-complete1 in the deterministic planning, unless severe
restrictions are applied to the search [1].
Developing automated methods for reasoning about plans and schedules and for
generating the plans has remained the part AI, which helped both the autonomous

1 Inthe theory of computational complexity, a decision problem is in complexity class PSPACE-

complete if it can be solved using a memory whose size is polynomial on the size of input (i.e.,
polynomial space), and if every other problem that can be solved in polynomial space can be
transformed into it in polynomial time.
© Springer Nature India Private Limited 2020 445
K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_15
446 15 Automated Planning

and human agents. There is always a need for planning when an agent wants to
control the evolution of its environment. An algorithm for a planning problem has
the following inputs:
1. possible courses of actions,
2. a predictive model that underlies the dynamics, and
3. a performance measure that evaluates the courses of actions.
The output or solution to a planning problem is one or more courses of actions that
satisfies the specified requirements for the minimum performance. Hence, a planning
problem involves deciding “what” actions to be performed and “when” they should
be performed. The “when” part of the problem has been traditionally called the
scheduling problem. The separation between “what” and “when” is motivated by
computational considerations. In fact the algorithm for automating scheduling has
been around for a long time, for example, in the areas of operations research as well
for automating the complete planning problems.
The classical scheduling methods are concerned with the planning problems where
resource constraints are static, and prespecified in numerical form. This is in con-
trast to the AI methods for scheduling that allow declarative specifications for both
symbolic and numeric resource constraints and handle dynamic changes in the con-
straints, i.e., while in execution. The Constraint Satisfaction Problems (CSFs) (see
Chap. 10) form the canonical backbone of most AI scheduling methods, and there
are several sophisticated heuristic search strategies that are common for such prob-
lems [11].
Although the classical panning model has been prevailing for a long time, a major-
ity of research in planning is toward planning in environments that are dynamic,
stochastic, and in a partially observable world. These requirements are not compat-
ible with the classic planning assumptions, which are for deterministic, static, and
fully observable world. To cope with the partially observable environments, gathering
of information is designed as part of the planning activity, and the existing classical
planning techniques are extended to accommodate interleaving of scheduling and
planing tasks. The environments that are stochastic are modeled using Markov deci-
sion processes, while the planning in such environments requires construction of
policies for the corresponding Markov decision processes [6].
All variants of domain-independent planning problem are known to be computation-
ally hard (P-SPACE complete or worst); efficiency can be improved only through
exploitation of problem distribution space and through the knowledge of domain
structure. For improving the efficiency of plan generation, formal planning models
are preferred. Following are the basic approaches for achieving efficiency enhance-
ment:
1. split both, domain and the problem into tractable sub-components using decom-
position techniques,
2. find out the control information using inductive and speedup learning techniques,
3. define abstractions for expected problem distribution, and
4. define language(s) to express domain-specific search control information.
15.1 Introduction 447

Another approach to planning is extending and complementing the planning

models that are based exclusively on techniques that define subgoals. A promising
approach is using constraint-logic-programming-type models that allow both: defin-
ing subgoals and constraint-satisfaction criteria. These work together in controlling
the planner, which results in handling action selection and scheduling of planning
problems through appropriate computational models. In addition to the above, plan
generation can also be carried out using network flow approaches, which are based
on a form of disjunctive projection of future states of the agent [9].
So far, both the classical and stochastic planning techniques are in common used
in the design of autonomous agents. The AI planning techniques are also common in
software domains such as database query planning and Internet browsing, however,
they have less impacted the problems that support industrial applications. Although
the industrial area has an enormous scope for planning problems—be it project
planning, process planning, or maintenance planning—the automation so far has
progressed to supporting scheduling, particularly concentrating on harder action-
selection problem, which are more concerned with the humans domain.
The type of plannings, which are incremental or interactive planning in nature,
are new areas in plannings. However, many planning applications involve reason-
ing with a variety of constraints that are not of temporal type, e.g., the problems
of path planning, assembly planning, and manufacturing problems need spatial and
geometric-reasoning capabilities. Due to the nature of these constraints, it is neither
possible to ignore or abstract away, nor it is advisable to encode them in a homo-
geneous representation. These conditions make it necessary to have an interaction
between the planner and the reasoners. Hence, understanding of the modalities of
interaction between a planner and the external reasoners, which may be both humans
or machines, is thus important [7].
Learning Outcomes of This Chapter:
1. Define the concept of a planning system and how it differs from classical search
techniques. [Familiarity]
2. Describe the differences between planning as search, operator-based planning,
and propositional planning, providing examples of domains where each is most
applicable. [Familiarity]

15.2 Automated Planning

Automated planning techniques are now commonly being applied in a number of

tasks, that include, robotics, process planning, web-based information gathering,
autonomous agents, and spacecraft mission control. A solution to a problem in auto-
mated planning can be described in terms of a sequence of steps that transforms
some initial description of the problem state, for example, the initial configuration of
a puzzle, into a description satisfying a specified goal criterion. The steps (transfor-
mations) are called operators and a problem is defined in terms of a set of operators
448 15 Automated Planning

and a language for describing problem states. The problem states correspond to
instantaneous descriptions of the world and operators correspond to actions that an
agent can perform to change the state of the world [3].
The planning is about how an agent achieves its goals, even the simplest ones an
agent must reason about the future. Since the goal is not achievable in a single step,
the number of steps to be carried out needs to be broken up into subtasks, and steps
for each needs to be the goal; what an agent will do in the next step also depends on
its past.
To complete each subtask, there are actions, which need to be carried out; each
action has a subsequent state as well as a preceding state. To be simple at start of this
subject, the following assumptions are made:
1. the actions are deterministic, i.e., the agent can determine the consequent of the
actions,
2. the world is fully observable, i.e., the agent can observe the correct state of the
world, and
3. the closed world assumption, i.e., the facts not described in the world are false.

15.3 The Basic Planning Problem

A basic planning problem usually comprises an initial world description, a description

of the goal world, and a set of actions (sometimes also called operators) that map
a world description to another. A solution is a sequence of actions leading from the
initial world description to the goal world description, referred to as a plan.
A deterministic action is a partial function from states to states.2 For example,
a robot can move block x onto y, if x has top-clear, y has top-clear, and obviously,
x = y. Also, all the alternatives are not available. A precondition of an action decides
about when the action can be carried out, and whether the resulting state due to an
action is the effect of the actions.
Definition 15.1 (Planning Problem) The relevant part of the world is in a certain
state, but managers or directors would like it to be in another state. The (abstract)
problem of how one should get from the current state of the world through a sequence
of actions to the desired goal state is a planning problem. 
AI planning techniques are techniques to search for a plan: forward planning is a
planning technique building a plan starting from the initial state; backward planning
starts from the goal states; and least-commitment planning constructs plans by adding
actions in a non-sequential order.
Ideally, to solve such planning problems, we would like to have a general planning-
problem solver. However, such an algorithm, solving all planning problems, can be
proven to be non-existing (that is, the general planning problem is undecidable). We

2 Partial function: Every state “(state, action)” pair does not necessarily result in a state.
15.3 The Basic Planning Problem 449

therefore, try to concentrate on a simplification of the general planning problem called

the classical planning problem. Although not all realistic problems can be modeled
as a classical planning problem, they can help to solve more complex problems.

15.3.1 The Classical Planning Problem

Classical planning problem is the simplest case of planning, where the environment
is static and deterministic, and the planner has complete knowledge about the current
state of the world. Two important issues are faced by the classical planners: 1. To
model the actions and changes, and 2. To organize search for plans (i.e., sequences
of actions) that are capable of achieving the goals. The logic programming and
nonmonotonic reasoning are frequently used in planning and search, however, many
implemented planners have used a variant of action model, called the STanford
Research Institute Problem Solver (STRIPS). This model represents the state of the
world in terms of state variables and their values, and actions as state-transforming
functions, which are deterministic in nature. Most of the early planners modeled the
state-space exploration of the world-states as a search, and the transitions between
states represented the actions.
The state exploration method worked suitable only for problems having small
search space, due to the complexity in space and time. Instead, the utility of manip-
ulating partial plans during search became popular, and this led to the design of
algorithms that search in the space of partial plans.
The classical planning problem is defined as follows. Given
1. a description of the known part of the initial state of the world (in a formal
language, usually propositional logic) denoted by I,
2. a description of the goal (i.e., a set of goal states), denoted by G, and
3. a description of the possible atomic actions (R (i.e., rule)) that can be performed,
modeled as state transformation functions,
determine a plan, i.e., a sequence of actions that transforms each of the states fitting
the initial configuration of the world into one of the goal states.Thus, classical plan-
ning problem is a tuple I , G, R. Consider the following example of planning the
“Transport a passenger by cab.”

Example 15.1 Classical planning problem of “Transporting by cab.”

Suppose that initially (i.e., in all states of the world that match the description I),
there is a cab at a location A, represented by a binary state variable cab(A), and a
passenger at a location B, represented by passgr(B). In each of the states described
by G the passenger should be at a location C, denoted by passgr(C). Furthermore,
suppose that there are three actions (move, load, unload) that can transform (some
part of) the state of the world.
450 15 Automated Planning

Table 15.1 Classical planning problem

State Transition Comment
0. I ; Initial state
1. move(A, B) ; cab moves from location A to B
2. load (passgr) ; passenger gets into cab
3. move(B, C) ; cab moves from location B to C
4. unload (passgr) ; passenger unloads from cab
5. G ; goal: passenger at location C

Following are the steps for actions:

1. The cab can move from one location to another: move(x, y) with x, y ∈ {A, B, C}.
This action requires that a priori cab(x) holds, and ensures that in the resulting
state ¬cab(x) and cab(y) hold, that is cab is not at place x as well not at the
place y.
2. The passenger can get into the cab: load (passgr). This action requires a priori
cab(x) and passgr(y) and x = y, and in the resulting state both ¬passgr(y) and
passgr(cab) (i.e., passenger in cab) should hold.
3. The passenger can get out of the cab: unload (). This action requires that cab
is at location x (cab(x)) and passenger in cab (passgr(cab)), and results in
¬passgr(cab) and passgr(x).
With I as the initial set of states, and G as a set of goal states, the sequence of
state transitions can be indicated as shown in Table 15.1. 

15.3.2 Agent Types

Agents are classified according to the techniques they employ in their decision mak-
ing:

1. Reactive agents: They base their next decision solely on their current sensory
input.
2. Planning agents: They base their course of action considering the anticipated
future situations, possibly as a result of their own actions.

Whether an agent should plan or it should be reactive, depends on the particular

situation it finds itself in. Consider the case where an agent has to plan a route from
one place to another. A reactive agent might use a compass to plot its course, whereas
a planning agent would consult a map. Clearly, the planning agent will come up with
the shortest route in most cases, as it will not be confronted with uncrossable rivers
and one-way hills. On the other hand, there are also situations where a reactive agent
can be at least as effective, for instance, if there are no maps to consult such as in
15.3 The Basic Planning Problem 451

Fig. 15.1 World states and s1

agent (robot) actions
a1 a2
s2 s3
a3 a4 a6
a5
s4

Table 15.2 Table for State Action Resulting state

mappings: State × Action →
State s1 a1 s2
s1 a2 s3
s2 a3 s4
s2 a4 s5
… … …

Fig. 15.2 Delivery robot Shyam,s

with service locations Coffee
Shop Office

mcc mc
Robo
Mail
Room AI Lab

a domain of planetary exploration, like Mars or Moon. Nevertheless, the ability to

plan ahead is invaluable in many domains.
Considering the states set as s1 , s2 , . . . , and set of actions as a1 , a2 , . . . , they can
be represented using a tree shown in Fig. 15.1 or explicitly by Table 15.2.
We consider the following example to better understand the actions and states
world, where agent is a robot, called Robo.

Example 15.2 A delivery Robot’s (Robo’s) planning.

A delivery robot shown in Fig. 15.2 is responsible for some jobs. It can pick up mail
from the mail room and deliver it to Shyam’s office, and also can pick up coffee from
the coffee room and can deliver it to Shyam’s office. The robot, called Robo, can also
reach these places by moving clockwise (mc) as well as moving counterclockwise
(mcc). Assume that the mail it handles is a postal mail (not the email). The domain
of the world is represented by the terminology described as follows [13].
Various locations are represented by the following symbols. Each of them can be
true/false.

cs: robo at coffee shop

off : robo at Shyam’s office
mr: robo at mail room
lab: robo at AI lab
452 15 Automated Planning

Following are the various states’ variables of the world, which can also be
true/false:
mw: Mail waiting in the mail room
rhc: Robo is holding coffee
swc: Shyam wants coffee
rhm: Robo is holding mail

Absence of a state is represented by its negation. For example, rhc indicates that
Robo is not holding coffee. Various actions performed by the Robo are

mc: Robo moves clockwise

mcc: Robo moves counterclockwise
puc: Pick up coffee
dc: Deliver coffee
pum: Pick up Mail
dm: Deliver Mail

Presence of an action symbol indicates that the action is true.

A state (of the world) may comprise many parameters or preconditions for the
action to take place at that state. For example, the state,

lab, rhc, swc, mw, rhm (15.1)

indicates that the Robo is in an AI lab, Robo has no coffee in hand, Shyam wants
coffee, mail is not waiting, and Robo holds the mail. Another state,

lab, rhc, swc, mw, rhm, (15.2)

indicates that the Robo is in the lab, Robo is holding coffee, Shyam is waiting for
coffee, mail is waiting in the mail room, and Robo is not holding the mail. Table 15.3
shows the transitions for certain states. For example, in the third row, we note that
after performing the action dm (deliver mail), in new state created, the state rhm
indicates that variable “Robo is holding mail” is false.

Table 15.3 Some mapping: State × Action → State, for Fig. 15.2
State Action Resulting state
lab, rhc, swc, mw, rhm mc mr, rhc, swc, mw, rhm
lab, rhc, swc, mw, rhm mcc off , rhc, swc, mw, rhm
off , rhc, swc, mw, rhm dm off , rhc, swc, mw, rhm
… … …
15.4 Forward Planning 453

15.4 Forward Planning

One of the simplest plannings is to treat the planning as a path planning problem in
the state-space graph. The nodes here are states, transitions are actions, and results
of actions are also states. A forward planner searches the state-space graph from
the start state for the goal state. Figure 15.3 shows the state-space graph for forward
planning with the start state as cs, rhc, swc, mw, rhm, with three transitions from
start state, corresponding to the actions: pick up coffee (puc), Robo moves clockwise
(mc), and Robo moves counterclockwise (mcc). If we choose the action puc, the next
state is cs, rhc, swc, mw, rhm. The new state indicates that the Robo still remains
facing the coffee shop; since it picked up coffee, the robot holding coffee is no more
false, mail waiting remains true (unchanged). The new states as a consequence of
the various actions are self-explanatory, and are similar to this description [2, 13].
Note that, being closed reasoning, in the world of actions we explicitly specify
each of the variable as True or False.
The branching factor in Fig. 15.3 is 3, and the search can be done in DFS or
BFS. Theoretically, since, the Robo can be at any of the four locations, and the other
four parameters in a state can be true/false, there are 4 × 2 × 2 × 2 × 2 = 64 total
possible states in the world. Obviously, all of these states are not possible to reach
in a graph search.
The representation above is simple and clear, but it is not suitable due to following
reasons:

• there are too many states to acquire, reason, and represent,

• small change in the requirements will need a major change in the model, for
example, if we need to have information about robot battery level to be added as
one of the parameters, the entire structure gets modified.

cs, rhc, swc, mw, rhm

puc mcc
mc
cs, rhc, swc, mw, rhm mr, rhc, swc, mw, rhm
mcc of f, rhc, swc, mw, rhm
mc
mr, rhc, swc, mw, rhm
of f, rhc, swc, mw, rhm mcc

dc mc mcc mc

of f, rhc, swc, mw, rhm lab, rhc, swc, mw, rhm cs, rhc, swc, mw, rhm

cs, rhc, swc, mw, rhm

lab, rhc, swc, mw, rhm

Fig. 15.3 State space for forward planning

454 15 Automated Planning

The improvement in complexity is possible, and can be based on the following

criteria: we note that the actions have a structure and that can be used to make
actions compact. We also note that a precondition of an action should be true before
the action takes place. For example, the action of the Robo to pick up the coffee (puc)
requires the precondition of “Robo’s location is coffee shop and Robo does not hold
the coffee”, which is expressed as cs ∧ rhc. This means that puc is not available at
other preconditions (or constraints) [11].

15.5 Partial-Order Planning

A partial-order planning is a search carried out by refining the partial plans through
addition of actions and orderings. Alternatively, a search may proceed through per-
forming abstract actions in the plan, by replacing fragments capable of carrying out
those actions—the planning is called hierarchical planning. It has been now well
understood that even the state-space search methods are nothing but other ways of
refining partial plans. This refining is done by growing the prefix or suffix of the plan
and different ways of refining a partial plan can be suitably interleaved.
The forward planner enforces a total ordering on actions at all the stages in the
planning process. The idea of partial ordering between actions only commits to an
ordering between actions when forced. A partial ordering is a “≤” relation, that is,
reflexive, transitive, and antisymmetric. It is the set of actions together with partial
ordering, representing a “before relation” on actions, such that any total ordering of
the actions, consistent with partial ordering, will solve the goal from the initial state.
That is, Act0 < act1 means action Act0 appears before the Act1 in the partial order.
A partial-order plan comprises the following sets:
1. Set of steps. Each step maps to some operator, except the start and end step. There
are no preconditions for start step and start state is its post-condition. Similarly,
the final step has the goals as its preconditions and has no post-conditions.
2. Set of orderings of steps. Each ordering is a pair of steps, in the form of previous
and next step. The start step is ordered first of all, and the finish step is ordered
after all the steps.
3. Set of causal links. Each causal link is a pair of steps and a proposition, which is
a post-condition of the first step and a precondition of the second. The first step
is always ordered before the second step.
In fact, a partial planner works as follows: begin with actions start and finish,
and with partial order start < finish. The planner maintains an agenda set of P, A
pairs, where, A is action in plan, and P is the precondition of A. First, G, finish is
chosen, such that G is the precondition for Goal. Then at each stage, a pair P, act1 
is chosen from the agenda, where P is the precondition for action act1 . Subsequent
to this, act0 is chosen to achieve P, which is either already in the plan or it is start to
achieve P, or it could be a new, that is, added to the plan. The act0 must occur before
15.5 Partial-Order Planning 455

act1 , that adds a new causal link. Any action to delete P must happen after act1 or
before act0 . If act0 is a new action, its preconditions are added to the agenda, and the
process continues till the agenda is empty.

15.6 Planning Languages

The STRIPS formulation for planning problems uses a simple and general format
for specifying operators with clear semantics. It employs propositional logic as a
language for describing states of the world, with a set of conditions and Boolean
variables to describe states. A complete assignment maps the set of conditions to
possible values. For example, truth values and a partial assignment maps a subset of
conditions to values, and a complete assignment is a state. An operator comprises
two partial assignments—first is preconditions or results, which decide the states
for applying the operator. The second is post-conditions, that decide the next state
resulting when an operator is applied in a particular start state. In addition, the implicit
frame axioms indicate that the value of any condition, which is not mentioned in an
operator’s post-conditions, is unchanged due to the application of an operator.
A planning problem instance consists of a set of operators, an Initial state, and a
Goal. The Goal may be a Boolean formula also. Generally, a solution is in the form
of a partially ordered multi-set of operators, which satisfies the following condition:
any total ordering consistent with the given partial order transforms the initial state
assignment into a new assignment that satisfies the Goal formula. The STRIPS for-
mulation provides the semantic foundations for many extensions, including those
handling external events, multiple agents, probabilistic transformations, and vari-
ants that allow the agent to observe aspects of the current state and choose actions
conditioned on observations.
The STRIPS is action-centric representation, which is based on the idea that most
things are not affected by single action. It specifies action, precondition, and effect.
For example, for the goal: “Robo to pick up coffee”, we write

precondition : cs ∧ rhc
effect : rhc.

The other features are unaffected in the above. The action of delivering coffee
(dc) is

precondition : off ∧ rhc

effect : rhc ∧ swc.

Apart from the above, you need to specify the initial states and goal.
The action or the rule has two formats:
456 15 Automated Planning

– Causal Rule, when a feature gets a new value, and

– Frame Rule, when a feature keeps its value.
We can demonstrate the change of Robo’s location (see Fig. 15.2), from current
RLoc0 to a new location RLoc1 on account of action mcc, mc, and one which is neither
mcc nor mc, as follows:

(RLoc1 = cs) ← (RLoc0 = off ) ∧ (Act = mcc)

(RLoc1 = cs) ← (RLoc0 = mr) ∧ (Act = mc)
(RLoc1 = cs) ← (RLoc0 = cs) ∧ (Act = mcc) ∧ (Act = mc). (15.3)

In the above formulas, the first two are causal rules, and the last one is the frame
rule. Similarly, the state “Robo holds coffee” in the resulting state would depend on
whether it was holding coffee in the previous state and its action:

rhc1 ← rhc0 ∧ Act = dc

rhc1 ← Act 0 = puc (15.4)

where the first is frame rule, and the second is the causal rule.

15.6.1 A General Planning Language

STRIPS is a problem-solving program implemented in LISP and has been used in

the application of robotic research; it is a member of the class of problem solvers that
search a space of world models to find a model through which the goal is achieved.
We assume that, for some world model, there exists a set of operators; each of such
operators transforms the world model into some other world model. Having this, the
task of the problem solver is to determine a composition of operators that transform
a given initial world model into one that satisfies the goal condition [3].
The primary objectives in robotics-based class of problems are rearranging phys-
ical objects, and navigation of robot—problems that require general world models,
and are more complex than those used in the solution of puzzles and games. Usually,
a list or simple matrix structures is adequate to represent a state of such problems.
The world model for a robot problem solver must comprise a large number of facts
and relations about the position of the robot, and about the positions and attributes
of objects, open spaces, and boundaries. In STRIPS, the world model is represented
by a set of well-formed formulas in the first-order predicate logic (FOPL) [4].
A solution is built using operators, which are the basic elements of general robotics
planning language. For a robot problem, each operator corresponds to an action
routine whose execution causes a robot to take that action. For example, we might
have a routine to push back if the robot touches the wall, or a routine to lift and object,
or to grip an object, and so on, there are a large number of routines that correspond
to actions of a robot.
15.6 Planning Languages 457

15.6.2 The Operation of STRIPS

The problem space for STRIPS language comprises the initial world model, a set
of available operators with their effects on the world model, and a goal statement.
The world model is represented by a set of well-formed formulas (wffs) of FOPL,
e.g., to describe a world model in which the robot is at location a and boxes B, C
are at locations b, c respectively, we would use the following wffs to express this
knowledge [4]:

ATR(a)
AT (B, b)
AT (C, c).

We might also use the wffs to express the general rule that an object u at place x is
not in a place y, i.e., an object cannot exists at two places.

(∀u ∀x ∀y){[AT (u, x) ∧ (x = y)] ⇒ ¬AT (u, y)}. (15.5)

We can represent a complex world model using the first-order predicate logic, and
can use standard theorem-proving programs to answer questions about the model.
The operators are grouped into families of operators, called schema. For example, an
operator goto for moving the robot from one point m on the floor to other point n is
a schema. For this, distinct operators (one for each pair of points) are grouped into a
family of goto operators, and goto(m, n) represents a move from the initial position
m to the final position n. The members of goto schema are goto(m, a1 ), goto(a1 , a2 ),
..., goto(ak , n). In STRIPS, specific constants will already have been chosen for the
operator parameters when an operator is applied to a world model.
First of all, it is necessary to determine whether or not there is an instance of an
operator schema applicable to the current world model. It is required that an instance
of the corresponding wffs (well-formed formulas) schema exists, and it logically
follows from the model. Each operator schema is defined by a description having
two parts: conditions under which the operator can be applied, and the effects of
application of that operator. The precondition for an operator schema is represented
as a wff. The effects of application of an operator are defined by a list of wffs that must
be added to the model, and a list of wffs that are no longer true, hence must be deleted.
As an example, consider the question of applying instances of the operator subschema
goto(m, b) to a world model containing the wff ATR(a). If the precondition wff
schema of goto(m, n) is ATR(m), then we note that the instance ATR(a) can be proved
from the world model. Thus, an applicable instance of goto(m, b) is goto(a, b).
It is important to understand the difference between the parameters in wff schema,
and existentially and universally quantified variables. These variables are used in
FOPL formula in theorem-proving programs (e.g., Resolution theorem) that would
handle wff schema. For example, in resolution, ∀x∀y goto(x, y) will have substitution
458 15 Automated Planning

{a/x, b/y}, to make it a clause, whereas in STRIPS there is a chain of pairs of (x, y)
to implement goto(x, y), hence it will require some modifications.
A goal statement is also in the form of a wff, e.g., a task of moving the boxes B
and C to locations d , e, respectively, might be expressed a goal as

AT (B, d ) ∧ AT (C, e).

In summary, the problem in state space for STRIPS comprises three entities:
1. Initial world model. It is a set of wffs to describe the present state of the world.
2. Set of operators. The operators and their description in the form of effects and
preconditions as wffs.
3. Goal. It is a condition in the form of a wff.
The problem is taken as solved when a world model that satisfies the goal wff is
deduced using the initial world model with the application of operators.

15.6.3 Search Strategy

A very simple problem-solving strategy is to apply all the applicable operators to

the initial world model and create a set of successor models. Then, continue to
apply all applicable operators to these successors and to their descendants generated
through DFS or BFS search, until a model is deduced having the goal formula as a
theorem. However, it is quite likely that in most real-word problems modeled using
this approach, the number of operators applicable to any given world model will be
too large. Hence, such a simple system would generate trees that are large in number
as well as size. Hence, such world models would be impractical.
An alternative strategy is to extract differences between current world model
and the goal model, with the objective as to how to move from one to another
with minimum changes. Then find out the operators that may reduce the differences
between these models. Once the relevant operators are found out, we try to solve the
subproblem that produces a world model to which it is applicable. If such a model
is found, then we apply the relevant operators and consider the original goal in the
resulting model.

15.7 Planning with Propositional Logic

Consider the graph/tree shown in Fig. 15.4, with s0 as start state, and a0 . . . an−1
as actions representing a path leading to the goal state sn . If s0 , a0 . . . an−1 , sn are
considered as propositional expressions, then Eq. [10],
15.7 Planning with Propositional Logic 459

Fig. 15.4 Propositional s0

planning
a0
a1

an−1
sn

p = s0 ∧ (a0 ∧ a1 ∧ · · · ∧ an−1 ) ∧ sn (15.6)

is a propositional expression representing the path to goal state from start state. If
there is only one goal, then p is the only path to goal, and all other propositions p
are false. Equation 15.6 can always be transformed into a CNF (conjunctive normal
form) or SAT expression. Thus, planning through propositional logic is to find a
satisfiability expression, comprising start state, path, and goal state, i.e., logical
sentence equal to

initial state ∧ all proposition action descriptions ∧ goal. (15.7)

In other words, a model that satisfies a sentence will assign true to all actions
that are part of a correct plan. If the planning problem is unsolvable then there is
no sentence that is satisfiable. The following example demonstrates planning with
propositional logic [12].

Example 15.3 Flight Planning.

Initial Plan: Let us assume that initially, the plane p1 is at DEL (Delhi) and p2 at CAL
(Calcutta). The actions should correspond to flying these planes so that the goal: “p1
at CAL, and p2 at DEL” is satisfied.
The initial state (0) is represented by

At(p1 , DEL)0 ∧ At(p2 , CAL)0 .

Since the propositional logic has no closed world, it is also necessary to show that
initially the planes p1 , p2 are not CAL, DEL, respectively, i.e.,

¬At(p1 , CAL)0 ∧ ¬At(p2 , DEL)0 .

The initial conditions correspond to start state s0 .

The goal also needs to be specified with a particular time step. Since it is not
known how many time steps it would consume to iterate the action, a worst case time
limit needs to be specified so that either the goal is reached within that time limit,
else the solution is terminated as failure.
460 15 Automated Planning

We test the following assertion for a goal at time T = 0 (at start),

At(p1 , CAL)0 ∧ At(p2 , DEL)0 .

If that fails, apply certain actions and again try it at time T = 1, and so on, up to
time T = Tmax . This is shown in Algorithm 15.1, where for every iteration, for the
next value of T (time), the problem is translated to SAT (satisfiability) problem using
the procedure translate-to-SAT. This results in CNF expression and a mapping of
the solution to the problem. The SAT-Solver procedure returns the assignment for the
above CNF expression. If the assignment is satisfying the solution (i.e., not null), the
solution is extracted for the present mapping, and assignment is returned as result.
If this does not happen, the procedure is iterated as per the Tmax iteration time [8].

Algorithm 15.1 Planning with Propositional Logic

1: INPU T : AplanningProblem;
2: Tmax : an upper limit for plan length
3: for T = 0 to Tmax do
4: cnf , mapping ← translate-to-SAT (problem, T )
5: assignment ← SAT -Solver(cnf )
6: if Assignment is not null then
7: Return Extract-solution(asignment, mapping)
8: end if
9: end for
10: Return Failure
11: End

15.7.1 Encoding Action Descriptions

We have a propositional symbol for each occurrence. For the plane p1 to be at CAL,
there is a proposition:

At(p1 , CAL) ⇔ (At(p1 , CAL)0 ∧ ¬fly(p1 , CAL, DEL))

∨ (At(p1 , DEL)0 ∧ fly(p1 , DEL, CAL)0 ) (15.8)

There ought to be a plan that tries to achieve the goal at T = 1. Now suppose
CNF is
Initial state ∧ successor state axioms ∧ goal1 ,

that is, goal is true at T = 1, we check and verify that

fly(p1 , DEL, CAL)0 ∧ fly(p2 , CAL, DEL)0

is the model, and other assignments are false.

15.7 Planning with Propositional Logic 461

15.7.2 Analysis

For the proposition fly having general form as fly(p, o1 , o2 ), with time T -steps, the
number of propositions represented by |p|, and objects (o1 , o2 , etc.) as O, the com-
plexity expression is given by,

fly(p, o1 , o2 ) ⇒ T × |p| × |o1 | × |o2 |

⇒ T × |p| × |O|2
⇒ T × |Act| × |O|p

where T is the number of time steps, p is arity of function (here it is 2), O is the
number of objects o1 , o2 , etc. We note that the complexity is exponential. The term
|Act| is for the action, like fly, i.e., how many predicates are there in total [8].

15.8 Planning Graphs

The planning graphs give better heuristics and consist of a sequence of levels, for
time steps in plan. Each level has literals (constant values) which have become true
because of the previous action, and each level has preconditions for the next action.
The planning graphs represent the actions as well as inactions. The following example
demonstrates the application of planning graphs [5, 12].

Example 15.4 Problem of the solution to “have a Pizza and eat Pizza”.

init(have(pizza))
Goal(have(pizza) ∧ eaten(pizza))
Action(eat(pizza)
Precond : have(pizza)
Effect : ¬have(pizza) ∧ eaten(pizza))
Action(cook(pizza)
Precond : ¬have(pizza)
Effect : have(pizza)).

The planning graph for these actions is shown in Fig. 15.5. The box action in the
figure indicates the mutual exclusions of actions.
In the planning graph, all the actions Ai at level i contain all actions that are
applicable at state Si , along with constraints saying which part of the actions cannot
be executed. Every state at level Si contains all literals that could result from any
choice of actions at Ai−1 , along with constraint saying which part of actions cannot
be executed.
462 15 Automated Planning

cook(pizza)
have(pizza) have(pizza) have(pizza)
¬have(pizza) ¬have(pizza)
eat(pizza) eat(pizza)
eaten(pizza) eaten(pizza)
¬eaten(pizza) ¬eaten(pizza) ¬eaten(pizza)

S0 A0 S1 A1 S2

Fig. 15.5 A planning graph for “To have pizza and eat pizza”

We note that graph plan provides lesser complexity because it does not require
choosing about all actions. It just records the impossibility of certain choices using
mutex, i.e., through mutual exclusion links. For example, when ¬have(pizza) is cho-
sen, the action have(pizza) is excluded. This results in the complexity as a function
of low polynomial actions and literals [11].

15.9 Hierarchical Task Network Planning

In the hierarchical task planning, each level of hierarchy is decomposed into smaller
levels. It is common for areas like military mission, administration, program devel-
opment, where a task is reduced to small number of activities at the next level, so that
the computational effort of arranging those activities is low. This results in reduction
in complexity to linear time from the original exponential [13].
Consider an example of “building a house”, where the task of house building can
be decomposed to acquiring land, preparation of design map, obtaining the NOC (no
objection certificate) from municipal corporation, arranging for a house loan, hiring
a builder, paying the builder, and so on, as shown in Fig. 15.6.

House
¬Money

Fig. 15.6 Decomposition of task of house building

15.9 Hierarchical Task Network Planning 463

We understand that some of these activities can be done in parallel, but not all.
Thus, there is a partial-order relation between those actions. The decomposition can
be expressed as decompose(a, d ), where action a is decomposed into a partial-order
plan d . Various activities for building a house as per the plan in Fig. 15.6 can be
formally described as follows.

Action(Buildhouse, Precond : BuyLand ,

Effect : House) (15.9)

Action(Buyland , Precond : Money,

Effect : Land ∧ ¬Money) (15.10)

Action(Getfiance, Precond : Goodcredit,

Effect : Money ∧ Mortgage) (15.11)

Action(Hirebuilder, Precondition : Nil,

Effect : contract) (15.12)

Action(Construction, Precond : NOC ∧ Builderhired ,

Effect : Housebuilt ∧ ¬NOC) (15.13)

Action(Paybuilder, Precond : Money ∧ Housebuilt,

Effect : ¬Money ∧ house ∧ ¬Contract) (15.14)

Decompose(steps : {A1 : Get NOC, A2 : Hirebuilder,

A3 : construction, A4 : Paybuilder} (15.15)

Orderings : {start ≺ A1 ≺ A3 ≺ A4 ≺ Finish;

Start ≺ A2 ≺ A3 }. (15.16)

Here Ai ≺ Aj indicates that activity Ai precedes the activity Aj .

The linking of states and activities through activities/resources for the house build-
ing plan can be expressed as follows:

Links : {start Land A1 , Start Money A4 , A1 NOC A3 ,

−−→ −−−→ −−→
A2 Contract A3 , A3 Housebuilt A4 , A4 House Finish,
−−−−−→ −−−−−−→ −−−→
A4 ¬Money Finish.} (15.17)
−−−−−→
464 15 Automated Planning

15.10 Multiagent Planning Systems

There are a number of good reasons for having multiple agents creating plans:

1. The agents may represent real-life entities, which may require the privacy and
autonomy also to be maintained.
2. Changing of an existing distributed system to a multiagent system is less costly
than to a centralized system .
3. Creating and maintaining plans locally allows for more efficient reaction, espe-
cially when communication is limited, and
4. Dividing a planning problem into smaller pieces and solving those in parallel turns
out to be many times more efficient. This is true particularly when the individual
planning problems are not tightly coupled.

In spite of many benefits of multiagent systems, following are the challenges in

developing multiagents planning:
1. How to put additional constraints upon the agents before planning, such that their
resulting plans can still be coordinated?
2. How to efficiently construct plans in a distributed environment?
3. How to make collaborative decisions when there are multiple options, and agents
have their own preferences for these options?
4. In what condition should a planning agent ask for more specific information to
the user?
5. How to find out the magnitude of privacy lost in the process of coordinating plans?

Definition 15.2 Multiagent planning problem.

In general, a multiagent planning problem is a problem of planning by and for a group

of agents. Except for centralized multiagent planning problems, each agent in such
a problem has in fact a private, individual planning problem. A typical individual
planning problem of an agent includes a set of operations with some costs attached,
and pre- and post-conditions that it can perform, and a set of goals (with reward
values), and the current (initial) state of this agent. The following statement captures
the concept of multiagent planning:

Multiagent planning = planning + coordination. (15.18)


The solution to a multiagent planning problem is a plan: a partially ordered
sequence of actions that, when executed successfully, results in a set of achieved
goals for some of the agents. Most techniques can deal with problems where the
actions and goals of the agents are 1. only weakly dependent upon each other, 2.
agents are cooperative, and 3. communication is reliable. However, in general a mul-
tiagent planning problem may encounter a lot variety of situations along these three
axes. Some characteristics are described below.
15.10 Multiagent Planning Systems 465

1. From independent to strongly related.

a. Independent: There are no shared resources, and no dependencies. For exam-
ple, a robot is lifting a box.
b. Strongly related: They have joint actions, and the resources are shared. For
example, car assembly is a case of strongly related.
2. From cooperative to self-interest agents: The participating agents interested in
optimizing their own utility is the case of self-interesting agents, while the agents
involved in supply chain management is an example of cooperative.
3. From no communication to reliable communication: In hostile environments
agents may not or cannot communicate during execution. This may require agents
to be equipped in advance, with or without some initial communication before the
execution starts. The robots rescuing people in disaster scenarios, or working on
a inter-planetary exploration mission, are examples of the first type, while agents
working in a supply chain management is an example of the second category.

15.11 Multiagent Planning Techniques

Multiagent planning techniques cover quite a range of solutions to different phases

of the problem. In general, the following phases can be distinguished in task sharing:
1. Allocate goals to agents.
2. Refine goals into subtasks.
3. Schedule subtasks by adding resource allocation including the agents, and timing
constraints.
4. Communicate planning choices (i.e., prior steps) to recognize and resolve con-
flicts.
5. Execute the plans.

Planning is a combination of phases 2 and 3 in the above, which are often inter-
leaved. Any of these steps could be performed by one agent or some subset. Not all
the above phases of general multiagent planning process need to be considered in
every multiagent planning problem. For example, there is no need for phase 1 if there
are no common or global goals for the multiagents. Some multiagents system may
combine different phases, for example, while constructing their plans agents may
coordinate with each other due to the combination of phases 2, 3, and 4. Alterna-
tively, robots may postpone coordination until the execution phase (i.e., combining
phase 4, 5). This may result, for example, when they unexpectedly encounter each
other while following their planned routes.
It is possible in general, to interleave any combination of five phases listed above,
depending on the nature of the planning problem, resulting in a wide variety of
possible problem sub-classes. In the following we present the phases in more detail.
466 15 Automated Planning

15.11.1 Goal and Task Allocation

Centralized methods often take care of the assignment of goals and tasks to agents
during planning. There are, however, many other methods to assign tasks in a more
distributed way, giving the agents a higher degree of autonomy and privacy. For
example, complex task allocation protocols may be used, or auctions and market
simulations can also be used.
One way to assign the tasks to agents is through the method of auction, which is
a way to assign a task to that agent who attaches (or bids) the highest value or the
lowest cost of performing it, called private value. A protocol called Vickrey auction
is a frequent example in multiagent systems, where each agent makes a closed bid,
and the task is assigned to the highest bidder, but not on the price of that bidder,
but for the price of the second highest bidder! The Vickery protocol has the good
property due to which the bidding agents will simply bid their true private values,
with no need for additional reasoning about it is worth for the others.
Economics and market simulations can also be the basis for allocation of resources
among agents. For example, it is shown how costs can be turned into a coordination
device. These methods are useful for task assignment (phase 2), and also for the
coordination of agents after plan construction (phase 5).
In concerning the value-oriented environments, these game-theoretical approaches
become more important where agents reason about cost of their decision-making (or
communication).

15.11.2 Goal and Task Refinement

Task assignment can be done through a single agent, using Hierarchical Task Net-
works or nonlinear planning. More than one planner with more sophisticated models
of temporal extent can be introduced, along with centralizing as well as combining
the phases 2 through 4.

15.11.3 Decentralized Planning

Instead of one agent planning for rest of the agents, the second and third phases
can be implemented through local planning by each of the agents. In principle, any
planning technique can be used in this condition, and different agents may even use
different techniques. Some of the approaches join the individual plannings (phases
2 and 3), along with the coordination of the plans (phase 4).
A distributed version of a planner can be used to integrate phases 1 through 4, to
plan for a single agent in parallel.
15.11 Multiagent Planning Techniques 467

In one setting, each agent has a partial knowledge of the plans of other agents
using some specialized plan representation techniques. In such kind of environment,
coordination is achieved as follows: If an agent A informs the other agent B about
part of its own plan, then agent B merges this information into its own partial global
plan. The agent B can then try to improve the global plan by, for example, eliminating
redundancy in the plan, and this improved plan is shown to the other agents, who
might reject, or accept, or modify it. This process may run concurrently with the
execution of the (first part of the) local plan.

15.11.4 Coordination After Planning

One of the tasks is the planning of coordination after plans are constructed on individ-
ual basis (phase 4), called plan merging . Its objective is construction of a joint plan
for a set of agents, given the individual plans of each of the participating agents. In
coordination planning, every pair of agents helps each other by changing the state of
the world such that the conditions of the other agent become satisfied. In this process,
changing the state of the world may be helpful to these two, but may also interfere
with the correct conditions of the remaining n − 2 agents, assuming a system of n
agents.
To specify the constraints on plans, one approach is to use propositional tem-
poral logic, which will ensure that only feasible states of the environment are
reached. A theorem prover algorithm generates a sequences of communication
actions, on receiving these constraints. In fact, these communication actions imple-
ment semaphores that guarantee that no event will fail. For resolving conflicts, restric-
tions are required to be introduced on individual plans in phase 3, which will also
ensure efficient merging.
A different approach to plan merging uses the distributed approach to improve
social welfare, based on the sum of the benefits of all agents. This approach uses a
process of group constraint aggregation, where agents construct an improved global
plan by voting for joint actions, in an incremental way. The agents even propose
algorithms to deal with insincere agents, and to interleave planning, coordination,
and execution [11].

15.12 Summary

When an agent is interested in controlling the evolution of its environment, there is a

need for planning. Thinking as an algorithm, a planning problem has an input in the
form of possible courses of actions, a predictive model for the required dynamics,
and a measure for performance to evaluate the courses of actions. The output or
solution of this algorithm is one or more courses of actions that satisfy the specified
requirements for performance.
468 15 Automated Planning

The agents are classified as reactive and planning agents.

The classical planning problem is to plan reaching the goal state(s) from the initial
state, for a given set of actions. A majority of research in planning is toward planning
in the environments that are dynamic, stochastic, and partially observable. To carry
out this, the existing classical planning techniques are extended to allow interleaving
of planning and scheduling.
Automated planning techniques are being applied in many domains, that include
robotics, process planning, web-based information gathering, autonomous agents,
and spacecraft mission controls. In automated planning, a solution to a problem can
be described in terms of a sequence of steps that transforms some initial description of
the problem state, for example, the initial configuration of a puzzle, into a description
satisfying a specified goal criterion. For a simple automated planning problem, called
classical planning problem, the following assumptions are made:
1. the actions are deterministic,
2. the world is fully observable, and
3. the closed world assumption.
Agents are classified according to the techniques they employ in decision-making:
1. Reactive agents, and
2. Planning agents.
Forward planning, one of the simplest planning, is path planning where nodes
are states, transitions are actions, and results of actions are also states. A forward
planner searches the state-space graph from start state for goal state. One problem
with forward planner is its time complexity, which is exponential.
A partial-order plan consists of (1) a set of steps, each mapping to an operator,
(2) a set of orderings, and (3) a set of causal links.
STRIPS is an action-centric language, based on the idea that most things are not
affected by a single action. It specifies action, precondition, and effect. The problem
space for STRIPS is defined by an initial world model, a set of operators, and a goal
condition state.
Planning through propositional logic is to find a satisfiability expression, which
can be given as

initial state ∧ all proposition action descriptions ∧ goal.

Other approach for automated planning is planning graphs, which consists of a

sequence of levels, for time steps in plan, representing actions as well as inactions.
In the areas such as military mission, administration, program development, where
a task is reduced to a small number of activities at the next level, such that the com-
putational effort of arranging those activities is low, the hierarchical task planning
using networks is preferred.
Splitting a planning problem into smaller problems and solving these in parallel
turns out to be more efficient, thus motivating the use of multiple agents for creating
plans. Multiagent planning covers the following phases:
15.12 Summary 469

1. Allocate goals.
2. Refine goals into subtasks.
3. Add resource including the agents.
4. Communicate planning choices.
5. Execute the plans.

Exercises

1. Consider the standard Towers of Hanoi problem with 3 pegs and 4 number of disks
(d1 , d2 , d3 , d4 , with d1 at the top). The disks are to be transferred from start-peg
to end-peg, using intermediate peg, one at a time such that at no time larger disk
comes over the smaller. The disk d1 is the smallest and d4 is the largest. Make
use of only 3-predicates: unary predicate: clear, and binary predicates: on and
smaller, and only one action: puton(x, y) needs to be used.
Write the domain of the problem, and make use of forward planning to plan the
solution to move all the 4 disks from start-peg to end-peg.
2. Given the 3-SAT problem:

(¬p1 ∨ p2 ∨ p3 ) ∧ (p1 ∨ ¬p2 ∨ p3 ) ∧ (p1 ∨ p2 ∨ ¬p3 ),

solve it using forward planning. (Hint: you need to assume some operators (i.e.,
actions) to assign the values to variables p1 . . . p3 .)
3. Given the Table T , and blocks A, B, C, D, having different positions on the table,
apply STRIPS to plan the solution of the following problem:
Initial state I as

clear(A), clear(B), clear(C), clear(D),

on(A, T ), on(B, T ), on(C, T ), on(D, T ),

i.e., the blocks A . . . D are on the table, and their tops are clear.
Final State F :
on(A, B), on(B, C), on(C, D), on(D, T ), clear(A).
Use the action puton(X , Y ), x = y, where X is a block A . . . D and Y is either
table T or block A . . . D. Give the forward planning to reach state G starting with
state I.
4. Use STRIPS for planning of the following problem: You are at home, and you
have money, and you are required to buy milk. Assume the necessary start and
goal states, actions, preconditions, and results for this planning job.
5. Give the STRIPS representations to actions: pick up mail and deliver mail (ref.
Fig. 15.2).
6. Suppose the robot (in Fig. 15.2) cannot carry both coffee and mail at the same
time. Make use of some constraints to provide the planning for this situation.
470 15 Automated Planning

Assume that the robot can carry a box in which it can place objects, so that it can
carry the box and the box can hold the mail and coffee.
7. Modify the problem in Fig. 15.2, so that the robot has the work of cleaning the
four rooms (mail room, office, coffee shop, lab). Assume that it will clean the
room only when the room is unclean, and will not consume more than one rotation
mcc or mc to reach any of these rooms.
8. Using the method of hierarchical task network planning, provide the automated
planning for the following problems:
a. Shopping grocery items from market.
b. Deliver a lecture of AI.
c. Robot path planning to cover the diagonal in a room.
9. Assume that you have three operators:
f1 : Precondition: a; effect: ¬a ∧ b
f2 : Precondition: a ∧ c; effect: ¬a ∧ b ∧ ¬c
f3 : Precondition: b ∧ c; effect: ¬c ∧ d
Show the first three layers (proposition, action, and proposition) of the graph plan
when the initial state is a ∧ c (a and c both are true). Include the mutual exclusions
and justify each of them.

References

1. Bborrajo D et al (2015) Progress in case-based planning. ACM Comput Surv 47(35):1–35

2. Bonet B, Geffner H (2001) Planning as heuristic search. Artifi Intell 129:5–33
3. Dean T (1996) Automated planning. ACM Comput Surv 28(1):85–88
4. Fikes RE, Nilsson NJ (1971) STRIPS: a new approach to the application of theorem proving
to problem solving. Artifi Intell 2:189–208
5. Jonsson P et al (2000) Towards efficient universal planning: a randomized approach. Artifi
Intell 117:1–29
6. Kaelbling LP et al (1998) Planning and acting in partially observable stochastic domains. Artifi
Intell 101:99–134
7. Kambhampati S (1995) AI planning: a prospectus on theory and applications. ACM Comput
Surv 27(3):334–336
8. Kautz H, Selman B (1992) Planning as satisfiability. In: Proceedings of the 10th European
conference on artificial intelligence (ECAI 92), Vienna, Austria
9. Leckie C, Zukerman I (1998) Inductive learning of search control rules for planning. Artifi
Intell 101:63–98
10. Melis E, Siekmann J (1999) Knowledge-based proof planning. Artifi Intell 115:65–105
11. Nareyek A (2005) Constraints and AI planning. Intell Syst 03(04):2005
12. Oh SC et al (2005) A comparative illustration of AI planning-based web services composition.
ACM SIGecom Exchanges 5(5):1–10
13. Russell SJ, Norvig P (2005) Artificial intelligence—a modern approach, 2nd edn, Pearson
Chapter 16
Intelligent Agents

Abstract The intelligent agents are being viewed as new theoretical models of com-
putation that more closely reflects current computing reality, aimed as new generation
models for complex and distributed systems. An agent system can work as a single
agent, or as a multiagent system. The intelligent agents have many applications—
they are used in software engineering, in buying and selling—like online sales, bids,
trading; the agents are also modeled for decision-making—with preferences and cri-
teria for making decisions. This chapter also presents the classification of agents,
agent system architecture, how the agents should coordinate among themselves, and
the formation of a coalition between agents. The multiagents communicate with each
other using agents’ communication languages which are oriented towards perform-
ing actions. Other categories of agents are mobile agents—programs which can be
moved to any far off place, and can communicate with the environment. The chapter
ends with chapter summary, and the set of exercises.

Keywords Intelligent agent · Mobile agent · Multiagents · Agents’ coordination ·

Cooperative agents · Agents’ coalition · Software agents

16.1 Introduction

An artificial agent or intelligent agent is a recent term in computer science, and specif-
ically in artificial intelligence. There is a number of definitions of agents. The agents
are viewed as a new theoretical model of computation, that reflects current com-
puting reality in a better (tangible) way than the existing model of Turing Machine.
They are being projected as a next-generation model to engineer the complex and
distributed systems.
Among many characterizations of agents, the following definition is most com-
mon: An agent is an encapsulated computer system, which is situated in some envi-
ronment and it is flexible and capable of autonomous action in that environment in
order to meet its desired goals. There are associated number of questions about this
definition that require further explanation, which becomes somewhat clear through
the extended definition of agents, in the following:

© Springer Nature India Private Limited 2020 471

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_16
472 16 Intelligent Agents

1. Agents are entities for problem-solving for clearly identifiable problems with
well-defined boundaries and interfaces;
2. They receive inputs related to the state of their environment through sensors, when
embedded in an environment, and act on the environment through effectors;
3. An agent is designed to fulfill a specific requirement, and it has a particular goal
to be achieved;
4. They have control over both their own behavior and over the internal state, a
property called autonomous;
5. The agents have flexible problem-solving behavior. They are both reactive (able
to respond in time to the changes occurring in their environment) and proactive
(able to act in anticipation of future goals).
Agents are also being used as a framework to bring together various AI’s sub-areas
to design and build intelligent systems. In spite of the intense interest of research
community and progress made in the area of agents, a number of fundamental ques-
tions about the nature and the use of the agent-oriented approach remain unanswered,
which are as follows:
• What are the fundamental concepts and notions of agent-based computing?
• What makes the agent-based approach a natural and powerful computational
model?
• What are the implications of agent-based computing, in the wider perspectives,
for AI and computer science in general?

Learning Outcomes of this Chapter

1. List the defining characteristics of an intelligent agent. [Familiarity]
2. Characterize and contrast the standard agent architectures. [Assessment]
3. Describe the applications of agent theory to domains such as software agents,
personal assistants, and believable agents. [Familiarity]
4. Describe the primary paradigms used by learning agents. [Familiarity]
5. Demonstrate using appropriate examples how multiagent systems support agent
interaction. [Usage]
6. Syntactic structure of agent languages. [Familiarity]

16.2 Classification of Agents

Although there is no universally accepted definition of an agent, however, as per the

most commonly used definitions, an agent is a proactive software component that
interacts with its environment, as well as it interacts with other agents on behalf of
its user, and reacts to the changes in its environment. A component is called agent if
it exhibits several of the properties given below.
16.2 Classification of Agents 473

Autonomous
It is the property of an agent, as per which it proactively initiates the activities as
per its goal. An agent has its own thread of control, can act on behalf of its user, and
without necessarily depending on the messages from other agents.
Mobile
An agent can move itself from one execution context to another. For this activity,
it can move its code, and carry on executing from the current point onward, or it
can start afresh. Other modes of execution can be, serialization of its code and state,
such that it may continue its execution in a new context, and at the same time, it may
retain the same old state and can continue to work.
Adaptable
An agent is adaptable to a new environment; its behavior can change after its deploy-
ment through its own learning, downloading new capabilities, and through user cus-
tomization.
Knowledgeable
A software agent has reasoning capability, due to which it can reason about the
acquired information, about the knowledge of other agents, its user, and about its
goals.
Collaborative
Some agents are called collaborative agents, which can communicate with other
agents and work in a cooperative manner. This collaboration can be formed either
in a static manner or a dynamic manner. The collection of such agents is called
multiagent system.
Persistent
The infrastructure used by the agents allows them to retain their knowledge and states
over an extended period of time. The agents have property of robustness, i.e., they
work correctly on the face of some failures at run time.
Many characteristics of Intelligent agents are result of capabilities like adaptabil-
ity, cooperation, and proactivity. see Fig. 16.1 shows agent taxonomy. The circles
correspond to general agent capabilities, and intersection corresponds to the agents
having either two or three capabilities. For example, an agent having the capability
of proactive and cooperative is called a collaborative agent.
Depending on the functions performed, agents can be classified in one of several
major categories.
Adaptive agents
They can learn from their previous experience, and can change how they should
behave in a given situation, and can also behave differently in given situations.
474 16 Intelligent Agents

Fig. 16.1 Agent taxonomy

Adaptive Cooperative

Smart
Personal Collaborative

Proactive

Cooperative agents
They can communicate with other agents, and their action can be according to the
results of the communication performed.
Proactive agents
These agents can initiate proactive actions, i.e., without any prompting from the user
or other agents.
Personal agents
They are proactive and can interact directly with a user. While interacting with the
user, they present some personality or character, can do monitoring and adapting to
the user’s activities, can learn the user’s style and preferences. They can automate
or simplify certain rote tasks. Many software tool-kits, e.g., Microsoft Agent, offer
software services set that support the presentation of software agents as interactive
personalities and includes natural language and animation capabilities.
Collaborative agents
These agents are proactive and cooperate with other agents. They communicate and
interact in groups, many times on behalf of a number of users or organizations, or
services. Multiple agents exchange messages to negotiate or share information. Some
of their applications are: online auctions, planning, negotiation, logistics, supply-
chain management, and telecommunication services.
Smart agents
The smart agents exhibit a combination of all capabilities, i.e., they are adaptive,
cooperate with other agents, and are proactive.
Mobile agents
These agents are sent to remote sites to collect the information, and forward it to
the central or any other location. Before sending the results to a specified location,
these agents can aggregate and analyze data or perform some local control. They are
typically implemented in any of the following languages: Java, Java-based compo-
nent technologies, VBScript, Perl, TCL, or Python. The data-intensive processing is
16.2 Classification of Agents 475

usually performed at the source, as this avoids the shipment of bandwidth consuming
raw data. Examples of such applications are network management agents, Internet
spiders, and NASA’s mobile agents for human planetary exploration, etc.

16.3 Multiagent Systems

The multiagent systems are used by organizations or people with different /conflicting
goals, and having proprietary information. In such systems, multiagent system is
required to handle their interactions. As an example, in a manufacturing scenario
where company A produces launching-pad of missiles but subcontracts to company
B to produce the missiles. To build the whole system comprising missiles launcher
and missile, the internals of both companies must be modeled. However, none of
these companies are ready to share the details with the other company. Possibly,
the two companies may reach to some agreement, or when not ready to share any
details due to protocols imposed by the government, multiagent system (MAS) can
be created, with one agent for each company, that represents the goals and interest
of each company [10].
As another example, consider a teaching time-table system for a college. This
domain requires, different agents to represent the interest of different people in the
college. Faculty wants their classes should be evenly distributed throughout the week,
with possibly all classes in as few rooms as possible, management wants that all the
resources be used fully, students want that no more than two/three theory classes be
held each day. Similarly, the technicians will have their own requirements. In such
a scenario, a multiagent system, where different constraints are handled by agents
separately, can create time table static/dynamic to best meet all the constraints.
The multiagent system creates parallelism by assigning different tasks to differ-
ent agents, hence making overall a fast response system. In addition, many agent’s
systems will have redundant agents, this helps in building robustness in the system.
This is possible because, the control and responsibilities are shared among the agents,
hence the system can tolerate failures of some of the agents, and still working cor-
rectly and efficiently. The areas of applications that requires graceful degradation at
the time of failure, instead of sudden failure, are suitable domains where multiagent
systems’ use is welcomed. However, if single entity or processor or agent controls
everything, the then entire system may crash if there is a single failure.
The multiagent systems have the benefit of scalability. Because they are inherently
modular, it is easier to add new systems to a multiagent system than to add new
capabilities to monolithic systems. Due to the flexibility available, it is easier to
program a multiagent system.
476 16 Intelligent Agents

Fig. 16.2 A general Environmrnt

single-agent framework
S = Sensors
E = effectors

S
Agent
E
Goals
Actions
Domain
knowledge

16.3.1 Single-Agent Framework

Though it might appear that a single-agent system might be simpler in concern

to dealing with a fixed complex task, however, the opposite is often true. In fact,
when control is distributed among number of agents, an individual agent can be
simpler. A general agent is a single-agent system, together with the environment,
and the interaction between agent and environment. An agent is itself part of the
environment, but generally, the agents are considered to have extra-environmental
components, which are independent entities, having their own goals, knowledge, and
actions. In a single-agent system, no other entities are recognized by the system (see
Fig. 16.2).

16.3.2 Multiagent Framework

Following are the taxonomies of multiagent systems:

1. Agent granularity, which can be course or fine.

2. Heterogeneity of agent’s knowledge, can be redundant or specialized.
3. Methods used for distributing control can be: benevolent or competitive, team or
hierarchical, static or shifting roles.
4. The agents can communicate among other agents, in blackboard or through mes-
sages, it can be low-level or high-level communication.

In a multiagent system, there are several agents which are capable of modeling
each other’s goals and actions. In a general multiagent system, there may be direct
interactions among the agents. The inter-agent communication is viewed separate
from communication with the environment. A major difference with single-agent
system is that, in multiagent systems, the environment dynamics can be determined
by the other agents also, which can affect the environment in an unpredictable way.
Thus, all multiagent systems can be treated as having dynamic environments.
16.3 Multiagent Systems 477

Fig. 16.3 A fully general

multiagent framework

The Fig. 16.3 shows multiagent environment, where each agent is part of the
environment, as well as can be modeled as a separate entity. There may be any
number of agents with different degrees of heterogeneity and with or without the
ability to communicate directly [10].

16.3.3 Multiagent Interactions

In agent-oriented view of the world, it has been found that most problems require the
participation of multiple agents to represent the distributed nature of the problems,
with multiple locations of control, and multiple competing interests. In addition,
the agents need to interact with each other to manage dependencies resulting from
their existence in a common environment, and to achieve their individual objectives.
Such interactions may vary greatly—from a simple information exchanges to, in a
more complex form—to request for particular action for coordination/cooperation,
or negotiation, or arranging interdependent activities.
Agent interactions are differing on two characteristics with respect to computa-
tional models, like networking and shared computing: 1. Agent-oriented interactions
are conceptualized as taking place at knowledge level—realized in terms of what
goals should be followed, by whom, and at what time. 2. The agents are flexible
problem solvers, operating in an environment that is partially observable, and agents
have partial control over it. Therefore, the interactions need to be also handled in a
similar and flexible manner.
The agents make use computational models to make run-time decisions about the
type and scope of their interactions, and also to initiate and respond to interactions
that were not anticipated at the time of design of the system [5].
478 16 Intelligent Agents

Agent
Agent level interactions
Organizational interactions

Visibility and influence

Environment

Fig. 16.4 Canonical view of an agent-based system

Agents usually act to achieve objectives as individuals, or as a group for some

larger problem-solving initiative. Hence, when agents interact with the environment,
there is a hidden organizational context in them, that defines the nature of the rela-
tionship between agents—as peers working together in a team or one may be the
manager of the other agents, hence may influence their subordinates’ behaviors.
Since agents are required to make decisions about the various types of interac-
tions at run time, there is a need for an explicit representation of organizational
relationships of agents. Often, these relationships are subject to frequent changes,
for example, the agents working on the computation of social interaction must take
care of existing relationships in social networks, and should also support the evolu-
tion of these relations. The evolution is due to the creation of new relations, and due
to the exit of members from these social networks. Looking at these examples, we
understand that the life span of these relationships can vary from just long enough
to deliver a particular service once to a permanent bond.
To cope with the dynamic scenario of variety and dynamics of relationships in
agents, their protocols needs to be devised to support organizational groups to be
formed and dismantled, there is need of specified mechanisms to ensure that these
groups act together in a coherent way, and also there is need of structures to charac-
terize macro behavior in a collective way.
The Fig. 16.4 shows the essential concepts of agent-based interactions.
The agents capable of having the features presented above are the Intelligent
agents (also called software agents). These agents are autonomous components,
have their own goals and beliefs. They are designed with the capability to reason
about their behavior: both present and future, and offer abundant scope for fast, and
incremental development of Web-based enterprise applications. The developers can
use these systems for a variety of complex and dynamic domains, which range from
e-commerce to research on planetary exploration systems.
16.4 Basic Architecture of Agent System 479

16.4 Basic Architecture of Agent System

Many complex and intelligent agents navigate on the Internet, collect the relevant
data and process them, perform various tasks including data analysis and data com-
munication, and make the decisions on behalf of their users. The present generation
of intelligent software agents can manage, organize, and communicate huge amounts
of data on behalf of their users. For example, the agents in e-commerce applications
can dynamically discover and compose e-services and mediate interactions. They
can also be used to serve as delegates to handle routine tasks, like, monitoring activ-
ities, set up contracts, win the bid, execute business processes, and can find the best
services [3].
Agents reside and execute in a conceptual and physical location called an agency
(see Fig. 16.5), which provides facilities for locating and messaging to mobile agents
and also to those that are detached agents. The agency also facilitates for collecting
knowledge about agents. The core of the agency is an agent platform, with component
model infrastructure, which provides the local services to agents. The core also
includes the proxies to access remote services like security, agent management,
communication, persistence, and naming. For mobile agents, the agent platform also
provides agent transport. Some additional services provided by most agent systems
are in the form of specialized agents that reside in some remote agency. There are
some standard service agents, like a broker, auctioneer, or community maker, which
augment the basic agents infrastructure. The agent platform along with service agents
monitor and control message exchange by detecting any violation of rules while they
engage in communication. The agent’s platform is the system’s core, however, in
addition to this, a component model infrastructure empowers the agents with local
services and proxy access to remote services.
An agent system comprises components with simple interfaces. The major part
of the system’s capability results from its loose coupling, which helps the agents to
interact dynamically through the exchange of messages asynchronously. For com-
munication with each other, the agents must follow some common and well-defined

Fig. 16.5 Agent system

architecture
480 16 Intelligent Agents

protocols. Communication of agents is through a language, called ACL (agents

communication language). It is a specialized declarative language, which defines
the structure and pattern of interaction between agents. This language is associated
with the component model, and partitions the messages into many parts, which are
relatively independent of each other. The commonly used message partitions are:
message type, addressing, context, content of the message, domain description, and
expected conversation patterns. Due to the message protocols and its partitioning of
the message, it is easier to dynamically extend the agents to new problem domains,
while the system checks conformance to expectations and allows the component
model infrastructure to manage messages and agents.
The agents interact among themselves using a set of vocabularies, called ontol-
ogy, which is designed for the application domain of the agents. The word set in
the vocabulary describe the things, their attributes, action performed, various rela-
tionships, meanings, and how the agent’s system use this vocabulary to structure the
interactions, and access the devices.

16.5 Agents’ Coordination

For performing complex tasks there is requirements to integrate a group of agents to

coordinate the activity. This is possible by a multiagent system, working in either of
the two modes, static or dynamic. The agents can coordinate amongst themselves,
and also with people. This coordination requires messaging between the agents, the
sequence of the messages may have many possible levels of choreography, depending
on how loosely or tightly the allowed interactions are controlled by the system. It
is common practice that, instead of directly programming as code to handle the
messages coordination, some graphical models or high-level declarative rules are
used. These models/rules make it easier to visualize how the agents interact. The
agent system can use explicit rules or models to monitor or enforce compliance,
which makes the programmer’s task simple [8].
It is not always the case that coordination would mean cooperation. For example,
an effective competitor will coordinate the decisions to maximize his/her advantages
against the opponent. This may be seen in the planning of product promotion by a
company to undercut a rival.
Various coordination strategies have emerged for computational agents. How-
ever, it is not possible to devise a coordination strategy that works equally well in
all situations. If any such strategy exists, it can be easily applied for an unlimited
number of constructs employed today, such as governments, corporations, markets,
teams, committees, professional societies, mailing groups, etc. Whatever strategy is
adopted, certain situations can stress it to be a breaking point. Any adopted coordi-
nation strategy must now be concerned about how to scale to increasingly complex
situations. To map the space of potential coordination strategies, we must find out
important dimensions along which they must scale and then evaluate their response
to complexities along with those dimensions.
16.5 Agents’ Coordination 481

16.5.1 Sharing Among Cooperative Agents

Benefits at a global level of an agent system are bound to improve if all agents
cooperate. However, cooperation among the agents is difficult to realize, particularly,
in situations when agents are self-interested. For example, if a number of agents are
trying to get the same resource, say download a specific file, the download speed is
bound to decrease. Instead of this approach, if they work on grouping a social decision
that is mutually beneficial, it will be good for all of them. Therefore, designing
mechanisms that promote cooperation among self-interested agents is important. In
fact, several game theory approaches have been found to be useful for the study of
cooperation in agents, e.g., the Prisoner’s Dilemma (PD) as a theoretical framework
(see Chap. 11), which is well known for this purpose. The PD can be useful for
understanding the role of local interactions to maintain cooperation among the agents.
It is based on the conflict of interest, i.e., between what is the best for the individual
(i.e., defection) and what is best for the group (i.e., cooperation). This creates a
situation of social dilemma. Therefore, specific mechanisms are required to evolve
cooperation to help the population to overcome this dilemma.
There are three basic approaches to prevent social dilemmas, and to promote and
stabilize cooperation as follows:
Coalition-based mechanism
The coalition-based approach is useful for establishing collaboration among agents,
with an individual having properties and objectives. These mechanisms use a tax
model, due to which agents can achieve cooperation when coalitions are formed
around some emerging leaders. To maintain coalitions, the leaders charge some tax
from their agents in favor of some benefit (e.g., guaranteed cooperation, protection
from cheaters, etc). The concept of coalition has been used in the game theory for a
long time, and has been proved useful in real-world economic scenarios. The dynamic
coalition formation model considers the grid topology of agents for cooperation
between them and makes use of spatial prisoner’s dilemma.
The coalitions facilitate cooperation between self-interested agents. The first
approach is: a leader of coalition is paid by the agents that are in the coalition.
The coalition leader also imposes its decision on the agents in the coalition to max-
imize cooperation. The decision making of each coalition is done in a centralized
manner by a single entity, called leading agent. The agents’ cooperation with their
coalition-mates also assumes some restriction in the collaboration.
The coalition-based approach is a clear example of the known trade-off between
the benefits versus the costs of collaboration (e.g., taxes). Therefore, this mechanism
is called dynamic coalition formation model, and also tax model.
Partner switching mechanisms
In most real-world situations, the network topology changes frequently. There is
an empirical evidence in the games on dynamic topologies that a partner switching
leads to cooperative behavior. A variant model of prisoner’s dilemma allows agents
482 16 Intelligent Agents

to either adjust their strategies or switch their defective partners, with the aim that
partner switching may help stabilize cooperation.
Self-governing institutions
The resource allocation in case of self-governing Institutions is modeled in a net-
work, based on a formal characterization of socio-economic principles. An agent
should autonomously decide how to behave with respect to coalition-mates and
agents outside its coalition. Although some mechanisms promote cooperation on
different network topologies, these networks are static.

16.5.2 Static Coalition Formation

There are two approaches for static coalition formation: 1. Optimization-based

approaches, that focus on finding an optimal coalition, and 2. Game theoretic
approaches. The later has applications in many real-world domains, like electronic
commerce, auctions, and general resource allocation scenarios. The game theoretic
approaches may also involve automated agents [8].
In coalitions with optimization objectives, the challenge in coalition formation is
generating coalition structure, which turns out to be an NP-complete problem as a
general case, hence the existing algorithms cannot generate solutions in a reasonable
time, even with the moderate size of the game (number of agents). Hence, finding an
optimal coalition can become intractable because the number of coalition structures
grows exponentially with a number of agents.
Among other goals, one goal of coalition formation is to improve cooperation
among the agents. The game theory approaches have been widely used to address
the issue of cooperation among agents. A class of coalition formation game, called
hedonic games, is a rich and versatile class for coalition formation, suited for both
static and theoretical aspects of coalition formation, and has the property of encap-
sulation in the stable matching scenarios. The major focus of hedonic games is
on critical stability for coalition structures, e.g., Nash stability, individual stability,
contractual individual stability, and core stability. These games also characterize
conditions under which the set of stable partitions is guaranteed to be non-empty.

16.5.3 Dynamic Coalition Formation

In formation of coalition in a dynamic environment, the agents constantly change the

coalition they belong to. In such a scenario, since optimality is possible with a very
small number of agents, computing of optimal coalition is either infeasible, or may
take a time longer than the lifetime of a coalition for any realistic number of agents.
Thus, the time constraint to find an optimal coalition prevents its use in a dynamic
multiagent system, where some agents have to decide if it is beneficial to them to
16.5 Agents’ Coordination 483

join other agents for a small amount of time. This time limitation is due to the fact
that for n number of agents, the total number of possible coalition structures to be
enumerated are of the order of O(n n ), which is too large unless n is a small number.
For large n, the computation cannot be carried out in realistic times. Hence, it is
necessary to make use of domain knowledge, with mathematical games and some
constraints to solve the problem of coalition formation in an efficient way, for the set
of agents of any specified characteristics [8].
To form a dynamic coalition, it is required to have decentralized procedures to
allow self-interested agents to negotiate the formation of coalitions, as well as to
divide the coalition payoffs. In the real-world scenarios, the agents may turn out
to be selfish and may focus on improving their own performance, but if they are
cooperative, the performance of the whole system will improve. Hence, the theory
of non-cooperative games (i.e., agents are selfish) is suitable to model the formation
of coalitions and their dynamics. The prisoners’ dilemma is a case of this category,
since the prisoners are considered as selfish (defecting is the dominant strategy) in
this game. A variant of this, called Iterated Prisoner’s Dilemma (IPD) game is widely
used to model various social and economic phenomena, and the cooperation among
agents. In the IPD, where a total number of rounds is random or unknown, sustained
cooperation strategies are likely to emerge.

16.5.4 Iterated Prisoner’s Dilemma Coalition Model

We present a model, where a graph (or network topology) representing population

is iterated, the nodes in the network represent agents, and edges represent relations
between agents. Such agents interact with their peers in the social neighborhood
(the agents to which they are linked), and play the game of Possessors-Traders (an
agent is either a possessor or a trader). Such agents not only cooperate or defect,
but have resources using which they can trade. The agents can form coalitions to
increase the cooperation level of the multiagent system. In such coalitions, group
decisions can result in mutually beneficial cooperation, that holds over some time.
The group decisions lead to, and is an indication of social behavior. In addition, the
agents’ neighborhood is not static, and can change partners through rewiring. Hence,
in addition to the trading of resources, each agent decides the following during the
game [8].

– To remain independent or to be part of a coalition, depending on which alternative

provides more payoffs.
– Whom to rewire with? As agents change their neighborhood, they rewire to
improve the benefits.

In addition, all the agents in a coalition behave like a unit, and all together decide
how they should behave with those in the coalition (called insiders), and those outside
the coalition (called outsiders). Hence, the decision about what is a coalition and what
484 16 Intelligent Agents

is its behavior, is an important criterion in the dynamics of the coalition system. In

their behavior the agents work in cycles: trading strategies, rewiring (changing of
coalition) strategies, and coalition strategies. The Algorithm 16.1 shows this cycle,
where x represents Payoffs.

Algorithm 16.1 One cycle of Agents-dynamics

1: x = trade-with-all-neighbors();
2: rewire(x);
3: revise coalition(x);

Agents use certain trading strategies to trade among the agents, based on some
model, which can be called by a general name property ownership and trade (POT)
model. The trading model is based on extension of the Iterated prisoner’s Dilemma,
in which agents can cooperate or defect the actions. The model of POT comprises
two types of players: 1. The Possessors P, who own the resources, and 2. Traders
T , who sell and buy resources.
The strategy of any agent pi ∈ P models the practice of ownership, but does not
trade. The behavior of pi depends on whether it owns a resource or not. If pi owns
a resource it acts as a defector, but if it does not, then it cooperates. This strategy
is shown in Algorithm 16.2, where owns( pi , r esour ce) indicates that possessor pi
owns some resource, de f ects( pi ) means pi defects, and cooperate( pi ) means pi
cooperates.

Algorithm 16.2 Possessor pi ’s Strategy

1: if owns( pi , r esour ce) then
2: de f ects( pi )
3: else
4: cooperates( pi )
5: end if

An agent t j ∈ T is trader, who is willing to sell or buy a resource when dealing

with a fellow trader tk ∈ T . For example, if some t j has a resource for selling, it will
try to get the maximum benefit by selling it. Whenever any pair of traders (t j , tk )
meet, the trader (say t j ) owning the resource values the resource at a random value y j
(but does not make it open), such that v < y j < V , and v, V ∈ R (real numbers). In
response, the buyer agent tk offers a value yk for the resource, such that v < yk < V .
If yk > y j then the buyer purchases the resource at some random value yl , so that
y j < yl ≤ yk . This is called trader T ’s strategy, with its logic given in Algorithm 16.3.
If the trader plays against who is not a trader, then it is the possessor.
16.5 Agents’ Coordination 485

Algorithm 16.3 Trader T ’s Strategy

1: if isT rader (t j ) AN D isT rader (tk ) then
2: if owns(t j , r esour ce) AN D v < y j < V then
3: Sell for y j
4: else
5: if owns(tk , r esour ce) AN D y j < yk then
6: Buy for yl AN D y j < yl ≤ yk
7: else
8: Behave as Possessor
9: end if
10: end if
11: else
12: Behave as Possessor
13: end if

16.5.5 Coalition Algorithm

A Algorithm 16.4 shows the basic strategy followed by agents either to join a coalition
or leave a coalition (change to a new one). If an agent ai has the poorest payoff among
all its neighboring agents after completion of the previous round of computations
(line 1), then ai makes a new coalition with some agent a j (line 4), who is free and
has the best payoff. If a j is already in some other coalition, then ai joins a j ’s coalition
(line 6). This rule also enables any agent to change from one coalition to another, in
case that agent receives poor payoffs in the former coalition [8].
In a dynamic network, agents form coalitions to behave as a unity. An agent can
belong to only one coalition at a time. All agents belonging to a coalition are not
required to be linked with each other, but behaves as a set to act together to maximize
their performance. However, an agent must have at least one link to some agent
belonging to its coalition. If that is not the case, it is an isolated agent, hence it must
be declared as an independent agent (lines 9–10, Algorithm 16.4). This connection
helps an agent to know its coalition information, strategy, share, and divide gains.
Again, if an agent changes link, it does not imply that it changes its coalition—it
simply rewires to change neighbors.
The agents that are in a coalition, must agree to some specific behavior to play
with insiders (agents in the coalition) and also with outsiders (agents outside the
coalition). We assume a flat coalition, i.e., there is no leader or central authority to
impose any policy. To decide the coalition behavior in this situation, each agent votes
for a strategy of either P or T (for possessors and traders) to play with insiders as
well as with outsiders (line 16, Algorithm 16.4).
To decide the vote, each agent uses a Learning Automata (LA), trained from its
trading history and payoffs (lines 12–15). The LA algorithm keeps two probability
models: I n Pr ob, and Out Pr ob. The model [I n Pr obT, I n Pr ob P] is used to assess
the strategy to play against insiders. Here, I n Pr obT is the probability of being inside
the coalition as a trader, and I n Pr ob P is the probability of being inside the coalition
as possessor. In a similar way, [Out Pr obT, Out Pr ob P] is to assess the strategy to
play against outsiders.
486 16 Intelligent Agents

Algorithm 16.4 Revise Coalition(Payoffs)

1: if poor est Payo f f _ f r om_neighbor s(ai ) then
2: a j = neighbor W ith Best Payo f f ()
3: if independent (a j ) then
4: Cr eat N ewCoalition(a j , ai )
5: else
6: J oinCoalition(a j , ai )
7: end if
8: end if
9: if I solated Agent (ai ) then
10: makeI ndependent (ai )
11: else
12: [I n Pr obT, I n Pr ob P] = U pdateI nsider s L A(Payo f f s)
13: [Out Pr obT, Out Pr ob P] = U pdateOutsider s L A(Payo f f s)
14: I n Action = I n ActionChoice(Pr obI nT, Pr obI n P)
15: Out Action = Out ActionChoice(Out Pr obT, Out Pr ob P)
16: V oteBest (I n Action, Out Acion)
17: end if

16.6 Agent-Based Approach to Software Engineering

In respect of software engineering, we view agents as next-generation components

and agent-oriented software engineering as an extension of conventional CBSE
(case-based software engineering). The developers can integrate different types of
agents, like, personal, mobile, and collaborative agents, to build agent-based enter-
prise systems, covering a wide problem domain area. To patrol the networks to find
available resources, special software is used, called Daemons.
Developers often use distributed objects, active objects, and components that can
be scripted to implement agents. The agents are often driven by goals and plans
instead of procedural code, they encapsulate business or domain knowledge. These
agents often differ more from each other by the knowledge they have and the roles they
play, than by the differences in their implementing classes and methods. The agents
are capable of using different mixes of adaptability, mobility, intelligence, ACL,
and even multiagent support. Either AI programming languages or conventional
programming languages can be used to implement the agents.
Next, we introduce the techniques for tackling complexity in software [5].
Decomposition
For tackling large problems, the basic technique is to divide the problem into smaller
chunks, such that it is better manageable. Each of these chunks is dealt with relative
isolation. Since this limits the designer’s scope, it helps to tackle the complexity of
the issues, because it requires to consider only a small portion of the problem at any
given time.
16.6 Agent-Based Approach to Software Engineering 487

Abstraction
The abstraction is a process of defining a simplified model of any system, such that
only the necessary and important details or properties and emphasized, while all
unnecessary details are suppressed.
Organization
The process of organization is concerned with the identification and managing the
relationships between various problem-solving components. Specifying and imple-
menting organizational relationships are helpful to tackle the complexity due to two
reasons: 1. It facilitates the grouping of a number of basic components, which are
collectively treated as a unit at a higher level for the purpose of analysis. 2. Due to
grouping the components as a unit, as well as to specify the relationships between
them, a number of components may work together (cooperate) to provide a particular
functionality.

16.7 Agents that Buy and Sell

The Software agents were used much earlier for the applications, like filtering infor-
mation, match people having similar or identical interests, and automating repetitive
behavior.
In the recent past, agents have found the applications in e-commerce to con-
duct business-to-business, business-to-consumer, and consumer-to-consumer trans-
actions. Consider an example of buying and selling, where a company willing to
place an order for procurement of stationery, assigns the tasks to agents to moni-
tor the quantity and usage patterns of paper within the company. It also launches
the buying agents when paper inventory is low. The Buying agents would typically
perform the following tasks, more or less, in order [6].

1. collect the information automatically about vendors and the required products
which best fulfills the needs of the company,
2. evaluate the various offers from the vendors,
3. make a decision about merchants and products that require further investigation,
4. negotiate the terms of transactions with these merchants, and finally
5. place orders and make automated payments.

There are several descriptive theories and models that seek to capture buying
behaviors, e.g., Nicosia model, Howard-Sheth model, and Engel-Blackwell model.
All these share six fundamental stages of the buying process:
1. Identification. The buyer can be motivated through product information, hence
he/she becomes aware of some unmet needs.
488 16 Intelligent Agents

2. Product brokering. The buying process comprises as its part, the Information
Retrieval to determine what to buy. IR consists of evaluation of product alterna-
tives based on the criteria provided by the buyer, whose result is a set of products,
called “consideration set.”
3. Merchant brokering. This activity combines the “consideration set” with
merchant-specific information to help customer to decide as from where to buy the
goods. This stage also comprises the evaluation of merchant alternatives based on
buyer-provided criteria. The later is typically, the price, warranty, delivery time,
availability, and reputation, which are not necessarily be in order, but varies case
to case.
4. Negotiation. This step considers how to settle on the terms of transition. Negoti-
ations vary in duration and complexity, for price and other attributes.
5. Purchase and delivery. This step signals either termination of the negotiation
stage or occurs some time afterward.
6. Product service and evaluation. This involves the post-purchase product service,
customer service, and evaluation of the satisfaction of overall buying experience.

In the present online buying and selling, many of the processes and criteria dis-
cussed above are in a common place.
The continuous running personalized autonomous agents are well suited to medi-
ate for consumer behaviors, like, information filtering, IR, personalized evaluations,
time-based interactions, and complex coordination. Many agents perform constraint-
based, and collaborative filtering. Many websites of online-shopping use rule-based
techniques to personalize the products offering for individual customers. Some web-
sites use agents to experiment data-mining techniques to discover patterns in cus-
tomers’ purchase behavior, exploit those behaviors for sales, and use these patterns
also to help customers to find other products that meet their true requirements.
The product alternatives are compared at the product brokering stage, whereas
the merchant alternatives are compared at the merchant brokering stage.

16.8 Modeling Agents as Decision Maker

For modeling agents as decision makers, it is necessary to have modeling methods that
use formal notions of mental state to represent and reason about agents. The mental
states may consist of mental attributes such as beliefs, knowledge, and references.
In multiagent systems, the success of one’s actions and plans are governed by the
actions of other agents. Thus, agents can help in constructing plans that are likely to
succeed. The mental level models can bring two informal properties: 1. They provide
an abstract way of representing agents, which is implementation-independent, and, 2.
These models are built using an intuitive approach, and use attributes, such as goals,
beliefs, and intentions. The abstract nature of models have the following practical
implications [1]:
16.8 Modeling Agents as Decision Maker 489

1. A single formalism can capture different agents, written in different languages,

and running on different hardware platforms,
2. There are no implementation details in abstract models, and
3. Fever lower-level details in abstract models result in faster computation.

16.8.1 Issues in Mental Level Modeling

In mental level modeling following are the central questions:

1. Structure. The structure holds the designer’s initial database of beliefs, goals,
intentions, which are manipulated by the agent.
2. Grounding. It is the base for the model construction process, which is essential
because we cannot directly observe the mental state of another agent.
3. Existence. Under what conditions a model will exist? Answer to this question
will be helpful in evaluating any proposal for mental level models. Therefore, it
is necessary to know, what assumptions are made, or biases we are making when
we model agents in this manner.
4. Choosing a model. How do we choose between different models that are consistent
with our data?

16.8.2 Model Structure

A model’s mental level structure consists of three key components: beliefs, references,
and decision criteria, which in order corresponds to accounting for the agent’s per-
ception about the world, its goals, and method of choosing actions under uncertainty,
respectively.
The agents’ belief help in establishing about which states of the world it considers,
are plausible. As an example, the possible worlds of interest may be about weather
conditions: rainy and non-rainy, and let the agent believes rainy to be plausible. In
fact, the agent’s preference indicates how much it likes each preference. The agent
may have two possible actions: take an umbrella along to protect from rain, and do
not take an umbrella along. The outcomes of these actions are shown in Table 16.1.
The agent’s preferences tell us how much significance it gives to these values. We
will prefer to use real numbers to describe these values, such that larger numbers
indicate better outcomes, as shown in Table 16.2 [1].
An agent would choose its action (take or not take umbrella) by applying its
decision criteria to the outcome of different actions in the world. A commonly used
decision criterion is maximin, where the action for the “best worst-case” outcome is
chosen. The best outcome out of (10, −4) and (−1, 8) is 10 and the worst is −4.
The best in the worst-case is −1, so the agent chooses the action “Do not take an
umbrella”. This is because the outcome worst −1 is better than worst −4.
490 16 Intelligent Agents

Table 16.1 Decision table for an agent

Action (↓), Worlds (→) Rainy Not-rainy
1. Take umbrella Dry, Heavy Dry, Heavy, Illogical
2. Do not take umbrella Wet, Light Dry, Light

Table 16.2 Table with Action (↓), Worlds Rainy Not-rainy

weighted outcomes (→)
1. Take umbrella 10 −4
2. Do not take −1 8
umbrella

Having gone through the above example, we are now in a position for grounding.
We can view the problem of describing a mental state of the agent as a CSP (Constraint
Satisfaction Problem). The state of the model is such that it should have generated
the observed behavior, and it is consistent with the background knowledge.
In the above example, the background knowledge is agents preferences, given in
the Table 16.2, and decision criteria is maximin. We observe that if the agent goes out
without an umbrella, it believes that “no rain will come”, for it is had other beliefs it
would have taken a different action.
Once an agent’s model is constructed, it can be used to predict its future behavior.
To give it a formal shape, we consider that an agent A , is described as a state machine,
with set of possible local states L A , a set of possible actions AA , and a program,
which we call its protocol PA . Thus an agent is a tuple,

A = L A , AA , PA  (16.1)

where PA : L A → AA . All the agents function with some environment, so we

assume L E as set of all states in the environment. The environment describes every
things external to the agent, which may possibly include other agents also. The com-
bined state of the whole system, i.e., both the agent and the environment are referred
to as global state, and represent by a pair (lA × lE ) ∈ L A × L E . We further assume
that environment does not perform actions, and agent’s actions are deterministic
functions of its state and the environment’s state. Thus, the set of possible worlds
will be only a subset S of the set of global states L A × L E . And, finally, a transition
function,
τ : (L A × L E ) × AA → (L A × L E ) (16.2)

maps a global state and an action to a new global state.

Example 16.1 An agent A to model a an air-conditioner’s thermostat control.
The modeling of agent A is shown in Table 16.3, which shows the results of transi-
tion function τ . It should have local states L A = {−, +}, where ‘−’ corresponds to
16.8 Modeling Agents as Decision Maker 491

Table 16.3 Transition table for thermostat agent

Worlds: (−, cold) (+, cold) (−, ok) (+, ok) (−, hot) (+, hot)
LA × LE →
Action: AA ↓
Turn-on (−, ok) (+, ok) (−, hot) (+, hot) (−, hot) (+, hot)
Turn-off (−, cold) (+, cold) (−, ok) (+, ok) (−, ok) (+, ok)

the state when the thermostat indicates that the temperature is less than the room tem-
perature, and ‘+’ for temperature greater than or equal to desired room temperature.
The thermostat’s protocol is given below.

State − +
Action Turn-on Turn-off

The thermostat’s actions are modeled as AA = {tur non, tur no f f }, and the envi-
ronment’s states are, L E = {cold, hot, ok}. For the sake of simplicity, we assume
that the possible world is L A × L E , which are displayed in the heading row in the
Table 16.3.
Given the set of possible worlds W , we can associate with each local state l of
the agent (the thermostat), a subset S, W (l), comprising of all worlds in which the
local state of the agent is l. 
In the above example, the effects of an action on the environment do not affect
the state of the thermostat. In addition, the static one-shot model assumes some
simplifications. The first assumption is that the room temperature is affected only
by the thermostat and not by external influences. Second assumption is that the
thermostat state’s actions do not affect its state.
It may be noted that, while the thermostat knows its local state, it knows nothing
about the room temperature. Consequently, we made all pairs of L A × L E possible,
including the (−, hot) as the possible world, indicating that thermostat’s local state
is indicating low temperature, while the environment state is hot. In one aspect, it
simplifies the system by assuming all possible worlds, while in other terms, it is
a blessing, as this is taken as a situation, where we assume that there may be a
measurement error in the thermostat [1].
Definition 16.1 (Belief) A belief assignment function B may be defined as B :
L A → (2 S − φ), so that for all l ∈ L A we have B(l) ⊆ W (l). The value B(l) is
referred as worlds plausible at l.
492 16 Intelligent Agents

16.8.3 Preferences

The beliefs make sense being part of a more detailed description of the agent’s mental
state, which has more associated aspects. One such aspect is the agent’s preference
order in the possible worlds, which can be taken as the agent’s desire. There are
various assumptions about the structure of the agent’s preferences, which consider a
total order on the set of possible worlds W . However, we may need a richer algebraic
structure, in some cases, e.g., one in which addition is defined. We use a value function
to represent the agent’s preferences.
Definition 16.2 (Value function) A value function is a function u : S → R. 
This numeric approach of representation of agent’s preferences is most convenient,
where a state s1 is at least as preferred as state s2 , iff u(s1 ) ≥ u(s2 ).
Considering the example of the thermostat (as agent), the goal of the agent is
to make room temperature ok. Thus, the thermostat/agent prefers any global state
in which the environment’s state is ok, over any global state which may be cold
or hot, and is indifferent between cold and hot. And, it is also indifferent between
the states, where the environment’s states are identical, i.e., (+, ok) and (−, ok).
This preference order over possible worlds can be represented by a value function.
The value function assigns zero to global state in that environment when the state is
either hot or cold, and assigns 1 where the environment’s state is ok. This outcome
is represented in Table 16.4, where ∗ stands for either − or +.
If the exact state of the world was known to the thermostat, it would have no trouble
in selecting proper action based on the value of its outcome. Considering the case of
value as cold, the Turn-on action would lead to the best outcome. However, when there
is uncertainty, the thermostat must compare vectors of plausible outcomes instead
of a single outcome. For example, for the belief assignment B(l) = {cold, ok}, the
plausible outcome of the action Turn-on is (1, 0), and of the action turn-off it is (0, 1).
Given the transition function τ , the belief assignment B, and an arbitrary, fixed
enumeration of elements of B(l), the plausible outcomes of a protocol P in l is a
tuple whose kth element is the value of the state generated by applying P starting at
the state of B(l).

Table 16.4 Global outcome Worlds (→) (∗, cold) (∗, hot) (∗, ok)
preference for an agent
Action (↓)
Turn-on 1 0 0
Turn-off 0 1 1
16.8 Modeling Agents as Decision Maker 493

Table 16.5 Table with Action (↓), Worlds Rainy Not-Rainy

weighted outcomes (→)
1. Take umbrella 10 −4
2. Do not take −1 8
umbrella

16.8.4 Decision Criteria

The values can be compared easily, however, it is not clear as how to compare the
plausible outcomes. Therefore, we choose some protocols. A strategy for making
choice under uncertainty is required that depends on the agent’s attitude towards
risk. The strategy can be represented by decision criteria, which is a function with a
set of plausible outcomes and returning a set of most preferred out of these. For this
we make use of maximin criteria discussed earlier. Hence, we reproduce the same
Table as 16.5.
Note that when both the worlds are plausible, the two plausible outcomes are
(10, −4) and (−1, 8). On the condition, the maximin criteria is used, the first action,
corresponding to “take umbrella” is the most preferred one. But, when the principle
of indifference is used, the plausible outcome “do not take umbrella” is preferred.
Accordingly, decision criteria can be defined as follows:
Definition 16.3 Decision criteria.
A decision criteria is a function:
 n  n
ρ: 2R → 2R − φ (16.3)
n∈N n∈N

that is, from sets

 of equal length tuples of reals, to sets of equal length tuples of reals,
so that U ∈ n∈N 2R − φ, we have that ρ(U ) ⊆ U (i.e., it returns a non-empty
n

subset of the argument set).

Note that the decision criteria can be used to compare tuples. For example, if
ρ{u, v} = {v}, then we say that v is more preferred than u.

16.9 Agent Communication Languages

The agents working together, irrespective of whether they are cooperating or compet-
ing, is called a multiagent system. These systems provide higher level of abstraction
than the traditional distributed computing. The abstractions are closer to the users’
expectations, and allow the designers a higher flexibility in determining the behavior.
For example, instead of hard-wiring a specific behavior into the agents, multiagent
494 16 Intelligent Agents

system designers design the agents with the capability to negotiate amongst them-
selves and find out the best course of action for a given situation. The ACLs (Agent
Communication Languages) must be flexible enough to accommodate abstractions
such as negotiations. But, the same flexibility makes it harder to succeed in under-
standing their semantics [9].
Due to this reason, we must examine many elements to arrive at he meaning of a
communication, which includes, type of meaning, perspective, basis (semantics or
pragmatics), context, and coverage, i.e., number of communication actions included.
The formal study of languages comprises three parts: 1. Syntax, which is con-
cerned with organizing the symbols to create the structure of language sentences, 2.
Semantics, which deals with what sense is denoted by the sentences and their parts,
and 3. Pragmatics, which is concerned with how the sentences are interpreted and
used. The combined meaning of a sentence is obtained due to semantics and pragmat-
ics. The pragmatics includes those considerations that are external to the language,
like, state of the agents, and the environment in which the text exists. Therefore,
the pragmatics can restrict, as to how the agents can relate to one another and how
they process the messages which are sent or received. In a situation when agents are
not fully cooperative or they cannot find out the implications, they cannot meet the
pragmatic requirements.
Semantics versus pragmatics
A perspective can be combined with a type of meaning, either personal or conven-
tional. In case of personal, the meaning of communication is based on intention
and interpretation of receiver/sender. The action, “purge this file” shall be taken by
the receiver as directive, whereas “This is an old file”, shall be taken as an asser-
tion. In Fig. 16.6, the inform construct is to give the information to the receiving
agent, the request construct requests for rain, and so on. In conventional meaning,
the meaning of communication actions is based on usage conventions. A language
is nothing but a system of conventions. Violating the idea of conventions, the tradi-
tional approaches go against the wisdom of having different labels for communication
actions. The language KQML (Knowledge Query Management Language) have all
acts as variants of tell, whereas for communication language Arcol, it is infom.

Fig. 16.6 An example of agent language

16.9 Agent Communication Languages 495

Context
In general, we do not understand a communication without context. Here, in agents, it
is agent’s physical or simulated environment, which becomes the context. For agents,
the social context is not as subtle as for humans, but they must understand what an
agent expects from others.
Coverage of communicative acts
When heterogeneous autonomous agents exchange information, the meaning of the
exchange is decided by communicative actions. All these actions fall into one of the
following categories:
• Assertive. This action is to inform. For example, “The door is shut.”
• Directive. This is for request, for example, “Shut the door”. It can also be used for
query, e.g., “Can I shut the door?”
• commissive. To promise something, e.g, “I will shut the door.”
• Prohibitive. It can ban something. For example, “Please do not shut the door”
• Declarative. It causes events in themselves. For example, “This information is
redundant.”
• Expressive. To express emotions and evaluations. “I wish that hurricane will stop.”
Communication actions can be represented in stylized forms like, “I hereby
request …” or “I hereby declare …”. The grammatical form emphasize that through
the language, you not only make statements but perform actions. The action by speak-
ing becomes the essence of communication. Figure 16.6 shows that all primitives of
this agent language are assertive or directive. In the agent language Arcol, one can
simulate commissiveness using other acts. All the acts can be reduced to the cate-
gory of assertive, but these categories have only restricted meanings. For example,
a request in Arcol language is the same as conveying to the receiver that the sensor
intends for it to perform the action.
Considering the code given in Fig. 16.6 for agent Avi, each communication act has
a challenge for language, which promotes mental agency. The traditional approaches
ignore whether Bob has really the capability to cause rain when it is requested or
allowed to do so, or whether it can stop the rain when it is prohibited from causing
the rain. Similar is the case for, whether Avi can make it rain when he promises; or
whether Avi has the authority to permit or prohibit any of Bob’s actions or to name
whether conditions.
Finally, the ACL approaches conclude that if Avi’s designer wants it to comply
with, then it does. This is quite unsatisfactory, because it means that agents do not
have any reasoning about their limitations.

16.9.1 Semantics of Agent Programs

A platform that supports the creation and deployment of multiple software agents
must have the capability to interoperate with a wide variety of custom-made, as
496 16 Intelligent Agents

Action Policy

Security Incoming
message

Meta
knowledge Outgoing
message

Action-
base
Function calls
Security
AGENT constraint

Action
constraint
Legacy data

Fig. 16.7 Architecture of IMPACT agent system

well as legacy software sources. What it requires for a software package S to be

considered as an agent program, is that it must come accompanied with tools to
augment, modify, and message S to another agent.
Figure 16.7 shows, the architecture of a system, called IMPACT, used for the
creation and deployment of multiple interactive intelligent agents. It was a joint
research project created by the collaborative work of some Universities. In IMPACT,
an agent comprises two parts as described below [2].
Software code
It is a program written in any programming language that supports a well-defined API
(application programming interface), which may be part of the code or developed
separately to augment the code. The program (S) may be represented as a pair,

S = (T S , F S ) (16.4)

where,

– T S is set of all data types manipulated by this program, and the set is closed under
all the subtypes, i.e., if τ is subtype of T S , then τ ∈ T S , and,
– F S is the set of all pre-defined functions of set S that are provided by the package’s
API.

In other words, S is a collection or hierarchy of objects classes in any standard

object data management language.
For example, in Oracle, the database may be viewed as S = (T S , F S ), where
T S comprises all data types (all attribute domains, tuple of different combinations
16.9 Agent Communication Languages 497

of domains, and relations on tuples). Whereas, F S is a set of all functions, i.e., all
relational operations: select, project, Cartesian products, join, union, etc.
At any given point of time t, the state of an agent will refer to a set O S (t) of objects
from the type T S , managed by its internal software code. An agent may change its
state by taking action, which may be triggered internally or by processing a message
received from other agents. But, an agent cannot directly change the state of another
agent, but can do so by issuing a request message to that agent.
Semantic wrapper
A semantic wrapper contains a large collection of semantic information. Following
are the typical contents of this information (see Fig. 16.7):

1. Service description. It is represented in some language, with flexibility to modify

it.
2. Message manager. It manages the data-structure associated with the message box,
and specifies and implements the policies.
3. Action module. It takes input of a new message consisting of an event. This
message is used to trigger zero or more actions. Thus, the action module requires:
(1) action base: the actions the agent may take in principle, conditions the agent
state must satisfy for the actions to execute, as well as the effects of those actions,
(2) Action requirements: conditions under which the agent is allowed or barred
from taking the actions, (3) Action policy: What actions to choose out of many?
4. Meta-knowledge module. It provides to the agent information about itself, as well
as about other existing agents in the world. This knowledge may include statistical
information on the reliability of other agents, the speed at which other agents can
provide the services, financial charges levied for such services. It also provides the
self- knowledge, like about its own performance, analysis of various operations
performed by itself.

16.9.2 Description Language for Interactive Agents

An Agent’s internal mechanism is based on languages, that describe the agent’s

behavior and its communication protocols. Examples are (1) Soar, a general cog-
nitive architecture for developing systems that exhibit intelligent behavior, and (2)
Knowledge Query and Manipulation Language (KQML), a language and protocol
for developing large-scale sharable and reusable knowledge-bases [4].
There is another language Q, which is used for describing interactions between
agents. Rather than depending on the internal mechanism, Q provides an interface
between computing professionals and scenario writers. Due to change in focus, from
internal mechanism to interaction, language’s syntax and semantics are quite differ-
ent. For example, agent accepting requests on or off, which have the standard mean-
ings. However, if agents received a move command, it may have different semantics,
like move fast, slow, as detailed by the semantics.
498 16 Intelligent Agents

Because the language Q is suited for interactions, it is used for scenario writing.
Example of primitives for interactions are cue and action. A cue is an event that
triggers interaction, while actions are requests to an agent which causes the change
in the environment. Unlike the programming languages, the language Q does not
define the semantics of cues and actions. Since different agents execute the cues
and actions in different ways, their semantics depend on corresponding agents. The
Example 16.2 demonstrates the cues (preceded with question mark) and actions
(preceded by exclamation mark).
Example 16.2 Cues and Actions.

(?hear "Hello" : from Tom}

(!walk :from class_room
:to library)
(!speak "Hello" : to Tom)
(?see library
:direction north)


In the above example, the following cues and actions are there:
1. Agent waits for Tom to say Hello (?hear),
2. Tom walks from the class room to library (!walk),
3. Agent says hello to Tom (!speak), and
4. Agent asks, do you see the library in north (?see).
The above are synchronous actions, and each one to be followed on completion
of the previous.
The asynchronous actions allow overlapped execution, like, in the Example 16.2,
walk can be asynchronous action, we can walk and speak, and so are agents. To
represent an action to be executed in asynchronous mode, we precede the action
with a double exclamation (!!), e.g., !!walk. For this, the agent may say hello to Tom,
just after it has begun the walk.
Example 16.3 Guarded commands.

(guard
((?hear "Hello" :from Tom)
(!speak "Hello" :to Tom) ... )
((?see library
:direction north)
(!walk :from class_room
:to library) ... )
(otherwise
(!send "I am still waiting"
:to Dickens) ... )))

16.9 Agent Communication Languages 499

Just like the common programming languages, the interaction language Q has
commands for conditional branching and recursive calls. Apart from this, its has
commands, called guarded commands for situations that require to observe multiple
cues at the same time. A guarded command combines the cues, actions, and forms.
After either of the cues becomes true, the guarded command evaluates the corre-
sponding form. If no cue is satisfied, it evaluates the otherwise clause, as shown
in Example 16.3.
In the above code, if any cue is encountered, e.g., “agent hears hello from Tom,”
then corresponding forms will be performed, i.e., agent says (replies) “hello to Tom.”
If no cue is observed by the guard command, it performs the otherwise clause,
and the agent sends the message “I am still waiting” to Dickens.
A collection of state transitions in the language Q constitutes a scenario. A sce-
nario defines each state as a guarded command, and it can include the conditions. A
program writer can draft scenarios in the form of simple state transitions, which can
describe fairly complex tasks. The scenarios can be invoked recursively.

16.10 Mobile Agents

All agents are not of the type, mobile agents. An agent sitting at a far off place can
communicate with its environment through older time mechanisms, RPC (remote
procedure call) and messaging. Such agents are called stationary agents, and executes
only on the system on which they begin execution (not moving, but stationary). If
such agents need information, which is not available on their systems, or they need
interaction with an agent (program) residing on other systems, they usually make
use of a communication mechanism, such as RPC or messaging.
However, a mobile agent is not bound to a particular system on which it begins
execution, but it is free to travel to other hosts in the network. Once created in one
execution environment, the mobile agent can transport its state (including data and
other information), and its code, to other execution environments in the network,
when everything is delivered there, it starts execution. The mobile agent is designed
with special ability, due to which it can transport itself from one system to another
system in the same network. This ability of the agent allows it to move to the system
containing an object with whom this agent wants to interact, then to take the advantage
of being in the same host or network, as an object. There is a number of benefits of
using mobile agents rather than doing the same job by remote procedure calls and
messaging. Some of the advantages are as follows.
Reduction in network traffic
Distributed systems need communication, involving interactions with multiple des-
tinations to perform a given task. This results to a large traffic in the entire network.
The mobile agents permit us to package a conversation and dispatch it to a destina-
tion host where interactions take place locally. This gets rid of the flow of raw data
in the network. When a large quantity of data is stored at remote hosts, that data is
500 16 Intelligent Agents

processed locally by the transported mobile agent rather than transferring over the
network. That is, computation is moved to the data center.
Overcome network latency
Critical real-time systems, such as nuclear reactors, and robots in manufacturing
processes are required to respond to act to changes in their environments. Controlling
such systems through a factory network introduces significant latency, due to many
reasons, like network being busy. For critical real-time systems, such latency is not
tolerable. Mobile agents offer a solution by moving the required programs and state
at the place where it is needed.
Encapsulates protocols
When data is exchanged in a distributed environment, each host owns the code, which
implements the protocols. However, when protocols change to add new features, it
becomes difficult to upgrade the protocol. Since mobile agents can move to remote
hosts to establish “channels”, this problem does not occur.
Execute asynchronously and autonomously
The mobile devices often need to rely on expensive and fragile network connections.
Tasks that require a continuously active connection between a mobile device and fixed
network are not economically, as well as technically feasible. To solve this, tasks are
dispatched into the network in the form of mobile agents, which can operate at their
ease, can move anywhere, where CPU resources and memory are abundant, and can
operate asynchronously and autonomously. The devices can reconnect at a later time
to collect back the agent.
Adapt dynamically
The mobile agents have capability to sense their execution environment, and can
react autonomously to changes. The mobile multiagent system can distribute the
agents geographically among the hosts in the network to perform any required task.
Robust and fault-tolerant
Mobile agents’ ability to react dynamically to unfavorable situations is useful to
build robust and fault-tolerant distributed systems. For example, if a host is being
shutdown, all agents executing on that machine are warned and given time to migrate
and continue their operation on another host in the network!

16.11 Social Level View of Multiagents

Since intelligence is mainly a social phenomenon, and it is due to the necessity of

social life, there is a need to construct socially intelligent systems to understand it, and
we need to build social entities to have intelligent systems. The society has adopted
a set of social laws, and each agent will be required to obey these laws, and will
16.11 Social Level View of Multiagents 501

assume that all other agents also follow the same. These laws, in one hand, constrain
the plans available to agents, and on the other hand, will guarantee certain behaviors
on the part of other agents. A social law may include communication protocol which
leads to rational deals with multiagents. The protocol may also include rationality
constraints for cooperation. The social law also, includes the rules, like those that
exist for humans for driving—left-drive or right-drive as it may be prevailing [7].
The idea of traffic rules for mobile robots highlights the important aspects of the
artificial social system approach. Rather than having a centralized controller or a
robot to continuously negotiate in order to avoid collisions, the better approach is
that robot should follow the traffic rules, “keep always to the left of the road”.
To consider the applicability of social laws, assume that there is a multi-robot
network, and we think of laws for agent mobilization in such systems. In such a
network, it is assumed that there is coordination among multiagent (i.e., robots). It
is formally defined using the following definition.
Definition 16.4 (Multi-robot network) A multi-robot network consists of a graph
G = (V, E), and set R of robots, and a strictly positive length function λ = E → R,
such that λ associates with each edge (u, v) of G, a distance which robot needs to
travel to go from u to v. We assume that there exists a clock such that a robot is at
some node or at some point between the nodes, at each point in the time scale. 
The action of a robot (agent) is direction and velocity. The velocity is a number
of distance units it passes in a unit time. The direction and velocity are decided by
the robot when it is a node in the graph. Also, a robot can observe another robot. The
robots need to meet the goals which arrive at them in a dynamic fashion. The goals
shall be met without collision to other robots. A collision may take place if they are
at the same node at the same point of time, or at the same step distance on edge at
the same time point. Based on these facts, it is possible to define some social laws.
Definition 16.5 (Social law for robot’s movement) Given a graph G = (V, E), the
social law for robot’s movement determines a subset A ⊆ E of edges in which robot
is allowed to move, and restricts the direction of movement along each edge of A,
and also restricts the velocity at which robots are allowed to move along each edge
e ∈ A. 
In the above definition, the social law is traffic law, which should guarantee that
each robot will be able to achieve its goals (say reaching to a destination node),
without any collision with other robots. This irrespective of what the other robots do.
Given a multi-robots system, a useful social law is one that guarantees non-
collision system, even if all the robots initially enter the graph G at arbitrary nodes,
with offset of at least one unit of time from each other, and they obey the social laws.
Such a system also guarantees that all the robots will reach their targets ultimately.
Modeling social actions
Design of social laws can be reduced to the problem of finding a route in a graph.
For this purpose consider that there is a simple graph G = (V, E), having no cycles,
502 16 Intelligent Agents

no parallel edges, no cut-vertex, and has at least three vertices. A graph with no
cut-vertex is called block.
Let N be the set positive integers, and f : V → N be a labeling of vertices in
G. For U ⊆ V , let min f (U ) and max f (U ), respectively, be the smallest and
largest labels of two vertices in U . An f-minimal vertex in U is any u ∈ U for which
f (u) = min f (U ), and similarly for f-maximal.
The labeling f as above indicates a directed graph G f = (V, A) on the same
vertex set as G, whose edge set A ⊆ E is obtained from E by removing each edge
(u, v) ∈ E with f (u) = f (v). Alternatively, orienting each edge (u, v) ∈ E with
f (u) < f (v), in the direction vu if v is f-maximal and u is f-minimal in the block
containing (u, v), and in the direction (u, v) otherwise.
Given this scenario, we have the following definition for routing a graph.
Definition 16.6 (Routing graph) A routing of a graph G = (V, E) is labeling f :
V → N of its vertices, for which the induced graph G f is strongly connected. A
routing under which there is unique f-minimal vertex r ∈ V , shall be called root.
Assigning the root is called rooting process. 
The routing of a graph can serve as a basis for useful social laws in multi-robot
systems. For this, the robots are required to enter the graph network from an f-
minimal vertex with an offset of one or more time unit from another. Additionally,
the robots are required to move only along the arcs of the graph G f induced by
the routing f . For this, a velocity function is defined as v : A → R as follows: for
e = (u, v) ∈ A, put v(e) = λ(e)/D if u is f-maximal, and v is f-minimal, and put
v(e) = λ(e)/( f (v) − f (u)) otherwise. It is mandatory that robots should move at
the calculated velocities.

16.12 Summary

We call a component an agent if it exhibits a combination of following characteristics:

autonomous, adaptable, knowledgeable, mobile, collaborative, persistent. Accord-
ingly, agents are classified as multiagents, autonomous, adaptable, collaborative,
proactive, personal, and mobile agents.
Agents have well-defined boundaries and interfaces, they are autonomous and
are capable of flexible, autonomous action in that environment in order to meet its
design objectives. An important benefit of multiagents is scalability. Since they are
inherently modular, it should be easier to add new agents to a multiagent system than
to add new capabilities to a monolithic system. They have flexible problem-solving
behavior, and they can be reactive or proactive. While parallelism is achieved using
multiagents, robustness and scalability are additional benefits.
In all cases of interactions, there are two major differences of agents when they
are compared with networked computing and shared computing: 1. agent-oriented
interactions take at knowledge level, and 2. operating in an environment that is
partially observable, the apparatus should make run-time decisions.
16.12 Summary 503

Software agents navigate on the Internet to collect relevant data, perform tasks,
and make decisions autonomously. They can transfer an enormous amount of data
on behalf of their users. Some agent systems also include standard service agents,
such as broker, auctionerr, or community maker.
Cooperation in multiagents is difficult when agents are self-interested, say, every-
one tries to download the same file at the same time, the speed will come down. Thus,
they need to communicate and cooperate. The cooperative agents should avoid the
situation of a prisoner’s dilemma. In cooperative agents coalitions need to be formed.
There are two approaches for this: (1) optimization-based, which finds an optimal
coalition, and (2) game theoretic approach, which has applications in many real-world
domains.
When agents are constantly changing coalitions, there is a need of formation of
dynamic coalitions. The total possible coalitions turn out to be of the order of O(n n )
for n number of agents. Hence, the number of agents should be small in number. To
study the coalitions phenomena, the agents are represented by nodes of a graph and
edges by the links indicating coalitions. The agents which are ready to provide the
resources are taken as sellers and those receiving are taken as buyers. This becomes
a structure to design a coalition algorithm.
Agents approach can also be applied to software engineering, where agents are
treated as next-generation components and this software engineering as case-based
software engineering. The complexity issues in software engineering can be tackled
through decomposition, abstraction, and organization. Agents can be assigned the
task of buying and selling. Software agents are used for filtering information that
matches people-to-people with similar interests, and automate the repetitive behavior.
There are theories to model the buying agents, which share the six fundamental
stages of buying processes: need identification, product brokering, merchant broker-
ing negotiations, purchase and delivery, product services and evaluation. For model-
ing agents as decision makers, formal notions of the mental attributes are used, such
as belief, knowledge, and references, accordingly, the modeling is called mental level
modeling.
When agents function together in cooperative or competitive mode, the multia-
gent system must provide the abstractions. Instead of providing specific behavior it
is designed flexible and can be coded using agent communication languages. The
languages have syntax, semantics, and pragmatics.
Many of the agents are mobile, and can just sit at a far place and communicate
with its environment, for example, through remote procedure calls or messaging. A
mobile agent has a feature that it can partly execute on one system, and can move to
another along with data, and can continue to execute the remaining part. Mobile
agents reduce network traffic, overcome network latency, encapsulate protocols,
execute asynchronously and autonomously, adapt dynamically, and have features
of robustness and fault tolerance.
The society has adopted a set of social laws, and each agent will be required to
obey these laws and will assume that all other agents also follow the same. These
laws, in one hand, constrain the plans available to agents, and on the other hand, will
guarantee certain behaviors on the part of other agents.
504 16 Intelligent Agents

Exercises

1. Label the following as an agent or not an agent. Explain your reasoning with
justification for each.

a. There is a program on a website to collect answers for a questionnaire.

b. Google’s web crawler, i.e., Googlebot.
c. A distributed IR (Information Retrieval) program to helps you locate Web
documents, you are interested in.
d. A program operating for a supermarket to automatically locate and bid for
the lowest food prices.
e. A mail-filtering program that removes SPAM messages in your e-mail
received in your account.
f. An Internet-wide multi-user game playing program.
g. A “chatterbot” program aimed to send messages to chat-rooms and try to
fool the people to make them believe that messages are coming from real
human beings.

2. In a multiagent system agent interact with the environment. How you can model
a situation where one agent modifies the environment and the other perceive it,
as a dynamic system?
3. How the architecture of a computer system is different from agent system? Give
the salient differences, and justify their significance.
4. A rat searches for food, and at the same time it has to save itself from its predators,
and expecting any such it either runs away or hides. For example, a single-agent
system model of a rat succeeds in protecting itself from predators as well as in
searching the food.
5. Explain the coordination and coalition functions between agents. How they differ
from each other.
6. Write the coalition algorithm in your own language.
7. Give an example of evidence of the prevailing use of agents in online buying from
the online stores.
8. Give a brief note of agent communication languages and compare them with other
high-level languages.

References

1. Brafman RI, Tennenholtz M (1997) Modeling agents as qualitative decision makers. Artif Intell
94:217–268
2. Eiter T et al (1999) Heterogeneous active agents, I: semantics. Artif Intell 108:179–255
3. Griss ML, Pour G (2001) Accelerating development with agent components. Computer 5:37–43
4. Ishida T (2002) Q: a scenario description language for interactive agents. Computer 11:42–47
5. Jennings NR (2000) On agent-based software engineering. Artif Intell 117:277–296
6. Maes P et al (1999) Agents that buy and sell. Commun ACM 42(3):81–91
References 505

7. Onn S, Tennenholtz M (1997) Determination of social laws for multi-agent mobilization. Artif
Intell 95:155–167
8. Peleteiro A (2014) Fostering cooperation through dynamic coalition formation and partner
switching. ACM Trans Auton Adapt Syst 9:1:1–1:31
9. Singh MP (1998) Agent communication languages: rethinking the principles. Computer 12:40–
47
10. Stone P, Veloso M (1997) Multiagent systems: a survey from a machine learning perspective.
CMU-CS-193 1-37
Chapter 17
Data Mining

Abstract Data mining, or knowledge discovery in databases, provides the tools to

sift through the vast data stores to find the trends, patterns, and correlations that
can guide strategic decision-making. The chapter highlights the major applications
of data mining, their perspectives, goals of data mining, evolution of data min-
ing algorithms—for transaction data, data streams, graph, and text-based data—and
classes of data mining algorithms—prediction methods, clustering, and association
rules. This is followed with cluster analysis, components of clustering task, pattern
representation and feature extraction, similarity measures, and partitional algorithms.
Data classification methods like decision trees and association rule mining are pre-
sented with worked examples. Sequential pattern mining algorithms are presented
with typical pattern mining and worked examples. The chapter concludes with sci-
entific applications of data mining, chapter summary, and list of practice exercises.

Keywords Data mining · Knowledge discovery · Goals of data mining · Data

mining algorithms · Transaction data · Data streams · Graph data · Prediction
methods · Clustering · Association rules · Rule mining · Cluster analysis · Pattern
representation · Feature extraction · Similarity measures · Partitioned algorithms ·
Scientific applications

17.1 Introduction

As Information Technology (IT) has progressed, there has been an abundant increase
in volumes of collected data in the recent past from all sorts of varieties. It is, therefore,
beyond the capabilities of humans to extract meaningful information from this vast
amount of data, and it has become necessary to develop algorithms which can extract
meaningful information from these vast stores of data. Searching for useful chunks
of information in the huge amounts of data is known as the field of data mining.
Data mining can be applied to all varieties of formats of data, including relational,
transaction, spatial databases, as well as large stores of unstructured data such as the
World Wide Web.

© Springer Nature India Private Limited 2020 507

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_17
508 17 Data Mining

The amount of data stored in digital form worldwide has on the average doubled
every 9 months, over many years, which is twice the rate for increase of computing
power, predicted by Moore’s law. This doubling of stored information, called storage
law, is one of the reasons of motivation for data mining. Irrespective of whether
this increasing volume of data will support exploration in commercial or scientific
activity, the data is potentially a valuable information [1].
It took many established organizations to accumulate large volumes of data about
their employees, suppliers, customers, products, and services. To perform the data
mining, also known as knowledge discovery in databases, organizations make use
of tools to sift through this vast amount of data to find the trends, patterns, and
correlations among the data sets, which can guide for strategic decision-making.
The traditional algorithms for data analysis assumed the input data sets of limited
number of records, hence the available memory was sufficient. Current databases,
however, are much larger to fit in the main memory of most computers. In addition,
to be efficient, the techniques for data mining applied to very large databases must
be highly scalable. An algorithm is considered as scalable if, given a fixed size of
main memory, its runtime increases linearly with the number of records in the input
database.
Data mining is concerned with the identification of patterns of important structures
in data; these structures may represent patterns, statistical or predictive models of
data, or some relationship among the parts of the data. In the context of data mining,
the terms like patterns, models, and relationships have definite meanings as follows:
a pattern is a compact summary of a subset of data (such as people who own a racer
car are likely to participate in car races); a data model can be model of an entire data
set, and it can be predictive. This model can be used to predict the future customers’
behaviors (such as whether the customer may buy a so and so product), based on the
historical data of interaction with some employees of the company. It can also be a
joint probability distribution with reference to some variables.
An algorithm can enumerate a large number of data patterns using a finite database,
but it is important to find out what data structure is suitable for a particular data set.
Identifying interesting structures and useful data patterns among the large number
of possibilities is the job of a good data mining algorithm, and it must do it fast
even for large data sets. Consider sales transactions in a database of an online store,
where some items’ transactions are more frequent than others, e.g., smartphones.
The variable values frequently occurring together in such databases could be used
to answer, say, which items were bought together from the same store. Such an
algorithm could also discover a pattern in the database in terms of demography, with
very good confidence.
Data Mining is a Science
The concept of unsupervised learning—from basic facts or axioms—has remained a
curiosity since long. In the present scenario, knowledge discovery engines are com-
monly used to extract general inferences from facts or using training data. Using more
structured approach, the statistical methods attempt to quantify the vast amounts of
17.1 Introduction 509

data by known and intuitively understood models. The approach of problem solution
through assembling knowledge from existing data sources is a radical shift from
traditional approaches of problem solution.
The nature of typical data sets, in respect of their size, diversity, high dimension-
ality, their distributed nature, and noisy contents, makes it challenging to decide the
formal specifications to any problem. This lack of control gives scope for solutions
that are over-fitting, have limited coverage with missing/incorrect data coupled with
high dimensionality. Once such problem is specified, the solution techniques deal
with the presentation and scalability of the problem, and complexity of the solution.
This complete process through which data mining makes its transitions is a science,
called data science.
Due to the emergence of WWW in the form of a large distributed data reposi-
tory, and the provision of large online databases that can be tapped for significant
commercial gain, interest of researchers and commercial organizations toward data
mining techniques has gone up exponentially. Many core mining techniques have
been explored by the researchers in the major domains, like classification, clustering,
rule associations, and time-series analysis. The work on data mining has focused on
scaling data mining algorithms to very large data sets. In the following sections of
this chapter, we shall discuss the algorithms concerning three classical data mining
problems: market-basket analysis, clustering, and classification.
Learning Outcomes of this Chapter:
1. Compare and contrast different uses of data mining as evidenced in both research
and application. [Assessment]
2. Explain the value of finding associations in market-basket data. [Familiarity]
3. Characterize the kinds of patterns that can be discovered by association rule
mining. [Assessment]
4. Describe how to extend a relational system to find patterns using association rules.
[Familiarity]
5. Evaluate different methodologies for effective application of data mining. [Assess-
ment]
6. Identify and characterize sources of noise and redundancy in presented data.
[Assessment]

17.2 Perspectives of Data Mining

The data mining algorithms employ a variety of models to characterize or evaluate

patterns. These models are from the areas of statistics, machine learning, databases,
and experimental algorithms. In addition, there are mathematical approaches that
used such approximation, and technical approaches such as dynamical systems. Fol-
lowing are the important perspectives used in the mining of data, which cover the
most domain areas of research in data mining [2, 3].
510 17 Data Mining

Induction
It is the most common perspective, which is based on the principle of proceeding from
specific to general, and answers the questions, like “Given ten specific examples of
good tourist destinations, find out the characteristics of a favorite tourist attraction?”
Thus, induction is typically implemented as a search through a space of possible
hypotheses. Such searches usually employ some special characteristics or aspect to
arrive at a good generalization, like “sand-dunes are favorite.”
Compression
Many a time, one set of data may correspond to a number of general concepts. In such
cases, mining techniques typically look for the most easily described pattern. This
principle is called Occam’s Razor, which effectively equates mining to compres-
sion, such that it becomes possible to describe the original data in a more compact
form using learned patterns, rather than exhaustively enumerating original data itself
requiring much larger size of space. The solid base to this issue are (1) feasibility
of models such as Minimum Description Length (MDL) principle and (2) compu-
tational learning theory. Most data mining systems make use of one of these views
of compression to establish the effectiveness of mining patterns. For example, if set
of a data is of size ten, and the number of patterns mined turns out to be 20 features
long, then this mining is not good at compression.

Definition 17.1 (Pattern) A pattern is a single data item of d dimensions, used by

a clustering algorithm. It is also called observation, or datum, or feature vector. It is
represented by x = (x1 , . . . , xd ). 

Definition 17.2 (Feature/attribute) Each of the scalar components xi of a pattern x

are called features (or attributes) of the pattern. 

Querying
The perspective of query comes from the databases, because most business data
reside in industrial and commercial databases and warehouses. Commercial database
experts take data mining as a form of query. For mining the data (which often is in
the form of text), it required to often enhance the expressiveness of the query, like
“find all the customers with similar transactions.” The other perspective of querying
is to find out suitable model for database, in place of relational, for data mining.
Approximation
This view of mining has an objective to find an accurate model of data, and introduce
some deliberate approximation in it to find out some hidden structures in the data.
One technique that has been found useful for this is latent semantic indexing, which
is useful in document retrieval. This technique makes use of transformations based
on linear algebra and approximations of matrices to locate hidden structures of word
usages, which means doing the searches beyond the simple keyword search.
17.2 Perspectives of Data Mining 511

Search
The search is related to induction, but its focus is on efficiency, and uses forward-
pruning patterns, e.g., frequent itemsets, to restrict the overall pattern space.
Beyond what has been discussed above, there are many other approaches to clas-
sify the task of data mining, which fall in various categories:
• based on the data they operate on, e.g., discrete data, labeled data, continuous data,
and time-series data;
• based on application domains, e.g., finance models, economic models, web-log
mining, and semi-structured models;
• based on their induced representations, e.g., association rules, decision trees, and
correlations.

17.3 Goals of Data Mining

The field of data mining aims to explore very large data sets efficiently, using methods
that are convenient, easy, and practical. However, this should be without extensive
training as well without a large work force. All the data mining applications have some
common goals, of identifying the patterns in the data, interpretation of these patterns,
and then perform the prediction or description either qualitatively or quantitatively
in general for all the data including those which may be generated in the near future.
Following are the major goals as well as the nature of applications of data mining.
Scaling analysis to large databases
The meaning of scalability is capability to handle large volumes of data which cannot
fit in memory of any computer. The objective is to abstract away the patterns from
the large databases, which provide information to mining algorithms to search for
the patterns.
Scaling to higher dimensional data and models
The normal statistical data analysis is a two-step process—at the first step we formu-
late some model, then use the data to fit to this model. However, for human beings,
formulation of a hypothesis that results in a model is not possible when the data set has
a very large number of variables, of the order of thousands, involving various demo-
graphics, text document analysis, retail transactions, and web browsing. Through
automated discovery from such a large volume of data, a model can be derived, and
can be used in lower dimensional spaces, as the problem can be understood much
easier at that level.
Automating search
The search requires enumeration of large data sets, creating hypotheses—the jobs
beyond human capacity. However, these are done by algorithms meant for this pur-
pose.
512 17 Data Mining

Finding pattern and models understandable

The most classical methods score on the models based on accuracy, i.e., how well the
model is useful to predicts the data, and on utility, i.e., on the magnitude of benefits
due to derived pattern. In addition to these, there are new measures in data mining,
like how well the model can be understood, on the novelty of pattern discovered by
it, and on how to further simplify the model.

17.4 Evolution of Data Mining Algorithms

Data mining is a data-driven field; here mining is concerned with real-world data
sets. In traditional data analysis, popular data sets are d-dimensional vectors of x
measurements on N number of objects, or N such objects having d number of mea-
surements (or attributes). These data are called multivariate data, and are represented
as an N × d matrix of data [4].
Some of the classical data mining and analysis problems associated with multi-
variate data are the following:

• Clustering: It is the process of learning a function, which can map x into a set of
categories, such that the categories are not known in advance.
• Classification: It is the process of learning a function (i.e., learning a mapping)
from x to y, where y is a categorical, or already defined scalar target variable of
interest.
• Density estimation: It is estimating the probability density function (PDF), for
vector x, i.e., p(x).
• Regression: The regression is the same as classification, except that table y takes
real values.

The dimension d of vectors x is significant in multivariate modeling. For example

in applications of text classification, and in clustering of gene expression data, d
can be as large as 104 . As per the density estimation theory, the amount of data
needed to reliably estimate a density function grows exponentially in d. However,
many predictive problems, like regression and classification, do not require a full
d-dimensional estimate of the PDF p(x), and instead rely on a simpler problem of
determining a conditional probability function p(y|x), where y is variable whose
value is required to be predicted.
The first tools used to model multivariate data are old modeling methods from the
domain of statistics and machine learning, e.g., logistic regression, linear regression,
discriminant analysis, and Näive Bayes. The new predictive models used in this field
are additive regression, decision trees, neural networks, as tools for more complex
data models. They are more flexible, however they sacrifice the interpreting ability.
17.4 Evolution of Data Mining Algorithms 513

17.4.1 Transactions Data

A common form of data to be mined in most business contexts is records of individual

“transactions” conducted by some individuals. Some of the common examples of
these transactions are as follows: while doing purchase of groceries in a store each
record is called a “basket”, and in case of an individual’s surfing a website each
record is a description of a page requested in the session. Applying a multivariate
concept we can view each of these record sets as sparse N × d matrix, where each of
the N rows in this matrix corresponds to an individual basket or session, and each of
the d columns corresponds to a particular item, and an entry (i, j) ∈ N × d is true
if item j was purchased by a customer, or it was requested as part of the session i by
the surfer, and it is false otherwise.
In actual cases, the parameters N and d can be very large. For example, a large
retail chain or online store like Amazon or Flipkart might record on the order of
a million baskets (i.e., N ) per week and may have 105 different items in its store
available for purchasing or downloading. That means 1011 entries of (i, j) cells.
These numbers are a challenge for the system to be computationally tractable and
suitable for statistical modeling. Considering these numbers for a store, we need
to compute a pairwise correlation matrix taking time of the order of O(N d 2 ) and
memory of O(d 2 ), which in numeric vales are 1016 and 1010 , respectively.
However, we note that the N × d matrix is sparse. Considering each of the N
customers buy only d = 10 items—the size of grocery basket—thus, only 10/105
entries or 0.01% of this matrix are nonzero. Hence, our objective is to use a small
subset of data, of the basket, called itemset I as information nugget (IN). An example
of itemset is a combination of items as “bread, butter, and porridge” in a basket of
grocery store.
There are many algorithms that can bring out all the frequent itemsets from a
sparse set of transaction data. A specific itemset I (I ∈ d) is called frequent if for
I , the relation f i > T holds, where f i is frequency (rows count in which all the
items in I were purchased), where T is some preselected threshold of rows (T ∈ N ).
Consider that f i = 10, and some preselected threshold is T × N = 0.00005 × 105 =
5. Obviously, the itemset I is a frequent itemset.
Other approach to data mining algorithms takes a statistical view of basket data,
as a density estimation problem instead of a search problem. This has a requirement
of an approach to find statistically significant itemset, i.e., an itemset I whose empir-
ical frequency deviates significantly from some baseline frequency. For example, a
Bayesian conditional probability-based approach can discover complex multi-items’
association, which has been ignored by others approaches.
We discussed that the transaction data is in the form of a sparse matrix N × d. This
form is in fact not a true picture of data—the real transactional data has significant
additional finer structures. This structure, in the case of retail items is that they are
arranged in some order, e.g., in the order of product hierarchies, and in the case of
the web pages, the order is, they are usually related to each other through hyperlinks.
Thus, columns of the sparse matrix of products as well of the web pages can have
514 17 Data Mining

themselves further more attributes, like “price” for itemsets and “contents” for web
pages, as well as inter-item relationships, like laptop and its cover, web page which
is the home page, etc. In addition, the rows of the transactional data (i.e., basket) can
have further attributes like purchased as general, weekends, monthly, or seasonally.

17.4.2 Data Streams

The transaction data, instead of all available at one time, are in fact arriving continu-
ously as a stream. As they flow by, they are available for mining only once. Similarly,
the web logs continue to grow as browsing continues, over time, resulting in a stream
of data. In such situations, the data miner’s interest is also to study the evolution of
the activity. The data streams give challenge, as to how to compute the aggregate
of transaction data. One possible approach is to use an incremental learning model,
such as classification tree. The further sections in this text discuss in more detail
about data streams.

17.4.3 Representation of Text-Based Data

The text-based data can be represented in the form of a graph. The N objects can
be represented as nodes in a graph, and edges can represent relationships among
the objects. Such “data graphs” can be used for representation of web pages, where
web pages are nodes and hyperlinks are edges of the graph. These graphs can be
represented by adjacency matrix, where nodes are labels for row and columns and
edges are entries in the cells of adjacency matrices. These matrices are however,
large and sparse. Like in graphs, some nodes have extremely high degree—outgoing
or incoming edges—while others may have degree of one only. If nodes of these are
sorted accordingly to their degrees, the result is a rule of the form,
1
degr ee α (17.1)
rank a
where a is called as “degree” of the exponent.
Use of matrices for the representation of graph has benefits that many classical
methods in linear algebra can make use of graphs for analyzing the properties of the
corresponding problems. For example, to discover from the connectivity information,
Google PageRank algorithm makes use of a recursive system of equations, that define
the importance of any page in terms of the importance of the pages pointing to it, as
well as how many such pages are pointing to it. The page rank of each page then can
be computed by solving these linear equations.
17.5 Classes of Data Mining Algorithms 515

17.5 Classes of Data Mining Algorithms

The process of data mining is increasingly being recognized as a key solution to

analyzing, digesting, and understanding the huge volume of digital data generated
through businesses, government activities, and through scientific research. Achiev-
ing this solution requires the scaling of mining algorithms to very large data sets.
Majority of the traditional database algorithms access the databases multiple times
or sometimes access these randomly. These are not possible when databases are very
large. At the same time we need to speed up the computation. The approaches used
for designing algorithms for handling such large databases are prediction methods,
clustering, and association rules [5].

17.5.1 Prediction Methods

The predictive modeling algorithms use data sets of training records as their input.
The goal of predictive modeling is to build a model that predicts some specified
attribute(s) value from the values of the other attributes. The predictive methods are
based on one or more of the following: Bayesian probability-based methods, decision
trees, neural networks, and support vector machines. We elaborate the two methods
here. The Bayesian approach and neural nets have been discussed in the previous
chapters.
Linear Classifiers
A model after the training procedure can be a rule set or the whole training set like
the nearest neighbor. In the following we will see that a straight line can be a model
as well. In an example, it has been found that the height and weight of persons are
available and medical experts have found that some of them are over-weight or under-
weight (see table 17.1). Accordingly, the data have been divided into two classes,
and are labeled as shown in Fig. 17.1. Consultation with experts may be expensive,
so we would like to construct a model from the available data. Then for any new
person, given the weight and height, this model could easily predict whether he/she
is over- or under-weight.
A model can be a rule like If weight ≥ 60, then over-weight.

Table 17.1 A training set

ID 1 2 3 4 5 6
Weight (kg) 50 60 70 70 80 90
Height (m) 1.6 1.7 1.9 1.5 1.7 1.6
Over-weighted No No No Yes Yes Yes
516 17 Data Mining

Fig. 17.1 A training set 200

190

170

Height
160

150

140
40 50 60 70 80 90
Weight

Fig. 17.2 A linear classifier 2.0

1.9

1.7
height

1.6

1.5

1.4
40 50 60 70 80 90
Weight

We note that this rule does not make much sense, because some tall people may be
thin even though their weights are more than 60. A better model may be the following
rule:
If weight/(height)2 ≥ 23, then over-weight.
The objective of classification is to identify a good model so that future predictions
are accurate. Here “weight” and “height” are called features or attributes. In statistics,
they are called variables. Each person is considered as a data instance (or a data
observation). Mathematically, we have x = [weight, height] as a data instance and
y = 1 or −1 as the label of each instance. Here, we have six training instances
x1 , . . . , x6 with corresponding class labels as y = [−1, −1, −1, 1, 1, 1]T (here −1
and 1 mean under-weight and over-weight, respectively).
As discussed above, we now concentrate on the linear classifier, and show that
a straight line can be a model also. Figure 17.2 shows that a line is separating the
training data of over-weight and under-weight persons, and this line can be expressed
by
0.2 × weight − 10 × height + 3 = 0. (17.2)

In general, such a line can be expressed by

17.5 Classes of Data Mining Algorithms 517

w T x + b = 0, (17.3)

where x = [weight, height]T , w = [0.2, −10]T , and b = 3. Then, for any new data
x, we check whether it is on the left or the right side of the line. That is,

i f w T x + b > 0 predict x as “over-weight”,

< 0 predict x as “under-weight”.

Support Vector Machines

The support vector machines (SVMs) (see page 418, Chap. 14, for details) are based
on simple principle of classification; they are powerful and popular approaches for
predictive modeling. They are successful in a number of applications, such as face
detection, handwriting recognition, text classification, and charmed quark detection.
SVMs are suitable for solving classification problems where attributes having
two possible values 0 and 1 are to be predicted. The classification using SVMs is
performed in a 2D space, where predictor attributes (i.e., 0, labeled with circles)
separate from those with dependent attribute (i.e., 1, labeled with boxes) as shown
in Fig. 17.3.
We can compute the optimal separating surface in SVMs by maximizing the
margin of separation between predictor and dependent attributes. This margin is the
distance between boundaries of the points of these two types of attributes (Fig. 17.3),
and is important as it is a measure of safety (robustness) in separating the two sets of
points, hence larger the margin, the better it is. As per the standard SVM formulation,
computation of optimal separating surface requires solving a quadratic optimization
problem [5].
Decision-Tree construction
Decision trees are specially attractive in data mining applications because any human
analysis can easily comprehend the models of decision trees. Further, the construction
of decision trees does not require any input parameters or prior knowledge about the
data. We start at the root node of the tree, and repeatedly choose a child node based

Fig. 17.3 Supported Vector

in

Machine-based classification
arg
M

w x=y+1
wx=y
w x=y−1
518 17 Data Mining

on the splitting criteria that evaluates a condition on the input record at the current
node. At the final node (leaf node), the process associates the record with the unique
leaf node of the tree.

17.5.2 Clustering

The process of clustering partitions a set of data, according to some similarity mea-
sure, into several groups such that “similar” records are in the same group, so that
each group represents a similar subpopulation in the data. As an example, each cluster
could be a group of customers, which has similar purchase histories or interactions or
some other factors or combinations (Sect. 17.6 discusses clustering in more details).
One technique to achieve scalability in clustering is to incrementally summarize
the regions of the data such that denser regions are summarized first. Because a cluster
corresponds to a denser region, the records within this region can be summarized
collectively through a summarized representation called “cluster feature” (CF). An
example of CF is a triple comprising of cluster centroid, cluster radius, and the
number of points in the cluster.
The CF-based approach is considered efficient due to two reasons: (1) they con-
sume lesser memory space as all the objects in a cluster are not required to be
maintained, and (2) if properly designed, they constitute sufficient information for
computing all intra-cluster and inter-cluster measurements required for making clus-
tering decisions. In addition, the distances between clusters, CFs, and radii of clusters,
and hence other properties of merged clusters, all can be computed quickly from the
CFs of individual clusters.
Some points in clusters can be discarded, while the others can be compressed as
defined below.

Definition 17.3 (Discardable point) A point is considered discardable if its mem-

bership can be ascertained with high confidence. 

Definition 17.4 (Compressible point) A point that is not discardable, but belongs
to a tight subcluster consisting of a set of points that always move between clusters
simultaneously, is called a compressible point. 
The CFs are also useful for scaling iterative clustering algorithms, like k-means
algorithm, and while doing so it identifies three types of points (records): sets of
discardable points, sets of compressible points, and a set of main-memory points. In
a cluster, only the CF of all the discardable points are retained, and the actual points
are discarded. Out of the remaining points, the compressible points are compressed,
and those still remaining, since they are neither discardable nor compressible, are
designated as main-memory records. The iterative cluster forming algorithm then
moves only the main-memory points and the CFs of compressible points, between
clusters until a criterion function is optimized, which concludes the formation of a
cluster.
17.5 Classes of Data Mining Algorithms 519

17.5.3 Association Rules

Under the association rules, we use a concept called market basket—a well-defined
business activity—which is a collection of items purchased by a customer in an
individual transaction. Such a transaction is possible due to a customer’s purchase
from, say, a grocery store, or an online purchase from a store such as Flipkart and
Amazon.
By performing business activities over time, the retailers usually accumulate huge
collection of transactions of performing a business activity. A common type of analy-
sis performed on the collections of such transactions’ database is to find sets of items,
or itemsets, that appear together in many transactions. A pattern usually required to
be extracted through this analysis consists of an itemset (market basket) and the cor-
responding number of transactions that contain it, with the objective that the business
can use knowledge of these patterns to improve the placement of items in a store
together, or can arrange the mail-order catalog pages, or web pages on a website, or
use this criteria to motivate potential customers with attractive offers who can buy
these itemsets.
The task of association rule mining is finding correlation between items in a
data set. The initial research in rule mining was largely motivated by the analysis
of market-basket data, the result of which allowed companies to better understand
purchasing behaviors and, as a consequence, better market audiences [5].

17.6 Data Clustering and Cluster Analysis

Data mining and clustering are both exploratory activities, hence clustering methods
are well suited for data mining. Clustering is often used as an important initial step
in several data mining processes. The data mining approaches that use clustering
methods are predictive modeling, database segmentation, and visualization of large
databases [6].
Data mining is often carried out on relational databases, in cases of transactions
that have well-defined structure and its columns can be used as features. However,
it is also carried out on very large unstructured databases like WWW, where the
contents are natural language text, mostly in HTML or XML format.
Clustering is in fact not a new field; it was being used in machine learning,
statistics, and biology. However, scalability was not a design goal in these fields,
and it was always assumed that complete a data set would fit in the main memory,
and focus remained on improving clustering quality and not scalability. Due to this
historical reason, majority of the clustering algorithms available today do not scale
to large data sets. Clustering methods are used in data mining for segmenting the
databases into homogeneous groups, that serve the purposes of data compression,
because now we are working with the clusters and not with individual items. It helps
to identify characteristics of subpopulations that are targeted for specific purposes
520 17 Data Mining

(e.g., marketing certain items aimed at specific section of the population). Clusters
in large databases can be visualized, to help the analysts in identifying groups and
subgroups that have similar characteristics.
By classification and grouping of objects based on common properties in some
meaningful way, we are able to analyze and describe the world. We human beings
are skilled enough at dividing the objects into groups (i.e., clusters) and can assign
particular objects to these groups. For example, even young children can label the
objects, as flowers, animals, trees, books, etc. Hence, clustering is an important cri-
teria for understanding the objects as different classes. Every new object is identified
based on whether it belongs to a so and so class or not. For example, if an object
x ∈ A, where A is a class, say apple, then we say “x is an apple.” Naturally, whether
x ∈ A or not depends on certain attributes of x.
The clustering process groups various data objects based only on information in
the data items, which describes the objects and their relationships to objects in the
same group. The goal of clustering is that the objects within a group be similar/related
to one another, and different/unrelated to objects in other groups. The clustering will
be better or more distinct, if there is greater similarity/ homogeneity within a group
and greater difference between the groups.
The following example demonstrates a sample database as clusters.

Example 17.1 Clusters of customers’ database based on three purchase behaviors:

quantity, unit price, and their combination.
From the given data set, for each customer we compute the total number of items
purchased and average price of all the items purchased. There are total 9 trans-
actions as shown in Table 17.2, which are distinguished across three clusters. The
clustering is based on common properties of items in each group/cluster, as follows:
customers in cluster 1 purchased a few high-priced items, customers in cluster 2 pur-
chased many high-priced items, and customers in cluster 3 purchased few low-priced
items. 

In clustering, we organize a collection of patterns into groups based on their

similarity. These patterns are usually represented as vectors of measurements, or

Table 17.2 Data groupings Cluster no. <Qty, unit price>

of similar objects
Cluster 1 <2, 1800>
<3, 2050>
<5, 2270>
Cluster 2 <15, 1800>
<18, 2200>
<12, 2380>
Cluster 3 <3, 250>
<4, 180>
<4, 200>
17.6 Data Clustering and Cluster Analysis 521

Fig. 17.4 Data clustering: a

Y x
x Y 44
x 3
input patterns, b clusters x x x x x x 4 5 4 3 3 3
formed xx 55
x x xx 4 4 33
x 4
x 2
x x x 1 1 2
x x x 1 1 2

X X
(a) (b)

points in a multidimensional space. Intuitively, the patterns in the same cluster are
more similar to each other than those in different clusters. Figure 17.4a, b depicts an
example of clustering based on patterns, where input patterns are shown in (a), and
the desired clusters formed are shown in part (b), where clusters belonging to the
same cluster are shown by identical labels. Note that the measurement of the patterns
in this case is (x, y) coordinate values.
One technique for clustering is called supervised learning, while other is called
unsupervised learning. In supervised technique, a collection of labels, i.e., pre-
classified patterns, are already provided, and the task is to label newly encountered
unlabeled patterns. The already provided labels, called training patterns, are used
to learn the descriptions of classes that in turn are used to label new unlabeled pat-
terns. In the case of unsupervised learning, the task is to group a given collection of
unlabeled patterns into some meaningful clusters. In fact, some kind of labels are
there, associated with the clusters this time also, but this category of labels are data
driven—obtained solely from the data, and not predefined.

17.6.1 Applications of Clustering

Clustering has applications in several fields, which require exploratory pattern-

analysis, grouping, decision-making, and machine learning. Some applications, as
examples, are given below.
Information Retrieval
The WWW comprises millions of pages of text, and a query to a web search engine,
like “colleges” would return hundreds of pages having relevance to the query. How-
ever, these results can be grouped (as clusters) based on the categories like graduate
programs, fees structures, intake, specializations, etc. Then each individual can be
explored by the user.
Biology
Biologists are applying clustering techniques to analyze large volume of genetic
information, for example, to find out the genes groups which have similar functions.
522 17 Data Mining

Business
In business and commerce, a large amount of information is collected on current and
potential customers, then clustering is performed on this information to segment the
customers into smaller number of groups so that additional analysis can be performed
on each group, for example, to predict marketing potential.
Similarly, there are applications for image segmentation, pattern classification,
and data mining, discussed in the following. In many such problems in such areas,
there is little prior information (or statistical models) available about the data, and
it is requirement that a decision-maker should make as few assumptions about the
data as possible. It is under these conditions the clustering approach is useful for
exploring the interrelationships among the data points to determine their structure.

17.6.2 General Utilities of Clustering

Clustering provides an abstraction from individual data objects to clusters comprising

these data objects. Thus, each cluster is representative of a data item and can be
called as a prototype for the data item. Having this, given a data item, it is possible to
determine the closest representative cluster of that data item. Based on this deductive
process, the following general utilities can be constructed using clustering techniques.
Summarization
Many data analysis methods, such as regression, have time complexity of O(n 2 ),
which makes it computationally difficult for large size of n, the number of objects.
However, instead of applying the required algorithm to the entire data set, if it is
applied on the representative clusters, the complexity will be far less, as it will be
decided by the number of clusters, and not the data items.
Nearest neighbors
To find the nearest data item neighbor of a ith element, it needs to be compared with
n − i elements, and for all n items it is O(n 2 ) complex. We know that if two data
items are in two different clusters, then they cannot be nearer than the corresponding
clusters. Thus, if n data items are in a set C clusters, then complexity to find the
nearest neighbor is only O(|C|), which may be far less than O(n 2 ).
Compression
Consider using prototype for representing data items, such that each data item is
represented by an index to its cluster. When similar data items are represented by
a single cluster holding only one data item, it is effective compression. When this
process is applied to sound, image, and video data, there is some loss of information,
which is acceptable, however, there is substantial reduction in size.
Example 17.2 Consider a group of 12 sales records each indicating sales price, and
have been sorted in ascending order as 5, 8, 11, 13, 15, 35, 45, 55, 72, 92, 201, and
215. It is required to partition these into three clusters.
17.6 Data Clustering and Cluster Analysis 523

Table 17.3 Simple clustering of data

Cluster 1 Cluster 2 Cluster 3
5, 8, 11, 13, 15 35, 45, 55, 72, 92 201, 215

Fig. 17.5 Data sets on

which centroids appear
failed

(a) (b)

The partitions finally formed are shown in Table 17.3.

This has been obtained using a simple clustering technique that partitions the data
along two largest gaps in the data sets. 

17.6.3 Traditional Clustering Methods

The partition-based technique, such as k-means partitioning, is based on optimizing a

given criterion that attempts to break a data set into k-clusters. This approach assumes
the cluster shapes as hyper-ellipsoidal, and the sizes of all clusters are assumed to
be same. However, it cannot find clusters that vary in size, as shown in Fig. 17.5.
However, the Density-Based Spatial Clustering of Applications with Noise
(DBscan) clustering technique can be used to construct clusters of arbitrary shapes
and sizes. This method defines a cluster to be a maximum-set of density connected
points—every core point in a cluster has got at least a minimum number of points
within a given radius. We can reach to all these points in a cluster from any point in
that cluster by traversing through a path of densely connected points, but the points
across different cluster cannot. Finally, this technique can be applied only when
density can be found out in advance, and it is uniform throughout the data set [6].

17.6.4 Clustering Process

A general pattern clustering process has the following steps:

1. pattern representation, which may also include feature selection and extraction,
2. defining proximity measures patterns specific to data domains,
524 17 Data Mining

3. grouping of patterns (clustering),

4. optionally, abstraction of data, and
5. optionally, assessment of output.
Figure 17.6 shows the first three steps from above, of a typical case of clustering.
The feedback path indicates that the grouping process output could affect feature
extraction and similarity computations in the next iteration.
The pattern representation depends on a number of criteria, these are available
patterns, classes and their number, feature types, and their scale for clustering algo-
rithm. All this information is not in the control of a programmer, hence feature
selection process helps in identifying the most effective subset of the original fea-
tures to use in clustering. Through feature extraction, one or more transformations
of input features is carried out to produce new salient features. Either the selection
or selection along with extraction can be used to obtain an appropriate set of features
to use in clustering [7].

Definition 17.5 (Pattern) A pattern set is denoted by X = {x1 , . . . , xn }. The ith

pattern in X is denoted by xi = (xi,1 , . . . , xi,d ). In many cases a pattern set to be
clustered is viewed as an n × d pattern matrix. 

Pattern proximity or closeness of one pattern to other is usually measured by a

distance function defined on a pair of patterns. A simple distance measuring func-
tion is Euclidean distance, which is often used to reflect dissimilarity between two
patterns—the more is the Euclidean distance between two patterns, the more is the
dissimilarity between them. And, if the Euclidean distance is zero, the patterns are
identical. Other similarity measures can be used to characterize the conceptual sim-
ilarity between patterns.
The grouping step can be carried out in many ways. The output of clustering
can be crisp or fuzzy. When clustering required at output is crisp (hard), the data is
partitioned into groups, whereas when it is fuzzy partition, each pattern has a variable
degree of membership of [0, 1], in each of the output clusters. In the other approach,
the algorithms based on hierarchical clustering produce a nested series of partitions
by merging or splitting of clusters based on similarity measures. In yet another way,
partitioning is performed such that, algorithms identify a partition that optimizes a
clustering criterion.

Pattern
represent- Patterns with
Patterns Feature ations Inter-pattern proximities Grouping Clusters
selection/
extraction similarity

feedback loop

Fig. 17.6 Stages in clustering

17.6 Data Clustering and Cluster Analysis 525

Definition 17.6 (Hard clustering) It is a technique that assigns a class label li to

each pattern xi , which identifies its class. The set of all labels for pattern set X is
L = {l1 , . . . , ln }, with li = {1, . . . , k}, where k is the number of clusters. 
Definition 17.7 (Fuzzy clustering) It is a clustering procedure that assigns to each
input pattern xi a fractional degree of membership f i, j in each output cluster j, for
all the k clusters. 
Data abstraction is aimed at extracting simple and compact representation of a data
set. Here simplicity is from the point of view of automatic analysis, so that a machine
can do the processing efficiently. Alternatively, the simplicity is due to being human
oriented, such that the representation is easy to understand and intuitively convincing
for humans. Usually, a data abstraction is a compact description of each cluster in
the form of cluster prototypes or as representative patterns such as the centroid.
All clustering algorithms produce clusters when presented with data, irrespective
of whether the data really contain clusters or not. It is not necessary that every set of
data contains some clusters. For example, the continuous sequence 1, 2, . . . , 100, in
no way represents a cluster. Only what we can have is all these numbers as clusters,
each of size one. So, it is important, how to evaluate the output of a clustering
algorithm? Apart from this, to evaluate the cluster quality, we want to know what
characterizes a “good” clustering result and a “poor” one? If the data does contain
clusters, some clustering algorithms may obtain “better” clusters than others.
Some of the criterias used to compare clustering algorithms are based on (1) the
manner in which clusters are formed, (2) data structure of the cluster, and 3. how
sensitive the clustering technique is to changes, which do not affect the data structure
of the cluster.

17.6.5 Pattern Representation and Feature Extraction

A better quality of pattern representation will result in a clustering that is simple and
easily understood; on the contrary, when it is a poor representation, it may result in a
complex clustering, even some times its true structure may be difficult to recognize.
For example, if a standard technique like Cartesian coordinates is used to represent
the patterns, many clustering algorithms are likely to fragment the data into two or
more clusters. But, if polar coordinate are used for representation of the clusters,
the radius coordinate causes tight clustering and a one-cluster solution can be easily
obtained.
A pattern can represent a physical object or an abstract notion. A physical object
can be a chair, table, book, house, etc., while an abstract notion can be, e.g., a style
of writing, attitude, belief, etc. Both of these can be represented in the form of
multidimensional vectors, one dimension for a one feature. The features of a pattern
can be either quantitative or qualitative. For example, if weight and color are the two
features used, then (black, 5) is the representation of a black object with 5 units of
weight, for degree of blackness.
526 17 Data Mining

The features can be classified as quantitative and qualitative.

1. Quantitative features:
a. discrete values;
b. continuous values;
c. interval values.
2. Qualitative features:
a. nominal or unordered (e.g., color);
b. ordinal (for temperature, e.g., cool or hot) or (for sound intensity, e.g., quiet
or loud).
Other representations making use of structured features are represented as tree
structures. In structured representations, a parent node represents a generalization of
its child nodes. For example, a parent node “4-wheeler” could be a generalization
of child nodes labeled as “cars”, “jeep”, and“tractor”. Further, the node “cars” could
be a generalization of car make, “Hundai”, “Tata”, “Maruti”, etc. The generalized
form of pattern representation, also called symbolic objects, are defined by logical
conjunctions of events. These events link values and features, where features can
take one or more values and all the objects are not required to be defined on the same
set of features.
It is often beneficial to isolate only the most discriminative and descriptive fea-
tures of the input set, and use these features exclusively in subsequent analysis.The
isolation of features can be through selection or extraction. A feature selection tech-
nique identifies a subset of the existing features for subsequent use, while feature
extraction technique computes new features from the original set to be used later.
In both these cases, the objective is to obtain better classification or computation
efficiency or at least one of these.

17.7 Clustering Algorithms

The selection of features in a pattern is an essential requirement in statistical pattern

recognition. However, in the clustering context, where it lacks class labels, the feature
selection is an ad hoc, but a necessity. As it lacks class labels, there can only be a trial-
and-error process for the selection of features. The resultant patterns are clustered,
and the output is evaluated using a validity index. There are some popular feature
extraction processes, like principal components analysis, which do not depend on
labeled data and can be used directly. The patterns having smaller number of features
are beneficial, as the output can be visually inspected by a human.
For clustering the objects/patterns, the first requirement is to find out similarities
between them, and more similar patterns are clubbed together to form clusters. In
the following, we discuss how to measure the similarities between patterns and some
standard algorithms for clustering.
17.7 Clustering Algorithms 527

17.7.1 Similarity Measures

Similarity is fundamental to the definition of a cluster. Hence, a measure of the sim-

ilarity between any two patterns drawn from the same feature space is important for
clustering procedures. Since there are many feature types and scales, the choice of
distance or proximity measure must be done carefully. It is usually common to com-
pute the dissimilarity (distance) between two patterns rather than the similarity [7].
Definition 17.8 (Distance measure) Given two objects O1 , O2 from the possible
universe of objects, the distance or dissimilarity between O1 and O2 is a real number
denoted by d(O1 , O2 ). 
Consider that A, B, and C are three objects or patterns, the following properties
hold for the distance measure for these objects:
d(A, B) = d(B, A) : by rule of Symmetry,
d(A, B) = 0, if and only if A = B : by Constancy of self-similarity,
d(A, B) ≥ 0 : by Positivity,
d(A, B) ≤ d(A, C) + d(C, B) : by rule of Triangular inequality.
The dissimilarity between two patterns is defined on the feature space using the
distance measure. Our focus shall be on distance metrics with continuous features.
The popular metric for continuous features is the Euclidean distance, expressed by


d
d2 (xi , x j ) = ( (xi,k − x j,k )2 )1/2
k=1
= xi − x j 2 . (17.4)

Equation (17.4) is a special case of the Minkowski’s metric, where p was taken
as 2, expressed by


d
d p (xi , x j ) = ( (xi,k − x j,k ) p )1/ p
k=1
= xi − x j  p . (17.5)

The approach based on Euclidean distance has an intuitive appeal, and the method
is commonly used to evaluate proximity of objects in 2 and 3D spaces. The method
works well when the data set comprises “isolated” or “compact” clusters. However, it
has a drawback, as there is a tendency of large-scaled feature to dominate the others
features. The solutions to this, is to incorporate normalization to the continuous
features (with a common range or variance) or some other weighting schemes.
Example 17.3 Consider a set of 2D data points as shown in Table 17.4, and given
a new data point, x = (2.5, 2.9) as a query, rank these database points based on
similarity with the query, using Euclidean distance.
528 17 Data Mining

Table 17.4 Original 2D data A1 A2

x1 1.9 1.7
x2 2.1 2.1
x3 2.6 3.0
x4 2.2 2.5
x5 1.8 2.0

Table 17.5 Euclidean Given data point Euclidean distance with x

distances
x1 1.341
x2 0.894
x3 0.141
x4 0.500
x5 1.140

Using Eq. (17.4), we compute the Euclidean distance for the 2D data points x1 , …,
x5 with respect to the query x = (2.5, 2.9). The result are shown in Table 17.5.

17.7.2 Nearest Neighbor Clustering

Since proximity between items plays an intuitive role in clustering, a method based
on the nearest neighbor distances can be an obvious choice as a basis of clustering
procedures.
An iterative algorithm assigns each unlabeled pattern to the cluster of its nearest
labeled neighbor pattern, with the condition that the distance to that nearest pattern
is below the given threshold.
This process continues until all the input patterns are labeled.
To grow the clusters from the nearest neighbor, a concept called mutual neighbor
distance, call it M Nd , can be used, which is expressed as

M Nd (xi , x j ) = Cn (x j , xi ) + Cn (xi , x j ). (17.6)

In the above, xi , x j are patterns, Cn (xi , x j ) represents the count of neighbor numbers
of x j with respect to xi . Figure 17.7 illustrates this concept. There are total six pat-
terns, P, Q, R, S, T, andU . In part (a), nearest neighbor of pattern P is Q, and Q’s
nearest neighbor is P. Also, Cn (P, Q) = Cn (Q, P) = 1, hence M Nd (P, Q) = 2. If
Cn (Q, R) = 1 but Cn (R, Q) = 2, then M Nd (Q, R) = Cn (Q, R) + Cn (R, Q) = 3.
Figure 17.7b is obtained from Fig. 17.7a by adding three new patterns S, T, andU .
Now, M Nd (Q, R) = 3, but M Nd (P, Q) = 5. Note that M Nd between P and Q has
increased from 2 to 5 due to additional three patterns S, T, andU , though the position
of P and Q have not changed.
17.7 Clustering Algorithms 529

X2 X2
R R

Q Q
U
P P
T S
X1 X1
(a) (b)

Fig. 17.7 Nearest neighbor clustering: a P and Q are more similar than P and R, b Q and R are
more similar than Q and P

(a) (b) (c) (d)

Fig. 17.8 Construction steps of k-nearest neighbor graph using original data: a original data, b 1-
c 2- d 3-nearest neighbor graphs

A general case of the nearest neighbor algorithm is k-nearest neighbor algorithm.

Figure 17.8 illustrates for k = 1, 2, and 3-nearest neighbor graphs for some simple
data set.
There are many advantages of representing data items using a k-nearest neigh-
bor groups (clusters). The far apart data items are completely disconnected in this
approach, and since the data items are connected with the nearer items such that the
weights on the edges in the graph are indicators of nearness, these weights are also
indicators of population density in the data items’ space. Since the items in sparser
and denser regions are modeled uniformly, the sparsity of the representation results
in algorithms that are computationally more efficient.

17.7.3 Partitional Algorithms

A partition-based clustering algorithm first obtains a single partition of the data,

without any structure. As the next step, clusters are produced by optimization of a
criterion function defined locally (over a subset of the patterns) or defined globally
(on the entire set of the patterns). Searching for a set of possible labeling for an
optimum value of a criterion is computationally expensive, and combinatorial in
nature. To simplify this, the algorithm is typically run multiple times with different
starting states, and the best configuration obtained from all of the runs is used as the
output clustering.
530 17 Data Mining

The above discussed algorithm has a problem of providing a choice of the desired
number of output clusters. However, it is computationally more efficient than hier-
archical clustering, when used for a large data set.
Squared Error Algorithms
The squared error function approach is the most intuitive concept for partitional
clustering, as it is ideally suited for compact and isolated clusters. For an input set of
X patterns, the squared error for clustering C , consisting K clusters (C1 , . . . , C K ),
can be expressed as


K 
mj
( j)
e2 (X , C ) =  xi − c j 2 . (17.7)
j=1 i=1

In Eq. (17.7), c j is centroid of the jth cluster in total K clusters formed, m j is the
( j)
number of patterns in the jth cluster, and xi is the ith pattern in the jth cluster.

Algorithm 17.1 Squared Error Clustering Algorithm

1: Select an initial partition X of patterns, with a fixed k number of clusters, and cluster centres
2: repeat
3: for each pattern xi ∈ X do
4: Find centroid c j (of cluster C j ) having minimum distance with pattern xi
5: C j = C j ∪ {xi }
6: Compute the new centroids (cluster centers) of all the clusters
7: end for
8: Merge and split clusters based on some heuristic criterion
9: until convergence is achieved
10: end

The steps of the squared error clustering algorithm are listed in Algorithm 17.1.
The repetition in the repeat …until loop continues until the convergence is achieved,
i.e., the cluster membership is stable.
k-means partitional clustering
The k-means is a partitional clustering technique that tries to find a k number of
clusters, the count is specified by the user. These are represented by their centroids.
It is the simplest and the most commonly used algorithm that uses thesquared error
criterion. The k-means algorithm starts with a random initial partition and keeps
reassigning the patterns to clusters based on the similarity between the pattern and
the cluster centers (centroid distances) until a convergence condition is reached. In
the process of clustering, there is no reassignment of any pattern from one cluster
to another, which gives it a property of linear time complexity. In other words, the
squared error decreases to some minimum threshold after some number of iterations.
Following are the major advantages of k-means algorithm. (1) It is easy to imple-
ment, (2) its time complexity is O(n), where n is the number of patterns. However, its
17.7 Clustering Algorithms 531

Fig. 17.9 The k-means X2

clustering is sensitive to
U V
initial partition

R
Q S T
P

X1

disadvantage is that it is sensitive to selection of the initial partition—if not properly

selected it may converge to a local minima of the criterion function value.
The following example demonstrates creation of clusters based on the k-means
algorithm.
Example 17.4 Using the k-means approach to perform partitioning.
Figure 17.9 shows 2D patterns P, Q, R, S, T, U, and V . The process is started with
initial patterns P, Q, and R. Around these, three (k = 3) clusters are to be con-
structed. We end up with the partition {{P}, {Q, R}, {S, T, U, V }}, where three clus-
ters are shown by ellipses. The squared error criterion value turns out to be much
larger for this partition. This will happen, for example, for the centroid versus the
patterns in the largest ellipse. Hence, we construct a better partition {{P, Q, R},
{S, T }, {U, V }}, where clusters are shown by rectangles. This grouping results in the
global minimum value of the squared error criterion function, for clustering com-
prising of k = 3 clusters. The correct three-cluster solution is obtained by choosing,
for example, P, S, and U as the initial cluster means, which will form the partition
as {{P, Q, R}, {S, T }, {U, V }} [7]. 

17.8 Comparison of Clustering Techniques

A collection of all the clusters corresponding to a given input data set is gener-
ally called as clustering. In the following discussions, we shall distinguish between
various types of clustering approaches such as partitional (i.e., un-nested) versus
hierarchical (i.e., nested), monothetic versus polythetic, and hard versus fuzzy [7].
Partitional versus Hierarchical
At the top level, distinction is made between hierarchical and partitional approaches
for clustering, where hierarchical-based methods produce a nested series of partitions.
However, the partitional (un-nested) methods produce only one partition. In the
hierarchical clustering, the features are used sequentially, whereas in the creation of
partitional clustering they are used simultaneously.
The hierarchical clustering algorithms produce a sequence of clusters, which
are nested in nature. They have a single all-inclusive cluster at the top, and single
532 17 Data Mining

X2
Cluster 3 U V

Similarity
Cluster 2
T
Cluster 1 S

Q
P R

P Q R S TU V
X1
(a) (b)

Fig. 17.10 a Data items in three clusters, b dendogram for clusters in (a)

point clusters at the bottom of the hierarchy. At the start, the algorithms, called
Agglomerative hierarchical algorithms, treat each data point as a separate cluster.
After this initial step, each further step merges two clusters that are most similar. For
demonstration of this, see Fig. 17.10b, from bottom towards top. For total n number
of data points at the beginning, each such step compares each node with each other
node, hence the worst case analysis is O(n 2 ). After each merger, the total number of
clusters reduce by one. The user can repeat these steps until the desired number of
clusters are obtained or the distance between two closet clusters goes above a certain
threshold. Since in the worst case n number of data points reduce to one cluster
in total n − 1 number of steps, the worst case time complexity of this algorithm is
O(n 3 ).
In some of the hierarchical methods, each cluster is represented by its centroid.
A centroid of a cluster is a data point, which is closest to the centre of that cluster. In
centroid-based methods, the distance between two clusters can be measured by how
similar the centroids of these clusters are. However, the centroid-based scheme of
distance calculation fails when some data points in a cluster are closer to the centroid
of other clusters than the centroid of their own cluster.
The working of a hierarchical clustering algorithm is illustrated in Fig. 17.10a. It
makes use of a 2D data set, in the form of seven patterns labeled as P, Q, R, S, T ,
U , and V . After merging these data sets, three clusters (1, 2, and 3) are obtained.
A hierarchical algorithm results in a nested grouping of patterns, which produces a
dendrogram, as illustrated in Fig. 17.10b. The similarity levels at which groupings
change are also marked in Fig. 17.10b; any two clusters can also be merged to form
a larger cluster based on the minimum distance criteria. In this example, clusters 2
and 3 can be merged to make a single bigger cluster.
A partitional clustering is created simply by dividing the set of input data objects
into non-overlapping subsets at the output such that each data object is in exactly
one subset, i.e., a cluster. A simple partitional (monothetic) clustering considers the
features sequentially to divide the given collection of patterns (see Fig. 17.11).
17.8 Comparison of Clustering Techniques 533

Fig. 17.11 Clustering based V

on monothetic partitioning X2
2 3 3
2 2 2
3 3
2 2 2
H1 3 3 3 3
3
1 1 H2
1 11 1 4
4
1 1 4 44 4 4
4
4 4
X1

Monothetic versus Polythetic

This criteria is related to the sequential versus simultaneous use of features for per-
forming the clustering. Most clustering algorithms are of the type polythetic, i.e., first,
all the features enter into the algorithm before the computation of distances between
patterns begin. Computation of distances, based on which clustering decisions are
made, are based on all these features, and not any specific feature. However, a mono-
thetic algorithm considers the features sequentially to divide the given collection of
patterns into clusters.
Figure 17.11 illustrates the the construction of monothetic type of clusters. It shows
that the collection is divided into two groups based on feature x1,i ∈ X 1 , which are
separated by a broken vertical line V . These two clusters are independently further
divided using the feature x2, j ∈ X 2 , indicated by broken lines H1 and H2 . The major
problem with this method is that, for a patterns’ dimensionality of d, it generates
2d clusters, which is exponential. For example, for even moderately large values
of d, (say d > 50), which is typical in Information Retrieval (IR) applications, the
number of clusters generated by this algorithm is 1015 —a division of data set into
uninterestingly small and fragmented clusters.
A monothetic class is defined in terms of features, such that these features are both
sufficient as well as necessary to identify the members of that class. For example,
people are clustered in age groups 0–25, 26–50, and greater than 50 years old. And, of
course there may be many other features existing for each individual in each cluster,
but one essential feature is the range of age, which is strictly followed in each group.
A broad set of criteria that are neither the compulsory requirement, nor are suf-
ficient, are used to define a polythetic class. A certain minimal number of defining
characteristics must be possessed by each member of the category, but it is not neces-
sary that some feature must compulsorily be found in each member of the category.
The distance between the members defines the membership in a class.
Incremental versus non-incremental
Incremental algorithms are useful when the pattern set to be used is very large, and
the constraints imposed on execution time, or memory space, or both of these affect
the nature of the algorithm. The traditional clustering algorithms were not designed
to work with large data sets, hence the feature of scalability was absent. However,
this feature is very much in need in most present day applications of large data sets.
534 17 Data Mining

The field of data mining has led to the development of clustering techniques that
minimize the number of scans through a pattern space, reduce the number of patterns
examined during execution, or reduce the data structures’ size in the algorithm. The
main advantage of incremental algorithms is that for them it is not required to store
the entire pattern matrix in the memory, to be used again and again. But, instead,
only part of that is stored at a time, which helps in reducing the total requirement
of space. These algorithms are generally not iterative, hence the execution time is
also small.
Algorithm 17.2 illustrates a typical incremental clustering algorithm. Let D0 be
the first data item, and C0 be the first cluster.

Algorithm 17.2 Incremental clustering Algorithm

1: C0 = {D0 }
2: d = 1 ; data item counter
3: k = 0 ; cluster counter
4: while there is data-item available in input do
5: Pickup the next data item Dd
6: Compute distance of Dd from all existing clusters’ centroids
7: Let m is minimum distance, from centroid of cluster Cm
8: if m greater than threshold then
9: k = k + 1 ; new cluster
10: Create new cluster Ck
11: Ck = Ck ∪ {Dd }
12: else
13: Cm = Cm ∪ {Dd }
14: end if
15: d = d + 1 ; next data-item
16: end while
17: end

Hard versus Fuzzy Clustering

A hard clustering algorithm assigns each input pattern to a single cluster in the output
produced. However, a fuzzy clustering method may assign a degree of membership
(belongingness) for each input pattern to several clusters in the output produced. A
fuzzy clustering can be converted to a hard clustering by assigning each pattern to
only one cluster, that which possesses the largest measure of membership of that
pattern, and the partial membership of other clusters is ignored.

17.9 Classification

The basic idea of the data classification problem can be simply described as follows:
given a training data with known labels or classes (e.g., as shown in Table 17.6), we
would like to learn a model, so that it can be used to predict data with unknown
labels. Let us consider that we have identified some customers through clustering
17.9 Classification 535

Table 17.6 Sample training database

Record ID Employment Age Salary Group
1 Self 30 30K C
2 Industry 35 40K C
3 Self 35 60K A
4 Self 30 70K A
5 Industry 35 40K C
6 Academia 50 70K D
7 Self 45 60K D
8 Academia 30 70K B
9 Industry 35 60K B

of the aggregated purchase information about the currently existing customers for a
certain company (see Table 17.2, page no. 520). Further, we have also acquired the
mailing list of potential customers out of these, with their demographic information.
As the next step, we would like to assign each person in the mailing list to one of
the three groups: A , B, and C, as shown in Table 17.6. The latter is for the purpose
of mailing them a catalog of items tailored to the individual’s buying patterns. This
task, data mining, makes use of historical information about current customers to
help in the prediction of cluster membership of new customers.
Let us assume that the training database with historical information has records
with attributes: salary, age, employment, and group. The goal is to build a model that
takes as input the predictor attributes and outputs a value for the dependent attribute.
When the dependent attribute is a numerical value, the problem is called regression,
otherwise it is a classification problem. In our present discussion the dependent
attributes (also called class labels) are A, B, and C, hence it is a classification
problem. The Table 17.6 is a sample training database with four predictor attributes:
salary, age, and employment, and group as dependent attribute.
There are many classification models for data mining applications. These include
the following: genetic algorithms, Bayesian networks, neural networks, log-linear
methods, statistical methods, and decision tables. The classification trees (i.e., deci-
sion trees) are popular due to the following reasons:
• Their representation is intuitive, which makes classification model easy to under-
stand,
• An analyst does not need to supply any input parameter for construction of decision
trees,
• The accuracy of prediction of decision trees is better than other classification
methods,
• It is possible to construct decision trees from very large training databases using
algorithms that are scalable and fast.
536 17 Data Mining

Decision Trees
The decision trees are tree structures whose leaves are classifications and their
branches are conjunctions of features that lead to classifications. Their significance
is because many data mining methods generate decision trees. The problem solution
methods learn these decision trees. One approach to learning a decision tree is to
split the example set into subsets, based on the value test of some attribute. This
process is repeated recursively on the subsets, with each split value becoming a sub-
tree root. The splitting stops when the subset becomes so small that further splitting
is not possible, i.e., the subset example contains only one classification. A split is
considered best if it produces the minimum number of classification in the subset.
That means, subsequent learning generates smaller subtrees, which will require less
further splitting, in effect reducing the number of steps for solution of Problem [8].
A decision tree algorithm used for the classification comprises two steps: tree
building and pruning. In the first step, most of the decision tree is grown top-down
in a greedy way. Starting with the root node, the database is examined by a method,
called “split selection”, that selects the split condition at each node. Then, the database
is partitioned and the procedure is applied recursively. In the pruning stage, the tree
constructed in the previous phase is pruned to control its size. The pruning methods
select the tree in a way that minimizes prediction errors. In some cases, decision-tree
construction algorithms separate the process of tree building and pruning, but other
algorithms interleave these processes to avoid unnecessary expansion of some nodes.
Algorithm 17.3 shows a sample tree building phase, node n where tree is to be
split. At the output, the algorithm provides a decision tree for the data partition D,
and new node value n which is the root of the decision tree.

Algorithm 17.3 Sample code for Tree building

1: % Input: Data-partition D, Node n, split selection criteria P
2: % Output: Decision tree for D, with it root as n
3: Build-Tree(n, D, P)
4: Apply P to D, and find splitting criteria for node n
5: if n splits then
6: Create its child nodes n 1 , n 2
7: partition D into D1 , D2 using split criteria
8: Build-Tree(n 1 , D1 , P)
9: Build-tree(n 2 , D2 , P)
10: end if

The choice of splitting criteria determines the equality of the decision tree. If the
training database does not fit into memory, we need a scalable data access method.
Many scalable algorithms are designed with a built-in feature, which ensures that
only a small set of statistics, like aggregate measures and counts, be sufficient to
implement the split selection method. Since aggregated data is far smaller in size
than the actual data, it is possible to construct the statistics in memory for each node
in a single scan over the corresponding database partition. These nodes satisfy the
17.9 Classification 537

Fig. 17.12 Sample decision Salary

tree for catalog mailing
≤ 50K > 50K

Group C Age

≤ 40 > 40

Employment Group D

Academia, Industry Self

Group B Group A

splitting criteria, such that in the process of recursively splitting we ultimately reach
the class label node at leaf. In a decision tree, each internal node is labeled with a
prediction attribute, called splitting attribute, and each leaf node is labeled with a
class label.

Example 17.5 Figure 17.12 shows the decision tree for a training data set as shown
in Table 17.6. The splitting attributes are salary, age, and employment, and the class
labels are Groups A, B, C, and D.

Every edge that originates from an internal node is labeled with a splitting predicate,
which involves only the node’s splitting attribute. The splitting predicate in the deci-
sion tree (Fig. 17.12) are ≤50K , >50K , ≤40, >40, Self, Academia, and Industry.
The splitting predicate has a property that any record will take a unique path from
the root to exactly one leaf node, that is the class label of that record. At a node, the
combined information about splitting attributes and splitting predicates is called the
splitting criterion. 

17.10 Association Rule Mining

Association rules are a set of significant correlations, frequent patterns, associations,

or causal structures from data sets found in various types of databases. Such databases
are transactional databases, relational databases, and other forms of data repositories.
Mining of association rules is capturing those correlations, patterns, and rules and
representing them in the form of some if …then rules. For example, given a set of
transactions, each transaction comprising a set of items, an association rule can be
an implication, X ⇒ Y , where X and Y are sets of items, indicating that presence
of itemset X implies the itemset Y . Consider that an insurance company finds a
strong correlation between two sets of policies X and Y of the form X ⇒ Y , which
may be an indicator that customers holding policy set X were also likely to hold
policy set Y , where X and Y may have one or more elements. Now the company
538 17 Data Mining

could more effectively target marketing the policy Y through those clients who hold
policy X but not Y , to motivate them to buy policy Y . In effect, the association rule
represents the knowledge about purchasing behavior of customers. The association
rule mining has been effectively applied to many different domains that include
market-basket analysis in commercial environments, astrophysics, crime prevention,
fluid dynamics, and counter terrorism, i.e., all the areas in which a relationship
between objects can be concluded as useful knowledge [5, 9].
Association rules analysis is a technique to uncover how items are associated with
each other. There are three commonly used ways to measure the association, which
indicate the strength of the association.
Definition 17.9 (Support) For a given set of data items, an association rule has
support s for some set of items X if s percent of transactions include all the items of
set X . 
In the sales transactions, for example, if we discover that sale of some items beyond
a certain proportion tend to have a significant impact on the total profits, we might
consider using that proportion as support threshold.
Definition 17.10 (Confidence) “Confidence” says how likely the itemset Y is pur-
chased when itemset X is purchased, i.e., to determine the value of X → Y . For a
given set of data items, an association rule has confidence c if c percent of transactions
that contain itemset X also contain itemset Y . 
“Confidence” is measured by the proportion of transactions with item X , in which
Y also appears. For the total number of transactions T , we compute c as

(X, Y )
c=
X
(X, Y )/T
=
X/T
suppor t (X, Y )
= .
suppor t (X )

In association rule mining, usually the goal is to discover all association rules
having the value of support and confidence greater than some user-specified minimum
threshold value [8].
Definition 17.11 (Lift) “Lift” (l) says how much it is likely that the itemset Y will be
purchased whenever itemset X is purchased, while controlling “lift” for how popular
the item Y is.
The lift is measured as
suppor t (X, Y )
l= . (17.8)
suppor t (X ) × suppor t (Y )

A lift value greater than 1 means item Y is likely to be bought if X is bought,

while its value less than 1 means Y is unlikely to be bought if X is bought.
17.10 Association Rule Mining 539

Table 17.7 Transactional database

Tran. ID Cust. ID Item name Price (in $) Date
101 201 Laptop 1500 8/20/2018
101 201 Tablet 300 8/20/2018
101 201 Smartphone 100 8/20/2018
102 201 Music system 500 8/25/2018
102 201 Smartphone 100 8/25/2018
103 202 Laptop 1500 8/30/2018
103 202 Music system 500 8/30/2018
103 202 Smartphone 100 8/30/2018

Example 17.6 Given the sales transactions in Table 17.7, find out the “support” for
the following:
1. Laptop,
2. Smartphone, and
3. Laptop and Smartphone
and “confidence” of a music system with respect to
1. Laptop,
2. Smartphone, and
3. Laptop and Smartphone.
There are three transactions in this table: numbers 101, 102, and 103, stored in
a relational database system. We note that out of the three transactions, two have
Laptop, three have Smartphone, and two have Laptop and Smartphone combined.
Accordingly, their “support” is 67, 100, and 67 percent, respectively.
In the second part, we are interested to compute confidence of “Laptop” → “Music
system”, i.e.,

suppor t (laptop, music system)

c=
suppor t (laptop)
1/3
=
2/3
= 50%.

The “lift” for the above is computed as follows.

suppor t (laptop, music system)

c=
suppor t (laptop) × suppor t (music system)
1/3
=
2/3 × 2/3
= 0.75.
540 17 Data Mining

The lift of 0.75 (< 1) indicates that item Y is unlikely to be bought by a customer
who is buying item X . 
Most of the algorithms for association rules mining comprise two distinct steps:
1. Find all sets of items with minimum support. The data are of the order of millions
of transactions, and a mining algorithm needs to count a large number of candi-
date itemsets to identify the frequent ones; hence this phase is computationally
expensive, and often time consuming.
2. Generation of inferences. The inferences or rules can be generated directly from
the frequent itemsets, and there is no need to scan the data sets again.
In the above two steps, most of the time is usually consumed by the first step, hence
the scalability becomes more important. The techniques for scalability can be divided
into two groups: 1) those techniques which reduce the total number of candidates to
be counted; and 2) those which make candidates’ counting more efficient. The first
category of techniques make use of the property—subsets of a frequent itemset must
also be frequent, and uses this concept as a pruning technique to reduce the number
of itemsets to be counted.
Variations of actual problem
The other approaches used in the data mining algorithms are focused on variations of
the actual problem. For example, for data sets and support values where the frequent
itemsets are very long—with n items there are 2n possible frequent subsets—this
makes finding all frequent itemsets an intractable problem. However, by looking
ahead, the set of maximal frequent itemsets can still be found efficiently. We can
make use of the following criteria for this: if an itemset is identified as infrequent, its
subsets are also infrequent, and hence none of its subsets need to be counted. A key
approach to this solution is to maximize the probability that itemsets counted due to
looking ahead are actually frequent, which requires estimation of some heuristics.
For this, a good heuristic can be to bias candidates’ generation such that most frequent
items appear in most of the candidate groups. This heuristic will make it more likely
for high-frequency items to be part of long and frequent itemsets.
Why at all the finding of frequent itemsets is a nontrivial problem? The first reason
is the number of customer transactions in the database, which can be so large that it
usually does not fit into the memory of a system. The second is the potential number
of frequent itemsets is exponential on the number of different items, although the
actual number of different itemsets can be much smaller. If a table has four different
itemsets, there can be 24 − 1 = 15 potential frequent itemsets. But, if minimum
support is taken as 60%, only five itemsets are actually frequent. Thus, there is need
of algorithms that are scalable with respect to the number of transactions, such that
they need to examine as few infrequent itemsets as possible.
17.10 Association Rule Mining 541

Nested Hash Tables

This is a new and efficient technique, which makes use of hash tables in a nested
manner to check as to which itemsets are contained in a transaction. This approach
is more effective when it is required to count shorter itemsets. A technique for longer
itemsets is database projection technique, given as Algorithm 17.4.

Algorithm 17.4 Database Projection algorithm for counting itemsets

1: Partition the candidate itemset into groups such that candidates in each group share a set of
common items.
2: Discards transactions Td , that do not include all the common items, from T . (T = all transactions)

3: Discards the common items from remaining set (T − Td ) (since they are known to be present.)
4: Discard the items not present in any of the candidates.
5: Count each of the candidate group.

This reduction in the number and size of the remaining transactions can result in
substantial improvements in the speed of counting.

17.11 Sequential Pattern Mining Algorithms

Sequential pattern mining is aimed at discovering patterns of frequent subsequences

in a sequence database. This database usually comprises a large number of sequences
of ordered events, with or without a notion of time; each such subsequence is con-
sidered as a database record. Such sequences commonly occur in any metric space,
either as a total ordering or partial ordering. There can be many examples of such
sequences: events in time, in codons, or nucleotide in amino acid, in computer net-
works, in website traversals, or characters in a text string are all examples, where
existence of a sequence may be significant, and the detection of frequent subse-
quences might be useful. Sequential pattern mining has emerged as a technology to
discover such subsequences [10].
The sequential pattern mining problem may be defined as follows: given a database
of sequences, such that each sequence is a list of transactions ordered by transaction
time, find out all sequential patterns with a user-specified minimum support value.
Here, a transaction consists of a set of items. The support of a pattern is the number
of data sequences that contain the pattern as a fraction of the total number of trans-
actions [11]. This parameter indicates a minimum number of sequences in which a
pattern must appear to be considered frequent. For example, if a user sets the mini-
mum support threshold to 2 sequences, the task of sequential pattern mining consists
of finding all subsequences appearing in at least 2 sequences of the input database.
542 17 Data Mining

17.11.1 Problem Statement

A sequential pattern mining problem can be stated using the following examples.
• Can we find out some strong correlations between users doing online purchasing
from a web page based on their behavior patterns versus the sequences of pages
visited and/or speed at which he/she browses a website pages?
• A market basket is a record with fields: customer ID, and list of all items, with
most of the item fields as False, and True only where the customer has bought
those items. Can we derive through analysis, more useful information about the
buying pattern of the a customer, if information related to time and sequence is
also included in the market basket?
• Can we recognize a user as a genuine user or hacker based on the suspicious
behavior observed due to analyzing a sequence of commands entered by that
user?
• Can we declare the elements of actions as “best practices”, based on the outcomes
and analysis of sequence of actions performed on any activity that results in the
given outcomes?
• Given the sequences of alerts and status by any system before it has failed, can we
determine those sequences set or subsequence set through some analysis which
are confirmatory predictions about the failure? That is, whether it is possible to
map those sequences to failure?
In simple terms, having many significant or important events occurring over the
scale of time, space, or some other metric, it is required to learn from the data by
considering the ordered sequences appearing in the data.

17.11.2 Notations for Sequential Pattern Mining

We want to have a formally defined notation to describe the problem of mining

sequential patterns. We base our discussions on market basket, consisting of customer
ID and list of items. Let the set of items in the form of literals be I = {i 1 , i 2 , . . . , i m }.
A non-empty unordered collection of items α = (i 1 , i 2 , . . . , i k ) is an event. However,
for ease of processing, and without any loss of generality, we assume them to be in
lexicographic order. A sequence (α1 → α2 →· · · → αl ) is an ordered list of events,
and a sequence with k-elements, where k = j |α j | is a k-sequence, i.e., the total
sum of events in that sequence. For example, (A → BC) is a 3-sequence, as there
are three events A, B, C in this sequence. A sequence α1 → α2 → · · · → αm 
is called as a subsequence of another sequence β1 → β2 . . . → βn , if there exist
integers i 1 < i 2 < · · · < i m such that α1 ⊆ βi1 , α2 ⊆ βi2 , . . . , αm ⊆ βim , and m ≤ n.
For example, the sequence A → BC is a subsequence of AE → F → BC D,
since A ⊆ AE and BC ⊆ BC D. Note that the order of events is also preserved in
this case.
17.11 Sequential Pattern Mining Algorithms 543

Consider that input sequences are stored in a database D, and input sequence is
an event with fields: sequence-id, event time, and the list of items. It is assumed that
event time is unique in every event, hence this field can be used as the event ID. We
define the support (the frequency) of a sequence, as the number of input sequences
in the database that contains the sequence αs and express it as σ (α, D).
The support can also calculated by a different method. For user-specified minimum
support value min_supp (a frequency), a sequence is frequent if the sequence occurs
more than min_supp times. We denote the frequent k-sequences as Fk . In addition, a
frequent sequence is called as maximal if it is not a subsequence of any other frequent
sequence. Further, a frequent sequence is called maximal if it is not a subsequence
of any other frequent sequence.
The above discussion can be termed as the problem definition for sequence mining
algorithms, whose data are transaction database or they are transaction data sets.

17.11.3 Typical Sequential Pattern Mining

We consider an example of customer’s retail transactions or purchase sequences

in a store, which shows for each customer the store items he/she purchased in
every week for a month’s duration. Out of these sequences of customer purchases,
each sequence can be represented using a schema [T ransaction I D/Customer I D,
ordered sequence of events], where each sequence may be a set of store items like
bread, butter, milk, vegetable, etc. Considering only two customers, a sequence in
the database may be [T 1, (br ead, butter ), (br ead, vegetable, butter ), (br ead),
(co f f ee, milk)]; [T 2, (milk), (co f f ee, br ead), (vegetbable, milk)]. We note
that customer having transaction ID T 1 made a purchase in each of the 4 weeks
in the month, while the customer with transaction ID T 2 made purchases during 3
weeks. There can be one or more items purchased in each of an event, i.e., during
each market visit. Hence, different sequences can have different lengths (T 1 has four
and T 2 has three), and each event in a sequence can have one or more items in its
set.
Sequence databases are mined using a sequential pattern mining algorithm, which
looks for repeated patterns, called frequent sequences. These patterns can be later
used to find associations between different items or events in their data, with end
use like marketing campaigns, reorganization of business, prediction, planning, etc.
Among the most common applications of sequential pattern mining, are web-based
applications, making use of WWW for e-commerce, business, trading, etc. Typ-
ical applications of web usage mining are the areas, like user modeling such as
web content personalization, prefetching and catching, reorganization of website,
e-commerce, and business intelligence.
Web usage mining is an application of sequential pattern mining, which confines
to finding user’s navigational patterns on WWW. These patterns are extracted from
web logs. Against the market-basket concept, where an event comprises the number
of items, the ordered events in web mining are each comprising only one item. This
is because a web user can access only one page at a time, and not many items in
544 17 Data Mining

an event. If, instead of an event time, a time window is considered for web mining,
then there can be more than one item (multiple pages), and in that case it becomes a
general case like market basket, having many items in an event, where an event is in
the form of a time window.

Example 17.7 Sequential pattern mining to raise promotional sale in a online store.

As discussed above, when a user is accessing web pages, such that only one page is
accessed at a time, it gives rise to a sequence database, which is equal to only one
item in each sequence’s ordered event list.
Let E = {a, b, c, d, e, f } be a set of events for six products’ web pages accessed
by different users in different order in some online sales application. The E also
indicates a set of six products.
We assume that web accesses database is web log, and for four users the web
access sequence is in the form of four records: [T 1, adbce]; [T 2, abecd f ];
[T 3, bab f aec]; [T 4, ab f ac]. From the web-log pattern mining of this web
sequence database, we find a frequent sequence abc, indicating that all the users
who visit product a’s web page http://www.prompt-sale.com/a.htm also immediately
visit product b’s web page http://www.prompt-sale.com/b.htm and then visit prod-
uct c’s page http://www.prompt-sale.com/c.htm (These are only ficticious names of
portals).
Having the patterns sequence of visiting the web pages in a certain order, e.g.,
the customers who visit the web page for product a will also visit web page b and
c, it is possible to increase the sales of product b and c. By placing promotional
discounts on product a’s web page, which is visited a number of times in sequence,
it is possible to increase the sale of other products b, and c. 
The web log can be maintained on the server where the web page is stored.
However, it can also be stored on the client side, or on a proxy server. Each choice
has its own advantages and disadvantages concerning finding the users’ relevant
patterns and navigational sessions.

17.11.4 Apriori-Based Algorithm

Apriori algorithm is used for mining of frequent itemsets, and for association rule
learning over transactional databases. The algorithm works based on the identification
of frequent individual items in the database and extending them to larger and larger
itemsets, as long as those itemsets appear frequently in the database. Apriori has the
property that “any subset of a frequent itemset must also be frequent.” This algorithm
makes several rounds of computation to compute frequent itemsets, such that in the
ith round it computes all frequent i-itemsets. A round in the Apriori algorithm has
the following three steps:
1. candidate generation,
2. candidate counting, and
17.11 Sequential Pattern Mining Algorithms 545

3. discarding unimportant candidates.

In the ith round of the candidate generation step, it generates a set of candidates
with i itemsets. Then, in the candidate counting step, it scans the transaction database
and counts the support of candidate itemsets. In the third step, candidates with support
lower than a user-specified minimum threshold are discarded, and frequent i itemsets
are retained [10].
In the first round, the Apriori algorithm generates a set of candidate itemsets each
containing all 1 itemsets, and counts their support value. Therefore, after the first
round, all frequent 1 itemsets are known. Similarly, the candidate itemsets generated
during the candidate generation step of round two are roughly all pairs of items.
Apriori algorithm reduces the set of candidate itemsets through pruning—a priori,
based on the knowledge about infrequent itemsets obtained from the previous rounds.
Those are the itemsets which cannot be frequent. The concept of pruning used here is
based on the general rule that “if an itemset is frequent”, all its subsets must also be
frequent. Hence, before we go to the candidates’ counting step, it is better to discard
every candidate itemset whose subset is infrequent [11].
The Apriori-based algorithm is presented as Algorithm 17.5. The task of the
algorithm is to find the set F of frequent sequences (also known as frequent sequential
patterns), for a given minimum support threshold min_supp.

Algorithm 17.5 Apriori-based algorithm

1: Start scanning of database in Apriori-based algorithm at k = 1
2: Let C1 = initial sequence generated in 1st generation (set of candidate 1-sequences)
3: repeat
4: k = k + 1 (next step)
5: Scan the entire database (In kth-iteration it finds frequent itemsets of size k)
6: Join Fk−1 frequent itemsets with itself to generate Ck (set of candidate sequences in the
kth-iteration).
7: Prune sequences in Ck which have no subsequences in Fk−1 (i.e., they are not large)
8: Create frequent itemsets Fk by adding all sequences from Ck with support ≥ min_supp
9: until no more candidate sequences left
10: end

The scanning of the database in an Apriori-based algorithm starts at k = 1. Let us

call the initial sequence generated in the first generation as C1 , the set of candidate
1-sequences (in kth generation as Ck ). Subsequently, the database is scanned by
the algorithm several times. Let, in kth-iteration it find frequent itemsets Fk , of
size k (k ≥ 2). Next, the algorithm performs Apriori-generate and join ( ap−gen
F1 ), to join k − 1 frequent itemsets in Fk−1 with itself to generate Ck , the set of
candidate sequences in the kth-iteration. It then prunes sequences in Ck which have
no subsequences in Fk−1 (i.e., which are not large), and creates Fk by adding all
sequences from Ck with support at least equal to min_supp. The above iteration
process goes on for k = 1, 2, . . . until there are no more candidate sequences left.
546 17 Data Mining

Example 17.8 Illustrating iteration steps in the Apriori algorithm.

Let the total number of items after scanning in the fist generation be 5, i.e., C1 =
{a : 6, b : 6, c : 3, d : 3, e : 5}, here every item has been represented with its support
in the form s : n, where s is the sequence, and n is the support count. Let us assume
that a minimum absolute support is min_supp = 4 (i.e., 67%). After pruning is over,
the list of frequent 1-sequences is F1 = {a : 6, b : 6, e : 5}. Using the join operation
), the Apriori has generated candidate sequences at k = 2 as, C2 = F1 
( ap−gen F1
= {aa : 3, ab : 5, ae : 4, ba : 3, bb : 4, be : 4, ea : 1, eb : 2, ee : 0}. Here, 
ap−gen
stands for Apriori “generate and set-join operation.” Note that sequences like ac, ad
are not considered as their subsequences: a, c and a, d are each pair not frequent,
and similar rules applied to the others cases. The support count in the sequences of
C2 is as per their frequency in the database.
Next, it prunes all sequences in C2 that do not have frequent subsequences (i.e.,
their subsequences do not appear in F1 ), and so on. The final resulting frequent set
is f s = Uk Fk = {a, b, e, ab, ae, bb, be, aba}, where Uk is the union of k number
of frequent sequences of Fk . The pruning procedure performed on itemsets in Ck
removes sequences that have infrequent subsequences. 

Example 17.9 Perform data mining using Apriori-based algorithm to find out the
frequent itemsets for the sales transaction data, for a minimum support of 60%.

Consider the transactional database shown in Table 17.7 (page number 539); for
convenience of reference it is reproduced as Table 17.8.
For min_supp value of 60%, an itemset is frequent if it is present in at least
two transactions. As per the Apriori algorithm, we perform many rounds through
the given database. In the first round all single items, which are candidate itemsets,
are counted during the candidate counting step. The frequent itemsets resulting in
1-sequence itemset are {Laptop(67%), Smartphone(100%), Music system(67%)}.
In the second round, the pairs are counted as frequent itemsets, and only those
pairs are candidates in which the individual item is frequent in round one, and they are
present in at least 60% of the transactions. As per these restrictions, the 2-sequence

Table 17.8 Transactional database

Tran. ID Cust. ID Item name Price (in $) Date
101 201 Laptop 1500 8/20/2018
101 201 Tablet 300 8/20/2018
101 201 Smartphone 100 8/20/2018
102 201 Music system 500 8/25/2018
102 201 Smartphone 100 8/25/2018
103 202 Laptop 1500 8/30/2018
103 202 Music system 500 8/30/2018
103 202 Smartphone 100 8/30/2018
17.11 Sequential Pattern Mining Algorithms 547

itemsets are {Laptop, Smartphone}(67%), {Music system, Smartphone}(67%). Note

that the pairs in which {Tablet} is an item are dropped, as it is not a frequent item in
the first round. However, it disqualifies as a pair also due to minimum support.
In round three, no itemset qualifies for 3-sequence. Hence, with respect to a
minimum support of 60%, the frequent itemsets in this sample database and their
support values are
{Laptop} 67%,
{Musicsystem} 67%,
{smar t phone} 100%,
{Laptop, smar t phone} 67%, and
{Musicsystem, smar t phone} 67%.
The Apriori algorithm counts the support of all frequent itemsets, as well of those
infrequent candidate itemsets that could not be eliminated during the pruning step.
These latter items sets are called negative border. Note that an item is in negative
border if it is infrequent, but its subsets are frequent. In the example above, the only
negative border itemset is {Laptop, Music system}. It is necessary that all subsets of
an itemset in the negative border are frequent, in the absence of that the item should
have been eliminated by the subset pruning step. 
We know that the total number of subsequences possible for n items is 2n , which
is exponential. Accordingly, as the number of sequences in the database becomes
larger, the Apriori techniques suffer exponential growth of candidate sequences dur-
ing execution, and consequently, suffer increased delays in mining.
The Apriori is a family of algorithms; they are better suited for discovering intra-
transaction associations and then to generate rules about the discovered associations.
However, the task of mining the sequences is actually discovering inter-transaction
associations, i.e., sequential patterns across the same or similar data. Hence, the
algorithm uses transactional databases as its data source.
One form of the Apriori algorithm has horizontal formatting, i.e., the original data
is sorted first by Customer_Id and then by Transaction-time, so that each customer’s
transactions appear together, but in the order of time stamp of the transaction. This
transformation results to the database, where the time stamps determine the order of
events. The mining is then performed on these database using the breadth-first search
(BFS) approach. Following are the steps of Apriori-based algorithm using horizontal
formatting.
1. Sort phase. This phase transforms the data set from the original database into
customer sequence database by sorting on the primary key as Customer_Id, sec-
ondary key as Transaction_Time.
2. L_itemset phase. This phase finds all large itemsets L (which meet the minimum
support).
3. Transformation phase. As there is a repeated requirement to find out which
of the long sequences are present in the customer sequence, each customer’s
sequence is transformed by replacing the corresponding transaction with the set of
L_itemsets contained in that transaction. The transactions without any L_itemsets
548 17 Data Mining

are dropped, as well as the customer sequences not containing any L_itemsets, but
these customer sequences are still considered as part of database for the purpose
of customers’ count.
4. Sequence phase. The sequence phase mines the set of L_itemsets to find the
frequent subsequences. The algorithm make multiple passes over the data, with
each pass beginning with a seed-set, for producing potentially large sequences
(i.e., candidates), and these candidates’ support is also calculated during this
pass. The sequences not meeting the minimum support criteria are pruned, and
the remaining become the seed-set for the next pass. The process starts with the
large 1-sequences, and terminates when either no candidates are generated or
none meet the minimum support.
5. Maximal Phase. The phase finds all maximal sequences among the set of large
sequences. The process is similar to finding all subsets of a given itemset, the
algorithm is similar as well.

Example 17.10 (Association rule mining) Consider the database having four trans-
actions as shown in Table 17.9. It is required to find the association rules if the
minimum support is 60% and minimum confidence is 80%.

We perform the following steps for the solution.

1. We note that item a appears in four transaction, hence, it has support of 100.
Accordingly, the items with frequency and support are (a, 4, 100%), (b, 4, 100%),
(c, 2, 50%), (d, 3, 75%), (e, 2, 50%), and ( f, 1, 25%).
2. Next, construct the itemsets having two items using the previous phase with
support ≥ 60%: (a, b, 4, 100%), (a, d, 3, 75%), and (b, d, 3, 75%).
3. Next, construct the itemsets having three items using the previous phase with
support ≥ 60% or more : (a, b, d, 3, 75%).
Next, we formulate rules and compute their confidence. For this, we take the items
from the previous phases with confidence at least 60% (see Table 17.10).
The rules with confidence less than 80% are pruned, and we are left with rules
shown in Table 17.11.

Table 17.9 Transactions Transaction_Id Itemset

database
T1 {a, b, d, f }
T2 {a, b, c, d, e}
T3 {a, b, c, e}
T4 {a, b, d}
17.12 Scientific Applications in Data Mining 549

Table 17.10 Rules with Rule Confidence c (%)

confidence ≥60% |b∩a|
a → b = P(b|a) = |a| = 4
4
100
b → a = P(a|b) = |a∩b|
|b| = 4
4
100
|d∩a|
a → d = P(d|a) = |a| = 43 75
d → a = P(a|d) = |a∩d|
|d| = 3
3
100
|d∩b|
b → d = P(d|b) = |b| = 43 75
d → b = P(b|d) = |b∩d|
|d| = 3
3
100
|d∩ab|
ab → d = P(d|ab) = |ab| = 3
4 75
d → ab = P(ab|d) = |ab∩d||d| =
3
3 100
|ad∩b|
ad → b = P(b|ad) = |ad| = 3
3 100
b → ad = P(ad|b) = |b∩ad||b| =
3
4 75
bd → a = P(a|bd) = |bd∩a|
|bd| =
3
2 100
|a∩bd|
a → bd = P(bd|a) = |a| = 3
4 75

Table 17.11 Rules with Rule Confidence c (%)

confidence ≥80% |b∩a|
a → b = P(b|a) = |a = 4
4
100
|a∩b|
b → a = P(a|b) = |b| = 44 100
d → a = P(a|d) = |a∩d|
|d| = 3
3
100
|b∩d|
d → b = P(b|d) = |d| = 3 3
100
d → ab = P(ab|d) = |ab∩d|
|d| =
3
3 100
ad → b = P(b|ad) = |ad∩b|
|ad| =
3
3 100
bd → a = P(a|bd) = |bd∩a|
|bd| =
3
2 100

17.12 Scientific Applications in Data Mining

The scientific applications have resulted in the accumulation of high-dimensional

data, data in the form of data streams, and temporal and spatial data. Following are
potential scientific applications of data mining [12].
Biomedical Engineering
Examples of data in biological sciences include genome DNA sequence sequencing
of organisms, as well as large macro molecules such as proteins, and RNA. Due to
the large size availability of this data, there is need to create systems for organizing,
storing, and dissemination. In addition, there is abundance of potential for automated
learning from these data sets. A number of robust machine learning and data mining
algorithms have emerged to take advantage of previous knowledge, and to create
new knowledge. Consequently, biology is changing from the field of research which
was dominated by—“formulation of hypothesis, conducting experiments, and eval-
uate results”, to more of a computational science, with attitude of “collecting and
550 17 Data Mining

storing data, mining new hypotheses, and confirm these with data or supplement by
experiment.” These results/data can be combined with the clinical data to achieve
better results and resolution for treatments.
Telecommunications
Due to the wide use of telecommunication in the world, vast quantities of high-
quality data is available already, mainly including call details collected at network
switches mainly for billing, which can be used for data mining tasks in detection
of toll fraud and marketing to consumers. Based on the data mining algorithms on
telecommunications data, the following benefits can be drawn:
• New architectures for networks. The new generation network infrastructures may
adapt to changes in traffic demands dynamically, to be achieved through data
mining techniques to understand and predict the network load.
• Mobility and micro-billing. Consumer-related activities may have separate billing
patterns.
• Mobile services. Data mining may enable customers for adaptive solutions, so that
they can get it through a few keystrokes.
• Security. Data may be collected and maintained through the records of billing,
travels, migration, to obtain national security-related information, to ensure the
national security.
Geospatial Data
The geographical data has scope and uses, which include digital data of all sorts,
which can be created, stored, processed, and disseminated by government and private
organizations, many of them through high-resolution remote sensing devices, data
collected through geographical positioning systems, and other devices, which are
position aware, like cellular phones.
The new pattern mining and clustering algorithms, which are highly scalable, are
useful to discover new and unexpected patterns, trends, and relationships embedded
in these data.
Climate Data and Earth’s Ecosystems
The climate data acquired through the terrestrial observation, Earth-observation satel-
lites, and ecosystem models provides lot of data, which can be mined for patterns
to discover future ecological problems and their management, including ecology
and even the health of the planet Earth. Such data consists of global snapshots of
the Earth, at typical intervals, with variables like atmosphere, land, and ocean (sea
surface temperature, precipitation), or accumulation of carbon.
There are two main components of Earth science data mining: (1) Modeling eco-
logical data, and (2) Design of efficient algorithms for discovering potential patterns
in these data. Through these patterns as well the existing patterns, it may be possible
to predict the effects, such as El Nino and tornado, etc.
17.13 Summary 551

17.13 Summary

The digital storage capacity has doubled over every 9 months—this is the phenomena,
which has necessitated the mining of data. Data mining provides tools, to sift through
vast amount of data accumulated by organizations over the years, to find the trends,
patterns, and correlations, which can be helpful in planning.
Data mining process is confronted with dimensionality, size, diversity, noise, and
distributed nature of data sets, which makes the formal specification of the problem,
a challenge. The solution techniques needs to deal with complexity, scalability, and
presentation. The entire process through which data mining makes its transitions
is called data science. The algorithms on data mining address three classical data
mining problems, i.e., clustering, market bucket analysis, and density estimation.
One important difference between the traditional databases, and data mining is the
size of data in later, such that data does not fit in the storage of computer, therefore, the
mining algorithm should be scalable. Data mining is concerned with the identification
of interesting data structure in the data, which may be patterns, statistical or predictive
model, or relation information. A model can be used to predict future customer
behaviors, e.g., the customer “A” may buy the so and products and goods.
The major goals of data mining are
• scaling analysis of large databases,
• scaling to higher dimensional and data models,
• automated search, and
• finding patterns and models understandable.
Five perspectives from the existing areas of machine learning (statistics, algorithms,
and databases) are used in data mining. These comprise induction, compression,
querying, approximation, and search.

The data mining techniques are also classified based on the following criteria:
• their induced representations, e.g., decision trees, association rules, and correla-
tions;
• the data on which they operate, e.g., continuous, discrete, labeled, time series; or
• application domains, e.g., finance, economic models, web-log mining, and semi-
structured models.
The common form of data for data mining is records of individuals, like transaction
record or record in the form of requested web page—the set of such records can
be viewed as N rows × d column matrix, and entry (i, j) is 1, for example, in a
transaction, if item j was purchased in the session i. The itemsets (d) in a row is
called basket. In one approach to data mining, the objective is to find statistically
significant itemset, that is, whose frequency is higher than some baseline frequency.
The transaction data is often in the form of a continuous stream, like continuous sales
transactions, or web logs. This gives a challenge on how to compute aggregate of the
data. The incremental learning algorithms are used in such cases.
552 17 Data Mining

Many of the traditional algorithms access the databases multiple times, or ran-
domly for analysis. This is not possible when databases are very large. The algorithm
used for such large databases are prediction methods, clustering, and association
rules. The clustering partitions a set of records according to some similarity func-
tion, into groups such that “similar” records are in the same group, an identification
of similar subpopulations in the data. Clustered records are represented by a sum-
marized feature called cluster feature (CF), which are efficient as they occupy less
space. This grouping is sufficient for inter-cluster and intra-cluster measurements. A
pattern x (which may be a feature vector, a datum, or observation) is a single data
item used by the clustering algorithms. It is typically a vector of d dimensions, repre-
sented as x = (x1 , . . . , xd ). In the feature vector x, the individual scalar components
xi are called features (or attributes).
Structured features can be represented using a tree data structure. The feature
selection techniques identify a subset of the existing features to be used later, while
feature extraction techniques compute new features from the original set. It is com-
mon to measure the dissimilarity between two patterns by measure of distance
between these objects, say, O1 and O2 , determined by d(O1 , O2 ).
The association rules, another method of classification, makes the use of mar-
ket basket—collection of items purchased by a customer—finds correlation between
itemsets, helping to understand the purchase behavior of customers. The cluster-
ing process groups the data objects using the information found in the data items
as well as the relation between data items. Clustering is also called unsupervised
classification, where it is required to group the given unlabeled patterns into mean-
ingful clusters. The supervised classification classifies the unlabeled patterns into
pre-classified patterns.
General applications of clustering are for pattern analysis, decision-making, and
machine learning. Typical applications are Information Retrieval (IR), analysis of
semantic information, analysis of business (sales) transactions, e.g., to find out poten-
tial customers, etc. The utilities like summarization, nearest neighbor, and compres-
sion can also be constructed based on techniques of data clustering. Typical pattern
clustering activity involves—pattern representation, feature selection and extraction,
pattern proximity measures, and data groupings (clustering).
Proximity or similarity measure is done by a distance function in Euclidean dis-
tance, where grouping can be based on many criterions, like soft (fuzzy) versus hard
partitioning, hierarchical versus partitional algorithms, etc. Patterns for clustering
algorithm may be represented using Cartesian coordinates or polar coordinates, and
a pattern can measure either a physical object, like chairs, or an abstract notion, like
style of writing. Different types of clustering are categories, which are distinguished
as having contrast characteristics, e.g., monothetic versus polythetic, hierarchical
(nested) versus partitional (un-nested), exclusive versus fuzzy, and complete versus
partial. The hierarchical clustering algorithm yields nested grouping of patterns in
the form of a dendogram, such that two clusters can be nested to form a larger cluster
based on minimum distance criterion.
17.13 Summary 553

When the patterns set to be used is very large such that constraints on memory size
and execution time affect the algorithm, we use the incremental algorithm. The latter
has advantage that it does not require the entire matrix to be stored in the memory,
hence space requirement is less. Since these algorithms are not iterative, the time
required is also less.
Since proximity between items plays an intuitive role of clustering, the nearest
neighbor approach is a useful basis of clustering procedures, which iteratively assigns
each unlabeled pattern to the cluster of the nearest labeled pattern provided that
the distance with that is below some prespecified threshold value. The partitional
technique of clustering produce clusters by optimizing a criterion function defined
on a subset of patterns or on all patterns. Its algorithm is run multiple number of times,
and every time with different starting points, and the best outcome is considered from
all the runs.
Many data mining approaches generate classification trees—a tree whose leaves
are classifications and their branches are conjunctions of features that lead to those
classifications. The decision-tree algorithm has two phases—tree building and tree
pruning. In the tree building phase, the algorithm examines the database using thesplit
selection method to compute the locally best criteria. This works recursively to build
trees.
We can mine the rules in the data, called association rules. The rules mining
algorithms capture the set of important correlations present in the database. For a
given set of transactions where each transaction is a set of items, an associative (or
association) rule can be an implication X ⇒ Y , e.g., in case of sales transactions,
this may indicate that those who purchase itemset X may also purchase itemset Y .
Therefore, for selling itemset Y , one may target the people whose bought X , and not
again who bought itemset Y . The association rule mining consists of two steps: (1)
find all itemsets with minimum support, and then, (2) generate the inferences from
these minimum support itemsets.
Sequential pattern mining technique can be described as given a database of
sequences where each sequence is a list of transactions ordered by transaction time,
discover all sequential patterns having a minimum support. A sequence in a sequential
pattern mining is an ordered list of events, where an event is denoted as i 1 , i 2 , . . . , i k ,
and i j is an item.
A sequential pattern mining algorithm mines the database of sequences, looking
for repeating patterns (known as frequent sequences). These sequences can be ana-
lyzed to find associations between different items or events in their data for purposes,
like marketing campaigns, business reorganization, prediction, and planning.
Web usage mining is an application of sequential pattern mining concerned with
finding user navigational patterns on the World Wide Web, by extracting knowledge
from web logs.
A commonly used family of algorithms, called Apriori algorithm, computes fre-
quent itemsets through several rounds, such that in a round i it computes all i-
itemsets’ frequent transactions. A round has two steps: 1) candidate generate, 2)
candidate counting, i.e., those having support greater than min_supp.
554 17 Data Mining

The scientific applications have resulted in accumulation of high-dimensional

data, stream data, and spatial and temporal data. Hence, they are most appropriate
fields for data mining algorithms. Some of the applications of data mining in scientific
fields are Biomedical engineering, Telecommunication, Geospatial data, Climate
data, and Earth ecosystems.

Exercises

1. How is data mining different from querying databases, like Oracle or MySql?
2. What were the trends in Information Technology (IT), which gave birth to the
field of data mining? Why did the field of data mining emerge so late compared
to databases?
3. Are the goals presented in this chapter identically applicable to all the data min-
ing domains, for example, mining of data generated due to collision in particle
accelerators versus online sales transactions versus Twitter data? Justify your
answer.
4. Given that the Apriori algorithm makes use of prior knowledge of subset support
properties,
a. Show that all non-empty subsets of a set of frequent items must also be frequent.
b. Show that the support of any non-empty subset R of itemset S must be at least
as large as the support of S.
5. The algorithms for frequent patterns mining consider only distinct items in a trans-
action (market basket or shopping basket). However, the multiple occurrences of
an item are common in a shopping basket, e.g., we often buy the things like 10
eggs, 3 breads, 2 kg dalia, 4 kg oil, 2 kg milk, etc., and this can be important
in transaction data analysis. Suggest an approach on how you will modify the
Apriori algorithm, or propose alternate method to efficiently consider multiple
occurrences of items?
6. Assume nonnegative prices of items in a store, and find out the nature of constraint
they represent in each of the following cases. Also suggest, how you will mine
the association rules in these.
a. At least one Sudoku game.
b. Itemsets, whose sum of prices is less than $250.
c. There is one free item, and other items, whose sum of prices is equal or more
than $300.
d. The average price of all the items is between $200 and $500.
7. Given a decision tree, you have the following options:
a. Convert the decision tree into rules, and then prune the resulting rules,
b. Prune the decision tree and then convert the pruned tree into rules.
Critically analyze both the approaches, and discuss their merits and demerits.
Exercises 555

8. Find out the worst case time complexity of decision-tree algorithm. Assume that
data set is D, each item has n number of attributes, and the number of training
tuples are |D|. (Hint. Answer is |D|.n. log |D|.)
9. It is required to cluster a given set of data into three clusters, where (x, y) represent
the location of the object, and the distance function is Euclidean distance. The
points are P1 (3, 12), P2 (3, 8), P3 (9, 5), Q 1 (4, 7), Q 2 (6, 4), Q 3 (7, 5), R1 (2, 3),
and R2 (5, 8). Use the k-means algorithm to show the three cluster centers after
the first round of execution, and after the final round of execution.

References

1. Fayyad U, Uthurysamy R (2002) Evolving data mining into solutions for insights. Commun
ACM 45(8):28–31
2. Han J et al (1999) Constraint-based multidimensional data mining. Computer 4:46–50
3. Ramakrishnan N, Ananth YG (1999) Data mining: from serendipity to science. Computer
8:34–37
4. Smyth P et al (2002) Data-driven evolution of data mining algorithms. Commun ACM
45(8):33–37
5. Bradley P et al (2002) Scaling mining algorithms to large databases. Commun ACM 45(8):38–
43
6. Karpis George et al (1999) Hierarchical clustering using dynamic modelling. Computer 4:68–
75
7. Jain AK et al (1999) Data clustering: a review. ACM Comput Surv 31(3):264–323
8. Ganti V et al (1999) Mining very large databases. Computer 8:38–45
9. Ceglar A, Roddick JF (2006) Association mining. ACM Comput Surv 38(2):1–46
10. Nizar R et al (2010) A taxonomy of sequential pattern mining algorithms. ACM Comput Surv
43:1:3.1–3.41
11. Carl H et al (2013) Sequential pattern mining—approaches and algorithms. ACM Comput Surv
45(2):19:1–19:39
12. Han J et al (2002) Emerging scientific applications in data mining. Commun ACM 45(8):54–58
Chapter 18
Information Retrieval

Abstract Information retrieval (IR) is the identification of documents or other units

of information in a collection that are relevant to a particular information need—a set
of questions to which someone would like to find an answer. This chapter presents the
basic strategies of IR in length along with their analysis, particularly emphasizing the
vector space and probabilistic models of IR, with worked examples in each category;
gives the detailed coverage to construction and maintenance of index, and its parallel
processing. The fuzzy logic-based retrieval, concept-based retrieval techniques, their
algorithms, and worked examples are presented; and Automatic Query Expansion
has been dealt with at length. Application of Bayesian networks, and inferences using
these have been demonstrated for IR. The newly emerged semantic web for futuristic
IR and its applications have been introduced; and the design aspects of distributed
IR suited for currently distributed information resources are treated in depth. The
chapter ends with the summary and a set of practice exercises.

Keywords Information retrieval · IR · Vector space model · Boolean model ·

Probabilistic model · Fuzzy-based IR · Concept-based IR · Automatic Query
Expansion · Indexing · Parallel index · Distributed index · Semantic web · Parallel
IR · Distributed IR · Query expansion · Bayesian networks

18.1 Introduction

The problem of Information Retrieval in the present context is highly relevant when
the volume of information generated is much more than the individuals can easily
digest. Consequently, it is extremely difficult to search, locate, and disseminate the
precisely desired information from the storage media, irrespective of whether it is
local or globally distributed. Ever-increasing information needs to be continually
added to the storage systems. This process has been fueled further by the arrival of
the Internet and the World Wide Web, as well as digital libraries, research publications
repositories, electronic editions of newspapers, journals, and magazines.
Given a set of documents or information collection and information need, Infor-
mation Retrieval (IR) identifies the documents or other units of information in the

© Springer Nature India Private Limited 2020 557

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_18
558 18 Information Retrieval

collection, which are relevant to that particular information need. The information
need is in the form of a set of questions one is interested to find an answer to. The aim
of IR is to locate the document or file which holds the desired information. Some-
times, it is also required to locate the actual position of the required information in
the document selected when the document is of a large size.
Here are some examples of IR tasks: finding an article in the Times of India that
discusses the freedom struggle of India; searching the recent postings in Twitter that
are related to a particular model of smart phones; finding the entries referring to
butterflies in an online encyclopedia, etc. [2].
The early IR methods were based on manually assigned keywords to the doc-
uments and required complicated Boolean queries. These methods were primarily
used by retrieval experts. However, in the 1970s, automatic indexing and natural lan-
guage queries gained popularity, and the IR facility became more and more available
to non-experts. The documents were commonly indexed by automatically consid-
ering all the terms in them as independent words. The documents were represented
using the set of these keywords, known as Bag-of-Words (BOW). The query format-
ting also became simplified in the form of short natural language formulation. This,
however, added noise in the documents’ representation, but the basic methodology
for indexing remained the same. Due to this, the non-expert users faced increasingly
more troubles, due to the “vocabulary problem”. This was because the keywords
chosen were often different than those used by the authors of the relevant docu-
ments. Due to this, systems’ recall 1 rates came down. In other cases, the contextual
differences between ambiguous keywords were not properly resolved through the
BOW’s approach, which reduced the precision of the results. These two problems
are generally referred to as synonymy and polysemy, respectively [16].
To solve the problem of polysemy,2 automatic Word Sense Disambiguation algo-
rithms helped to disambiguate the documents’ contents as well as the query. The
disambiguation algorithms make use of resources like WordNet Thesaurus, corpora
text, or co-occurrence data to find the possible senses of a word and map word
occurrences to the correct sense. The disambiguated senses are used in indexing and
in query processing, which would help to retrieve only the documents that match
the correct sense. Unfortunately, sometimes the inaccuracy and errors of automatic
disambiguation are more dangerous than not using the disambiguation at all [5].
The models of Information retrieval assume each document described by a set
of representative keywords, called index terms. An index term is a word from a
document, which is supposed to represent the semantics of the document. In general
the index terms are nouns, as nouns carry maximum meaning of a sentence. Each
of the index terms in a document is not equally important; some of the index terms
describe the given document better than others. An index term which appears in every
document in a given set of documents is of no use. However, a term which appears
in only, say, five documents out of thousands of documents should distinctly identify

1 Precision and recall are parameters which represent the performance of any IR system.
2 Polysemy is an ambiguity in an individual word or phrase, such that the word can be used (in
different contexts) to express two or more different meanings.
18.1 Introduction 559

those documents. Accordingly, an index term should be assigned some numerical

weight to capture this effect. If ki is an index term, d j is a document then wi, j ≥ 0 is
the weight associated with the (keyword, document) pair (ki , d j ).
Consider that there are a number of documents, and a query is put to retrieve the
documents relevant to this query. Ideally, the retrieval algorithm must return all the
documents that are relevant to that query, i.e., retrieved documents (say Ret) are the
same set as the relevant documents (say, Rel) in the entire set of documents. However,
in a realistic situation, systems may retrieve many non-relevant documents also along
with the relevant ones. To measure the effectiveness of retrieval, two ratios, precision
and recall, are used. The precision is the ratio of the number of relevant documents
retrieved to the total number of documents retrieved, and it provides the quality of the
answer set. However, this does not consider the total number of relevant documents
available in the original set. A system may have 90% precision if nine documents
are relevant out of total ten that it has retrieved, and there are a total of 100 relevant
documents in the store. Hence, the total relevant documents available also matters.
The recall is ratio of number of relevant documents retrieved to the total number of
relevant documents in the entire collection (see Fig. 18.1).
Intuitively, to keep precision high, one need to be overly careful (assuming that
human is itself an algorithm). In that case too a few documents will get retrieved.
But, in that process many useful documents would be left behind, making recall very
low. On the other hand, if one puts efforts to achieve high recall, i.e., trying to retrieve
the maximum number of relevant documents out of the total relevant count (a case
of too much aggressiveness), many non-relevant documents may also be retrieved,
resulting in poor precision. A typical such scenario is shown in Fig. 18.2.

All Documents

Retrieved
Documents
Relevant
Documents
Revevant
Ret
retrieved
Documents Rel

Relevant retrieved |Rel∩Ret|

P recision = T otal Retrieved = Ret
Relevant retrieved |Rel∩Ret|
Recall = T otal Relevant = Rel

Fig. 18.1 Precision and Recall

560 18 Information Retrieval

Fig. 18.2 Typical relation

1.0 Optimal
between Precision and
Recall

Precision
0.5

Typical

0.0
0.0 0.5 1.0
Recall

Learning Outcomes of This Chapter:

1. Explain basic information storage and retrieval concepts. [Familiarity]
2. Describe what issues are specific to efficient information retrieval. [Familiarity]
3. Give applications of alternative search strategies and explain why the particular
search strategy is appropriate for the application. [Assessment]
4. Design and implement a small to medium size information storage and retrieval
system, or digital library. [Usage]
5. Describe some of the technical solutions to the problems related to archiving and
preserving information in a digital library. [Familiarity]
6. Generate an index file for a collection of resources. [Usage]
7. Explain the role of an inverted index in locating a document in a collection.
[Familiarity]
8. Explain how stemming and stop words affect indexing. [Familiarity]
9. Describe key challenges in web crawling, e.g., detecting duplicate documents,
determining the crawling frontier. [Familiarity]

18.2 Retrieval Strategies

Consider a set of documents D1 , D2 , . . . , Dn and query Q, a retrieval strategy is

an algorithm for information retrieval that assigns a similarity measure sim(Q, Di )
for i = 1, 2, . . . , n, between query Q and document set D. Following are the most
common retrieval strategies [14].
• Boolean Indexing. A Boolean query results in ranking based on some score
assigned to the terms. This can be achieved by associating a weight with each
query term, which can be used to compute the similarity coefficient.
• Vector Space Model. In this model, the document Di and query Q are each repre-
sented as vectors in the term space. The model computes the similarity sim(Q, Di )
between two vectors.
• Probabilistic Model. A probability of relevance of a document to a query is based on
the likelihood that a term (i.e., query term) will appear in a relevant document. This
18.2 Retrieval Strategies 561

probability is computed for each term in the collection of documents. For terms
that match between a query and document, the similarity measure is computed as
a combination of probabilities of each of the matching terms. This similarity is a
measure of relevance of the query to the document.
• Inference networks. A Bayesian network is used to infer the relevance of a doc-
ument to a query, the later is a measure of similarity between the query and the
document.
• Fuzzy set-based retrieval. In this IR approach, a document is mapped to a fuzzy
set.3 For example, rain/0.9 versus rain/0.2 shows that in the first case, the meaning
of the keyword rain indicates heavy rain, while in the second set it shows a very
mild rain [7].

18.3 Boolean Model of IR System

The Boolean model for IR is a simple information retrieval model, based on set theory
and Boolean algebra. Since the concept of set is quite intuitive, the Boolean model
provides the framework that is easy to grasp by a common user of an IR system,
as well as simple to implement. Also, the Boolean queries in the form of Boolean
expressions, are precise in their semantics [14].
The Boolean model considers that the index terms, which are the represen-
tation of documents, are either present or absent in a document. Consequently,
the index terms’ weights are taken as binary (1 for presence and 0 for absence),
i.e., for a document Di the index term t j is either present or absent in Di , i.e.,
weight wi j ∈ {0, 1}. A query Q is composed of index terms connected through
the binary operators ∧, ∨, ¬ (and, or, not). A query can be expressed as con-
junctive normal form (CNF) or disjunctive normal form (DNF). For example the
query “Q = ta ∧ (tb ∨ ¬tc )” = “(ta ∧ tb ) ∨ (ta ∧ ¬tc )” can be written in DNF as
−
→
Q = (1, 0, 0) ∨ (1, 1, 0) ∨ (1, 1, 1), where each term is a binary weighted vector
corresponding to the tuple (ta , tb , tc ). Figure 18.3 represent this query using a Venn
diagram.
IR is a process of matching the patterns in the user inquiry with the patterns in the
prospective text documents. If the inquiry words are taken as a set of words X , and
the text document words are taken as a set of words Y , then IR is nothing but finding
the binary relation X × Y . However, words may appear in different morphological
forms. It is, therefore, necessary that—before the matching is performed, the words
in the inquiry as well as the words in the text document are reduced to their basic
form—called stem words, by a process called stemming.
Consider that X is a Boolean set inquiry (or query) keywords, Y is a Boolean set
representing the keywords in the document. With this, a binary relation from X to Y
can be represented as

3 A set that contained not only the elements but a number associated with each element that indicates

the strength of the membership of the term.

562 18 Information Retrieval

Fig. 18.3 Conjunctive ta tb

components of query
Q = ta ∧ (tb ∨ ¬tc )
(1,1,0)
(1,0,0)
(1,1,1)

tc

R : X × Y ∈ {0, 1}. (18.1)

If for every x ∈ X , there is a corresponding y ∈ Y , then we say R(x, y) = 1, i.e.,

x is related to y, otherwise R(x, y) = 0.
Since in a Boolean model a query term x is either related to a document term y
(i.e., R(x, y) = 1), or not related to y (R(x, y) = 0), there are sharp boundaries of
relations, hence a Boolean-based model is also called Crisp set -based model. The
following example demonstrates crisp set-based IR.

Example 18.1 Crisp set-based Information Retrieval.

Consider a set of text documents Y consisting of documents y1 , y2 , ..., yn , as potential

documents to be searched for the keywords x1 , x2 , ..., xm in the inquiry set X . Assume
that n = 2, and y1 , y2 are as given below: (quoted from, “Einstein—The Life and
Times,” by Ronald W. Clark).
y1 : “Thus the new concept of the subatomic world was even by 1920 beginning
to produce a gulf. Bohr, Born, and a number of Einstein’s other contemporaries, as
well as the many of younger men who were in great part responsible for the new idea
readily jumped the gap. Einstein stayed where he was. Therefore, the scene in many
ways paralleled that into which he has launched his theory of relativity two decades
earlier. But then he had been in the iconoclastic vanguard; now he took up station
with the small conservative rearguard.”
y2 : “Plank, the man of honor who had yet not signed the manifesto of 93, had in
fact for the first time done as much to keep Einstein in Berlin as he had done to bring
him there in 1914. His letter, which, in Einstein’s words, had induced him ....”
Let the inquiry be “Einstein’s Scientific Theory of Relativity”; and therefore the
corresponding set of keywords in the inquiry is X = {x1 = “Einstein”, x2 = “Scien-
tific”, x3 = “Theory of Relativity” }, and document set Y consists of documents y1 ,
y2 , therefore, Y = {y1 , y2 }.
In crisp set-based IR, it is required to find R—a subset of X × Y , i.e., R (“Ein-
stein”, y1 ), R (“Scientific”, y1 ), R (“Theory of Relativity”, y1 ), R (“Einstein”, y2 ),
R (“Scientific”, y2 ), R (“Theory of Relativity”, y2 ). The Boolean relation R for the
example under consideration is given in Table 18.1.
It may be seen that while y1 has two matchings, y2 has only one matching. Hence,
Table 18.1 shows that the relative relevance of y1 with respect to y2 , given by the sum
of matching counts in y1 divided by the sum of matching counts in y2 is 2. For larger
18.3 Boolean Model of IR System 563

Table 18.1 Crisp set relation Query (X ) Document (Y )

between query and document
y1 y2
x1 1 1
x2 0 0
x3 1 0

number of documents: y1 , y2 , …, yn , the relevance can be computed in the similar

manner for a specified inquiry. Once the relevance is computed, the IR system lists
the documents in the order of their relevance.
In the classical binary logic, there is either 100% match for an index term in the
inquiry and corresponding term in the document, so R(xi , yi ) = 1; or there is no
match at all with R(xi , yi ) = 0. However, the real-life situations are often different.
It may happen that the two terms from index and document which are being matched
are two forms of the same word, e.g., real and reality, phrase and phrases, exact and
exactness, etc. In all these cases the crisp logic returns zero value of relevance. 

18.4 Vector Space Model

The vector space model determines the measure of similarity (sim(Q, Di )) between
a query and a document using a vector that represents query Q, and document Di .
The model is based on the intuitive notion that meaning of a document is conveyed
by the words present in the document. Figure 18.4 represents the concepts of vectors
for a query Q and documents D1 , D2 , . . . , Dn . Here t is the total number of terms
in the collection [12].

Fig. 18.4 Vector space

model Query
Q q1 , q2 , . . . , qt

Document D11 D12 . . . , D1t

D1

Document
D21 D22 . . . , D2t
D2

Document
Dn Dn1 Dn2 . . . , Dnt
564 18 Information Retrieval

The vector space model is based on a method which compares how close the query
vector is to the document vector. The traditional method of determining the closeness
between two vectors is the angle between them, where the angle is computed using
dot products of two vectors. In our context here, the similarity coefficient is used
instead of the angle. If a term is present in the vector, 1 is placed else 0 is placed in the
corresponding position in the vector. To account for multiple occurrences of a term in
the document, frequency of that term is used instead of merely its presence/absence.
Thus, for a query a, b, two documents’ vectors, 1, 0, and 5, 1 indicate that in
the first terms a and b have frequency of 1 and 0, respectively, and for the second
these frequencies are 5 and 1, respectively.
The similarity between query and documents can be computed as the distance
between the query and each of the document vectors. If a document has the same
vector as the query, their distance is minimum and have the highest similarity.
Instead of specifying the list of terms in a query, a user is often interested to
indicate that certain terms are more important than others. One approach to this
is that user indicates a higher weight to specific terms by manually specifying the
weight. The other approach is automatically assigning the weight equal to the number
of times (frequency of the term) a term appears in a document. But, if a common term
appears in every document, it cannot be used to distinguish a rare document from the
rest. For example, if a term appears only in very few documents, say two documents
comprising such a term out of a thousand documents, then that term is an important
criterion to distinguish those two documents as relevant to that term. In this case,
the similarity is proportional to the ratio of the total number of documents (d) to the
documents count (say d f j , the document frequency) which have the occurrence of
this rare term. The ratio is called inverse document frequency, id f .
The important terms encountered in the above discussion are defined and listed
below as
t : Terms. It is the number of distinct terms (keywords) in the entire collection of
documents.
t f i j : Term frequency. It is the number of occurrences of the term ti in the document
Dj.
d f j : Document frequency. It is the number of documents that contain the term.
id f j : Inverse document frequency. It is equal to log ddf j , where d is the total number
of documents.
Each document is represented as a vector, with a total t number of components,
and each entry in the vector corresponds to a distinct term from the entire collection
set. In addition, each component in a vector is filled with weights computed for the
corresponding term. This weight is automatically assigned based on how frequently
the term has occurred in the entire documents collection, and another weight is based
on how often the term has appeared in the particular document. The weight of a term
in a document increases, the more often it appears in that document, and decreases,
the more often it appears in other documents.
The weight of a term in a document is defined as a combination of term frequency
t f , and the inverse document frequency id f . Each term has a position in the vector,
18.4 Vector Space Model 565

if it is present in the document, that position is marked with its weight, else marked as
weight 0. To compute the weight (dwi j ) of a term ti in a document D j , the following
equation of t f.id f is used:
dwi j = t f i j × id f j . (18.2)

When an information retrieval system (actually it retrieves the document carrying

the needed information) is used to query a document collection using a query of n
number of terms, the system computes one document vector, dw1 j , dw2 j , . . . , dwn j 
of size n for each of the documents j, and these vectors’ components are filled
with weights as discussed above. Similarly, a query vector, qw1 , qw2 , . . . , qwn  is
computed for the terms found in the query. Note that, size of the query vector is also
n. The similarity measure between query Q and a document D j is defined as the dot
product of two vectors.


n
sim(Q, D j ) = qwi × dwi j . (18.3)
i=1

18.5 Indexing

The vector space model and other retrieval strategies make use of an inverted index
file structure to avoid the length of search in the keywords of every document for
which relevancy is to be established. Instead of searching into a document, an inverted
index is generated in advance for all the keywords in the document set. An entry for
each of the n terms (t1 , ..., tn ) is stored in an index structure, like the one shown in
Fig. 18.5. For each term ti , a pointer references to a linked list, which contains an
entry for each document containing this term as well as the term frequency in that
document is present. For example in row i, there are entries for (d j , t f i j ) forming
a connected list for term ti . The d j is a document in which term frequency is t f i j
for the term ti . The figure shows that the term ti has a frequency of 7 in document
d3 . This indexing structure has the advantage that the retrieval system can search the
term quickly, as well as the documents in which the term appears [18].

18.5.1 Index Construction

A key challenge in the construction of an index is the size of the data involved in
the index itself (see Fig. 18.5), which is a dynamic data structure typically used for
cross-reference generation, but cannot be kept together in the memory of a typical
system due to its size. The task to be performed here is a matrix transposition,
given that the documents’ terms’ matrix is very sparse. Such matrices are not so
easy to manipulate directly as an array. Therefore, the index construction makes use
566 18 Information Retrieval

Fig. 18.5 Inverted index file t1 (d5 , 2)

t2
t3

ti (dj , tfij ) (d3 , 7)

Inverted List

tt (d7 , 1)

of index compression techniques, together with distribute-comparison-based sorting

techniques.
Algorithm 18.1 is a simple in-memory inversion algorithm. The key idea in this
algorithm is that the first pass through the documents to be indexed collects terms
frequency (t f i j ) information, which is sufficient for the construction of an inverted
index. The index is stored in the memory in the form of a template. The second pass
places the pointers, shown by arrows, at their correct positions in the template as
shown in Fig. 18.5.

Algorithm 18.1 To build an inverted index using the in-memory technique

1: Pass I: Make an initial pass over the collection of documents.
2: For each term ti , count its term frequency t f i j in each document d j , and determine the upper
bound u ti in bytes, on the length  of the inverted list for ti .
3: Allocate an in-memory array of ti u ti bytes, and, for each term ti , create a pointer cti to the
start of a corresponding block of u ti bytes.
4: Pass II. Process the collection of documents a second time.
5: For each document d j , and for each term ti ∈ d j , append a code representing d j , t f i j  at cti ,
and update cti .
6: Pass over in memory Index: Make sequential pass over these index, for each term ti , copy the
t f i j representations of the d j , t f i j  pointers from the allocated u ti bytes to the inverted file, and
compress if required.

It is possible to extend the in-memory technique to data collection technique. In

this technique, the index size may exceed the memory size. This is possible by laying
off the index skeleton on the disk, while partial sequences of indexes are created in
the memory, and each one is transferred to the disk in a skip-sequential manner into
a template in a large file. Using this extended method, and using compression, it is
possible to create indexes of the size of terabytes using a memory of a moderate size
of 4–8 GB—a size common in the present time.
18.5 Indexing 567

Parallel Processing of Index

For parallel processing of indexes, they can be constructed in parallel and can be
merged after regular intervals (see Algorithm 18.2). The merge-based inversion tech-
nique reads the documents and indexes them in the memory until a fixed capacity
is achieved. Every inverted list is represented using a structure, which can grow as
more information about the index terms become available. For this structure, dynamic
resizable arrays are most appropriate. As soon as the memory is full to its predefined
capacity, the indexes are flushed out to disk in a single run, such that the inverted
lists are stored in the disk file in a lexicographic order. This lexicographic order later
becomes useful for the sequential merging of the indexes. Since the runs of these
subindexes are never used to answer a query, there is no need to store their vocabulary
in an explicit structure, hence each run can be written at the head of its inverted list.
Once a run is written into the hard disk, it is fully deleted from the memory so that
the construction of the next run begins with initially empty vocabulary.

Algorithm 18.2 To build an inverted index using Merge-based technique

1: while all the documents are not processed do
2: Initialize an in-memory index, using a dynamic structure for the vocabularies and a static
coding scheme for inverted lists; store lists either in dynamically resized array or in linked
blocks.
3: Read documents and insert d, t f  pointers into the in-memory index, continuing until all
allocated memory is consumed.
4: Flush this temporary index to disk, including its vocabulary.
5: end while
6: Merge all the set of partial indexes to form a single index, compressing the inverted lists if
required.

When all the documents have been processed, the runs available in the disk are
merged to get the final index. The merging process builds the final vocabulary on the
fly. Since the runs from the disk are read (into a buffer) for merging, a sufficiently
large size of the buffer will reduce the disk accesses, hence making the process further
faster. However, if the free disk space is limited, the final index can be written back
into the RAM at the space occupied by the runs, progressively as they are processed,
which is helpful in representing the final inverted lists more efficiently. The latter
becomes possible because the final index is typically smaller than the runs, hence
the vocabulary information is not duplicated.
The index construction using the merge-based approach is common and practical
for data collections of all sizes. It is scalable, and operates efficiently in a memory
size as small as 100 MB. The overheads of a disk space can be limited to a small
fraction of the final index, as it requires only one parsing pass over the data, and the
method can be extended from keyword indexing to phrase indexing.
568 18 Information Retrieval

18.5.2 Index Maintenance

Inserting a new document into the text collections amounts to inserting a few bytes to
the end of every inverted list depending on how much the terms in that document are,
i.e., size of the document. Since a document may consist of 100s–1000s of terms, such
insertion of terms requires fetching and extending 100s of inverted lists, and it may
require 10–20 s in the worst cases, to rebuild the inverted list for the new collection
of documents. However, the merge-based inversion approach can index thousands
of documents per second, making this method almost 10,000-times faster [12].
For fast insertion of terms for indexing into the inverted list, the disk resident
part of the list be not accessed frequently, else it will reduce the overall speed. The
practical solution is to amortize the update cost over a sequence of insertions. There
are many properties of text databases, which allow strategies for cost amortization.
The new documents are not immediately available for searching, if they are searchable
they can be made available through a temporary in-memory index—that is, the last
subindex in the merging.
There are three broad categories available for updating the index as the new
documents get added into the collection, they are rebuild from scratch, merge an
existing index with an index of new documents, and incremental update.
Rebuild
There are applications where the index is not updated at all, but it is rebuilt from
scratch at some regular intervals. The presence of new documents is established
through crawling, and an update is not immediately needed for these documents.
This approach is considered economical even for gigabytes of data, where rebuilding
takes just a few minutes, rather than updating the existing index, which anyway will
require fetching the document as well as the index to be updated.
Intermittent Merge
Number of inverted lists can be maintained for the documents collection, in the
memory of a system. Having the lists in memory, it is easy (less complex) to insert new
documents into these lists when the new documents are discovered. This is carried
out using Algorithm 18.2 discussed on page 567, through the process of merging the
indexes. Alternatively, the existing documents’ index remains in a standard inverted
file, and new documents are indexed as an inverted file data structure in the memory.
With this, the two indexes can share a common vocabulary. In this process, both the
indexes are made available to queries, and the result of any query is the union of the
query results from both of these inverted indexes. When the size of the in-memory
index crosses a threshold size, it is merged into the index file on the disk.
Other approaches for indexing use incremental update, or choice of alternative
strategies for indexing as a combination, suffix arrays, wavelet trees, Bayesian infer-
ences, predicate-based indexing, and probabilistic indexing [14].
18.6 Probabilistic Retrieval Model 569

18.6 Probabilistic Retrieval Model

The probabilistic model of retrieval computes the similarity measure (sim(q, di ))

between the query q ∈ Q and a document di ∈ D as the probability that di is relevant
to q, where D is the set of documents in the collection, and Q is the set of queries.
The probabilistic retrieval method estimates the term weight based on how often the
term appears in the relevant but does not appear in the non-relevant documents. The
term weight is calculated using the probability ranking principle, which is based on
the assumption that optimal performance is observed when documents are ranked
on their relevance to the query. In the approach used, probabilities are first assigned
to components of the query and then each of these is used as evidence in computing
the probability that a given document is relevant to the query [8].
Each term in the query is also assigned a weight corresponding to the probability
that the document term matched with the query will retrieve a relevant document. The
weights of the query terms are aggregated to obtain the final measure of relevance.
A probability-based information retrieval system ranks the documents in decreasing
order of probability of relevance to the user’s information needs. Following are
essential preconditions for this probabilistic retrieval model:
• Retrieval accuracy is dependent on how accurately the query and document have
been represented, and does not directly depend on the documents and the queries,
• The representation of documents and queries may not be accurate due to a variety
of uncertainties prevailing in the method of representation itself.
Figure 18.6 shows the conceptual probabilistic model for IR, where event space is
represented by Q × D, such that Q = {q1 , q2 , q3 , . . .} is a set of queries representing
the information need, and D = {d1 , d2 , d3 , . . . , } is a finite set of documents. Each
query qi and document d j is in the form of descriptors, where qi and d j are set of
terms (i.e., keywords). The descriptor is a binary-valued vector, and each element in
that corresponds to a term. Every query is taken as a unique event, i.e., two identical
queries at different times are treated as different events. We assume that G is a set
of possible relevance judgments for the Cartesian product of documents’ set D and
queries set Q. Let the relevance relationship be r , between the query set and document
set, in the form of a mapping r : Q × D → G. In case of Boolean IR, a document
is either relevant to a query or not, hence for any query qi and document d j , there is
r (qi , d j ) → {0, 1}.

Fig. 18.6 Conceptual αQ Q βQ

Q Q
probabilistic model of IR

H
G

αD βD
D
D D
570 18 Information Retrieval

In fact, an IR system does not directly handle the documents and the queries,
but handles them indirectly, in the forms of their representations. For example, a
document is represented in the form of index terms, and a query is represented
as a Boolean expression comprising terms and Boolean operators (see page 562
for more detail). Let us assume that Q and D are representations of queries and
documents, respectively. These representations have a mapping from the original
query and documents through some functions, which are expressed as α Q : Q → Q
and α D : D → D . Hence, if there are two different documents, but represented
with the identical set of index terms, then they will be mapped onto the identical
representation.
Further mapping is introduced from the representation (Q , and D ) to object
descriptions, to make the models more general. This is done by supplementing a
weight to the index term forms of the queries’ and documents’ representations. The
weight is a real number. Let us assume that these object descriptions, for query set
and document set, respectively, are Q and D , and the mappings as β Q : Q → Q
and β D : D → D , respectively. Due to the introduction of weight the new mapping
shows a more accurate relevance relation between the query and its descriptor set,
and similar is the case for the document set to their descriptors. Therefore, the more
correct value of the relevance function r is r : Q × D → H .
For a submitted query qi ∈ Q to the IR system, the documents d j ∈ D are ranked
according to the decreasing order of r (qi , d j ), such that the document with the
highest rank (of relevance) is at the top. The job of an IR system that ranks the
documents in the order of their relevancy for a query qi is to calculate relevance and
rank every document d j ∈ D . Often, for the sake of simplicity, the description and
representation are treated the same, and both are represented as the form of set of
terms [4].

18.7 Fuzzy Logic-Based IR

The fuzzy retrieval technique is based on the fuzzy set theory and fuzzy logic—an
extension of the classical set theory. This fuzzy retrieval technique is based on the
concept that the word matching between the inquiry word set and the text word set
should not be limited to the perfect matching with the stem words. But, since the
words in inquiry also match with their synonyms in the text documents, the matching
should be graded, depending on the degree or level of matching in the range from
0 to 1. The extremes of this range, a special case in fuzzy logic, corresponds to the
Boolean matching. Fuzzy logic is more realistic than the Boolean logic, simply due
to the fact that it considers the exact as well as vague matching, the latter being more
frequently encountered in the real world [7].
In the relations over fuzzy sets, the elements of two sets have a degree of asso-
ciation as a form of relation rather than simply—related (binary 1) or not related
(binary 0). The degree of association ranges from 0 to 1, where 0 indicates the total
18.7 Fuzzy Logic-Based IR 571

absence of relation and a 1 indicates the total presence of the relation; therefore, a
fuzzy relation between query keywords set and of document’s keywords is

R : X × Y ∈ [0, 1] (18.4)

and the range of R(x, y) varies from 0 to 1 depending on how close the y, (y ∈ Y )
is associated with x, (x ∈ X ).
Information Retrieval Using Fuzzy Sets
The membership value R(xi , y j ) specifies, for each xi ∈ X , y j ∈ Y , the grade of
relevance of index term xi with the document y j . The grade of relevance depends on
many factors [3]:
1. position of term y j in the text document, if the document is a research article,
and y j appears in the list of keywords, the abstract, or in the conclusion part, the
relevance is higher;
2. frequency of occurrence of y j in the document;
3. xi and y j are terms formed from the same basic stem word; and
4. y j is synonym of xi —the proximity of the meaning of xi and y j decide value of
R(xi , y j )’s closeness with 1.
The criteria (1) and (2) above are user-defined and they can be programmed in
the implementation according to the user needs. The stem word criteria (3) requires
a data structure similar to the one shown below in Fig. 18.7, which helps to locate
the stem word for a given word, and then the stem word is substituted in the original
text before the retrieval technique is applied to it [4].
Other, more often used approach for stemming is through some stemming algo-
rithm, e.g., Porter’s stemmer, which takes the benefit of certain patterns in the words
to obtain the stem word. For example, for the words with “ing” at the end, the stem
word can be obtained by removing the “ing” part, like “book” from “booking”. Also,
there are other features, like removing “ed” at the end of the past tense of a verb; we
obtain its stem word, say “book” and “look” from “booked” and “looked”, respec-
tively. Along with this, there are some more complex patterns, like “goose” from
“geese”, etc.

Fig. 18.7 Data Structure for Keyword1

finding stem words Stemword1
Keyword2 ....
.... ....
Stemwordm
Keywordn
572 18 Information Retrieval

Another important relation for IR based on criteria (iv) above is the fuzzy the-
saurus, which plays a pivotal role for FIR. The fuzzy thesaurus shows the relationship
between the pairs of words based on their centrality or degree of relevance. The struc-
ture of a fuzzy thesaurus (T ) is

W C1W C2R D

where W C stands for word category and R D is the degree of relationship between
the words W C1 and W C2. A typical examples for this can be as follows.

attraction, love, 0.8

studious, hardworking, 0.9
war, crime, 0.7.

A relation xi , x j , 1.0 shows that xi is a perfect synonym of x j . The fuzzy

thesaurus can be manually constructed, or can be generated from the lexicons. Tran-
sitivity relationship can be applied by computing the missing relationship degrees
from the existing ones. The thesaurus, say T , is a reflexive fuzzy relation, defined
over X 2 . For each pair of index terms (xi , x j ) ∈ X 2 , the T (xi , x j ) expresses the
degree of association of x j with xi , such that the degree to which the meaning of the
index term x j is compatible with the meaning of the index term xi . The objective
of this relation is to deal with the problem of synonyms among the index terms, for
example a document’s term is a synonym of query term or vice versa. The relation
helps to identify the relevant documents which otherwise would not be selected in
the absence of a perfect match between the keywords in the user inquiry and those
in the text document.
Different approaches can be used for the construction of fuzzy thesaurus. For
example, experts in the domain of text can be asked to identify, in a given set of
index terms, the pairs of words whose meaning they consider are associated, and
provide the degree of association for each pair. In Fuzzy Information Retrieval (FIR)
an inquiry can be expressed in the form of a fuzzy set (say Q) based on the index
term X . Then, by composing Q with the fuzzy thesaurus T , we obtain a new fuzzy
set on X , say A—which represents the augmented inquiry, i.e.,

A = Q ◦ T, (18.5)

where “◦” is called max-min composition operator, such that

A(x j ) = max-min[Q(xi ), T (xi , x j )]. (18.6)

Here, xi ∈ X , for all x j ∈ X . The retrieved documents, expressed by a fuzzy set F

defined over Y , are then obtained by composing the augmented inquiry, expressed
by the fuzzy set A, with the relevance relation R, i.e.,
18.7 Fuzzy Logic-Based IR 573

F = A∗R (18.7)

where “∗” is a matching operator, which evaluates the degree of fuzzy matching by
multiplying the fuzzy membership of the augmented inquiry with the fuzzy mem-
bership of the corresponding words in the text. Finally, the relevance measure of the
text with the inquiry under consideration is obtained by summing all the values of
fuzzy matching for the text.

Example 18.2 Fuzzy logic-based Information Retrieval.

Let the terms be xi , i = 1, 6, representing keywords—“Einstein”, “scientific”, “The-

ory of Relativity”, “Bohr”, “subatomic”, and “New idea”, respectively. Let the given
inquiry be Q = “Einstein’s Scientific Theory of Relativity”, and the vector represen-
tation of corresponding fuzzy inquiry be

x1 x2 x3
Q = [1 0.6 0.8] (18.8)

where 1, 0.6, and 0.8 are called the centralities of x1 (Einstein’s), x2 (Scientific), x3
(Theory of Relativity), respectively. The centrality indicates the presence of certain
qualities, whose computations are modeled as a computation of fuzzy membership
degree. The relevant part of the fuzzy thesaurus T , restricted to the support of Q, is
given by the matrix:
⎡ ⎤
x1 x2 x3 x4 x5 x6
⎢ x1 1 .6 .9 0 .1 0 ⎥
T =⎢ ⎣ x2 .6 1 .8 .6 .5 .8 ⎦ .
⎥ (18.9)
x3 .9 .8 1 0 0 .7

Note that terms’ pairs (x1 , x1 ), (x2 , x2 ), and (x3 , x3 ) has each a fuzzy matching
of 1. On using (18.5), and the data given by Eqs. (18.8) and (18.9), we get


A = 1 .8 .9 .6 .5 .7 . (18.10)

The values in augmented query A are obtained as follows: we show the compu-
tation for third element (A3 ), i.e., 0.9 as

A3 =max(min(1 × 0.9), min(0.6 × 0.8), min(0.8 × 1))

=max(0.9, 0.6, 0.8)
=0.9.

Assume that relevance relation R (i.,e., R(xi , y j )) is given by the matrix,

574 18 Information Retrieval
⎡ ⎤
y1 y2
⎢ x1 .7 .3 ⎥
⎢ ⎥
⎢ x2 .3 .1 ⎥
⎢ ⎥
R=⎢
⎢ x3 .6 .1 ⎥
⎥, (18.11)
⎢ x4 .6 0⎥
⎢ ⎥
⎣ x5 .6 .3 ⎦
x6 .6 .1

where y1 , y2 are the documents related to index terms xi , i = 1, . . . , 6. Using Eq.

(18.7) and data given by (18.10) and (18.11), we get fuzzy relevance relation F as


F = A ∗ R = 2.56 0.69 . (18.12)

Equation (18.12) shows that the relative relevance of y1 with respect to y2 is

2.56/0.69, i.e., 3.7. Thus, the result appears to be more realistic in comparison to the
result obtained using crisp (binary) logic. This fact is also supported by the contents of
texts y1 and y2 . Once the FIR system lists the documents with their relevance values,
the user can now decide whether to inspect all the retrieved documents supported by
the fuzzy set F or to inspect only some of the documents depending on the degree
of association of the document with the index terms. 
The use of fuzzy set theory for IR shows that the fuzzy relevance relation and
fuzzy thesaurus are more expressive than their crisp set counterparts. Also, since the
degree of association is returned along with the retrieved documents, it helps the
user to decide the order in which the documents can be viewed, particularly when
the documents are in large number.
The FIR promises a higher potential for cross-language text processing and IR.
Every language and its semantics have a close association with the culture in which it
has its roots, and therefore, exact matching terms for any language are not possible in
other languages. In fact a degree of relevance or, only fuzzy relation exists between the
matching words of the two or more languages. We have planned to work on the cross-
language areas, which include English, and Sanskrit-based Indian languages [3].

18.8 Concept-Based IR

The keyword-based approach we discussed above is also called the BOW (Bag-of-
Words) approach. The concept-based information retrieval makes use of semantic
concepts for the representation of the documents and queries, instead of (or in addition
to) keywords; and the approach performs the retrieval in the concept space of query
and documents. Hence, this retrieval model is less dependent on the specific terms
(keywords) used, and yields a match even when the same notion is described by a
different set of terms in the query or in the target document, or in both. This helps
in eliminating the problem of synonymy. Since more relevant documents are likely
18.8 Concept-Based IR 575

to be retrieved from the store, the recall rate increases. In the similar way, if the
concepts chosen for the words, particularly the ambiguous words in the query and
document are accurate, the non-relevant documents that were retrieved with the BOW
approach could be eliminated from the results. Since the non-relevant documents are
lesser in the retrieved documents now, it increases the precision rate. Note that in
the keyword-based approach of IR, non-relevant documents’ share increases in the
retrieved documents when the keywords have multiple meanings, e.g., “bank” (of
river and of money). The problem of many senses of a word is called polysemy.
The concept-based methods can be characterized by the following three
parameters:
1. Concept representation. The concept-based IR makes use of real-life concepts
that closely resemble human perception. The concepts are based on the language
of the text of documents and queries.
2. Mapping method used. It is the method used for mapping the natural language text
to concepts. In fact, the ideal mechanism is the manual approach, which builds a
hand-crafted ontology of concepts along with a list of words to be assigned to each
concept. However, this approach is inefficient and time consuming. The automatic
mapping between concepts and words can also be done through machine learning,
but with loss of accuracy.
3. Use in IR. The concepts are used during the indexing and retrieval phases of the
documents. A simpler but less-accurate method applies the concept analysis in
one stage only. For example, it is better to apply the concept analysis in the query
expansion rather than document retrieval.
A large and diverse knowledge repository, like Wikipedia, can create a powerful
concept ontology, well-suited for concept-based IR. A wide range and exhaustive cov-
erage of topics in Wikipedia’s diversity, coupled with automatic ontology-building
capability, can be used for the purpose of highly fine-grained ontology. Additionally,
the inverted index language provides a mapping from massive natural language text
terms of the entire Wikipedia collections, to the concepts as per the context and sense
of the text terms. This entire process produces a powerful classifier that automatically
maps any given text fragment to its concept ontology [9].

18.8.1 Concept-Based Indexing

A concept-based IR algorithm maps documents and queries individually to the

Wikipedia concept space. The indexing and retrieval are performed in this space
only. In the Wikipedia-based Semantic Analysis (SA), the semantics of a word are
described by a vector that stores the word’s association strengths to Wikipedia-based
−
→
concepts. A concept is in the form of a concept vector Fd , which is generated from a
single Wikipedia article d, as a vector of words appearing in that article. The article
is weighted by t f.id f score. Once these concept vectors are generated, an inverted
index is created for the purpose of mapping back each word to the concepts it is
576 18 Information Retrieval

Concept Concept Vectors

market Bazaar (0.72) Economy (0.5) ....

Economy
trade Economy(0.61) Trade union (0.5) ....
The ... trade..
...market... ....
trade... money property Real estate(0.8) Theft(0.55)
..property ...

....

....
....
Wikipedia articles wordn < ci , wi > < cj , wj > ....

Fig. 18.8 Semantic analysis from Wikipedia articles

associated with. Thus, each word appearing in the Wikipedia corpus can be seen as
triggering one or more concepts c j it points to the inverted index. An attached weight
w j with the concept represents the degree of association between that word and the
concept. This process is illustrated in Fig. 18.8, which shows how the semantic anal-
ysis is carried out of a Wikipedia article.4 The articles and the words in the articles are
processed to build a weighted inverted index, which represents each word of article
as a vector in the space of all Wikipedia concepts, i.e., articles in the Wikipedia itself.
In the concept-based indexing, each document (i.e., article) in the corpus is
mapped to a vector of weighted concepts, represented by c1 , w1 , c2 , w2 , ...,
cn , wn . Like Bag-of-Words (BOWs) vectors, the concept-based vectors are also
sparse, hence concept weights are zero for a large majority of the Wikipedia con-
cepts. Since every word in the document to be indexed may be related to a large
number of concepts, and a document containing a collection of words is likely to be
related to an even larger number of words, indexing an entire list of related concepts
for every document is not feasible. Therefore, only the concepts having the highest
weights (for example, “Bazaar”, “Economy”, etc., in Fig. 18.8) are used for indexing.
In a sorted representation of the weighted vector, this subset of concepts is simply
its prefix.
It more difficult to map the long documents in full, into the concept space. For
example, a small part of a long document might be relevant to a given query, but
the semantics of this small part are not likely to be fully represented in the concepts
vector of the complete document. Note that, a somewhat similar problem also exists
in the Bag-of-Words approach, where the term frequency (t f ) value is normalized
to account for the documents of varying lengths.
Due to the averaging effect of the representation of longer text fragments, and due
to the practical limitation to use only a small subset of the representation of concepts,
the challenge in the concept-based retrieval technique is even greater. Because the
concepts generated for a subset of the larger document, where the subset is relevant

4A Wikipedia article is an article about some topic, for example, we find articles in the collections
of Wikipedia, like websites, Internet, WWW, etc.
18.8 Concept-Based IR 577

and the remaining document consists of non-relevant topics, the representation of

the latter need to be pruned out of the index vector. This is necessary as otherwise
the concepts weights in the overall document concepts vector might be too low to
show any significance of the retrieval results.
Semantic Analysis-Based Indexing Algorithm
An algorithm suitable for indexing larger size documents, based on semantic analysis
and using the inverted index, is given as Algorithm 18.3. This algorithm indexes a
corpus D of documents using semantic analysis concepts, by trimming of semantic
analysis vector to s as the first concepts. Each document d ∈ D is represented by
−
→ −
→
a concept vector Fd . For each concept ci , wi  ∈ Fd , the corresponding d, wi  is
added into the inverted inverted index. Here, ci is the concept and wi is the concept
weight.
To overcome the problem of mapping each large document d ∈ D in full into
concept space, d is divided into smaller fixed length (size l) overlapping passages
set Pd . Then each passage p ∈ Pd is represented separately by its own generated set
−
→
of concepts Pd .
Each passage p is indexed and can be retrieved as a stand-alone unit of informa-
−
→
tion. For this, all concepts ci , wi  ∈ F p corresponding to a passage p, the passage
is ranked separately as an independent unit along with its relevance in its parent
document (d), shown in the algorithm by “add p, wi  to I nv I ndex[ci ]”. These
concepts are indexed in a standard IR inverted index, with each concept having a
unique identifier in the form of a token. The score wi , associated with each concept
ci in the vector, is used as the token weight, which is equivalent to term frequency
t f , in standard text indexing.

Algorithm 18.3 Semantic analysis-based indexing in an inverted index

1: Procedure SA-Indexing(D, s, l) {Indexes corpus D using SA concepts; trims SA vector to s as
first concept; then segments document to passages, each of length l.}
2: for all d ∈ D do
−
→
3: Fd ← S A(d, s)
−
→
4: for all ci , wi  ∈ Fd do
5: addd, wi  to I nv I ndex[ci ]
6: end for
7: Pd ← Divide-Into-Passages(d, l)
8: for all p ∈ Pd do
−
→
9: F p ← S A( p, s)
−
→
10: for all ci , wi  ∈ F p do
11: add p, wi  to I nv I ndex[ci ]
12: end for
13: end for
14: end for
578 18 Information Retrieval

18.8.2 Retrieval Algorithms

Once a query is received by an IR system, it is converted into a concept vector by

the retrieval algorithm. The representation method used is identical to the one we
discussed above, for documents and passages during the indexing. The indexes of
full documents and passages, as evidences, are to be combined for ranking. This
combination is performed by retrieving both set of results, and then by summing
each document’s full score with the score of the best performing passage in it. In the
next phase, documents are sorted on a combined score, i.e., sum, and the top-scoring
documents are output. All the above steps are described in the retrieval algorithm 18.4.
The algorithm works as follows:
1. Retrieve the query results, which are based on SA concept, for query − →q , as well
as the cutoff concept vector at s;
2. Retrieve results for query −
→
q using the combined results;
3. Score the document’s match to the query using the standard inverted index func-
tion InvIndex-score().

Algorithm 18.4 SA-based Retrieval

1: Procedure SA-Retrieval(− →
q , s) {Retrieve the SA concept-based results for query −
→
q , and cutoff
concept vector at s}
−
→
2: Fq ← S A(− →
q , s)
−
→
3: return DocsPass-Retrieve( Fq )
4: Procedure Docs Pass-Retrieve(− →
q ) {Retrieve results for query −
→
q from the combined results;
Score the document’s match to the query using the standard inverted index function I nv I ndex-
scor e().}
5: for all d ∈ D do
6: Wd ← I nv I ndex-Scor e(− →
q , d)
7: for all p ∈ P ASS AG E S(d) do
8: W p ← I nv I ndex-Scor e(−
→q , p)
9: end for
10: Wd ← Wd + max W p
11: end for
12: return ranked list according to Wd

The retrieval algorithm 18.4 has a single parameter s that controls the query
concept vector’s cutoff. The value of s can be chosen the same as that during the
indexing, however it is not necessary. If the entire corpus is indexed with large cutoff
values, the resultant costs for computation as well as storage will be high, hence
it is not advisable to index the entire corpus with large cutoffs. Since the query is
much smaller in size, there are no such costs as that for text corpus, hence a finer
representation will be beneficial. That can be achieved using a higher value of s.
18.9 Automatic Query Expansion in IR 579

18.9 Automatic Query Expansion in IR

Most IR systems, and particularly the web search engines, have a standard user
interface comprising of an input box to accept from a user, a query in the form of
keywords. The submitted keywords are matched against the collection of index terms
to find the documents that contain those keywords. The results are then sorted by
various methods. When there are many topic-specific keywords in the user’s query,
which accurately describe user’s information need, the system is likely to return good
matches for the query. However, when this query is short—comprising 2–3 words,
as the case usually is—there is likely ambiguity in the language of the query, then
this simple retrieval model is sensitive to errors [1].
The most serious issue in retrieval effectiveness is the term mismatch problem, i.e.,
indexers and the users do often not use the same keywords. For example, a document
uses “tv” while user submits the query with keyword “television”. This is known
as the vocabulary problem. This problem gets compounded due to polysemy, i.e.,
the same word with different/multiple meanings, such as Java (name of language,
and also name of a place), and also due to synonymy, i.e., different words with the
identical or similar meanings, such as “tv” and “television”. Synonym words, along
with word inflections (like in plural forms, “book” versus “books”) may result in a
failure to retrieve relevant documents. This may decrease the recall. The problem of
polysemy may cause retrieval of erroneous or non-relevant documents, thus causing
a decrease in precision.
Several approaches have been proposed, to deal with the vocabulary problem;
some of them are the following:

Interactive query refinement,

Relevance feedback,
Word Sense Disambiguation,
Query expansion, and
Search results clustering.

The technique of query expansion is one of the most natural and successful tech-
niques, using which the original query is expanded with other words that best capture
the user’s actual intent, or it simply produces a more useful query—a query that is
more likely to retrieve relevant documents.
As the use of search engines increased over time, the size/length of the user’s
query has also increased. In the year 2009, an average query’s length was 2.3 words.
However, there has been an increase in the number of long queries per user or
per session of interaction, of five or more words; the most common queries are
still those of one, two, and three words. When a query is short, the vocabulary
problem is more serious, because the shortage of query terms limits the scope of
handling synonymy. At the same time, the reduced resizing of data makes the effects
of polysemy more severe. This required the need and scope of Automatic Query
Expansion (AQE). Over the years a number of AQE techniques have increased that
employ sophisticated methods for finding new features related to query terms. Today,
580 18 Information Retrieval

there are firm theoretical foundations, and a better understanding of the utility and
limitations of AQEs. For example, what the critical parameters are that affect the
performance of the IR system, what types of AQEs are useful and what not, etc.
Along with the AQE, the basic techniques are being increasingly used in con-
junction with mechanisms to increase their effectiveness, like method combinations,
selection of information source dynamically for expansion, and discriminating poli-
cies for the application of methods. These advances have been supported by many
experimental findings. The AQE methods have gained their popularity due to the
evaluation results obtained at the Text REtrieval Conference series (TREC).
Document Ranking using AQE
The IR systems including search engines depend on computing the importance of
terms occurring in the query and documents to determine the relevance of docu-
ment(s) to the queries. The commonly used indicator of relevance is the similarity
measure between the query and the document. Considering that a query is repre-
sented by q and a document by d, the similarity measure between them, sim(q, d),
is expressed by

sim(q, d) = wt,q × wt,d (18.13)
t∈(q∩d)

where wt,q and wt,d are the weights of term t in a query q and a document d,
respectively, according to some weighting criteria adopted. The weight of a term t
in a document is typically proportional to the term frequency (t f ) in that document
and to the inverse document frequency (id f ). The purpose of id f is to diminish the
effect of very frequent terms like “the” for which we are not interested to find the
relevance. But we want to increase the effect of terms, which occur rarely, in a few
documents only, for example, “blue” and “bird”, which might be occurring only in
a few documents. If N is the total number of documents and n is the number of
documents in which the rare term occurs, the id f for this term is log( Nn ). Note that
if a term like “the” occurs in every document, then id f is zero, hence it will not
contribute to the similarity measure.
The similarity computation between a document and query representation,
expressed in Eq. 18.13, can be easily modified by abstracting away from the original
underlying weighting model, so as to work for the query expansion. In this revised
model, the basic input to AQE module is 1. original query q and, 2. source of data
from which to compute the weight and the expansion terms. The output of the AQE
module is the query formed (say, q ) due to the expansion of query terms, and their
associated weights w . These new weighted query terms q , w  are used for com-
puting the similarity between the query q and the original document d. The new
similarity computation formula is expressed by

sim(q , d) =
wt,q × wt,d . (18.14)
t∈(q ∩d)
18.9 Automatic Query Expansion in IR 581

The commonly used data source for generating new query terms is the collection
itself into which we are searching, and sometimes it is the thesaurus to get the
synonyms of the keywords. The simplest approach to weight the query expansion
terms is to use the same weighting function that is being used by the ranking system.
When more complex features in the terms, like phrases, are used for query expansion,
the underlying system must be able to handle such features. An example is “Jodhpur
departmental store”, which though a phrase, can be treated as a single term, for the
purpose of indexing.
In the following we address some new areas, in addition to document ranking,
where AQE is used heavily; these are: question answering, multimedia IR, informa-
tion filtering, and cross-language IR. This introduction is followed by pointers to
more recent applications [1].
Question Answering
The goal of question answering (QA) is to provide concise responses (instead of
full-length documents) to certain types of natural language questions, such as “Who
built the Taj Mahal?” Like the Document Ranking (DR), the QA is faced with a
fundamental challenge of mismatch between the vocabularies of the question and
answer.
To improve on the early stage of QA, which is document retrieval, a common
strategy used is to expand the original question terms with terms that are expected to
appear in documents containing answers to this question. One important goal of QA
translation is to learn the associations between question words and the answer words,
which may be synonyms to the words in question. For example, for the question,
“Where Taj Mahal is located?,” the answer part embedded in a document may be
“Taj Mahal is in Agra,” or “Taj Mahal is situated in the city of Agra.” Here, the words
“situated” and “exists” are the synonyms of the word “located”. Different resources
for AQE for QA include using lexical ontologies like WordNet—a lexical database
of synonyms (called synsets), and semantic parsing of questions based on roles, and
other criterias [7].
Multimedia Information Retrieval
With the widespread use of digital media and digital libraries, the requirements of
searching the multimedia documents like image, speech, and video have become
important. Up till recently, the multimedia IR systems performed only text-based
search over the media metadata, like annotations and captions, which are surrounding
the html/xml descriptions. However, when these metadata are absent, this method is
not suitable. Hence, the IR relies on some form of multimedia content analysis,
which is implemented in combination with the AQE techniques. For example, a
transcription is produced by an automatic speech recognition system for spoken
document retrieval systems, and this is augmented with related terms, in advance to
raising the query. Since automatic speech transcriptions often contain mistakes, this
form of a document expansion is very useful for spoken document retrieval.
582 18 Information Retrieval

For image retrieval systems, a typical approach is making use of query examples
having visual features, like colors, textures, and shapes. The query is iteratively
refined through a relevance feedback.
For video retrieval systems, both the documents and queries are mostly multi-
modal, i.e., they have both textual as well as visual aspects. An expanded text query
is usually compared against the textual description of the visual concepts, and the
matched concepts are used for visual refinement.
Information Filtering
Information filtering (IF) is different from IR. It removes redundant or unwanted
information from an information stream prior to presentation to a human user. Its
main goal is the management of the information overload and increment of the
semantic signal-to-noise ratio. The documents arrive continuously and the user’s
information needs evolve over time, as per the experience of the user. Some examples
of information filtering are electronic news, blogs, e-commerce, and e-mails. There
are two main approaches to IF: 1. collaborative IF, which is based on the preferences
of like-minded users, and the other is 2. content-based IF. However, these techniques
are said to bear a strong conceptual similarity to IR, because the user profile can be
modeled as a query and the data stream can be treated as a set of collection of
documents. The user profiles (i.e., queries) are learned using relevance feedback
techniques, or other forms of query expansion, such as those based on similar users.
Cross-Language Information Retrieval
The Cross-Language Information Retrieval (CLIR) is concerned with retrieving the
documents written in a language different than the language of the users’ query.
Earlier methods of such retrieval consisted of translating a query into the docu-
ments’ language, and then using the standard techniques of IR. The query translation
in these approaches was performed using machine-readable bilingual dictionaries,
through machine translation, or using parallel corpora. However, no such translation
is absolutely correct, and regardless of the translating resource used, there are usually
limitations due to insufficient coverage, for example, there are terms which cannot
be translated or do not require the translation, and due to the translation ambiguity
from the source language to the target language.
To reduce the errors introduced due to translation, one standard technique is to
use query expansion, so that even when the translation does not contain errors, use of
semantically similar terms yields better results than those due to the literal translation
of terms only. The query expansion can be applied either before the translation of
query, or after, or even can be applied at both the places. It has been found that query
expansion, before the translation of a query, yields better results than doing it after
the translation. The expansion done at both the places provided even better results.
18.9 Automatic Query Expansion in IR 583

Inter- Ranked
mediate candidate Expansion Reformulated
features Query
Features Features
Query
Feature Query
Data Feature -
processing Generation Selection Reformu-
and Ranking lation
Data
Source(s)

Fig. 18.9 Automatic Query Expansion process

18.9.1 Working of AQE

The Automatic Query Expansion (AQE) is performed in a number of steps, the major
steps are
1. Preprocessing the source text,
2. Generating expansion terms (features) and ranking them,
3. Selection of expansion features, and
4. Reformulation of Query.
Figure 18.9 shows these steps of AQE.
The objective of preprocessing the data source data is to transform the raw data
source used for expanding the user query into such a format so that it can be processed
more efficiently by the subsequent steps. The preprocessing task performs extraction
of intermediate features, followed by the construction of data structures for easy
access and manipulation of such features. The preprocessing is usually independent
of a particular user query to be expanded, but it is specific to the type of data source
and expansion method used [1].
To compute the initial retrieval run, it is necessary to index the documents collec-
tion, and then run the query against this collection’s index. The process of indexing
consists of, in order, the following steps:
1. text extraction from documents, which are in a certain format, like HTML, PDF,
MS Word, where there is format information as well as text inside them,
2. tokenization of the extracted text,
3. stop-word removal from the tokenized text (removal of common words such as
articles and prepositions),
4. stemming (reduction of inflected or derivational words to their root form), e.g.,
reduce “tokenize, building, training” into “token, build, train”, etc.
5. word weighting (score is assigned to each token such that the weight reflects the
importance of the tokens in each document).
We take an example of a short HTML text fragment to illustrate the weights
associated with tokens.

<p><b> An automatic query expansion </b>

increases the query’s semantics.</p>
584 18 Information Retrieval

In the above HTML document, first, text is extracted, then stop words “the” and
“an” are removed, then it is stemmed using Porter’s stemmer, and weight is assigned
to each word based on their frequency. Finally, we obtain the text representation as
follows:

automat 0.16, queri 0.33, ex pan 0.16, incr ease 0.16, semantic 0.16.

This is an example of a very small document, however, it gives an understanding

that each document can be represented as a set of weighted terms, such that the total
weight of the document is 1. The index is created in the form of a complementary
inverted index file, which maps terms to documents at the time of query. To reach the
location of index terms in the document faster, the system may also store the terms’
locations to provide proximity-based search.
The original query is preprocessed to remove the stop words and/or extract impor-
tant terms to be expanded. In the second stage of AQE, the system generates and
ranks the candidate expansion terms; most query expansion methods choose only a
small proportion of the expansion features (i.e., terms) to be added into the query.
Input to this stage of AQE is the original query and the transformed data source, and
the output is a set of expansion features, with/without scores.
Once the ranking of the candidate features is carried out, the top elements are
selected for query expansion. Selection of these top elements is performed on an
individual basis, without regard to mutual dependencies between the expansion fea-
tures. Usually, only a limited number of features are selected for expansion such
that 1. resulting query is not bulky, thus helpful for processing faster, and 2. retrieval
effectiveness of a small set of good terms is not necessarily less successful than the
effectiveness we get by adding all candidates’ expansion features. The addition of
expansion features will also be helpful in reducing the noise.
Sometimes, the feature scores are interpreted as probabilities. In that case, only
the terms having a probability greater than a certain threshold are selected for con-
sideration.
The last stage of AQE is query reformulation, which describes the expanded query
that will be submitted to the IR system. The description means the assignment of
proper weight to each feature that is part of the expanded query—the process called
query re-weighting.
The total time required for an AQE is the sum of two factors, 1. cost of generating
expansion features, and 2. increase in the cost for the evaluation of the expanded query
(due to its size), against the documents collection. In practice, the second factor is a
more critical one. Consider the architecture (data structure) of most ranking systems,
which are based on the inverted (linked) lists of N elements, one for each term in the
collection. Here, each inverted list specifies the documents in which the particular
term occurs, along with a pre-computed score for each term. At the time of query
processing, the system retrieves the inverted list of every query term, and updates
the score accumulators of the documents present in each list. The execution time of
a ranked query is almost linearly proportional to the number of terms in the query;
18.9 Automatic Query Expansion in IR 585

this is because the query terms are processed one at a time. The AQE runs with sizes
of practical interest, for example 10–20 word queries were found to be much slower
(by a factor of ten), than the original queries of 3–4 words.

18.9.2 Related Techniques for Query Processing

The mismatch of words between the query and documents for relevant documents is
an issue in IR for a long time. In the following section, we discuss AQE with respect
to alternative strategies in reference to the vocabulary problem [1].
Interactive Query Refinement
In interactive query refinement, the system provides several suggestions for reforma-
tion of the query, and it is the user who selects the best choice out of that. With respect
to the computations required to be performed, in Interactive Query Refinement (IQE)
versus AQE, the first two stages are common with both, i.e., data acquisition and can-
didate feature extraction. The IQE does not follow the steps of feature selection and
query reformulation of AQE. One of the best-known examples of Interactive Query
Refinement is the suggestion of complete query, which offers real-time hints to user
to complete a search query. This happens when a user progresses in typing the query
in the inbox, like we note in many search engines, including Google. The IQE has
better potential for superior results than AQE, but generally requires higher expertise
on part of the user. Looking from the usability point of view, an IQE provides the
user with better control over query processing than the AQE.
Relevance Feedback
The relevance feedback takes two inputs: 1. results initially returned due to the given
query, and 2. feedback provided by the user about whether those results are relevant
or not. Based on these two inputs, the system submits a new query to the search engine
provided that previous results returned were not considered useful for fulfilling the
need of the user.
The features in the assessed documents are used to adjust the weights of terms
in the original query and/or for adding words to the query. The relevance feedback
has the effect of reinforcing the system’s original decision. This is done by the user
by modifying the expanded query to look closer at the retrieved relevant documents.
However, the AQE tries to form a better match with the user’s existing intentions, and
does not give the user a second chance to rethink, support/reject the results produced
by the first query. The specific data sources using which the expansion features are
generated in the relevance feedback may be more reliable than the sources generally
used by AQE. In the relevance feedback, the user must assess the relevance of the
documents, requiring a user to be better trained.
The relevance feedback has directly inspired one of the most popular AQE tech-
niques, called pseudo-relevance feedback, which has also provided foundations for
modeling query reformulation in a variety of AQE approaches.
586 18 Information Retrieval

Word Sense Disambiguation in IR

The Word Sense Disambiguation (WSD) is the ability of the system to correctly
identify the senses of words in context to the remaining (surrounding) text in a
document. This identification is carried out in a computational manner. WSD is a
natural and well-known approach to the vocabulary problem in IR—usually, even if
we do not know the meaning of a word, for example, in a newspaper, we are still able
to understand the sentence, as well as able to discover the meaning of that unknown
word due to its context words in a sentence [5].
Early work of WSD concentrated on representing words using their dictionary
definitions, or using the WordNet Synsets. But, many experiments suggested that this
straightforward technique is not effective in IR, at least as long as the selection of
the correct sense from the resource is flawed. For example, if precision does result in
a good value, say greater than 75%. However, more sophisticated methods in AQE
still used the WordNet resource [11].
Instead of depending on short predefined lists of senses, using a corpus is found to
be more convenient as an evidence for performing the Word Sense Disambiguation.
Due to its nature of the process, it may be called as word sense induction.
In one approach based on the corpus, the context of every occurrence of a word
in the corpus is identified and similar contexts are clustered together to help in
determining the word senses. This method can provide a maximum disambiguation
rate of 90%, hence can be used successfully with an IR system. With the reliance of
this method on corpus-based analysis, this approach is similar in spirit to the global
AQE techniques.
In another corpus-based WSD technique, a metaphor of small words is applied to
word co-occurrence graphs, due to which it is capable of discovering low-frequency
senses (senses which are not very common), that are as low as 1%.
Further, in the context of a query to web search engines, where the query is too
small, it may be too difficult to disambiguate the word senses in it in the absence of
sufficient context available in the query. Therefore, longer queries due to their higher
contexts are likely to be helpful in performing the disambiguation, and consequently
to produce better results. As a whole, we note that the application of WSD to IR
presents the challenges of computational nature, there are limitations of effectiveness
as well.
Search Results Clustering
The objective of Search Results Clustering (SRC) is that for the new queries, exactly
similar to some previous queries by users, the IR system should store the previous
results in some compact forms, so that it can provide the results directly, without per-
forming any search process. For this, the SRC organizes the search results topic-wise,
which allows direct access to the documents relevant to the user queries, making over-
all IR far faster. In contrast to the standard clustering techniques, the SRC algorithms
try to optimize the clustering structures, as well as the quality of cluster labels. This
is because, a cluster with a poor description (labels) is likely to be entirely omitted by
the user, even though it may be pointing to a group of strongly related and relevant
documents.
18.10 Using Bayesian Networks for IR 587

18.10 Using Bayesian Networks for IR

A Bayesian network is an annotated directed graph, which can be used to encode a

probabilistic relationship among the distinctions of interest in an uncertain reasoning
problem. The representation rigorously describes these relationships, and can provide
a human-oriented qualitative structure which facilitates a communication between
a user and the system based on a probabilistic model. As the computing power
is available chiefly even in small systems, the modeling tools based on Bayesian
networks are abundantly used in real-world applications, e.g., in forecasting, fault
diagnosis, sensor fusion, anti-virus software, automated vision, and manufacturing
control [10], [13].

18.10.1 Representation of Document and Query

For understanding the basics of Bayesian networks as well as Naive Bayes, the reader
may refer to the previous chapters (section 12.4, page no. 344, and section no. 14.7,
page no. 428). Using the conditional probability, the probability that a document dk
is relevant to the query q j can be expressed as P(R | q j , dk ). An accurate definition
of probability of relevance depends on the definition of relevance itself. The term
relevance is to some extent a subjective entity, and depends on many variables,
which are functions of documents, user’s information need, and the user itself. A
perfect retrieval is not achievable in true sense. However, it is possible to define an
optimal retrieval for the probabilistic model for IR. This optimal retrieval can be
proved theoretically with respect to representations (or descriptions) of documents
and information needs [4].
Let us assume that collections of queries and documents are described by a set of
index terms. Let T = {t1 , t2 , ..., tn } be the set of terms in the documents’ collection,
and a query q j and document dk are taken as subsets of terms in T . For the sake of
retrieval, each document is described by the presence/absence of these index terms.
Therefore, any document dk can be represented using a binary vector:
−
→
x = (x1 , x2 , . . . , xn ), (18.15)

where any term xi = 1 if ti ∈ dk , and for ti ∈ / dk , the xi = 0. A query q j is also

represented in similar way. The basic task of a relevance model-based IR system is
to compute the probability that a given document is relevant. This can be achieved
by estimating the probability P(R | q j , dk ), for every document dk in the collection.
Since relevancy in every document is computed for a single query, the term q j being
common, and can be dropped, and the relevancy can be expressed using the Bayes
theorem as
588 18 Information Retrieval

P(−
→
x | R)P(R)
P(R | −
→
x)= . (18.16)
P(−
→
x)

In the above,
P(R | −→x ) is called posterior probability—the probability of relevance, given that
the document is − →x,
−
→
P( x | R) is called likelyhood function or probability of evidence, which is the
probability of randomly selecting the document of description − →x from the set R
of relevant documents,
P(R) is prior probability of relevance, i.e., the probability that a randomly selected
document from the entire collection is relevant, and
P(−→x ) is probability that a randomly selected document has a description − →x . It is
determined as a joint probability distribution of the n terms in the collection [6].
Equation (18.16) can be expressed in simple language as

Posterior pr obabilit y ∝ likeli hood × prior pr obabilit y.

18.10.2 Bayes Probabilistic Inference Model

The Bayes probabilistic inference makes use of a network, which is an extension to

the probability-based IR. The network is a Directed Acyclic Graph (DAG), in which
nodes represent propositional variables or constants, and the edges represent the
dependency relations between the propositions [10].
If p and q are two propositions, and there is a relation of implication from p
to q, i.e., the first proposition “causes” the second, then p is cause and q is effect,
and is represented by p → q. In the DAG, p and q are nodes, and there is an edge
from a node marked as p to node marked as q. A link matrix is stored at node q,
which specifies the probability P( p|q) for all possible values of variables p and q.
The expression P( p|q) is the expression for the probability of occurrence of event
p given that q has already occurred. In the model we take q node as evidence, and
it stands for a query. In a scenario, a node has more than one parent (say p1 , p2 ,
...), the link matrix will indicate the dependence of query node q on all the parents.
The query node q now characterizes the dependence relationship between itself and
all the nodes p1 , p2 , ..., which are potential causes. This is illustrated in Fig. 18.10.
Using the Bayes theorem, the conditional probability expression can be expanded as

P(q| p).P( p)
P( p|q) = . (18.17)
P(q)

When the set of prior probabilities P( p) are given for the root of a DAG that
represents the document, the network can be used to compute the probability of
belief associated with all the remaining nodes. Figure 18.10 shows a document di at
18.10 Using Bayesian Networks for IR 589

Fig. 18.10 Basic Inference di

Network Model for IR
t1 t2 tj tn

AN D q1 q2 AN D

root, corresponding keywords t1 , . . . , tn , and submitted queries q1 , q2 . Note that, t1 ,

..., tn are the representation of document di [17].
Through the inference network, the random variables are associated with the
documents, index terms, and user queries. Multiple evidences of query terms in the
document’s representation for a given query are conjuncted to estimate the probability
that the document satisfies the user’s information need. For example, in Fig. 18.10,
query q1 is conjunct (ANDed) of terms t1 , t2 , t j , and q2 = t2 ∧ t j ∧ tn . Thus, variables
associated with document di represent the event that the document is observed. The
index terms/document variables are represented as nodes of the DAG, and the edges,
which are directed from document nodes to index terms nodes indicate that the
observation of document results in an improved belief on its term nodes.
Further, a random variable associated with the user’s query node models the
event that information need expressed in the form of user’s query has been met. The
dependency in the form of direction arrows indicate that belief in the query node
is function of beliefs in the nodes that correspond to the query terms. In Fig. 18.10,
document di comprises t1 , t2 , . . . , t j , tn as its index terms. Similarly, the query q1
comprises the query terms t1 , t2 , . . . , t j , hence, q1 = t1 ∧ t2 ∧ t j , and q2 = t2 ∧ tn .
From this Bayes inference model for IR, we note that a set of edges pointing
to a node represents the probabilistic dependence between that node and its par-
ents. Through its structure, the Bayes network represents the conditional dependence
relation among the variables in the network. These dependence relations provide a
framework for retrieving the information [4, 6].

18.10.3 Bayes Inference Algorithm

For IR using Bayes inference, a user specifies one or more topics of interest while
keeping in mind some document features, the latter are to be used as evidence for
topics of interest mentioned above. The task of IR using Bayesian inference network
is documented as Algorithm 18.5, which requires building an inference network for
the representation of query terms and document features (i.e., terms), and computa-
tion of posterior probabilities (P( p|q) in Eq. 18.17) based on the prior probability
of the document [6, 10].
590 18 Information Retrieval

Algorithm 18.5 Bayesian Inference-based IR

1: Construct the network representation of query
2: {Steps to score all the documents}
3: for all documents do
4: Extract the features {t1 , t2 , ...} from document
5: Label features in network
6: Compute posterior probabilities, P( p|q)
7: end for
8: Rank the documents set in order of posterior probabilities

Fig. 18.11 Two-Level t1 t2

Bayesian network model for
IR
f11 f12 f1m f21 f22 f2m

In this algorithm, steps 4, 8 are IR routines, whereas steps 5, 6 are routines to

draw inferences.
Figure 18.11 shows the term-weighting architecture of Bayesian network, which
indicates that there are topics of interest, shown as t1 , t2 , and there are features to be
examined, f 11 , ..., f 1m , f 21 , ..., f 2m . The features’ set and topics set shown here are
different, but they can be the same also. The occurrence of the topics t1 , t2 on graph
nodes represent the event that the document is related to topics t1 , t2 . Whereas, the
nodes corresponding to the features represent the events, for example, the features
f 11 , ..., f 1m represent the event that these features are present in the document d1 .
The network structure in Fig. 18.11 is based on the following assumptions con-
cerning the Bayesian probability:
1. Given the topics {ti , t2 , ...} from the document (the document is relevant to these),
the presence or absence of any feature does not imply about the presence or
absence of some other feature. In other words, it is assumed that there are no
dependency relations between the features.
2. Given that the document relevant to one topic does not affect one’s belief about
the relevance or non-relevance of that document to any other topic.
In the above, the first assumption specifies the conditional independence of fea-
tures, when the topic is already given. It is called as binary independence. The absence
of arcs between feature nodes in Fig. 18.11 is an explicit indication of independence
between features. Given these conditions and the Bayes network in Fig. 18.11, we
can draw some important conclusion. That is, the assumption “1” will not be valid,
if the query includes features that are identical or closely related, like synonyms,
because in that case the features are not independent.
Now, for the network in Fig. 18.11, we define two sets of probabilities, given
below:
(a) P(t1 ), P(t2 ), .., called prior probabilities, that a document is relevant to topics
t1 , t2 , ..., etc. If we are discussing one document only, then it is P(t1 , t2 , ..).
18.10 Using Bayesian Networks for IR 591

(b) The conditional probability P( f i j |ti ) of each feature f i j is defined as the prob-
ability that feature f i j is present in the document, given that this document is
relevant to topic ti .
Given the above, the task of IR is to compute the posterior probability P(ti | f i1 ,
. . . , f im ), which is the probability that the document is relevant to ti , given that we
have observed the presence or absence of all of the features f i j , called evidences.
The Bayes rule can be directly applied for this computation. The network topology
shown in Fig. 18.11 is called Bayes inference and, can be expressed by

P(ti ).P( f i1 , . . . , f im | ti )
P(ti | f i1 , . . . , f im ) = . (18.18)
P( f i1 . . . . , f im )

Generally, we are not interested in absolute numerical values of the posterior prob-
abilities, but want to just rank the documents by the posteriors. Thus, we can eliminate
the denominator term P( f i1 . . . , f im ) in this equation as long as this denominator
remains the same, with varying ti ,s. Further, we can simplify this Bayes rule to a
linear decision rule given in Eq. 18.19, where I ( f i j ) is an indicator variable that
equals to 1.0 only if f i j is present in the document and 0.0 otherwise, and w is a
coefficient corresponding to a specific (feature, topic) pair:

g(ti | f i1 , . . . , f im ) = I ( f i j ) × w( f i j , ti ). (18.19)
j

A careful choice of w results in a ranking of documents in descending order of

g(), which turns out to be in the same order as that of ranking them in decreasing
order of the posterior probabilities. However, since g() does not include the priors
probabilities (P(ti )) of the topics ti , which is indicator of the relative rarity of the
topics, one cannot compare a document’s strength of relevance to one topic with
respect to its strength of relevance to a different topic.
The coefficients w can be interpreted as weights corresponding to each feature f i j
and term ti . Similarly, the function g() can be interpreted as the sum of weights of
the features f i j that are present in the document, which is relevant to topic ti . Hence,
this method is known as “term weighing”.
In the Bayes network topology shown in Fig. 18.11, the query corresponding to
each topic, e.g., t1 , is represented by its own subnetwork, which is shown disconnected
from the subnetworks of other topics’ queries. Hence, these isolated models fail to
represent the possible relationships between the topics, for example between t1 , t2 .
Therefore, it is difficult to acquire consistent, feature-conditional probabilities, as
well as find out the probabilities by combining the topics.
592 18 Information Retrieval

Fig. 18.12 Information t1 t2

retrieval model with two
related topics
f1 f2 f3 f4 fm

Fig. 18.13 Multi-topic

S
query as a single
compound-topic query

f1 f2 f3 f4 fm

18.10.4 Representing Dependent Topics

In the previous section, we discussed that, a document relevant to one topic does
not affect one’s belief about the relevance or non-relevance of that document to any
other topic.
An IR network topology that removes this assumption, and explicitly represents
the relationships among different topics is shown in Fig. 18.12. We note that the
Bayesian inference problem becomes more completed with multiple topics.
The addition of a relationship between topics in Fig. 18.12 requires two changes in
specifying the network probabilities. In the first one, we must specify what are those
topics (t1 , t2 , ...), and then compute the strength of the relevance between them. In
the second step, we must compute the probability of each feature ( f i j ), conditioned
on each combination of its parent topic. Having done in this manner, it gives rise to
a number of probabilities, which are combinatorial in size.
However, there is an approach to simplify the multiple-topic network, such that it
looks like a single-topic network, and its range is only all the possible present/absence
combinations of topics t1 , t2 , ..., etc. Thus, in a way we have translated the topology
from one form to another. The modified topology of Fig. 18.12 is shown in Fig. 18.13,
where node S represents the compound topic. An advantage of this representation is
that the same set of formulas (18.18) and (18.19) for the computation of probabilities
can be used now also. The disadvantage of this approach is that the compound query
shall contain 2n states for n parent states, which will complicate the computations.

18.11 Semantic IR on the Web

The semantic web is a vision of the future WWW. It is an extension of the current
web, where information is given with well-defined meanings, that will better enable
the computers so that computers and people can work in cooperation. However, a
universal implementation of the semantic web—a full substitution for the existing
web—is still far away from reality. Therefore, it could be useful to have a system
that analyzes the documents from a contextual point of view for more accurate
18.11 Semantic IR on the Web 593

retrieval. The semantic IR (SIR) on web is based on computing semantic relations to

evaluate the relevance of documents with a query in a given context, and makes use
of structures like lexical chains, semantic networks, and ontologies. The semantic-
based approach is context-driven, where keywords (topics/terms) in documents are
processed in the context of the information in which they are retrieved. This will
help solve the semantic ambiguity so that the retrieval is accurate, and as per the true
need of the user [15].
In the recent times, the information and knowledge representation using ontologies
have acquired great importance, as it is found to be suitable for strategic requirements.
These strategies are intrinsically independent on information codification, which
helps to isolate the information, as well as to recover, organize, and integrate it with
respect to its content.
Following are the definitions of ontologies.

Definition 18.1 Ontology. An explicit and formal specification of shared conceptu-

alization is called an ontology. It is an abstract model of specified reality, such that
the components are clearly identified. The terms in definition are further clarified as
follows:
– Explicit means type of concepts used and the constraints on them are well defined,
– Formal refers to the ontology property of being machine-readable, and
– Shared means, a property of ontology of capturing consensual knowledge, accepted
by a group of persons. 

Definition 18.2 Ontology (Definition-2). An ontology defines the basic terms and
their relations consisting the vocabulary of a topic, as well as the rules for combining
terms and relations to define extensions to the vocabulary. 

The above definition also indicates a path to be followed in order to construct an

ontology:
1. first identify the basic terms and their mutual relations;
2. agreement on the rules that arrange them;
3. defining of terms, and relations among concepts.
From the above perspective it is clear that an ontology does not include just the
terms that are explicitly defined in it, but keeps provision to derive new terms using
defined rules and properties. Also, the ontology can be viewed as a “set of terms and
relations between them, which denote the concepts used in a domain.”
The concept of semantic relatedness refers to the relations between words and
concepts that are in the practice, or those based on the perceptions. There can be
several metrics to measure the semantic relatedness of words. Some of the approaches
to these metrics are as follows.
• Thesaurus-based metrics. These metrics make use of thesaurus where words are
related to concepts, and each word is referred to a category by an index structure.
594 18 Information Retrieval

• Dictionary-based metrics. Dictionaries are linguistic information sources of our

knowledge about the world; they form a knowledge base in which headwords are
defined using other headwords and/or their derivatives.
• Semantic network-based metrics. These metrics use semantic networks—graphs
in which the nodes are the concepts, and the arcs between nodes represent relations
between concepts. Number of edges’ links between terms (nodes/concepts) in the
semantic network, without loops, is a measure of conceptional distance between
terms (nodes).
A sequence of related words in a text is called lexical chain (a linguistic structure),
which may span short distances (adjacent words or sentences) or long distances,
covering the entire text. Computing a lexical chains helps in the identification of the
main topics of a document. The semantic relatedness measures use lexical chains to
perform their computations, the lexical chains are used for IR and related areas, and
to explore structure of texts as well. The lexical chains have been also used to index
video-conference transcriptions by topic, construction of typical IR system and text
segmentation systems, and for automatic generation of hypertext links.
The strength of a relation between words that connect different fragments of the
text is measured by their cohesion, and the cohesion between lexical units of text is
called lexical cohesion—the most common type of cohesion. The lexical cohesion
can be expressed by repetitions of relations like synonym, hyponym, or by other
linguistic relations between words, such as whole-part and object-property.
Semantic IR Systems
Use of ontologies in IR consists of an approach that identifies important concepts in
documents using criteria of semantic relatedness and co-occurrence; this is followed
by disambiguation of them using an external general purpose ontology (e.g., Word-
Net). On matching the ontology with a document provides a set of scored concept
senses (nodes) with weighted links, called semantic representation of the document.
The steps for the process of Semantic IR (SIR) are as follows:
1. For improving the web searches, only important information is selected from
the user query, which is helpful in extracting information from documents;
2. The user’s query or a phrase expressed in natural language is sent to a lexical
processing module;
3. The boundaries of words and phrases are detected through tokenization. The
system labels the words using some tagger like Brill’s tagger;
4. A phrase parser splits every phrase into several members, as nouns and verbs;
5. The stop words are removed, and the system uses some keywords to represent
the main concept of the phrase;
6. The remaining process steps of SIR, like Word Sense Disambiguation (WSD),
query expansion, and post-processing, have been already discussed in this
chapter.
18.11 Semantic IR on the Web 595

To solve the problem of polysemy,5 the concept of “word sense” from WordNet is
used, which helps the user interacting with the system to associate with every concept
a list of terms semantically are related to.

18.12 Distributed IR

When the size of data is very large, or when it is required to support high query vol-
umes, single machine is not enough to support the load of IR, even when the number
of the enhancements discussed above have been used. For example, even when the
Internet was not so common, in mid-2004, Google search engine processed more
than 200 million queries a day against more than 20 GB of crawled data, and it used
over 20,000 computers. The requirement in the present time is hundred times bigger.
For handling large size of data loads, a combination of distribution and replication is
necessary. The distribution means, document collections and their indexes are split
across multiple machines, so that answers to every query is synthesized from many
collections of components. The word replication means mirroring, which involves
making enough identical copies of the system so that the required query load can be
handled with acceptable minimum response time [18].
Document-Distributed Architectures
A very simple distributed document’s architecture is to partition the collection and
allocate one sub-collection to each of the separate processors (see Fig. 18.14). To
make use of distributed document architecture, a local index is built and main-
tained for each sub-collection, and when a query arrives, it is passed to every sub-
collection to search and evaluate in parallel against every local index. The sets of sub-
collections’ answers are then merged in some way to provide an overall answer. The
main advantages of such document’s partitioning system is that, collection growth
is handled by designing one of the hosts in the form of dynamic collection, such that
only this host needs to rebuild its index. The parts of the process that are computa-
tionally expensive are distributed equally across all the hosts in the computer cluster.
These parts are searching the index, updating individual indexes, computation of
weights, and id f for documents, etc.
Figure 18.14 shows a simple inverted index of a document-distributed retrieval
system. It shows two index partitions: 1. a term-based index partition, and 2.
document-based index partition. Elements in a inverted list have format document
number, term-frequency, for example 2, 2 in second row, second column indicates
that the term “quick” has frequency 2 in the document number 2, while 3, 1 in the
same row, third column indicates that the term “quick” has frequency 1 in the docu-
ment number 3. Each of the two dashed regions in Fig. 18.14 show one component
of a document-distributed retrieval system; with this, one processor indexes all terms

5 Polysemy: A single term with several meanings.

596 18 Information Retrieval

Document number

1 2 3 4 5 6
< 6, 1 >
the
quick < 2, 2 > < 3, 1 >

brown < 6, 1 >

Vocabulary

fox < 4, 4 >

jumped < 2, 1 >

...

over < 4, 2 >

a < 6, 1 >
Term-based
lazy < 1, 3 > < 2, 2 > < 3, 3 > < 4, 1 > < 5, 3 > < 6, 2 > index partition
dog < 1, 1 > < 3, 1 >
today < 4, 3 >

Document-based
index partition

Fig. 18.14 Partition and distribute index across a cluster of machines

that appear in the first two documents of the collection, creating a document-based
partition.
Term-Distributed Architectures
An alternative strategy for partitioning the index is term-based partitioning, where
index is split into components by partitioning the vocabulary, with one possible
partition shown by the dotted horizontal split in Fig. 18.14. Each processor has full
information about a subset of the terms, i.e., all the necessary things to handle a
query. Hence, only the relevant subset of the processors needs to respond to a query.
The term partitioning has the advantage that it requires fewer disk seek and transfer
operations during query evaluation than document partitioning because each term’s
inverted list is still stored contiguously on a single machine rather than being split in
fragments across multiple machines.
On the other hand, each of these disk transfer operations involve more data.
Mainly because, in a term-partitioned arrangement we are discussing, the majority
of the processing load is on the coordinating machine; the experiments have shown
that it can easily become a bottleneck and starve the other processors of work.
Term-Based Versus Document-Based Partitioning
Compared to the term-based partitioning, the document-based partitioning typically
results in a better balance of workload and achieves higher throughput for queries.
18.12 Distributed IR 597

Also, the document-based partitioning (document distribution) allows for more nat-
ural index construction and for document insertion. On the other side, for term-
partitioned index, index construction involves first of all, distributing the documents
and building the document-partitioned index. Then, once the vocabulary split has
been agreed upon by the processors, the index fragments are exchanged between all
pairs of processors.
The document-based partition also has the practical advantage of providing the
search service even when one of the hosts is offline for some reason, because any
answers not resident on that machine are still available to the system. For exam-
ple, there are 10 machines, m 1 ...m 10 and 100 documents d1 ...d100 , with document
distribution as follows:
machine m 1 : d1 ...d10 ,
...
machine m 10 : d91 ...d100 .
Now, if machine m 5 is offline, the queries for all the documents, except d51 ...d60 ,
is answered. Whereas, in the term-based distribution, since index terms of every doc-
ument are uniformly distributed on all the processors, the query cannot be answered
even if one processor is offline. However, in the term distribution case, if any machine
is offline or idle, it is immediately noticeable because it will affect queries belonging
to almost every document.
Google indexing uses document-based partitioning, with massive replication and
redundancy at all the levels, for example at machine level, cluster level, and indi-
vidual level. In addition, the document partitioning remains effective even if the
collaborating systems are independent and unable to exchange their index data. The
distributed system makes use of a meta-searcher, using which the final result answer
list is synthesized from the possibly overlapping answer sets provided by a range of
different services.

18.13 Summary

Information retrieval (IR) process is the identification of documents or other units of

information in a collection that are relevant to particular information needs expressed
through queries for which people are interested to find answers. The IR models
consider that each document is described by a set of representative keywords, called
index terms—a word from a document that represents the semantics of the document.
In general the index terms are nouns. If ti is an index term, d j is a document then
wi, j ≥ 0 is the weight associated with the pair (ti , d j ), or single entity wi j .
Two ratios, precision and recall are used to measure the effectiveness of an IR
system. Precision is the ratio of the number of relevant documents retrieved to the
total number of documents retrieved, i.e., what fraction of the retrieved documents
are relevant, and recall is the ratio of the number of relevant documents retrieved to
598 18 Information Retrieval

the total number of relevant documents in the entire collection, i.e., what fraction of
the relevant documents have been retrieved.
The common retrieval strategies are Boolean Model, Vector Space Model, Prob-
abilistic Model, Inference networks, and Fuzzy set-based retrieval.
The Boolean model is a simple retrieval model based on Set Theory and Boolean
Algebra, Vector Space Model computes the measure of similarity (sim(q, d j ))
between query q ∈ Q and a document d j ∈ D, where q and d j are vectors for query
and document, and the Q, D are sets for queries and documents, respectively.
The probabilistic retrieval model computes the similarity measure (sim(q, d j ))
between the query q and a document d j as the probability that document d j is relevant
to q. This model estimates a query term’s weight on how often the term appears or
does not appear in relevant and non-relevant documents, respectively. One class of
probabilistic approach for IR is Bayesian networks. These networks are annotated
directed graphs encoding probabilistic relationship among distinctions of interest, in
an uncertain reasoning problem.
The fuzzy retrieval technique is based on fuzzy set theory and fuzzy logic—an
extension of the classical set theory. The word matching between the query set and
the text word is not limited to the perfect matching, but, the matching is graded,
depending on the degree or level of matching in the range from 0 to 1. In fuzzy set
-based retrieval, a membership value R(xi , yi ) specifies for each xi ∈ X , yi ∈ Y , the
grade of relevance of index term xi with the document yi .
In a concept-based information retrieval, queries and documents are represented
using semantic concepts, instead of (or in addition to) keywords, and they perform
retrieval in that concept space. This results in a retrieval model that is less dependent
on the specific terms used, and yields matches even when the same sense is described
by different terms in the query and target documents.
The retrieval models make use of an inverted index file structure. Instead of search-
ing into a document, an inverted index is generated in advance for all the keywords
in the document set. For each term, a pointer references to a linked list, which con-
tains an entry for each document containing this term as well the term frequency
in that document. The single key problem with this index is that volume of data
involved cannot be held in the main memory. To solve this problem, the indexes are
constructed in parallel and can be merged after a regular intervals.
Yet, there is another method for IR that uses an approach based on the measure of
semantic relatedness, applied to evaluate the relevance of a document with respect to
a query in a given context. The approach makes use of structures like lexical chains,
ontologies, and semantic networks. The semantic approach implements a context-
based system, such that keywords are processed in the context of the information
from which they are retrieved. This approach helps in solving semantic ambiguity,
and results in giving a more accurate retrieval, that is based on the real-world interests
of the user.
In most information retrievals, user queries are short and the natural language is
inherently ambiguous, consequently, the retrieval is prone to errors and omissions.
The most critical language issues are polysemy and synonymy. To resolve this, the
18.13 Summary 599

query is expanded by appending the synonyms of query terms into the query, called
Automatic Query Expansion (AQE).
When large volumes of data sets are involved or when the query volumes are high,
one machine may be inadequate to support the users’ query load, even when the
various enhancements and optimizations are carried out. For handling heavy load of
users’ queries, a combination of distribution and replication is required. Distribution
means, the document collection and their indexes are split across multiple machines
(servers) and that answers to the query as a whole must be synthesized from the
various collection components. Replication (or mirroring) involves making enough
identical copies of the system so that the required query load can be handled at speed
and accuracy.

Exercises

1. Show how the vector space model can be modeled using an inference network.
2. Consider a documents collection made of 100 documents. Given a query q, the
set of documents relevant to the users is D ∗ = {d4 , d15 , d34 , d56 , d98 }. An IR
system retrieves the following documents D = {d4 , d15 , d35 , d56 , d66 , d88 , d95 }

a. Compute the number of True-Negatives, True-Positives, False-Negatives,

False-Positives.
b. Compute Precision, Recall, and F-measure.

3. Consider an IR scenario in the following: It has been found in some hospital,

results of blood tests taken on a specific day are unreliable for diabetic patients
due to equipment malfunction. The hospital uses an IR system to identify these
patients. Suppose the collection of patients’ records contains 10, 000 documents,
500 of that are relevant to the query. The system returns 350 documents, 225 of
that are relevant to the query. Answer the following for this scenario:
a. Calculate the precision and recall for this system.
b. Based on your results from above, explain how well would you say about the
working of hospital’s IR system.
c. Knowing about the precision-recall trade-off, what is likely to happen if an IR
system is tuned to aim for 100% precision?
d. Knowing about the precision-recall trade-off, what is likely to happen if an IR
system is tuned to aim for 100% recall?
e. For the trade-off given scenario, which measure do you think is more important,
precision or recall? Why?
4. You are looking for information on “Economic growth in India” in a large doc-
ument collection, during the period of last 3 years. You decide to search using
the terms: India, banks, growth, economy, business, agriculture, using an IR
system, which recommends three possible documents given below with term
frequencies.
600 18 Information Retrieval

Term Economy India Growth Banks Business Agriculture

Document- 15 10 3 4 2 9
1
Document- 0 0 9 8 7 8
2
Document- 4 2 4 4 6 10
3

There is no additional information about the documents. Make use of each of

the following models to find out the relevancy of the documents to the query.
a. Boolean model
b. Vector space model
c. t f.id f model
5. Take any three small documents of size, approximately 100 words.
a. Build a matrix of an inverted index for these documents, in the format shown
in Fig. 18.5.
b. Weight terms by their presence/absence (binary), and also and by t f × id f
(with estimated IDFs).
c. Compute the memory requirements for this inverted index. Make necessary
assumptions for character size, pointer size, etc.
d. Construct a suitable query, and calculate document–query similarity, for the
following scenarios:
i Cosine (with normalization)
ii Inner product (i.e., cosine without normalization)
iii Does the normalization has any effect? Justify.
6. Consider that we submit the queries to search engines for searching the needed
information on WWW.
a. Does the search process use a stop-word list?
b. Can you search “The”, “The a”, “An a”, etc.? Justify.
c. Is it a practice to search the above terms?
d. Does the search process use stemming?
e. Are there different results for two queries “Human body”, “Humanly body”.
Justify your answer.
f. Does it normalize words to lower case?
7. “Having the knowledge of the sense of a query term may help a document retrieval
system, especially for short queries.” Why it is not true for longer queries?
8. Comment on the validity of following statements for Boolean model:
a. “Stemming does not lower the precision of a Boolean retrieval system.”
b. “Stemming does not lower recall of a Boolean retrieval system.”
9. Answer the following in brief:
Exercises 601

a. Why is the idf of a term always finite?

b. What is the idf of a term that occurs in every document?
c. What is the idf of a term that appears in one document only?
d. What is the idf of a term that appears in no document?
10. Answer the following in brief:
a. Name three criteria for evaluating a search engine.
b. What is an easy way to maximize the recall of a search engine?
c. What is an easy way to maximize the precision of a search engine?
11. What is the difference between clustering and classification? How can they be
used in a complete IR system?
12. Discuss the merits and demerits of following, suggest as to which one will provide
better response time?
a. Document-distributed architecture.
b. Term-distributed architecture.

References

1. Carpineto C, Romano G (2012). A Survey of Automatic Query Expansion in Information

Retrieval. ACM Comput. Surv. 44(1): 50.https://doi.org/10.1145/2071389.2071390
2. Chowdhary K R, Bansal VS (2001) Current trends in information retrieval. In: The 4th Inter-
national Conference of Asian Digital Libraries, Dec. 10–12, 2001 Bangalore, pp. 306–319
3. Chowdhary K R, Bansal VS (2003) Fuzzy Logic-based information retrieval. In: Conference
proceedings on algorithms and artificial systems, Allied Publishers Pvt. Ltd. pp 297–307. ISBN
81-7764-403-3
4. Chowdhary KR (2004) Natural language processing for word sense disambiguation and infor-
mation extraction. PhD Thesis, J.N.V. University, Jodhpur, May 2004
5. Chowdhary KR (2005) Word sense disambiguation. J Comput Sci 1(1):30–37
6. Chowdhary KR (2008) Information retrieval from digital libraries using probabilistic-
possibilistic inferences. In: IR@INFLIBNET INFLIBNET’s Convention Proceedings CAL-
IBER 2008 Allahabad, http://ir.inflibnet.ac.in/handle/1944/1225
7. Chowdhary KR, Bansal VS (2006) Information extraction from natural language texts. J Institut
Eng (India), 87:14–19
8. Chowdhary KR, Bansal VS (2011) Information retrieval using probability and belief the-
ory. International conference emerging trends in networks and computer communications
(ETNCC). https://doi.org/10.1109/ETNCC.2011.5958513
9. Egozi O et al (2011).Concept-based information retrieval using explicit semantic analysis.
ACM Trans Informat Syst, 29(2):8.1–8.34. https://doi.org/10.1145/1961209.1961211
10. Fung R, DelFavero B (1995) Applying Bayesian networks to information retrieval. Commun
ACM 38(3):42–49
11. Miller GA (1995) WordNet: a lexical database for English. Commun ACM 38(11):39–41.
https://doi.org/10.1145/219717.219748
12. Grossman DA, Ophir F (2008) Information retrieval-algorithms and heuristics, 2nd edn.
Springer
13. Heckerman D et al (1995) Real-world applications of Bayesian networks. Commun ACM
38(3):24–26
14. Recardo BY, Berthier RN (1999) Modern information retrieval. Addison Wesley-ACM Press
602 18 Information Retrieval

15. Rinaldi AM (2009) An ontology-driven approach for semantic information retrieval on the
web. Trans Internet Technol 9(3):10:1–10:24. https://doi.org/10.1145/1552291.1552293
16. Smith LC (1976) Artificial intelligence in information retrieval systems. Informat Process
Manage 12:189–222. Pergamon Press
17. Wright (1921) Correlation and causation. Agric Res 20:557–585
18. Zobel J, Moffat A (2006) Inverted files for text search engines. ACM Comput Surv 38(2):1–56
Chapter 19
Natural Language Processing

Abstract The abundant volume of natural language text in the connected world,
though having a large content of knowledge, but it is becoming increasingly difficult
to disseminate it by a human to discover the knowledge/wisdom in it, specifically
within any given time limits. The automated NLP is aimed to do this job effectively
and with accuracy, like a human does it (for a limited of amount text). This chapter
presents the challenges of NLP, progress so far made in this field, NLP applications,
components of NLP, and grammar of English language—the way machine requires
it. In addition, covers the specific areas like probabilistic parsing, ambiguities and
their resolution, information extraction, discourse analysis, NL question-answering,
commonsense interfaces, commonsense thinking and reasoning, causal-diversity, and
various tools for NLP. Finally, the chapter summary, and a set of relevant exercises
are presented.

Keywords Natural language processing · Challenges of NLP · Natural language

parsing · Probabilistic parsing · NL ambiguities · Ambiguity resolution ·
Commonsense interfaces · Commonsense thinking · Commonsense reasoning ·
Discourse analysis · Question-answering · Causal-diversity · NLP tools

19.1 Introduction

A language is a set of finite length sentences, constructed using a finite alphabet

set, or in terms of language syntax, they are constructed using a finite vocabulary
of symbols. Since the alphabet set is finite, as well as the length of the sentences,
the set of sentences (in a language) is also finite. For example, if alphabet set is of
size two, and length of sentences is ten, there can be only 1024 maximum number
of sentences possible. However, in many of our theoretical studies we consider the
language as infinite, hence size of some sentences may also be infinitely large. For
our study of languages, we usually consider the languages without any bound, i.e.,
infinite, but for specific languages, which are of practical use, we limit our study
to finite set. This is because, the machines we would like to use for processing the

© Springer Nature India Private Limited 2020 603

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_19
604 19 Natural Language Processing

languages, are finite machines, they have finite memory, and have finite processing
power.
When we are interested in an infinite language set, say L, we can use the finite
devices called generating grammars to investigate the structure of L. In that sce-
nario, the theory of language will contain specification of a class of functions from
which grammars for a particular language may be drawn. Natural language process-
ing (NLP) is a collection of computational techniques for automatic analysis and
representation of human languages, motivated by theory. However, the automatic
analysis of text, at par with humans, requires a far deeper understanding of natural
language by machines, which is still far from reality. There are many examples of
NLP, like, online information retrieval, aggregation, and question-answering, have
been mainly based on algorithms relying on the textual representation of web pages,
as well NLP to some extent. Such algorithms are very good at retrieving texts (IR),
splitting it into parts, checking the spellings, and and word-level analysis, but not
successful for analysis at sentence and paragraph level. Hence, when it comes to the
question of interpreting sentences and extracting meaningful information, however,
the capabilities of these algorithms are still very limited.
NLP, in general, requires high-level symbolic capabilities, which includes the
following:
Access and acquisition of lexical, semantic, and episodic characteristics,
Creation and propagation of dynamic bindings,
Manipulation of constituent recursive structures,
Coordination of many processing and learning modules,
Identification of basic language constructs (e.g., objects and actions) and,
Representation of abstract concepts.
All the above capabilities are needed to shift from mere NLP to what is usually
called as natural language understanding. The present approaches to NLP are based
on the syntactic representation (also called syntactic structures) of text, i.e., relying
on the word co-occurrence frequencies. Such algorithms have the limitation that they
can process only based on the information they can see in the text being processed,
but cannot consider the background information what we humans do. For example,
when we say that “Sachin Tendulkar is a good batsman,” we understand this sentence
due to abundant information we have in our brain about the game of cricket, and bout
the successes of “Sachin Tendulkar” in many cricket matches. However, the present
algorithms do not have all this background with them, hence their understanding is
limited about any given text.
As human text processors, being humans we do not have such limitations, as
every word coming across in the text activates a cascade of semantically related
concepts, relevant episodes, and sensory experiences, all these enable the completion
of complex NLP tasks. These tasks are, for example, word sense disambiguation,
textual entailment, and semantic role labeling, all in a quick and effortless way.
Many new computational models are making attempts to bridge the cognitive gap
by emulating the processes recognized as being part of the human brain, and used
for language processing by humans. These approaches depend on semantic features
19.1 Introduction 605

that cannot be explicitly expressed through the text. The computational models are
useful for theoretical purposes, e.g., scientific studies, such as exploring the nature
of linguistic communication and its properties, as well for practical and industrial
applications, such as enabling effective human–machine communications.
Challenges of NLP
Developing a program that understands natural language is a difficult problem. Major-
ity of natural languages are large, they contain infinitely many sentences. Also there
is much ambiguity in natural language. Many words have several meanings, such as
can, bear, fly, orange, and the same sentences many times have different meanings
in different contexts. Due to this, creation of programs that understands a natural
language is a challenging task.
The syntax of a language helps us to decide how the words are being combined
to make larger meanings. For example, in the sentence “the dealer sold the merchant
a dog,” it is important to be clear about what is sold to whom. Some of the common
examples are the following:
I saw the Golden gate bridge flying into San Francisco.
(Is the bridge flying?)
I ate dinner with a friend.
I ate dinner with a fork.
Can companies litter the environment
(Is this a statement or question?)
Finally, assuming that we have overcome the problem at the previous levels, we
must create an internal representation, and then, some how use the information in an
appropriate way. This is the level of semantics and pragmatics. Here too the ambi-
guity is prevalent. Consider the following sentences.

Jack went to store. He found the milk in aisle three. He paid for it and left.

Here the problem is deciding whether “it” in the sentence refers to aisle or three,
the milk, or even the store.
The most important part in the above is what is internal representation, so that
these ambiguities in understanding the sentence do not occur and machine under-
stands the way a human being understands the sentences.

Learning Outcomes of this Chapter:

1. Define and contrast deterministic and stochastic grammars, providing examples
to show the adequacy of each. [Assessment]
2. Simulate, apply, or implement classic and stochastic algorithms for parsing natural
language. [Usage]
3. Identify the challenges of representing meaning. [Familiarity]
4. List the advantages of using standard corpora. Identify examples of current cor-
pora for a variety of NLP tasks. [Familiarity]
606 19 Natural Language Processing

5. Identify techniques for information retrieval, language translation, and text clas-
sification. [Familiarity]
6. Tools for NLP. [Usage]

19.2 Progress in NLP

Right from its start, NLP focused on these tasks: natural language translation, infor-
mation retrieval, information extraction, text summarization, question answering,
topic modeling, and the recent one on opinion mining.
Syntax Versus Semantics
The NLP advances that took place after the 1960s paid attention on syntax, and
partly on the semantics, but it was more of syntax-driven processing [1]. However,
the semantic problems of Natural Language and its processing were obvious require-
ments from the beginning itself, but strategy adopted was to tackle the syntax first,
for the more direct applications. Many works on NLP recognized the need for exter-
nal knowledge for interpretation and reasoning on input language (Minsky, 1968),
who explicitly emphasized semantics, e.g., the general-purpose semantics with case
structures for representation and semantically driven processing [17].
First-Order Predicate
One of the most popular Natural Language representation technique used is FOPL
(First-Order Predicate Logic), which is a deduction-based logic consisting of axioms
and inference rules. The FOPL can be used to construct relationally rich predicates
formulas, using quantification. The inferences can be drawn using formulas, called
knowledge base, with the application of resolution principle. The FOPL supports
the expressions of syntactic, semantic and, to a certain extent pragmatic. The syn-
tax expresses a way of grouping of symbols, so that they are considered properly
formed. The semantics specifies the meaning of these well-formed expressions. The
pragmatics is more difficult to implement, which specifies to make use of context-
related information to provide better correlations between different semantics. The
later is important in tasks like word sense disambiguation.
Default Logic
The Default Logic was proposed [15] to formalize the default assumptions, e.g., “all
birds fly.” However, problems arose when default logic formalized the facts that were
true in the majority of cases but false in the exceptional cases, i.e., “exceptions to
these general rules,” e.g.’ “penguins do not fly” [15].
Production Rules
The production rules based model is another popular model for representation of Nat-
ural Languages (Chomsky, 1956). This model keeps ongoing assertions in a volatile
working memory. The production rules comprise as their premises the set of condi-
tions and consequent as a set of actions (i.e., IF <conditions> THEN <actions>).
19.2 Progress in NLP 607

The basic operations in a production-based system comprise, in order, three steps:

1. Recognize, 2. Resolve the conflict, and 3. Act. These three operations repeat in
order as cycle until there are no more rules left in the working memory on which
these rules can be applied. The first step (recognize) identified the set of rules (called
conflict set) whose premises conditions are satisfied by the current working mem-
ory. The second step (resolve conflict) looks into the conflict set and selects a set of
suitable rules to execute. The suitability is with the objective of efficiency, such that
the task can be completed in shortest time. The “act” step is the third step, which
simply executes the actions and updates the working memory. The production rules
are created in modular forms for ease of writing as well for optimum use of memory
and processing requirements.
OWL
A new model for NLP is Ontology Web Language (OWL), which is an XML-based
vocabulary that extends the Resource Description Framework (RDF) language and
provides a more comprehensive set for representation of ontology. Some examples of
ontology representations are: definition of classes and relations between them, prop-
erties of the classes, and constraints on relations and on properties of the classes. The
subject–predicate–object model is supported by RDF, which makes assertions about
the resource. The RDF-based engines are commonly used for checking semantic
inconsistency, which helps to improve on the ontology classification.
Networks
The networks as tools have been commonly used for NLP, where inferencing is used
to be the primary goal. For example, the Bayesian belief networks (Pearl, 1985) is the
tool for expressing joint probability distribution over many interrelated hypotheses.
The variable in such networks are represented using directed acyclic graph (DAG),
the edges are causal connections between any two variables, such that the truth of
the causing variable directly influences the caused variable. A Bayesian network
represents the subjective degree of confidence. The representation explores the role
of prior knowledge and combines the pieces of evidence of the likelihood of the
events.
For computation of joint probability distribution of the belief network, we need to
follow the following steps: for each variable q represented through belief network,
there is a need to know the probability P(q| p1 , ..., pn ), where p1 , ..., pn are the
parents of q. In a large network it requires such computations at every node, making
it difficult to determine the joint probability distribution of the entire network. It will
require maintenance of probability table at each node in the belief network, which is
again a challenging task for large-scale information processing problems.
Finally, the Bayesian networks also have limited expressiveness in NLP, as it is
no better than that of proposition logic. Due to these factors, semantic networks are
more often used in NLP [13].
Semantics Networks
A semantic network is a graphical notation for representing using interconnected
nodes and arcs (Sowa, 1987), which is quite different from belief networks. The
608 19 Natural Language Processing

network’s focus is on relationships between a concept and a newly defined sub-type,

represented using Isa keyword. Such a network structure provides generalization,
which supports the rule of inheritance. The latter has the property of the copying the
properties defined for super-type to all the sub-types, there is no need to define the
properties again for the sub-types. The information represented through the semantic
networks is assumed to be true.
Apart from the inheritance-based networks, another kind of semantic networks
is the assertional network. This network is meant to assert propositions, and the
information it contains is assumed to be contingently true. However, the truth of
contingent is not implemented through default logic, but it is based on application
of human’s commonsense. The proposition also has sufficient reason, such that the
reason entails the proposition, e.g., “the rock is hot” with the sufficient reasons being
“the sun is shining on the rock” and “whatever the sun shines on is warm” [19].
The concept of semantic network came in the early 1960s, which was further
developed by Marvin Minsky, till the 1980s. The base of semantic network is human
mind, where it is hypothesized that the human intelligence results from a vast majority
of things, and not from any single perfect principle. As per Minsky, the human mind
(or even animals’) is made of many little things, which he termed as “agents”, but
each one of them is mindless, when considered as an independent element. But,
these agents lead to true intelligence when they work collectively, and responsible
for performing functions, like, remembering, generalizing, comparing, analogizing,
exemplifying, predicting, simplifying, etc.
The theory of human cognition by Minsky gave birth to many research projects’
attempts, that built commonsense knowledge bases for NLP tasks. The major repre-
sentative projects related to cognition are as follows.
Cyc—a Large Knowledge Based System [6],
WordNet—a Large Knowledge Based System [8],
Thought-Treasure—Natural Language Processing with Thought-Treasure [11],
and
Open Mind Common Sense project—a second-generation commonsense database
[18].

19.3 Applications of NLP

The importance of NLP is due to the fact that there is a huge amount of data in
WWW, at least 20 billion pages, and that can be used as a big resource, provided that
important information can be found from these through NLP. Some of the well-known
uses of this data are possible through the following applications [4]:
Indexing and searching large texts,
Information retrieval (IR),
Classification of text into categories,
Information extraction (IE),
Automatic language translation,
19.3 Applications of NLP 609

Automatic summarization of texts,

Question-answering (QA),
Knowledge acquisition,
Text generations/dialogues.
Some of the above domains of NLP will be discussed in this chapter, and we will
get the exposure sufficient enough to solve many problems of NLP however, the field
of NLP is so vast that even a book is not sufficient to present in detail all the available
techniques.

19.4 Components of Natural Language Processing

The NLP is the subject of computational linguistics—the study of computer sys-

tems for understanding and generating natural language. The linguistics has two
branches—computational linguistics and theoretical linguistics. The computational
linguistics has been concerned with developing algorithms for handling a useful
range of natural language as input. While the theoretical linguistics has focused
primarily on one aspect of language performance, grammatical competence—how
people accept some sentences as correctly following grammatical rules and oth-
ers as ungrammatical. They are concerned with language universals—principals of
grammar which apply to all natural languages [16].
Computational linguistics is concerned with the study of natural language analysis
and language generation. Further, the language analysis is divided into two domains,
namely sentence analysis, and discourse and dialogue structure. Much more is known
about the processing of individual sentences than about the determination of discourse
structure. Any analysis of discourse structure requires a prerequisite as an analysis of
the meaning of individual sentences. However, it is a fact that for many applications,
thorough analysis of discourse is not mandatory, and the sentences can be understood
without that [14].
The sentence analysis is further divided into syntax analysis and semantic analysis.
The overall objective of sentence analysis is to determine what a sentence “means”.
In practice, this involves translating the natural language input into a language with
simple semantics, for example, formal logic, or into a database command language.
In most systems, the first stage is syntax analysis. Figure 19.1 shows the relations
among different components of NLP.

19.4.1 Syntax Analysis

The syntax analysis of a language performs two main functions in analyzing natural
language text, which are as follows:
• Determining Structure. Given the sentences, the syntax analysis should identify
the subject and object of each verb, and determine what each modifying word is,
610 19 Natural Language Processing

Linguistics

Computational Theoretical
Linguistics Linguistics

Language Language
Analysis Generation

Sentence Discourse and

Analysis dialogue structure

Syntax Semantic
Analysis Analysis

Fig. 19.1 Components of NLP

and what phrase it modifies. This is achieved by assigning a tree structure to the
input, during the process called as parsing.

• Regularizing the syntax structure. Subsequence processing, i.e., semantic analysis

gets simplified if a large number of possible input sentences are mapped into a
small number of structures. For example, some material in sentences (enclosed in
brackets in the example below) can be omitted, but still maintaining the meaning
of these sentences unaltered.
“He talks faster than John [talks].”
“John ate cake and Mary [ate] Cookies.”
In addition, if the structures are appropriately chosen, operator–operand (e.g.,
verb–subject, verb–object) relations should be clearly evident in the output of the
syntactic stage. This kind of regularization is achieved, when passive voice sentence
is converted into active voice, for example.
“Those grapes were crushed by me,”
converted to
“I crushed those grapes.”
Some NLP type have chosen to omit syntactic regularization all together, and
have semantic component operate directly on the full variety of sentence structures.
In such systems, the syntactic regularization is subsumed within the semantic analysis
process, which will, however, require more complex semantic rules.
The syntax analysis is performed through grammars (phrase structure grammar)
using the process of parsing, and through transformational grammars, discussed in
the following sections.
19.4 Components of Natural Language Processing 611

19.4.2 Semantic Analysis

The objective of semantic analysis is to determine what a sentence means. The seman-
tics seeks the conditions under which the sentence is true. Almost equivalently, we
can say, what are the inference rules for the sentences of the language. Characterizing
the semantics of interrogative and imperative sentences are more problematic.
Why the meaning of a sentence is important as long as a sentence is correct or
syntactically correct? Here is one reason. Consider that we build a natural language
system, our aim is not only to ensure that sentence is correct, but usually we want the
system to do some thing in response to input, like, retrieve data, move a robot arm,
etc. In general, this will mean translating the natural language input into the formal
language of a database retrieval system, a robot command system, etc. In comparison
with natural language, these formal languages will have the following properties:
• unambiguous,
• simple rules of interpretation and inference, and
• logical structure determined by the form of the sentence.
Thus, there is a significant similarity between the tasks that translate into such a
practical language, and those translate into a logical formalism.
Formal Languages for Meaning Representation
One method for meaning representation is propositional Logic. Using this logic,
given that “If it is raining then ground is wet, and It is raining,” we can infer that
“ground is wet.” We observe that as long as we preserve the general structure of the
argument, including the “if ... then ...” connective in the first part, we may change
the sentences making up the argument and still have a valid argument. For example,
“If it is night then Ram is sleeping, and It is night,” we may conclude that “Ram is
sleeping.” And, the overall argument is a valid argument.
The above rule is represented in predicate logic as
Given that: P → Q, and
P,
we may conclude: Q.
The above rule is called Modus Ponens. Similar to propositional logic, the rule of
Modus Ponens also exists in Predicate Logic also. For details of these logics, refer
to chapters two and three in this book.
The other approaches for semantics are Semantic networks, and Conceptual
dependency, which are presented in details in the chapter seven of this book.

19.4.3 Discourse Analysis

So far we have discussed the structure and meaning of individual sentences, but what
is solution is the meaning of the entire text. Because, the information conveyed by a
612 19 Natural Language Processing

text is clearly more than sum of its parts—more than the meaning of its individual
sentences. For example, if a text tells a story, describes a procedure, or presents an
argument, we must understand the connections between the component sentences
in order to have fully understood the story. Discourse analysis is the study of these
connections between the sentences. Since these connections are implicit in the text,
identifying them may be a difficult task.
As a simple example of the challenge of discourse analysis, just consider the
following brief description of a naval encounter:
Just before the dawn, the Vikrant sighted the an unidentifiable blue-ship and fired two tor-
pedoes. It sank swiftly, leaving one survivor.

The problem we face is, what stands for “it”? There are four candidates in the
first sentence: dawn, Vikrant, blue-ship, and torpedo. The semantic analysis helps to
exclude the “dawn” as a meaning for “it”, and number agreement excludes “torpe-
does”. But, still leaves two candidates Vikrant, and blue-Ship, which are both ships.
Our syntax and semantic analysis tools will not enable us to resolve between these
two, as which stands for “it”. To get the noun for “it”, we must understand the causal
relationship between the firing of torpedoes and sinking of ship. Since Vikrant fired
torpedoes, it must have fired on the blue-ship (because Vikrant is not supposed to
fire on itself !). Hence, “it” stands for blue-ship.
Since much of the information conveyed by this text is implicit, we cannot claim
that the text is adequately understood unless the implicit information is recovered.
The role of discourse analysis is to recover this information.
As we have discussed that discourse is a multi-sentence text, discourse comes in
many forms, describe a scene, give instructions, or present an argument. The tools for
discourse analysis are Frames, Scripts, and Conceptual dependencies, which were
covered in detail in chapter seven in this book.

19.5 Grammars

The grammar of a language can be viewed as a theory of the structure of that lan-
guage. Like any mathematical theory, the theory of grammar is based on a a set of
finite number of observed sentences, but it projects this to an infinite set of grammat-
ical sentences. This becomes possible by creating general laws (grammatical rules),
framed in terms of such hypothetical constructs as the particular phonemes, words,
and phrases, of the spoken language under analysis. A properly formulated grammar
should be able to determine unambiguously the set of all grammatical sentences in
that language [2].
Let us assume that for most languages there are certain clear examples of grammat-
ical and ungrammatical sentences. Consider the following sentences of in English:
1. John ate a sandwich,
2. Sandwich ate a John.
19.5 Grammars 613

Suppose a large corpus-x does not contain either of these sentences. In this scenario
can we say, grammar that is determined for corpus-x will direct the corpus to include
the first sentence and exclude second. Even though such simple cases maybe not
adequate to confirm that all correct sentences of that language will be tested as
correct and all wrongs will be rejected, but still it becomes a strong test.
The first step toward the linguistic analysis of a language is to provide a finite
system of representation for its sentences. By the grammar of a language L we mean
a device that produces all of the strings that are sentences of L and no other strings.

19.5.1 Phrase Structure

The description of phrase structure comprises the grouping of words into phrases,
which are then grouped into smaller constituent phrases and so on, until the ultimate
constituents (generally morphemes) are reached. The phrases are typically classified
as noun phrases (NP), verb phrases (VP), etc. For example, the sentence “the man
took the book,” might be analyzed as shown in Fig. 19.2.
Evidently, description of sentences in such manner permit considerable simpli-
fication compared tor the word-by-word model. This is because the composition of
a complex class of expressions such as NP, can be stated just once in the grammar,
and this class can be used as a building block at various stages in the construction of
sentences. We now question, what form of grammar corresponds to this conception
of linguistic structure? We are going to get the answer in the following.

19.5.2 Phrase Structure Grammars

A phrase structure grammar is defined using a finite vocabulary (alphabet) Σ, a finite

set V of variables, a finite set P (rewrite rules) of productions of the form X → Y ,
where X is a string in V , and Y in V ∪ Σ. Hence, a grammar is G = (V, Σ, S, P),
where Σ is set of terminal symbols, which appear at the end of generation, S is
start symbol (for sentence). The corresponding language of G is L(G). Given this
grammar, we say that a string β ∈ (V ∪ Σ)∗ follows from α ∈ (V ∪ Σ)∗ , then we
write it as α → β. Also, if we are able to generate a sentence w from start symbol

Fig. 19.2 A sentence’s

the man took the book
structure
Verb NP
NP
VP

Sentence
614 19 Natural Language Processing

S, then we can write it as

S ⇒ α1 ⇒ α2 ⇒ · · · ⇒ αn = w. (19.1)

A derivation of a sentence is thus roughly analogous to a proof with V taken as

axiom system and P as the rules of inferences. We say that L is a derivable language
if L is the set of strings that are derivable from grammar G, and say that L is a
terminal language if it is the set of terminal strings from some system (V, Σ, S, P).
As a simple example of a system of this form, consider the following small part
of English grammar:

S → NP VP
V P → V er b N P
N P → the man | the book
V er b → took (19.2)

Among the derivations from (19.2) we have, in particular:

D1 : S → N P V P
→ N P V er b N P
→ the man V er b N P
→ the man took N P
→ the man took the book

Also, we have:

D2 : S → N P V P
→ N P V er b N P
→ the man V er b N P
→ the man V er b the book
→ the man took the book.

These derivations D1 and D2 are evidently equivalent; they differ only in the
order in which the rules are applied. We can represent this equivalence graphically
by constructing diagrams that correspond, in an obvious way, to derivations. Both
D1 and D2 reduce to the diagram as shown in Fig. 19.3. The diagram gives the phrase
structure of terminal sentence “the man took the book.” Note that “the man”, “took”,
“the book”, and “took the book” are phrases as these are traceable back to some
nodes. However, “man took” is not a phrase as it is not traceable back to any node.

Example 19.1 Consider that various tuples of a grammar are given as follows:
19.5 Grammars 615

Fig. 19.3 Phrase structure S

of “the man took the book”

NP VP

the man Verb NP

took the book

V = {S, N P, N , V P, V, Ar t}
Σ = {boy, icecr eam, dog, bites, likes, ate, the, a}
P = {S → N P V P,
N P → N,
N P → Ar t N ,
V P → V N P,
N → boy | icecr eam | dog,
V → ate | likes | bites,
Ar t → the | a}

Using the above grammar, we can generate the following sentences:

The dog bites boy.
Boy bites the dog.
Boy ate Icecream.
The dog bites the boy.
Note that, to generate any sentence, rules from the set P are applied sequentially
starting from the first rule. However, a grammar does not guarantee the generation
of meaningful sentences, but generate only those that are structurally correct as per
the rules of the grammar. 

Structural Representation
It is convenient to represent the sentences as tree or a graph to help expose the
structure of the constituent parts, called parts-of-speech (POS). For example, the
sentence, “The boy ate an Icecream” can be represented as a tree shown in Fig. 19.4.
For the purpose of computation, a tree must also be represented as a record, or a
list, or some similar data structure. For example, the tree above in Fig. 19.4 can be
represented as a list:

(S (NP ((Art the)

(N boy))
(VP (V ate) (NP (Art an) (N Icecream)))))
616 19 Natural Language Processing

Fig. 19.4 A Syntax Tree of S

“The boy ate an Icecream”

NP VP

Art N V NP

Art N

the boy ate an icecream

A more extensive English grammar can be obtained with the addition of other
constituencies such as prepositional phrases (PP), prepositions (Prep), adjectives
(ADJ), determiners (DET), adverbs (ADV), auxiliary verbs (AUX), and many other
features. Correspondingly, the additional rewrite rules are as follows:

P P → Pr ep N P
V P → V ADV | V P P | V N P P P | AU X V N P
Det → Ar t AD J | Ar t

These extensions allow the increase in complexity of the sentences generated

through this expanded grammar, along with its expression power, as in the following
sentences.
The(Art) cruel(Adj) man(N) locked(V) the(Art) dog(N) in(prep) the(Art) cage(N).
The(Art) laborious(Adj) man(N) worked(V) to(Aux) make(V) extra(Adj) money(N).

19.6 Classification of Grammars

The grammars are classified in many ways: when they are based on their generating
rules they are called Chomskian grammars; when a grammar transforms the passive-
voice sentences to active voice, it is transformational grammar; and when a sentence
generated by a grammar has more than one meaning(sense), it is called ambiguous
grammar. The following describes the theory of each in brief.

19.6.1 Chomsky Hierarchy of Grammars

In fact, it is not always possible to formally characterize the natural languages with
a simple grammar as discussed above. The grammars are defined by Chomsky Hier-
archy, as type 0, 1, 2, 3, with simplest as type-3 and most complex as type-0. The
19.6 Classification of Grammars 617

type-3, called Regular Grammars have rewrite rules as:

S → aS
S → a.

We note that the left-hand side is a variable symbol, and right hand is only terminal
symbol, or a terminal followed with variable symbol.
The type-2 grammars, called context-free grammars (CFG) have left-hand side
symbol of variable, and right-hand side has any combination of variables and terminal
symbols. Following all the productions belong to type-2 grammars, where S, A, B
are variable symbols; a, b are terminals, and α is any combination of variables and
terminals.

S → aS | aSb | a B | a AB | α
A→a
B→b

The typical rewrite rules for type-0, called unrestricted grammar and type-1 called
context-sensitive grammar (CSG), are are follows:

S → aS | a AB
AB → B A
a A → ab | aa
α→β

In the above rules, α ∈ (Σ ∪ V )∗ ∪ V , and β ∈ (Σ ∪ V )∗ . In all these rules,

left-hand sides are both variables and terminals and same applies for right-hand
symbols. The only difference between type-0 and type-1 grammars is that, if α → β
is production, then in type-1, |α| ≤ |β|.
Since, the grammars are called unrestricted, context-sensitive, context-free, and
regular grammars, the corresponding languages are also called unrestricted, context-
sensitive, context-free, and regular languages, respectively. The natural languages
are mostly based on the type-2 languages, as the type-0 and type-1 are not much
understood yet, and are difficult to implement.

19.6.2 Transformational Grammars

Transformational Grammar involves the use of defined operations called transforma-

tions, to produce new sentences from existing ones. The grammars discussed above
618 19 Natural Language Processing

produce different structures for different sentences, even though they have the same
meaning. For example,
Ram gave Shyam a book.
A book was given by Ram to Shyam.
In the above, the subject and object roles in second sentence are switched. In the
first, subject is “Ram” and object is “Book”, while in second sentence they are other
way round. This is an undesirable feature for machine processing of a language. In
fact, sentences having the same meaning should map to the same internal structures.
However, by adding some extra components, we can produce a single representa-
tion for sentences having the same meaning, through a series of transformations. This
extended grammar is called Transformational grammar. In addition, semantic and
phonological components are components, which are added as new, helps in inter-
preting the output of the syntactic components, as meaning and sound sequences.
These transformations are tree manipulation rules, which are taken from dictionary,
where words contain semantic featuring each of the lexica.
Using transformational generative grammar, a sentence is analyzed in two stages:
(1) basic structure of the sentence is analyzed to determine the grammatical con-
stitutional parts, which provides the structure of the sentence. (2) This structure is
transformed into another form, where a deeper semantic structure is determined.
The application of transformations is to produce a change from passive voice
form of the sentence into active voice, which changes a question to declarative form,
handle negations, and provide subject–verb agreement. Figure 19.5a–c shows the
three stages of conversion, from passive voice to active voice of a sentence.
However, the transformational grammars are now rarely used as computational
models for language processing [14].

Fig. 19.5 Transformational S

Grammar
VP
NP
V PP

Aux V Prep NP

Shyam is taught by Ram

(a)

S S

NP VP
NP VP
V NP Passive
V N

Ram teach shyam by Ram teaches Shyam

(b) (c)
19.6 Classification of Grammars 619

19.6.3 Ambiguous Grammars

An ambiguous grammar is a context-free grammar for which there exists a sen-

tence that can have more than one parse tree, while an unambiguous grammar is a
context-free grammar for which every sentence has a unique parse-tree. For com-
puter programming languages, the reference grammar is often ambiguous, due to
issues such as the dangling “Else” problem. If present, some of the ambiguities can
be resolved by adding precedence rules or other context-sensitive parsing rules, so
the overall phrase grammar is unambiguous. Note that, to find out if general grammar
is ambiguous, is a NP-hard problem [4].
The ambiguous grammars have more than one parse-tree for the same sentence.
For example, consider the sentence “He drove down the hill in the Car.” A process
for constructing a parse-tree is grouping the words to realize the structure in the
sentence. The parse-tree given in Fig. 19.6a for this sentence shows the grouping
of words and Fig. 19.6b shows the parse-tree for this sentence. Similarly, Fig. 19.7a

Fig. 19.6 a Grouping the words, b Parse-tree I: “He drove down the hill in the car”

Fig. 19.7 a Grouping the words, b Parse-tree II: “He drove down the hill in the car”
620 19 Natural Language Processing

shows another grouping for the same sentence and the corresponding parse-tree
is shown in Fig. 19.7b. Since this sentence can be generated through two different
parse-trees, the sentence is ambiguous, and hence its grammar is also ambiguous.

19.7 Prepositions in Applications

Prepositions, as well as prepositional phrases (PPs) are important to the precise

understanding of any language, which may be useful for certain applications, e.g., in
Information Extraction. However, they have been ignored on the grounds of being
syntactically promiscuous, and semantically having no significant meaning, hence
they were considered in the rank of “stop words” [20].
As a part of its argument structure, property of a preposition being subcategorized
or specified by the governor is called selection property. The governor is usually a
verb. An example of selection preposition is, “with”, like, in the phrase, dispense
with introductions. Here dispense is verb and introductions is the object of this verb.
But this object is realized in a prepositional phrase, which is headed by with. Con-
ventionally, the prepositions of selection type are uniquely specified. The following
are other examples with well-defined structures.
chuckleover/at
It is clear from the above examples the use of prepositions have no definite seman-
tics.
The selection proposition is useful in NLP applications; it operates at the syntax-
semantics interface, i.e., those application which openly translates surface strings
onto semantic representations, or vice versa. Examples of such translation are: Infor-
mation Extraction, and Machine Translation using some form of interlingua.
PP Attachment
To find a PP attachment is nothing but to find the governor for a given PP. Consider
the sentence,
Su jataeats 1 kheer withspoon.

This sentence1 consists a syntactic ambiguity, as in one case the PP “with spoon” is
governed by noun kheer (i.e., as part of the NP kheer with spoon (see Fig. 19.8a).
As an alternative sense the PP “with spoon”, is governed by the verb eats (i.e., as
a modifier of the verb, as indicated in Fig. 19.8b. Of these, the later case of verb
attachment (i.e., Fig. 19.8b) is, of course, the correct analysis.
The much interest in PP attachment comes from the PP being a common phe-
nomenon when parsing the languages such as English, which is a major cause of
parser errors. The ambiguity due to PP is called attachment ambiguity.
The syntactic preferences do not provide the solutions in dealing with PP attach-
ment, and they are not very effective for predicting the difference in PP attachment

1 Kheer: A sweet dish, like porridge, popular among the Indians.

19.7 Prepositions in Applications 621

S S

NP VP
NP VP
N V NP
N
Shakuntala eats V NP PP
N PP
Shakuntala eats NP
N P
kheer NP
P
kheer with N
with N
spoon
spoon

(a) (b)

Fig. 19.8 The prepositional phrase “with spoon,” being governed by: a noun kheer, b verb eats

between noun attachment and verb attachment. This difficulty led to a shift from
syntactic approaches in the 1980s toward AI-based approaches, the later used world
knowledge for resolving PP attachment ambiguity. Figure 19.8a shows that, due to
the availability of the knowledge that spoon is an eating implement, it would suggest
a preference for verb attachment. In a similar way, the knowledge that they are not
a foodstuff, it would suggest a preference against noun attachment.
An algorithm that resolves the attachment ambiguity, defines constraints that are
derived from semantic and world knowledge. These constraints are used for com-
posing the meaning of the child unit, which is attached to the meaning of each of the
possible parent’s syntactic units. This would be helpful for attaching the child unit to
the parent units (see Fig. 19.8). These constraints are called selectional constraints,
and usually represented in the form of the permissible range of fillers in the slots in
frames that represent the meanings. The potential filler, which is meaning of a child
unit, is compared against the constraint by a fuzzy match function. A popular way to
do this fuzzy match is to compute a weighted distance between the two meanings in
a semantic or ontological network of concepts. The closer the two are in the network,
higher is the score assigned to the particular choice.

19.8 Natural Language Parsing

Parsing is carried out to compute the structural description of a sentence. This struc-
tural description is assigned by the grammar of the language, and as a precondition
it is assumed that sentence is well-formed, i.e., grammatically correct. The process
of parsing consists of at least the following activities:
• Mathematical characterization of derivations in a grammar, and those associated
with parsing algorithms.
622 19 Natural Language Processing

• Computing the time and space complexities of parsing algorithms, with length of
input as the length of the input sentence.
• Comparing different grammar formalism and showing equivalences among them
whenever possible.
• Combining grammatical and statistical information for improving the efficiency
of parser, and ranking the parsers.
The structural descriptions provided by a grammar depend on the formalism to
which the grammar belongs. For the well-known context-free grammar (CFG), the
structural description is phrase structure grammar.

19.8.1 Parsing with CFGs

A parse-tree describe the structure of a sentence, and is also the record of the sequence
of steps of derivation of the sentence. The parse-trees are useful due to following
reasons:
• In carrying out semantic analysis, the parsing is an important intermediate stage,
• Grammatically checking the sentence, and
• The parsing is useful in following:
– Mechanical translation,
– Question answering, and
– Information Extraction.
A syntactic parser can be thought of as searching through the space of all possible
parse-trees to find the correct parse-tree for the given sentence. Before we go through
the steps of parsing, let us consider the following rules for grammar.

S → N P V P ;a start symbol for sentence can be replaced by

noun phrase followed by verb phrase
S → Aux N P V P ;Aux stands for auxiliary, e.g., do, does
S →VP
N P → Det N om ;Det is determiners, Nom is for Nomial
N om → N oun N om
N om → N
N P → pr oper -N ;for proper noun
VP →V
VP → V NP
V P → V P P P ;PP is prepositional phrase
P P → Pr ep N P ;Prep is preposition, PP is prepositional phrase
19.8 Natural Language Parsing 623

Fig. 19.9 Parsing tree of S

sentence: “Book that flight”
VP

V NP

det Nom

Book that flight

Det → a | an | the
N → book | f light | meal
V → book | include | pr oper
Aux → Does
pr ep → f r om | to | on
Pr oper -N → Mumbai
N om → N om P P

The parse-tree for the sentence “Book that flight” is shown in Fig. 19.9.

Example 19.2 Set of production rules (P) and parsed sentences for different forms
of verb-phrase.

S → N P V P ; I prefer a morning flight

V P → V N P ; prefer a morning flight
V P → V N P P P ; leaves Bombay in the morning
V P → V P P ; leaving on Tuesday
P P → pr ep N P ; from New Delhi.

A NP can be location, date, time, or others. Following are other examples of parts of
speech (POS).

N → f lights | br eeze | tri p | mor ning

V → is | pr e f er | like | need | want | f ly
Det → a | an | the | this | these | those
624 19 Natural Language Processing

Fig. 19.10 Parse-Tree of S

sentence: “I prefer morning
flight”
NP VP

Pronoun V PP

Prep NP

N
I prefer morning flight

Pr onoun → me | I | you | it
Pr oper -N → Mumbai | Delhi | I ndia | U S A
Ad j → cheapest | non-stop | f ir st | latest | other
Pr ep → f r om | to | on | near
Con j → and | or | but

The following examples show the substitution rules and with values for each parts-
of-speech to be substituted.

N P → Pr onoun(I ) | pr oper -N (Mumbai) | det N omial

(a flight) | N ( f light)
V P → V (do) | V N P(want a flight) | V N P P P
(leaves Delhi in Morning)
P P → Pr ep N P(from Delhi)

Making use of above rules, Fig. 19.10 demonstrates the parsing of sentence “I prefer
morning flight.”

19.8.2 Sentence-Level Constructions

A sentences can be classified as declarative, imperative, and pragmatic, as follows.

• Declarative Sentences. These sentences have structure: S → N P V P.

• Imperative Sentences: These sentences begin with “VP”. For example, “Show
the lowest fare,” “List all the scores,” etc. Following are the production rules for
imperative sentences.
19.8 Natural Language Parsing 625

S →VP
VP → V NP

The other substitutions for verb are already discussion in above.

• Pragmatic Sentences: The examples of pragmatic sentences are the following:

Can you give me the some information?

Do all these flights have stops?
What flights do you have from Delhi to Mumbai?
What Airlines fly from Delhi?

The pragmatic sentences are governed by substitution rule

S → Aux N P V P.

Corresponding to the word at begin of a sentence e.g., “What”, the production

rule is

W h-N P → W hat.

Hence, for the sentence,

“What flights do you have from Delhi to Mumbai?”,

can be generated by the production rule,

S → Wh-NP Aux NP VP.

Many times, the longer sentences are conjuncted together using connectives, e.g.,
I will fly to Delhi and Mumbai. The corresponding rule is

VP → VP and VP.

19.8.3 Top-Down Parsing

In top-down parsing, the searching is carried out from the root node, substitutions
are carried out at every step, and the progressing sentence is compared with the input
text sentence to determine whether the sentence generated progressively matches
with the original. Figure 19.11 demonstrates the steps of top-down parsing for the
sentence “Book that flight”.
To carry out the top-down parsing, we expand the tree at each level as shown in
the figure. At each level, the trees whose leaves fail to match the input sentence,
626 19 Natural Language Processing

Fig. 19.11 Top-down parsing for the sentence, “Book that flight”

are rejected, leaving behind the trees that represent the successful parses. Going this
way, we ultimately construct the original sentence: “Book that flight.”
Figure 19.11 shows that, first, root node is created as S, then in step (2), it is
expanded to generate three possible sub-trees,

S → N P V P,
S → Aux N P V P,
S → V P.

First of these is expanded at level 2, in step (3), by rule “NP → det Nom”, which,
we know will not match. Neither, when NP is expanded by NP → PropN in the next
sub-tree. So, we need to expand the next choice “S → Aux NP VP” of stage (2).
This also does not work. Finally, what works is as follows:

S →VP
VP → V NP
V → Book
N P → Det N
Det → that
N → f light.

And, ultimately generates the sentence, “Book that flight.”

The bottom-up parsing is other way round, starting from the sentence “Book that
flight,” and reducing it to start symbol S. Every time we perform a reduction, we
reduce the right-hand side of the production A → α, i.e., α by a variable A.
19.8 Natural Language Parsing 627

19.8.4 Probabilistic Parsing

In the discussion above, we drew a sharp line between a correct and an incorrect
parses, which was based on whether a terminal node either matched or did not match
the next word in the sentence. According to this parsing, a phrase is either acceptable
or not. There are situations under which we can relax these requirements, e.g., when
we cannot afford to try alternatives exhaustively. The examples are: analysis of con-
nected speech, text segmentation, and identification of words can never be done with
complete certainty. At best, one can say that certain sound is more probable than
the other. Consequently, one may associate a fractional number with each terminal
node, indicating the probability or quality of nodes below that, in respect of forming
grammatically correct sentence [5].
In the probabilistic parsing, the non-terminal nodes will be assigned some values,
which are sophisticated enough to realize that syntactic and semantic restrictions are
taken care of. Hence, a parser that aims to deliver the best analysis even if every
analysis violates some constraint, must associate a measure of being grammatical
(i.e., acceptable) with the analysis of portions of the sentence. The sentence ultimately
associates a measure with the analyses of the complete sentence. As a matter of rule,
it is possible to generate an analysis of every sentence with a nonzero acceptability
or matching probability, and then select the one having the best analysis.
One simple parser of this type is “best-first” parser, which is based on the modified
form of standard top-down serial parsing algorithm for context-free grammars. The
standard algorithm tries to generate one possible parse-tree until it gets stuck (i.e.,
on generation a terminal node which does not match the next word in the sentence).
In case of stuck, it “backs up” to try another alternative. A best-procedure, is like
a best-first search that tries all alternatives in parallel. A measure in the numerical
value is associated with each alternative path that indicates the likelihood, that this
analysis matches the sentence processed so far, and it can be extended to the complete
sentence analysis. At each progressive step, the path with the highest likelihood is
extended. In the process, if the “measure” of current path falls below that of some
other path, the parser shifts its attention to that of other paths.
A CFG (context-free grammar) consists of
terminal words w1 , w2 , ..., w V ,
non-terminals N 1 , N 2 , ..., N n ,
start symbol N 1 , and
production rules N i → α j ,
where α j is sequence of terminals and non-terminals. Given this, we can define a
generative PCFG (probabilistic context-free grammar) as
terminal words w1 , w2 , ..., w V ,
non-terminals N 1 , N 2 , ..., N n ,
start symbol N 1 , and
production rules N i → α j ,
628 19 Natural Language Processing

where α j is sequence of terminals and non-terminals. And can define the rule prob-
abilities as 
∀i P(N i → α j ) = 1, (19.3)
j

which shows that for each set of productions having same left-side variable (N i ), the
sum of probabilities is unity.
We consider that sentence is represented as sequence of words w1 w2 ...wm , and wab
= wa ...wb is a subsequence. Let non-terminal N i dominates subsequence wa ...wb is
represented by Nab i
, .i.e., Ni is root of sub-tree having children sequence as wa ...wb .
∗
Let N ⇒ α. We represent the probability of sentence w1n as
i


P(w1n ) = P(w1n,t ) (19.4)
t

where t is parse-tree of sentence w1n .

Example 19.3 Construct a correct parse-tree for the sentence “I saw an astronomer
with telescope,” for the following PCFG:

Rule Probability Rule Probability

S → NP VP 1.0 NP → NP PP 0.20
P P → Pr ep N P 1.0 NP → N 0.45
VP → V NP 0.7 N P → Ar t N 0.35
V P → V P PP 0.3 N → telescope 0.25
Pr ep → with 1.0 N → astr onomer 0.15
Ar t → an 1.0 N→I 0.60
V → saw 1.0

In this problem, the symbols are as follows:

• Terminals: I, saw, an, astronomer, with, telescope.
• Non-terminals: S, NP, VP, PP, V, Prep, Art, N
• Start Symbol: S
The sentence above can be generated using two different parse-trees, say T1 (see
Fig. 19.12) and t2 (see Fig. 19.13).
Probability for parse-tree t1 is the product of all the probabilities of rules used,
which starting from top, and going down through all the levels is given by

P(t1 ) =1.0 × 0.45 × 0.7 × 0.60 × 1.0 × 0.20 × 0.35 × 1.0 × 1.0 × 0.45 × 0.25
=0.001488375
19.8 Natural Language Parsing 629

Fig. 19.12 Probabilistic S

parse-tree t1 for the sentence
“I saw an astronomer with
telescope”
NP VP

N V NP

I saw NP PP

Art N P rep NP

an with N
astronomer
telescope

Fig. 19.13 Probabilistic S

parse-tree t2 for the sentence
“I saw an astronomer with
telescope” NP VP

N VP PP

I V NP P rep NP

saw Art N with N

an astronomer telescope

Similarly, the probability for parse-tree t2 is given by

P(t2 ) =1.0 × 0.45 × 0.3 × 0.60 × 1.0 × 0.35 × 1.0 × 0.45 × 1.0 × 0.15 × 0.25
=0.00047840625

Naturally, the parse-tree t1 is correct, as the probability of its construction is higher.

The tree t1 indicates that the observer (I) is seeing an astronomer carrying a telescope,
while t2 has semantics which indicates that observer (I) is seeing an astronomer with
the help of telescope, which obviously is incorrect.
However, the accuracy of the probability of each tree depends on the accuracy of
the probabilities of rules used. The probabilities associated with the rules are made
630 19 Natural Language Processing

available from the statistics of semantics from a large corpus of natural language
text. 
The probabilistic parsing provides a solution to ambiguity in language and gram-
mar, as it is conclusive from the above example. But, this is not necessarily complete,
but in partial. The other benefit is that it gives a probabilistic language model.

19.9 Information Extraction

We consider the following small text, which we would like to use for Information
Extraction.
Firm XYZ is a full service advertising agency specializing in direct and interactive market-
ing. Located in Bigtown CA, Firm XYZ is looking for an Assistant Account Manager to help
manage and coordinate interactive marketing initiatives for a marquee automotive account.
Experience in online marketing, automotive and/or the advertising field is a plus. Assistant
Account Manager responsibilities ensures smooth implementation of programs and initia-
tives, helps manage the delivery of projects and key client deliverables …Compensation:
50, 000 − 80, 000, Hiring Organization: Firm XYZ.

Given the above text, the extracted information may be

This information can be loaded into a database and can be queried any number of
times to find useful information, at a convenience [4].
The general architecture of an IE (Information Extraction) system is “a cascade
of transducers or modules such that, at each step structure is added to the document
and, sometimes, it filter relevant information by application of rules” (see Fig. 19.14).

INDUSTRY Advertising
POSITION Assistant Account Manager
LOCATION Bigtown, CA
COMPANY Firm XYZ
SALARY 50,000–80,000

The majority of current systems follow this general architecture, although specific
systems are characterized by their own set of modules. In general, the combination
of such modules allow of the functionalities for IE system, discussed below [21].

19.9.1 Document Preprocessing

Preprocessing of the documents can be achieved by a variety of modules such as:

text segmenters, filters, tokenizers, lexical analyzers, stemmers, and disambiguators
(POS taggers, semantic taggers, etc.)
19.9 Information Extraction 631

Text filtering

Lexicon Lexical analysis

Named entity recognition

Partial Parsing

Concepts
hierarchy Pattern extraction Pattern Base

Resolving co-references

Discourse Analysis

Generating output Output Format

Fig. 19.14 Typical architecture of an IE system

The most relevant document processing activity to IE is Named-Entity recogni-

tion, i.e., recognition of proper nouns. The process of named-entity recognition may
be performed using finite-state transducers (FST), together with dictionary lookup.
These dictionaries are domain-specific, or they are terminological databases.

19.9.2 Syntactic Parsing and Semantic Interpretation

For performing the Information Extraction (IE), a traditional architecture based on

Natural Language understanding was used. This method needed full parsing of sen-
tences, and then followed by a semantic interpretation of syntactic structure obtained
through parsing. The discourse analysis was carried out as the final step.
A new approach to IE is radically different from traditional, where only the con-
cepts that are within the scope of extraction, are required to be found out in the
documents. This leads to simplification of syntactic and semantic analysis due to
more restricted, deterministic, and collaborative process. The new approach has the
strategy that replaces the traditional parsing, interpretation, and discourse (module)
632 19 Natural Language Processing

with a simple phrasal parser (parsing based on phrases). The later finds the local
phrases. In addition, the discourse module does the job of event pattern matching,
and merging the of templates. The above approach is useful because, the full pars-
ing is expensive, not robust, hence produces ambiguous results, and cannot manage
off-vocabulary conditions.
The current IE systems make use of partial parsing, such that the process of finding
the constituent partial structure of a sentence consists one or more cascaded steps
onto the text fragments. The generated constituents are tagged as parts-of-speech, like
noun/verb, or phrases like prepositional or others. After parsing of the constituents,
the system resolves domain-specific dependencies based on the semantic restrictions
imposed by the extraction environment. The dependencies are resolved using the
following two alternatives.
Pattern Matching
Majority of IE systems follow the approach based on pattern matching, called named
extraction patterns or IE rules, to resolve the dependencies. In this approach, simpli-
fication of the syntax helps in reducing the semantic processing, and that leads ulti-
mately to simplify the pattern matching task. The use of patterns is scenario-specific
to recognize both, modifier and argument dependencies between constituents.
Grammatical relations
For representation of the grammatical relations, a graph is constructed (similar to
dependency grammars) using the general rules of interpretation, such that previously
detected chunks are constructed as nodes, and relations among these nodes are the
labeled arcs. Such graphs are useful for reasoning about IE. There are three different
types of grammatical relations for IE, as follows.
• First type is defined in the form of general relations of subject, object, and modifier.
• Second type is specialized modifier relations, which are specified as temporal
relations, and relations concerning to locations.
• Third type is mediated relations, i.e., they are mediated by prepositional phrases.

19.9.3 Discourse Analysis

The task of the majority of IE systems is to represent the extracted information

from the text in the form of partially filled templates or in some logical forms. Once
represented/stored in a specified format, this information can be queried later, like we
query a database. Since such information may be incomplete some times, it will result
in partially incomplete templates. To recover the missing parts of the information to
some extent, merging procedures are used to merge the partial templates, to explore
the recovery of missing information.
However, when working with logical forms the IE systems can use traditional
semantic interpretation methods in this phase.
19.9 Information Extraction 633

19.9.4 Output Template Generation

The output template generation phase is the final stage of IE systems. This stage
maps the extracted pieces of information to the required output format. Due to the
domain-specific restrictions in the output structure, some inferences can be drawn in
this phase. The inference drawing can happen in the following situations:
1. a value from a predefined set is taken as it is by the output slots;
2. a forced instantiation of output slots;
3. when an extracted information generates multiple output templates, it results in
inference;
4. some times, the normalization of the output slots produces inferences, e. g., when
date in the coded form is normalized, it produces, say day or day of month or
week, etc. Similarly, when products list is normalized we may get name of items
instead of its code, etc.

19.10 NL-Question Answering

The Natural Language Question Answering (QA) derives the common features from
NLP (Natural Language Processing) and Information Retrieval (IR). The QA (i.e.,
NL-QA) promises to deliver “answers”, unlike the “hits” by IR. Most of QA is
focused questions based on facts like “When did first human land on Moon?”, “Who
invented the paper clip?”, and “How far is the Moon from earth?” These are called
“factoid” questions, which can be typically answered using the named entities such
as people, organizations, measures, dates, locations, etc. [7].
The commonly used approach for answering the factoid questions of open-domain
requires the technique of named-entity recognition, together with IR. Typically, major
steps of a “traditional” ontology-driven question answering system that is primarily
based on IR and named-entity recognition technologies, are shown in Fig. 19.15.
Usually, the large ontologies drive the process of QA, which relates the question
types to semantic classes of answers. The detailed description of steps for QA is as
follows:
1. Question Analysis. The first step of QA focus is on the analysis of the question,
where knowledge resources are explored to find out the expected semantic type
of the answer.
2. Document retrieval. Next, in the document retrieval phase, the candidates’ docu-
ments containing the terms from the questions are retrieved. This retrieval is often
done using off-the-shelf IR engines.
3. Identification of passages. In the next stage, the system identifies passages within
these restricted candidate documents that contain a concentration of terms from
the question–these regions are likely to contain the answer.
4. Answer extraction. Finally, in the answer extraction stage, named-entity recog-
nizers identify candidates of the correct semantic type [9].
634 19 Natural Language Processing

Fig. 19.15 Major steps of a Question

QA system

Question Analysis

Document Retrieval

Passage Retrieval

Answer Extraction

However, there are challenges of the QA approach discussed above. One is due
to complex many-to-many mapping between question types and answer types. As
an example, the answer to a “who” question can be name of a person (e.g., “Who
invented the light bulb?”) or an organization’s name (e.g., “Who won the World Series
in 2004?”), and there are other possibilities also. On the other side, different question
types may map onto the single semantic answer type. For example, the questions –
“Where was the world book fair held in 1900?” and “What city hosted the world
book fair in 1900?”, would map to the same answer. To overcome these challenges
of multiple questions with single answer, a system must have elaborate ontological
resources which explicitly encode semantic relationships between questions and
answers. More advanced techniques such as abductive inferencing, feedback loops,
and fuzzy matching of syntactic relations are used, but the factoid question answering
is still driven by information retrieval and named-entity recognition technologies, and
only on large ontologies.

19.10.1 Data Redundancy Based Approach

The data redundancy based approach depends on statistical regularities, using which
it is possible to extract “easy” answers to factoid questions from web. For answering
a question, the system makes some connection between the question and the passages
containing the answers. Having done this connection at the lexical level, the rest of
the job is simple. The meaning of the lexical connection between the question and
likely answer carrying passage is the presence of large degree of overlap between
these texts. But, in reality, it is not the case very often. This is because, the richness
of natural language, and its expressive power provides the facility to create varying
19.10 NL-Question Answering 635

textual forms which have the same meaning (semantics). But, in that case also, there
is a good degree of semantics resemblance between the question and answer carrying
text. To appreciate this scenario, consider the following question, with two different
answers, which though lexically different but have the same semantics.
When did Sikkim become the state of India?
1. Sikkim became a state on 16 May 1975.
2. Sikkim was admitted to India on May 16, 1975.
In the above cases, both the passages contain correct answers, however, it seems
obvious that for computer, it would be easier to extract the answer from passage 1
than from passage 2. We note that the task of answering the factoid questions will
be far easier for the computer if the answer passage carries the same words as that
in the question.
From the point of view of text processing algorithms, answers can be expressed
in a variety of different ways. Within text collection, it is possible that the answer
to above question lies in passage 2. Though in that text, the answer is not obviously
stated. In such cases, since the answers share few words common with the question,
a sophisticated natural language processing may be required to find out the rela-
tion between them. These processing may typically comprise the following types of
domains [7].
Recognizing syntactic alternations,
Collapsing of paraphrases,
Resolving the anaphora,2
Making commonsense inferences, and
Performing date calculations.

19.10.2 Structured Descriptive Grammar-Based QA

A special grammar, called SDG (Structured Description Grammar) can be used for
information representation and extraction. The basic approach here is data redun-
dancy based. In this, a text sentence (S) is mapped into a transition graph or state
diagram, as shown in Fig. 19.16. It shows seven types of structurally different sen-
tences labeled as 1, 2, …, 7, based on the number of wh-pronouns and their positions,
which can be mapped into this structure. The structure of each sentence explicitly
indicates the position of interrogating wh-pronouns, who, what, where, when and
why. The positions of these wh-pronouns are invariant [3].
Following examples demonstrate that English-language sentences structures map
into the transition diagram of SDG. It can be easily verified that the sentences (a) to
(d) are of types 7, 7, 3, 4 respectively.

2 Anaphora: making use of a pronoun or similar word instead of repeating a word used earlier.
636 19 Natural Language Processing

S 1
Who What Where When Why

Sentence types
2
3

When Where Why 4

5
6
7

Fig. 19.16 Transition diagram of sentence structure

(a) Akbar | followed a liberal policy for religion.

Who | What
(b) Jahangir | Married Nur Jahan.
Who | What
(c) Red fort | is located | at Delhi.
Who | What | Where
(d) Shah Jahan | built Taj Mahal | during his rule | at Agra | in the memory of queen
Mumtaj.
Who | What | When | Where | Why
It may be noted that the structures of the sentences in SDG are simple English-
language sentences. Further, in SDG, i) a question is allowed to carry more than
one subquestion, as a sentence has more than one position for wh-pronouns, ii) sub-
answers for a question which are distributed in a single document can be aggregated
to fill up the slot-like structures in the transition diagram to generate a single answer,
and iii) sub-answers for a question which are distributed in multiple passages can also
be aggregated to fill up the slot-like structures in the transition diagram to generate
a single answer.

19.11 Commonsense-Based Interfaces

To make our computers easier to use, it is required to make them understand the
meaning of what we tell them. Usually these efforts are failed because meaning is
not one thing but a combination of many. This is because, the activities of human
thought engage an enormous collection of different structures and processes [10].
In human understanding, what some thing X means to us depends on how rep-
resentations of X (in our brain) connects to, that is, having the links to other things
we know. If we understand something in only one way then we know it very little,
because when something goes wrong, we are left no links to connect that object
(sense) with others. But if we use several features, each integrated with its set of
related pieces of knowledge, then if one of them fails we can switch to another. This
we implement by turning our ideas around in our mind to examine the other pieces
19.11 Commonsense-Based Interfaces 637

of knowledge from different perspectives, and carry on this process until we find one
that works.
When our goal is that our computers must understand us, we will need to equip
them with this kind of facility, i.e., they should connect each object/action (sense) to
many other senses. For example, the sense “cat” is connected with other senses: color,
size, height, no. of legs, tails, eyes, ears, leopard shape, etc. Those ways computers
must be equipped with knowledge, like human.
Present Limitations of Computers
For computers to understand like humans, there should be a program with common
sense like humans. This is a difficult task, because a typical program has only one
way to deal with a problem, so if something goes wrong, the program is totally
stuck. However, as humans, we search for alternatives in the event of failure of one
approach. The limitation with present-day computers is that they always start from
scratch. To make the computers of more worth, we need to supply them a great library
of commonsense knowledge, which are common even with young children.
The present-day computers are not designed with commonsense in them, nor they
have capability to learn from experience. The programs can solve difficult problems
in specialized subjects, but even today, computers cannot do the things, which even
young children do so easily. The current programs behaves in a limited and inflexible
way. Some of the critical differences between the capabilities of computers and
human are distinctly explained below.
Vitalist
Computers can do only what they are programmed to do so. The computers have
been programmed by people to speak, but in fact the computers do not know what
those words mean. The meaning is an intuitive thing, and it cannot be reduced to
zeros and ones. The present computers can only do logical things, and meanings are
not necessarily logical.
Humanist
The machines are not humanist, as machines have no goals or hopes, nor any fears or
phobias. They do not even know that exist, nor they have a sense of accomplishment.
Logician
Since we want that computers should have commonsense, it is more important to
understand as what the commonsense is? We need to define it more clearly and
precisely. That policy seems alright, but it is wrong when it comes to psychology.
Of course, we do not like to be imprecise, but some times, the definitions are not
sufficient. So instead, we will take a different approach, and try to design (as opposed
to defining) machines that can do the things we want.
638 19 Natural Language Processing

19.11.1 Commonsense Thinking

When we say “commonsense thought,” we are indicating to things that most people
can accomplish, even not knowing many times that they are doing them. Thus, when
we hear a sentence like: “Bill told the waiter he wanted some chips,” we will infer
all possible inferences. Some of them are [10] the following:
– The word “he” means Bill, and not the waiter.
– This event possibly took place in a restaurant, where Bill was a customer, and
waiter was close to him at the time. The waiter was working there, waiting for
Bill’s meal order at that time.
– Bill wants potato chips, not memory or IC chips. No count of chips is mentioned.
But, it may be 20–30, and in thousands.
– Bill will start eating the chips soon after he gets them.
– Bill and waiter are both human beings, Bill is old enough to talk, and the waiter is
old enough to work.
– This event happened after the date of invention of potato chips (i.e., year 1853).
– It is Bill’s assumption that waiter also infers all those the abovementioned things.
Every child learns to use thousands of words at a very young age, but no computer
knows even the meaning of some of those, so no computer can even understand a
casual conversation. For example, if you mention about “string,” any child would
know what it means, because it knows that many places are there it can be used, for
example, it can be used to tie a cart and pull it. With this, the child would also know
things like these:
– An object can be pulled by a string, and cannot be pushed by it.
– It is not good for eating, and you cannot construct a cart using a string.
– Before we put a string into a box, the box needs to be opened.
– And so, on.
Let us find out the size of networks of commonsense knowledge, i.e., network of, say
1000 words, each word having links with other knowledge structures, in different
ways. Some links are for objects, while others are for processes. Each such link will,
in turn, lead to other links (all in a semantic network), so that whenever some process
gets stuck, we usually can find some alternative. Language is not the only feature with
such abilities, but each expert skill a person possesses, there must be other similar
order of structures, say, for vision, for hearing, for perception, for speech, and for all
kinds of physical manipulations, and also for various kinds of social knowledge. So,
it may be that we possess millions/billions of units of knowledge.

19.11.2 Components of Commonsense Reasoning

The commonsense thinking we do every day involves a large number of hard-earned

ideas, which includes masses of factual knowledge about the problems we are trying
19.11 Commonsense-Based Interfaces 639

to solve. Not only the learning, but we must also adopt efficient ways to retrieve
and apply the relevant knowledge. Many processes gets engaged in the activities of
imagining, planing, predicting, and deciding, which also make use of many excep-
tions and rules. For these, it requires knowledge about how to think, i.e., how to
organize and control those processes, also even if the representations are different,
it can describe the same situation. In addition, there must be a facility to convert the
new experiences into suitable memory structures to represent them. Some of the pos-
sible structures are: property lists, frames, frame-arrays, database query languages,
explanation-based reasoning, logic programming, rule-based systems, semantic net-
works, scripts, and stories.
The first step in knowledge representation is to select a representation. Using any spe-
cific representation will shortly lead to limitations, which may quickly accumulate,
leading the reasoning to ultimately halt. It is usually required to use several different
representations for each fragment of the commonsense knowledge about each idea,
thing, or relationship. We swiftly change from one method of representation to other
methods. And, that depends on which methods are better than other methods for
solving problems.
Finally, the following three are the important capabilities we need for character-
izing the problems for making use of commonsense reasoning.
Negative Expertise
This deals with ways through which we recognize each of the methods as when they
are going to fail. If it becomes possible to recognize the way particular things went
wrong, it becomes a clue for deciding what action should be taken next. By knowing
the way how each of the methods is likely to fail, it can be used to control the higher
activity, e.g., by brain we control the mental activity.
Knowledge Retrieval
Retrieving the relevant information from the commonsense knowledge networks
requires appropriate methods to identify, as what problems or situations in the past
are having maximum resemblance to the context of present problem. This means,
the systems need ways to describe what we are trying to do, and then reason about
those descriptions. This is based on the skillful use of analogies.
Self-reflection
Our computers, when implementing commonsense reasoning, are required to keep
records which describe acts and thinking they have recently done, so that they are
able to reflect on the results of what they tried to do. Also, they must be aware of
what they are doing. This is what we call with human as consciousness.
How the programs like playing chess or proving theorem are different from com-
monsense reasoning? Is the reasoning process in a child’s brain for fixing blocks
of different colors is a game, simpler than that of chess or theorem proving? The
answer is No. There is a fundamental difference between these programs, like for
chess game or theorem proving versus playing these block games. In fact, the pro-
grams of expert systems, like chess or theorem prover, require much less knowledge
640 19 Natural Language Processing

skills. However, the children make use of thousands of skills, they manipulate all
kinds of things, irrespective of their texture and shapes; they stack the things up and
then knock down then, and learn the dynamics of how the stack got scattered. Even
for building a small toy house of blocks, one needs to mix and match the knowledge
of many kinds: that is, about shapes and colors, speed, space and time, support and
balance, stress and strain, and the economics of self-management.
Do computers lack in learning all these things? Not likely. These limitations persist
because we have learned to program only in certain ways. Our decades-old approach
to solve some problem X has been: find the best way to represent X , find the best
way to represent the knowledge needed to solve X , and find the best way to solve
X , and so on! These steps lead to write specialized programs that cope with solving
only that type of problems—called brittle programs! This has resulted to millions of
specialized programs for solving only that kind of problems, such as playing chess,
playing cards, or designing a bridge, banking, etc. In the computer programs, price
for best was sacrificed in limiting the resources.
To make the machines deal with commonsense reasoning, we must use multiple
ways to represent the knowledge, acquire huge knowledge in that, and find common-
sense ways to reason with this. Consider the following example.
Mary was invited to Jack’s party.
She wondered if he would like a kite.
What leads us to infer that Mary was thinking to take a kite as a gift, but there
was no mention of “birthday” or “present” in the first sentence. There should be a
suitable representation of phrase “invited to party”, which could help to infer that
she is thinking of a kite as a suitable gift for Jack.

19.11.3 Representation Structures

It is not well known as which structure of representation is best suitable for any
specific purpose. On the contrary, the brain represents a problem in several ways,
as well as represents the required knowledge. When one method fails to solve a
problem, we quickly switch over to another approach.
Now, to reply to the above question, we do a small analysis. Consider that we
are traveling by electric train, and there is a possibility of it being halted in any of
the stations due to electricity failure. Assuming that there are 10 railway stations,
and it may stuck at any one of them. Therefore, it is not difficult to plan for food
and shelter, as there can be a maximum 10 different solutions. So, given a cause of
halting of train, there are a maximum of 10 possible effects.
Consider another example, a possibility of meteorite strike on earth of sufficient
size, and depending on where it strikes, there are different rescue operations and
strategies needed depending on where it strikes. Naturally, for a single cause, there
are hundreds or thousands of possible different measures needed(effects), and even
more. Naturally, this second problem is more difficult to solve.
19.11 Commonsense-Based Interfaces 641

Fig. 19.17 Causal-diversity

matrix

The cause-effect matrix (shown in Fig. 19.17) illustrates the relation “cause ×
e f f ect” to reasoning process, which can be used to arrive at the solution. In the
cases, when the causes are only a few, and each cause having a small number of
effects as shown in the top-left corner cells of the matrix, such the problems are easy
to solve even by exhaustive search methods, or some times there is no need to search,
but simply answer is recalled [10].
When the causes are many, each with a small effect, then statistical methods and
neural networks may work well, as indicated by the top-right corner cell of the matrix.
But, such systems are likely to breakdown if those causes have different characters
that interact in hard-to-predict ways.
The symbolic or logical reasoning, shown at the bottom-left corner cell of the
matrix, can work well when the causes are few, but corresponding effects are in large
number, except that the search may exponential in the case of sequential processes.
It is a rare situation, when many causes correspond to a large number of effects,
indicated in the right-bottom of the matrix, such systems are interactable, unless
the system is linear. Sometimes it is possible to arrive at solutions by reformulating
those difficult problems by switching to different representations, that emphasize
fewer and more pertinent features, so that we can work with simpler descriptions.
Except in the right corners of the matrix, there are multiple causes with mod-
est number of effects, the heuristic programs succeed in these conditions, using
knowledge combined with controllable search. This is the region of “classical AI”
research—the causes and effects are both moderate. Consider the adjacent cells to
the bottom and to the right of classical AI. In these regions, analytical methods are
usually not helpful, but it may be possible to use use our knowledge and experi-
ence to construct and apply analogies, plans, or case-based reasoning methods. Such
methods may not work perfectly, but they are frequently useful for practical purposes.
The conclusion of the above discussion is that there cannot be a single type of
approach possible for all types of problems. Accordingly, one should not seek a
uniform way to represent commonsense knowledge. In fact, there is frequent need to
use several representations when we face a difficult problem and then we will need
additional knowledge about it. The causal-diversity method may help, but eventually
642 19 Natural Language Processing

Fig. 19.18 Architecture of

representations Story-like scripts
Frame arrays,
Picture-frames
Semantic Nets
Neural Networks

it must be replaced by more resourceful, knowledge based systems that can generate
new representations.
In fact, there is no best way to represent the knowledge. The present limitations of
machine intelligence are largely due to our quest for seeking unified theories capable
of reasoning well in all situations. The versatility human can emerge from a large-
scale architecture of representation, where each of several different representations
can help overcome the deficiencies of other representations.
Consequently, the commonsense knowledge in the machines be built, which rep-
resents knowledge about so many things like strings, houses, clothing, roads, books;
in other words, everything that most children know. For such commonsense knowl-
edge base, we will need ways to link each unit of knowledge to the use, or functions
that each unit knowledge can serve. Figure 19.18 shows some typical representation
levels, which have been covered in previous chapters.

19.12 Tools for NLP

There are a number of tools that exist, either as open-source or research tools created
by some research laboratories, as well as closed source for natural language process-
ing, and speech processing. These tools have built-in functions to perform a number
of commonly used tasks, which can be directly called, or using these scripts can be
written to perform more complex jobs of NLP and speech processing. For example,
for NLP, they can tokenize the given text, can do stemming and POS (parts of speech
tagging), can find out word frequencies in given documents, parsing of NL sentences,
etc. These inputs can help to compute, for example, t f × id f , which can be helpful
in IR (information retrieval), IE (information extraction), text classification, etc. In
the following part, we discuss some such tools, which are either open source or they
can be obtained on request from respective research laboratories.

19.12.1 NLTK

The NLTK (Natural Language Toolkit) is a collection of Python libraries and pro-
grams for symbolic and statistical natural language processing. It is suited to practi-
tioners as well those who are learning natural language processing (NLP), those who
19.12 Tools for NLP 643

are conducting research in NLP or areas close to NLP, like, empirical linguistics,
cognitive science, AI, IR, and machine learning.
The NLTK has been also used as a teaching tool, as a study tool for individuals,
and as a prototyping and building platform for research systems. Python language has
been chosen due to its shallow learning curve, its transparent syntax and semantics,
and due to its extraordinary capability for handling strings. Python is an interpreted
language, which facilitates interactive exploration. It is an object-oriented language,
allows data and methods to be encapsulated and reused easily. Python is also available
with an extensive standard library, that includes tools for speech processing, natural
language processing, numerical processing, and graphical programming [12].
Consider that tasks of stemming and parts-of-speech (POS) tagging are indepen-
dent, and both operate on sequences of tokens. If the stemming task is done first,
the information required for POS tagging is lost. If tagging task is performed first,
the stemming process must be able to skip over the tags. If these two tasks are done
independent of each other, it becomes difficult to align the resultant texts. Hence, as
the combinations of tasks increase, it becomes extremely difficult to manage the data.
To address this problem, NLTK version 1.4 onward comes with a new architecture
where tokens are based on Python’s native dictionary datatype, such that the tokens
can have an arbitrary number of named properties. The Tag and Stem are examples
of these properties. The NLTK allows for even whole sentence and document to
be represented as a single token with Sub-tokens attribute that holds sequences of
smaller tokens.
A parse-tree can also be treated as a token, which has special property/attribute
of Children. The benefit of this type of architecture in NLTK is that it unifies many
different data types, and allows distinct tasks to be run independently. Of course, this
architecture comes with an overhead for programmers, because the program needs
to keep track of a growing number of property names.

19.12.2 NLTK Examples

Example 19.4 Tokenization of natural language text.

The following commands in Python, with NLP tool NLTK installed, demonstrate
the tokenization of a given text into sentence tokens and word tokens. Since there is
only one sentence, the sentence token is one only.
$ python
Python 2.7.14 (default, Sep 23 2017, 22:06:14)
[GCC 7.2.0] on linux2
Type "help", "copyright", "credits" or "license" for more
information.
>>> from nltk.tokenize import sent_tokenize, word_tokenize
>>> text="Fundamentals of Artificial Intelligence."
>>> print(sent_tokenize(text))
644 19 Natural Language Processing

[’Fundamentals of Artificial Intelligence.’]

>>> print(word_tokenize(text))
[’Fundamentals’, ’of’, ’Artificial’, ’Intelligence’, ’.’]
>>>


Example 19.5 Stemming of given set of words.
Following are the commands for stemming a set of words to reduce them to their
stem words. This makes use of the Porter Stemmer algorithm.

>>> from nltk.stem import PorterStemmer

>>> ps=PorterStemmer()
>>> words=["python", "pythoning", "pythonize", "pythonly"]
>>> for w in words: print(ps.stem(w))
...
python
python
python
pythonli
>>>


The parts-of-speech tagging (grammatical tagging) or disambiguation of word-
category, is a process of marking-up word in a text (corpus) corresponding to a
particular POS. This is carried out based on its definition as well as its context.3
Various POS in the English language are: noun, verb, adjective, adverb, pronoun,
preposition, conjunction, and interjection.
The POS tagging is carried out as part of computational linguistics using some
algorithms. These algorithms associate discrete terms, as well as hidden parts of
speech, in accordance to a set of descriptive tags. POS-tagging algorithms fall into
two distinctive categories: rule-based and stochastic based. For example, Brill’s
tagger, one of the first and most widely used English POS taggers, makes use of
rule-based algorithms. The following example demonstrates the POS tagging using
NLTK.
Example 19.6 Parts-of-speech tagging.

>>> import nltk

>>> text=nltk.word_tokenize("Part of speech tagging and POS
tagger")

3 Context: Relationship with adjacent and related words in a sentence, or phrase, or a paragraph.
19.12 Tools for NLP 645

>>> text
[’Part’, ’of’, ’speech’, ’tagging’, ’and’, ’POS’, ’tagger’]
>>> nltk.pos_tag(text)
[(’part’, ’NN’), (’of’, ’IN’), (’speech’, ’NN’),
(’tagging’, ’NN’), (’and’, ’CC’), (’POS’, ’NNP’), (’tagger’,
’NN’)]

19.13 Summary

Natural language processing (NLP) is an academic, and technology-based research

domain comprising a range of computational techniques for representation and auto-
matic analysis of human languages—a field that is motivated by theory. Automatic
analysis of text requires a deep understanding of natural language by machines.
However, we are still far away from machines that have this capability. To develop a
program that can understand that natural language is a challenge, because most of the
natural languages are large and complex, which can have infinitely large number of
sentences. In addition, there is ambiguity in natural languages, as many words have
more than one meaning, such as can, bear, fly, orange, and many others. That gives
different meanings to the same sentence, when used in different contexts. Due to this
problem, creating programs that understand the NL correctly is a difficult task.
Right from its start, the field of NLP focused on areas like machine translation, IR,
IE, question answering, text summarization, topic modeling, and more recently, on
opinion mining. Although the semantic problems and the actual requirements of NLP
were apparent right from the beginning. However, for the more direct applicability
of machine learning techniques, the strategy adopted was to tackle syntax first. This
was due to the fact that semantics was considered a more challenging problem.
One of the early representation strategies in NL was FOPL (First-Order Predicate
Logic)—a deductive system comprising axioms and rules of inferences. The FOPL
supported to good degree of handling the syntax, but limited order of semantics and
pragmatics. The syntax is concerned with the well-formedness of the expressions,
while semantics specifies what the well-formed expressions mean. The handling of
pragmatics is more challenging, they specify how the contextual information can be
used to provide a better correlation between different semantics. Properly specifying
of pragmatics is essential for the tasks such as WSD (Word-Sens Disambiguation).
Production rule based language model is one of the popular models for the descrip-
tion of Natural language (Chomsky, 1956). Other important models for NLP are
Ontology Web Language (OWL), which uses XML-based vocabulary for ontology
representation, e.g., definition of classes and relation between them, classes’ proper-
ties, and constraints on the properties, and constraints on the relations. Network-based
models are also common for NLP, like Bayesian networks, which provides joint prob-
ability distribution over many interrelated hypotheses, and semantic networks—a
646 19 Natural Language Processing

graphical notation for representing knowledge in patterns of interrelated intercon-

nected nodes and arcs.
Some of the common applications of NLP are: Classification of text into cate-
gories, Index and search large texts, Automatic translation, Information extraction,
Automatic summarization, Question answering, Knowledge acquisition, and Text
generations/dialogues.
Syntax of natural language is checked by generating using its grammar, called
phrase structure grammar. The generating process is a derivation, similar to rigorous
proof. Different grammars are classified as per Chomsky hierarchy of grammars,
called type 0, 1, 2, and 3, with last one most simple, and first as the most complex.
A sentence of natural language is represented by a syntax-tree, which shows the
relations between various constituents of the sentence. Every sentence is supposed
to have a corresponding one and only one syntax tree. If a sentence can be generated
using more than one syntax-tree, then the language of that sentence as well the
grammar are said to be ambiguous—having more than one meaning of the sentence.
Parsing a sentence is the process of computing the structural description of the
sentence assigned by the grammar, assuming that the sentence is well-formed. Pars-
ing consists of : Mathematical characterization of derivations in the grammar using
a specified algorithm. A parse-tree describe the structure of a sentence, and is also
the record of the history of derivation of the sentence. The parse-trees are useful
for grammar checking of the sentence, which is an important intermediate stage in
semantic analysis, and it plays an important role in applications of language transla-
tion, question answering, and information extraction.
Parsing can be top-down—a sentence is generated using start symbol, or it can
be bottom-up—a sentence is reduced to start symbol through a sequence of steps.
Further, the parsing can be deterministic—each production rule has equal weight, in
comparison to probabilistic parsing—there is a probability weight associated with
each production rule. Usually, in parsing, a sentence is split into NP (noun phrase),
and VP (verb phrase). More extensive English grammar is obtained by with the
addition of other constituents, such as PP (prepositional phrase), ADJ (adjectives),
DET (determiners), ADV (adverb), AUX (auxiliary verb), etc.
Information extraction (IE) systems extract the important information from natural
text, and load into databases, that can be queried later a number of times to find useful
information, and at convenience.
The general architecture of an IE system is defined as “a cascade of transducers
or modules that, at each step, add structure to the documents and, sometimes, filter
relevant information, by means of applying rules.” Most existing IE systems are based
on partial parsing. Generally, the process of finding constituents consists of using a
cascade of one or more parsing steps against fragments. The resulting constituents
are tagged as noun, verb, prepositional phrases, etc.
Syntactical parsing for IE defines some grammatical relations (subject, object, and
modifier), and some specialized modifier relations (temporal, and location). This is
made possible using a dependency graph built following general rules of interpre-
tation for the grammatical relations, with previously detected chunks as nodes, and
relations. The other jobs performed by IE are: discourse analysis, output template
generation, and NL-language question answering.
19.13 Summary 647

Question-answering is possible using a special format of grammar called struc-

tured descriptive grammar. It is possible to map a sentence with a sequence of wh-
pronouns (who, what, when,...), and consequently, determine the value of a missing
wh-pronoun, thus, answering a question.
Preprocessing, is a common phase in NLP applications discussed above. It is car-
ried out on documents using the program modules such as: text zoners, segmenters,
filters, tokenizers, lexical analyzers, disambiguators (POS taggers, and semantic tag-
gers), stemmers, etc.
To make our computers to understand us, we need to equip them with adequate
knowledge. To help this work, the computer must know what our jobs are. To enter-
tain us they will need to know what their audiences like or need. This requires to
create commonsense reasoning in computers. However, it is not well known as which
structure of representation is best suitable for any specific purpose. On the contrary,
the brain represents a problem in several ways, as well as represents the required
knowledge. When one method fails to solve a problem, we quickly switch over to
another approach. That means, commonsense reasoning is that we should not seek
one uniform way to represent commonsense knowledge. We will frequently need to
use several representations when we face a difficult problem and then we will need
additional knowledge about it.

Exercises

1. What are the challenges of NLP?

2. Give one example of the following ambiguities:
a. Phonetic
b. Syntactic
c. Pragmatic
3. What are the applications of NLP?
4. Develop the parse-tree to generate the sentence “Rajan slept on the bench” using
following rewrite rules:
5. Draw the tree for the following phrases:
a. after 5 pm.
b. on Tuesday.
c. From Delhi.
d. Any delay at Mumbai.
6. Draw the tree structures for the following sentences:
a. I would like to fly on Air India.
b. I need to fly between Delhi and Mumbai.
c. Please repeat again.
648 19 Natural Language Processing

Fig. 19.19 Parse-tree

7. Convert the following passive voice to active voice. Construct the necessary
trees. Also write the steps.
The village was looted by dacoits.

S → NP VP
NP → N
N P → Det N
V P → V PP
P P → Pr ep N P
N → Ra jan | bench
Det → the
pr ep → on

8. Given the parse-tree in Fig. 19.19, construct the grammar for this.
9. Construct the grammars and parse-tree for the following sentences.
a. The boy who was sleeping was awakened.
b. The boy who was sleeping on the table was awakened.
c. Jack slept on the table.
10. Construct the parse-trees and resolve the ambiguities in the following sentences
using “selectional constraint”. Also, specify whether the ambiguities are syntac-
tic, semantic, or some other?
a. “He saw the man with the horse.”
b. “he saw the man with gun.”
c. “He saw the man with binocular.”
11. What are the different types of ambiguities in natural language, like English?
References 649

References

1. Cambria E, White B (2014) Jumping NLP curves: a review of natural language processing
research. IEEE Computat Intell Mag 5:48–57
2. Chomsky N (1956) Three models for the description of language. IRE Trans Inform Theory
2(3):113–124
3. Chowdhary KR, Bansal VS (2006) Information Extraction from Natural Language Texts. J of
the Institution of Engineers(India), 87:14–19
4. Chowdhary KR (2004) Natural Language Processing for Word Sense Disambiguation and
Information Extraction. PhD Thesis, JNV University, Jodhpur (India)
5. Jurafsky D, Martin J H (2002) Speech and Language Processing - An Intro to Natural Lan-
guage Processing, Computational Linguistics, and Speech Recognition. Pearson Education
Asia, ISBN 81-7808-594-1
6. Lenat D, Guha R (1989) Building Large Knowledge-Based Systems: Representation and Infer-
ence. The Cyc Project. Addison-Wesley, Boston, MA
7. Lin J (2007) An Exploration of the Principles Underlying Redundancy-Based Factoid Question
Answering. ACM Trans on Inf Sys 25(2):1–55
8. Miller AM (1995) Wordnet: A Lexical Database for English. Communications of the ACM
38(11):39–41
9. Minsky M (1968) Semantic Information Processing. MA, MIT Press, Cambridge
10. Minsky M (2000) Commonsense-based Interfaces. Communications of the ACM 43(8):67–73
11. Mueller E (1998) Natural Language Processing with Thought-Treasure. Erik T. Mueller, New
York
12. https://www.nltk.org/, Cited 19 Dec 2017
13. Pearl J (1985) Bayesian networks: A model of self-activated memory for evidential reasoning.
UCLA comput Sci, Irvine, CA: Tech Rep CSD-850017
14. Ralph G (1994) Computational Linguistics - An Introduction, Studies in Natural Language
Processing, Cambridge Univ Press
15. Reiter R (1980) A logic for default reasoning. Artificial Intelligence 13:81–132
16. Ronald C et al (1997) Survey of the state of art in human language technology – Studies in
Natural Language Processing, Cambridge Univ Press
17. Schank R (1975) Conceptual Information Processing. Elsevier Sc Inc, Amsterdam, The Nether-
lands
18. Singh P (2002) The open mind common sense project. http://www.kurzweilai.net/ Cited 19
Dec 2017
19. Sowa J (1987) Semantic networks. Encyclopedia of Artificial Intelligence, S. Shapiro edn,
Wiley, New York
20. Baldwin T (2009) et al (2009) Prepositions in Applications: A Survey and Introduction to the
Special Issue. Computational Linguistics, Vol 35, 2:119–150
21. Turmo J et al (2006) Adaptive Information Extraction. ACM Computing Surveys 38(2):1–47
Chapter 20
Automatic Speech Recognition

Abstract There are basically two application modes for automatic speech recog-
nition (ASR): using speech as spoken input or as knowledge source. Spoken input
addresses applications like dictation systems and navigation (transactional) systems.
Using speech as a knowledge source has applications like multimedia indexing sys-
tems. The chapter presents the stages of speech recognition process, resources of
ASR, role and functions of speech engine—like, Julius speech recognition engine,
voice-over web resources, ASR algorithms, language model and acoustic models—
like HMM (hidden Markov models). Many open-source tools like—Kaldi speech
recognition toolkit, CMU-Sphinx, HTK, and Deep speech tools’ introduction, and
guidelines for their usages are presented. These tools have interfaces with high-level
languages like C/C++ and Python. The is followed with chapter summary and set of
exercises.

Keywords Automatic speech recognition · ASR · Multimedia indexing · ASR

resources · Language model · Acoustic model · Julius · Kaldi · CMU-Sphinx ·
HTK · Deep tools

20.1 Introduction

Speech has long been viewed as the future of computer interfaces, promising signif-
icant improvements in ease of use over the traditional keyboard and mouse. There
are basically two application modes that exist for speech recognition: 1. Speech as
spoken input to computers, and 2. The speech is used as data or knowledge source.
The first application mode comprises the potential applications as, dictation systems,
navigation, and transactional systems.
In the application of dictation, a system transcribes the spoken words into written
text, and for dictating letters, reports, business correspondence, or e-mail messages,
to the computer/machine.
The speech can be used in the form of commands to navigate around the
applications, for example, selection of main application, then its one of the sub-
applications, and sub-sub-application, till you reach to the command to execute the
final application.

© Springer Nature India Private Limited 2020 651

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_20
652 20 Automatic Speech Recognition

The speech can be used for transactional applications, i.e., to use speech in the
form of command or command sequence to cause a transaction to be performed. For
example, the speech-based transactions can be purchase of stock, book a flight ticket,
reserve an itinerary for tour, or doing the fund transfer.
The second mode of speech application, i.e., speech as a knowledge source has
applications like meeting capture, and knowledge management. These applications
are basically multimedia indexing systems that use speech recognition to transcribe
words verbatim from an audio file into text. Subsequently, the IR (Information
Retrieval) techniques are applied to the transcript to create an index with time offsets
into the audio. Users can access the index using text keywords to search a collection
of audio or video documents.
We understand both written and spoken language, and also know that skill of
reading is learned much later. We now focus on spoken language. The problem
of understanding the spoken language can be divided into major parts discussed
below [6].
Phonological
The phonological processing step relates sounds to the words we recognize. Phone
is smallest unit of sound, and the phones are aggregated into word sounds.
Morphological
This is lexical knowledge, which relates to word construction from basic units called
morphemes. A morpheme is the smallest unit of meaning, for example, the construc-
tion of friendly from friend and ly.
Syntactic
It is knowledge about how the words are organized to construct meaningful and
correct sentences.
Pragmatics
It is a high-level knowledge about how to use sentences in different contexts and how
the contexts effect the meanings of the sentences.
World
It is useful in understanding the sentence and carry out the conversation. It includes
the other person’s beliefs and goals.
Speech recognition by machine is important, as many problems get solved if our
computer/laptop can recognize the spoken works. We know, in the present days it has
become possible to do search in some of the search engines by speaking rather than
entering the keywords. In the following, we discuss some of the potential applications
of automatic speech recognition (ASR).
Learning Outcomes of this Chapter:
1. Distinguish the goals of sound recognition, speech recognition, and speaker recog-
nition and identify how the raw audio signal will be handled differently in each
of these cases. [Familiarity]
20.1 Introduction 653

2. Implement a feature-extraction algorithm on real data, e.g., an edge or corner

detector for images describing a short slice of audio signal. [Usage]
3. Language model and acoustic models for speech recognition. [usage]
4. Implement an algorithm combining features into higher level percepts, e.g.,
phoneme hypotheses from an audio signal. [Usage]
5. Evaluate the performance of the underlying feature extraction, relative to at least
one alternative possible approach (whether implemented or not) in its contribution
to the classification task. [Assessment]
6. Tools for speech recognition system. [Usage]

20.2 Automatic Speech Recognition Resources

At the simplest level, the programs that are speech driven can be characterized by the
words or phrases we say to a given application and how that application interprets
them. An application’s active vocabulary is what it listens for, determines what it
understands. The speech recognition process makes use of a speech engine, which
is language-independent, and what it recognizes can be from several domains. A
domain comprises a vocabulary set, pronunciation models, word-usage models that
are associated with specific speech applications. An acoustic component is also
present in the speech recognition engine, as part of voice models used by the speech
engine during the recognition. The voice models can be speaker independent, or may
be unique to the speaker.
Figure 20.1 shows the resources used by a typical speech engine during the process
of speech recognition. The domain-specific resources can vary dynamically during

Fig. 20.1 Speech

German
recognition resources
English

World History domain

Human genome domain

CS domain
Vocabularies
Pronunciations
Word-usage models

User voice
Speaker-independent
Voice-model

Speech in Speech Words, commands,

engine
or index terms out
654 20 Automatic Speech Recognition

a given recognition session. The vocabulary is one of the domain-specific resources.

Some of the major applications are as follows [7].
1. Dictation application. It transcribes spoken input directly into the document’s
text content.
2. Transaction application. It facilitates a dialogue leading to a transaction, and
3. Multimedia indexing application. This application can generate words, which act
as index terms into the multimedia.

As far as application development is concerned, speech engines typically offer a

combination of programmable APIs (Application Programming Interfaces) and tools
that are helpful to create and define vocabularies and pronunciations for the words
they contain. A transactional application may use a smaller, task-specific vocabulary
of a few hundred words, but a dictation or multimedia indexing application may use
a predefined large vocabulary, some times as large as 100,000 words or so.
A small size of vocabularies is enough for some applications. However, they pose
usability problems in the system as their size grows. This is because, the system
requires a strict enumeration of the phrases, which must be recognizable by any of
the given states in the application. To overcome this limitation, speech grammars
for specific tasks are defined in the transaction-based applications, which provide an
extension to the single words or phrases a vocabulary supports. The speech grammars
are helpful in constructing structured collection of words and phrases bound together
by rules that define the set of speech streams the speech engine can recognize at a
given time. For example, using these grammars, the developers can define a grammar
that permits flexible ways of speaking a date, a dollar currency, a number, etc. The
prompts that cue users on what they can say next, are an important aspect of defining
and using grammar. Further, the speech grammars can serve as a critical component
of enabling the voice over the Web, discussed in the next section.

20.3 Voice Web

To have voice facility over the web, a group of industry organizations, which included
AT&T, IBM, Lucent, and Motorola, had established the VoiceXML Forum, in March
1999 to develop and promote a new language—the Voice extensible Markup Lan-
guage (http://www.w3.org/Voice/). The VoiceXML forum was established with the
objective, to bring the content delivery to interactive voice response applications,
using Web-based developments [7, 8].
The VoiceXML (or VXML) is a digital document standard for specifying inter-
active media and voice dialogues between humans and machines. This language
has applications for developing audio and voice response applications, such as for
banking systems, customer services, and for automated customer service portals.
VoiceXML applications are commonly used in industries and segments of com-
merce, some of the examples are as follows.
20.3 Voice Web 655

Speech synthesis,
Recognize spoken and touch-tone key input,
Digitize audio,
Order inquiry,
Package tracking,
Flight tracking,
Voice access to email,
Emergency notification,
Audio news magazines,
Can record spoken input,
Like voice dialing, and
Directory assistance.

VoiceXML applications are developed and deployed in all the above major fields.
These applications are analogous to how a web browser interprets and visually renders
the Hypertext Markup Language (HTML) it receives from a web server. In similar
ways, the VoiceXML documents are interpreted by a voice browser, and users interact
with voice browsers via the public network like Internet. Like HTML, the VoiceXML
text provides tags that instruct the voice browser to provide the functions of automatic
speech recognition, speech synthesis, dialogue management, and audio playback.
Figure 20.2 shows the architecture of VoiceXML model, which makes use of
standard client–server paradigm that integrates voice and data services. A voice
service consists of sequence of voice interaction dialogues between a user and an
implementation platform. A document server, which can be external to the imple-
mentation platform, provide the dialogues. The overall service logic is maintained
by the document servers or web servers, which also perform database and legacy
system operations, and produce dialogues.
A VoiceXML document specifies each interaction dialogue that is conducted by
the VoiceXML interpreter. The following is an example of a VoiceXML document:

Fig. 20.2 VoiceXML

architectural model Document
server
Request
Document

Interpreter for Interpreter

VoiceXML Context

Imprementation
platform
656 20 Automatic Speech Recognition

<vxml version="3.0" xmlns="https://www.w3.org/TR/voicexml30/">

<form>
<block>
<prompt>
This is a VXML Code!
</prompt>
</block>
</form>
</vxml>

A user’s input affects dialogue interpretation, and the system collects this infor-
mation in the form of requests, which it submits to a document server. The later can
reply with another VoiceXML document to continue the user’s session with other dia-
logues. The grammar-based recognition vocabularies are commonly used to support
voice services in VoiceXML applications.

20.4 Speech Recognition Algorithms

The initial attempts for speech recognition were targeted to use expert knowledge of
speech production and perception processes. Soon it was found that such knowledge
was inadequate for capturing the complexities of continuous speech. Currently, the
statistical modeling techniques trained using hours of speech have provided most
speech recognition advancements [4].
The process of speech recognition starts with a sampled speech signal. This signal
has a good deal of redundancy because the physical constraints on the articulators
that produce speech—the glottis, tongue, lips, and so on—prevent them from mov-
ing quickly. Consequently, the ASR (Automatic Speech Recognition) system can
compress information by extracting a sequence of acoustic feature vectors from the
signal.
Typically, the system extracts a single multidimensional feature vector every 10
ms that consists of 39 parameters. These feature vectors, which contain information
about the local frequency content in the speech signal, are called acoustic observa-
tions because they represent the quantities the ASR system actually observes. The
system attempts to infer the spoken word sequence that could have produced the
observed acoustic sequence.
To simplify the design, we assume that speaker’s vocabulary is known to the ASR
system. Having adopted this approach, it is helpful in restricting the search for the
possible word sequences only within the words listed in the ASR lexicon. The ASR
lists the vocabulary and provides phonemes—a set of basic units of words, which
are usually individual speech sounds to pronounce each word.
The commercially available lexicons usually include tens of thousands of words.
The length of the word sequence uttered by the speaker is not necessarily be known,
for the same word by different speakers, as well as by the same speaker at two dif-
20.4 Speech Recognition Algorithms 657

ferent times. Consider that length of the word sequence is N . If V is taken as the size
of the lexicon, the ASR system can hypothesize V N possible word sequences. The
language constraints dictate that these word sequences are not equally likely to occur.
For example, the word sequence “please call me” is more likely to occur than the
sequence “please me call.” In addition, the acoustic feature vectors extracted from
the speech signal can provide important clues about the phonemes which produced
them. The sequence of phonemes that corresponds to the acoustics observations, can
imply the word sequences that could have produced the sequence of these sounds.
Hence, the acoustic observations experienced provide an important source of infor-
mation that can help further narrow down the space of possible word sequences.
The ASR systems use the acoustic observations information to compute the prob-
ability that these observed acoustic feature vectors have been produced when the
speaker uttered a particular word sequence. Essentially, the system efficiently com-
putes these probabilities and outputs the most probable sequence of words as a
decoded hypothesis.
The most successful speech recognition systems of today, use a generative proba-
bilistic model, shown as Eq. 20.1. The speech recognizer tries to find the probability
of word sequence ŵ1N (of N words) that maximizes the word sequence’s probabil-
ity, having given some observed acoustic sequence y1T . This approach makes use of
Bayes’ theorem to compute the conditional probability of p(w1N ) given y1T . When
Bayes equation is expanded in second line of equation (20.1), it ignores the denomina-
tor term ( p(y1T )), common for all possible word sequences. This equation maximizes
the product of two terms: the probability of the acoustic observations given the word
sequence ( p(y1T | w1N )) and the probability of the word sequence itself ( p(w1N )).

ŵ1N = argmax p(w1N | y1T )

  
w1N

= argmax p(y1T | w1N ) p(w1N ). (20.1)

  
w1N

Figure 20.3 shows the process described by Eq. 20.1 as a block diagram. The
lexicon, language model, and acoustic model components construct hypotheses for

Fig. 20.3 Speech recognition system block diagram

658 20 Automatic Speech Recognition

interpreting a speech sample. Block 1, extracts multidimensional features from the

sampled speech signal. In Block 5, hypothesis search is carried out, where the search
hypothesizes a probable word sequence based on the observation of features, as well
the input from three models—lexicon, language, and acoustic. The other components
drive the hypothesis search as follows:
– Lexicon in Block 2 defines the possible words that the search can hypothesize,
where each word is a linear sequence of phonemes;
– Language model of Block 3 models the linguistic structure (sequence of words
i.e., p(w1N )), but does not contain any knowledge about the relationship between
the feature vectors and the words, and
– The acoustic model in Block 4 models the relationship between the feature vectors
and the phonemes ( p(y1T |w1N )), which might have produced the sounds.
Getting the best performance for feature extraction and hypothesis searches
requires customizing the ASR system for individual speakers. The following section
explains the hypothesis in detail.

20.5 Hypothesis Search in ASR

Three basic components comprise the hypothesis search: a lexicon, a language model,
and an acoustic model. Each one is described in detail in the following [4].

20.5.1 Lexicon

A typical lexicon is shown in Table 20.1 with each lexicon’s possible pronunciations
constructed from phonemes. English language has 44 phonemes. Despite there being
just 26 letters in the language, there are 44 unique sounds (phonemes). These sounds
are helpful in distinguishing one word or meaning from another. An individual word
can have multiple pronunciations, for example, the word “the” has two pronunciations
as shown in Table 20.1. These multiple pronunciations complicate the process of
recognition. The hypothesis search chooses the lexicon on the basis of task, trading
off vocabulary size with word coverage. Although a search can easily find phonetic
representations for commonly used words in various sources, task-dependent jargon
often requires writing out pronunciations by hand.

20.5.2 Language Model

The search for the most likely word sequence corresponding to the speech features
sampled, requires the computation of terms, p(y1T |w1N ) and p(w1N ) in Eq. 20.1. The
20.5 Hypothesis Search in ASR 659

Table 20.1 Typical lexicon Lexicon Phonetic representation

The dhah
The dhiy
Cat kaet
Pig pihg
Two tuw

term p(w1N ) is called the language model. The computation requires the assignment of
probability to a sequence of words w1N . A simplest way we can imagine to determine
such a probability, is to compute the relative frequencies of different word sequences,
like we discussed earlier, that “Please call me” is more probable than “Please me
call.” Note that, the total number of different sequences can grow exponentially
with the length of the sequence, making this approach computationally infeasible.
Therefore, there is a need of approximations.
A typical approximation used assumes the probability of current word in the
sequence as depending on previous two words only, called 2-grams, against n-grams.
When this is considered, the computation can approximate the probability of the word
sequence as follows:


i=N
p(w1N ) ≈ p(w1 ) p(w2 |w1 ) p(wi |wi−1 , wi−2 ). (20.2)
i=3

The term p(wi |wi−1 , wi−2 ) can be estimated through computation by counting
the relative frequencies of word trigrams, or triplets:
N (wi , wi−1 , wi−2 )
p(wi |wi−1 , wi−2 ) ≈ . (20.3)
N (wi−1 , wi−2 )

In the above, N is the associated event’s relative frequency. Typically, training

such a language model requires using hundreds of millions of words to estimate
p(wi |wi−1 , wi−2 ) for different word sequences. Even then, many trigrams do not
occur in the training text, so the computation must smooth the probability estimates
to avoid zeros in the probability assignment.

20.5.3 Acoustic Models

An acoustic model computes the probability of feature vector sequences (y1T ) under
the assumption that a particular word sequence (w1N ) produced the vectors. In other
words, an acoustic model is P(y1T |w1N ).
660 20 Automatic Speech Recognition

Fig. 20.4 Hidden Markov p(y/1) p(y/2) p(y/3)

model for a phoneme
p1 p2 p3

s1 1 − p1 s2 1 − p2 s3 1 − p3

Due to inherently stochastic nature of the speech, a speaker never utters a word
exactly the same way twice. The variation in a word’s or phoneme’s pronunciation
manifests itself in two ways: duration and spectral contents, also known as acoustic
observations. In addition, a particular phoneme’s spectral content are effected due
to phonemes in surrounding context, a phenomenon called co-articulation effect. It
is, therefore, necessary that acoustic models used should take into account these co-
articulation effects. One of the popular acoustic models is based on HMM (Hidden
Markov Model).
Hidden Markov Models
A hidden Markov model offers a natural choice for modeling speech’s stochastic
aspects. HMMs function as probabilistic finite-state machines—the model has a set
of states, and its topology specifies the allowed transitions between them. At every
time frame, an HMM makes a probabilistic transition from one state to another and
emits a feature vector with each transition.
Figure 20.4 represents a phoneme’s transitions using a HMM. The transitions in
the waveform of a phoneme correspond to state transitions in the HMM. We may
think of a HMM as a finite automata with transitions governed by probabilities.
Accordingly, we take a set of state transition probabilities in the HMM as, p1 , p2 ,
and p3 , due to which the possible transitions between the states of the HMM are
governed. The probabilities specify the probabilities of going from one state at time
t to another state at time t + 1. The feature vectors emitted while making a particular
transition in the speech waveform, represent the spectral characteristics of the speech
at that point. These feature vectors vary corresponding to varying pronunciations of
the phoneme. A probability distribution or probability density function can model
this variation. In Fig. 20.4, the functions p(y|1), p(y|2), p(y|3), could be different
for different transitions. Typically, these distributions are modeled as parametric
distributions, which are a mixture of multidimensional Gaussian distributions.
The HMM in Fig. 20.4 has three states, representing the pronunciation of a
phoneme starting at state s1 . Then, the phoneme corresponds to a sequence of tran-
sitions, and terminating at state s3 . Duration of a phoneme is equal to the number
of time frames required to complete the transition sequence. The transition proba-
bilities p1 ... p3 implicitly indicate probability distribution that governs the duration
of the phoneme. If any of these transitions exhibits high self-loop1 probabilities,
the model spends more time in that state, consequently consuming a longer time

1 forexample, the word “speech” can be also pronounced as “spee...ech”, repeating the sound of
’e’, which creates a self-loop.
20.5 Hypothesis Search in ASR 661

to go from the first to the third state. Note that, how time duration a self-loop may
repeat, in unknown and varies from speaker to speaker. However, a self-loop (a state)
may repeat on itself few times, typically 2-5. However, some words’ phoneme(s)
may have exceptionally long loop, for example, chanting of Aum.2 The probability
density functions associated with these three transitions govern the feature vector
output.
A fundamental task required to be performed through an HMM is computation
of the likelihood that it produces a given sequence of acoustic feature vectors. For
example, assume that the system extracted T feature vectors from speech corre-
sponding to the pronunciation of a single phoneme, now the system seeks to infer
which phoneme from a set of, say, 50 was spoken, given these feature vectors. The
procedure for inferring the phoneme first assumes that the ith phoneme was spoken,
then finds the likelihood that the HMM for this phoneme produced the observed
feature vectors. The system then hypothesizes that the spoken phoneme model is the
one which has the highest likelihood of matching the observed sequence of feature
vectors.
If the sequence of HMM states is known, we can easily compute the probability
of a set of feature vectors. For this, the system computes the likelihood of the tth
feature vector yt using the probability density function for the HMM state at time t.
Having done this, the likelihood that set of all the T feature vectors has occurred is
simply the product of all the individual likelihoods yt . Usually, it is not possible to
know the actual sequence of state transitions, for the computation of likelihood, all
possible state sequences are summed. Given that the HMM dependencies are local, it
is possible to derive efficient formulas for performing these calculations recursively.
Parameter Estimation
It is necessary that in advance to using of an HMMs to compute the likelihood
values of feature vector sequences, the HMMs needs to be trained to estimate the
model’s parameters. The training process requires the availability of a large volume
of training data in the form of mappings of “spoken word sequences” versus “fea-
ture vectors” extracted from the corresponding speech signals. The commonly used
process to find a particular correspondence is, estimation of maximum likelihood
(ml) function (θ̂ml ). Given that a correct word sequence is known corresponding to
the feature vector sequence, the maximum likelihood computation process tries to
choose those HMM parameters, which maximize the likelihood of training feature
vector. The computation of feature vectors also keeps target for obtaining the correct
word sequence. Consider that y1T is the representation for stream of T acoustic obser-
vations, and let w1N represents the correct word sequence; for these, the maximum
likelihood function estimate represented by θ̂ml is,

θ̂ml = argmax log[ pθ (y1T |w1N )]. (20.4)

  
θ

2 chanting of Aum is a common practice during meditation and yoga (IPA:/5wm/), where sound of

IPA ’‘m’ is repeated.

662 20 Automatic Speech Recognition

In the beginning, an HMM is constructed for a correct word sequence to start

the training process. Then, for each next word, the HMM is constructed by con-
catenating the HMMs for the phonemes that constitute the next word. The word
HMMs are concatenated to construct the HMM for the complete utterance. As an
example, the words “we” and “were” have corresponding phonemes as “W IY” and
“W ER”, respectively. Hence, HMM for the utterance “we were” would consist of
the concatenation of HMMs of four phonemes “W IY W ER”.
The training phase of an HMM assumes that it is possible to obtain the acoustic
observations y1T by the system, by traversing HMM from initial state to final state
in total T time-frames. However, we know that system cannot trace the actual state
sequence, e.g., due to the loops. Therefore, ml estimation assumes that this state
sequence is hidden, and thus, what is best possible is to average out all the state
sequence values. The system can express the maximization of Eq. 20.4 in terms of
the HMM’s hidden states st at time t, as follows:


T 
argmax pθ (st |y1T ) log[ pθ̂ (yt |st )]. (20.5)
  
t=1 st
θ

An iterative process is used to solve Eq. 20.5, with each iteration having two steps:
1. An expectation step, and 2. A maximization step. The expectation step computes
the posterior probability pθ (st |y1T ), or count of a state. The posterior probability
is conditioned on all the acoustic observations. The system makes use of current
parameter estimate of HMM, and the computation is performed using forward–
backward algorithm. The parameter θ̂ is chosen by the maximization step to maximize
Eq. 20.5. For Gaussian nature of the probability function, the computation can derive
closed-form expressions for this step [4].

20.6 Automatic Speech Recognition Tools

Before we proceed to understand some of the commonly available tools for speech
recognition, let us try to understand some of the important terminology of speech
recognition.
Automatic speech recognition (ASR) or simply the speech recognition, or
computer-based speech recognition, is a process of converting the speech into a
sequence of words, using some algorithms which have been implemented as pro-
grams. A standard way of doing this is to first split the utterances in the speech
waveform with respect to certain parameters of speech recognition. Some of these
are: presence of voice activity, duration, pitch, voice quality, voice intensity, S/N
(signal to noise) ratio, and the strength of Lombard effect.3 This is followed with

3 Lombard effect: Involuntary tendency of speakers to increase their vocal effort particularly when
speaking in loud background to enhance the audibility of their voice. Due to the Lombard effect, not
20.6 Automatic Speech Recognition Tools 663

recognition of each utterance. The important factors to be considered during the

recognition are detailed in the following paragraphs.
Concept of Features
The parameters or values in a complete waveform signal for a given speech is very
large. Therefore, to optimize it, a given speech is divided into a large number of
frames, such that each frame is of small length, typically 10 milliseconds. Each
such frame is used to extract 39 different features, i.e., 39 different numerical values
representing speech optimized to 1 frame. These values of one frame are called feature
vector. Each such vector that corresponds to a speech segment of ten milliseconds is
numerical representation of the speech of that duration.
Concept of Model
We need a mathematical model, called concept model, that has a representation to
gather common attributes of spoken words. Such a model needs to be evaluated for
various characteristics, like, how adaptive it is for the changing situations, how well
this model fits into practice, and how well it can be configured?
Concept of Matching Process
It would require a lot of time for comparing the feature vectors with all the models.
Hence, the search should be optimized for choosing the best matching variant.

20.6.1 Automatic Speech Recognition Engine

Julius is a high- performance continuous speech recognition software, based on word

N -grams. It is able to perform recognition at the sentence level with a vocabulary of
the order of tens of thousands of words. Julius can realize high-speed speech recog-
nition on ordinary laptop/PC, can perform the speech recognition at near real time,
and can achieve typically a recognition rate of greater than 90% using a vocabulary
of 20, 000-words, for dictation tasks. Julius is multipurpose, i.e., by recombining the
pronunciation dictionary, language, and acoustic models, one can build task-specific
systems. Its code is open source, so one can recompile the system for other platforms
or alter the code for specific needs [2].
Figure 20.5 shows the structure of Julius speech recognition system. The language
model of Julius uses N -gram mode, and context-dependent HMM (Hidden Markov
Model) is used as Acoustic model. As shown in the figure, the input speech is pro-
cessed through two passes: first pass is 2-gram frame synchronous beam-search (a
high-speed approximate search), and second pass is 3-gram N -best stack decoding,
which is a high precision technique. It can do online recognition using PC/laptop’s

only the loudness increases, but also the other acoustic features such as pitch, rate, and duration of
syllables. The Lombard effect also results in an increase in the signal-to-noise ratio of the speaker’s
signal.
664 20 Automatic Speech Recognition

Context-dependent HMM Acoustic

(cross word approx.) (no approx.) Model

Julius

Frame synchronous Stack decoding

beam search search
Input word Word
(1-best) trellis (N-best)
speech sequence
Index

word word Language

2-gram lexicon 3-gram Model

Fig. 20.5 Julius speech recognition system

microphone or can use any audio device. The first pass of Julius segments input with
short pauses, and the second pass sequentially decodes these segments and slots to
the results. During the first pass, when a short pause has the maximum likelihood at
a certain point of time, a break is placed at that point and second pass is executed
on that utterance segment. Due to this process, word constraints are preserved as the
context within a utterance segment, and first pass may continue over to the next utter-
ance. Using the above sequence of steps, an input speech file with multiple sentences
can be decoded.

20.6.2 Tools for ASR

Speech Recognition is available in English, and many other languages, and this
feature is common in most smart phones, laptops, and PCs. In the following, we
discuss basic tools available as open source.
Kaldi Speech Recognition Toolkit
Kaldi is an open-source speech recognition toolkit, written in C++ language, and
works under the Apache platform.
It is a finite-state transducer (FST) based framework, with linear algebra support.
Figure 20.6 shows different components of Kaldi: the library modules are grouped
together, which depend on two types of libraries, 1. linear algebra libraries (the
numerical algebra libraries) and 2. OpenFST (finite-state framework). These two
external libraries are also freely available as open source [3, 5].
Access to the library functionalities is provided through command-line tools writ-
ten in C++. These tools are called from a scripting language, for building and running
a speech recognizer. Each tool has a very specific functionality with a small set of
command-line arguments: for example, there are separate commands to be executed
20.6 Automatic Speech Recognition Tools 665

Fig. 20.6 Kaldi speech External Libraries

recognition toolkit
BLAS/LAPACK OpenFST

Kaldi C++ Library

Kaldi C++ Executables

(Shell) Scripts

for accumulating statistics, summing accumulators, and updating a GMM-based

(Gaussian Mixture Models) acoustic model. For language modeling (LM), Kaldi
uses FST-based framework, hence, in principle it can use any language model that
can be represented as FST.
CMU-Sphinx
Since its release as an open-source code in 1999, CMU-Sphinx provides a platform
for building speech recognition applications. It is used in desktop control software,
telephony platforms, intelligent houses, computer-assisted language learning tools,
information retrieval, and mobile applications. Traditionally, CMU-Sphinx provides
support for low-resource and underdeveloped languages. It is a speech recognition
toolkit with tools to build speech applications, which makes use of technologies
such as C, cross-platform, HMM (Hidden Markov Models), JavaScript, and Python.
CMU-Sphinx contains a number of packages for different tasks and applications.
The following is the list:
Sphinx4—adjustable, modifiable recognizer written in Java, and
Sphinxtrain—acoustic model training tools.
Pocketsphinx—lightweight recognizer library written in C, and
Sphinxbase—support library required by Pocketsphinx.
Deep Speech Tool
It is a simple end-to-end deep learning based speech system, which when combined
with a language model, achieves higher performance than traditional methods on hard
speech recognition tasks. The deep tool is realized by training a large recurrent neural
network (RNN) that uses multiple GPUs and thousands of hours of data. Due to which
the system learns directly from data, and there is no need for specialized components
for speaker adaption, like in other systems, neither it needs noise filtering. The more
traditional systems use acoustic models and Hidden Markov Models(HMM) [1].
The RNN, which is core of this system, is trained to speech spectrograms to gener-
ate English text transcriptions. A training set χ = {(x (1) , y (1) ), (x (2) , y (2) ), ...} is used
to sample a single utterance x and a label y, where each utterance x (i) is a time series of
length T (i) . Here every time slice is a vector of audio features, xt(i) , t = 1, ..., T (i) . The
role of RNN is to convert the input sequence x into a sequence of character probabil-
ities for the transcription y, with ŷt = p(ct |x), where ct ∈ {a, b, c, ..., x, space, }.
666 20 Automatic Speech Recognition

The RNN has five layers of hidden units, such that for an input x, the hidden
units at layer l are denoted as h (l) , with input as h (0) . The first three layers are not
kept as recurrent layers. At each time t, for the first layer, the output depends on the
spectrogram frame xt along with a context of C frames on each side. The remaining
non-recurrent layers operate on independent data for each time step. Thus, for each
time t, the first three layers are computed by

h (l) (l) (l−1)

t = g(W h t + b(l) ), (20.6)

where g(z) = min{max{0, z}, 20} is the clipped rectified-linear activation function
and W (l) , b(l) are weight matrix and bias parameters for layer l.

HTK Tool

The hidden Markov model toolKit (HTK) is a portable toolkit for building and manip-
ulating hidden Markov models. It is primarily used for speech recognition research,
and speech synthesis, as well for character recognition and DNA sequencing. HTK
consists of set of library modules and tools available in ANSI C source form. These
tools provide sophisticated facilities for speech analysis, HMM training, testing, and
result analysis.
The statistical speech models use here the context-dependent hidden Markov
models. Probabilities of word sequences is based on N-gram, which finds the most
probable word sequence using language model (refer Eqs. 20.1, 20.2) and acoustic
model (refer Eq. 20.1).

20.7 Summary

Automatic speech recognition (ASR) is another domain of human–machine inter-

face, where machine is to recognize human speech, that is, transform some kind of
frequency signal to text. This is a complex process, and requires many processes like
phonological, morphological, syntactic, pragmatics, and world. Speech recognition
is considered as the future of computer interface. There are basically two application
modes for speech recognition: 1. Using speech as input, or 2. As data or knowledge.
The application of “speech as input” addresses applications like dictation systems,
navigation or transaction systems (like purchasing stocks). Using speech as knowl-
edge has applications like meeting capture.
An application of speech recognition can be implemented using “Voice extensi-
ble Markup Language” (http://www.w3.org/Voice/). Developers can use VoiceXML
to create audio dialogues that feature synthesized speech, recognition of spoken
and touch-tone key input, digitized audio, recording of spoken input, telephony,
and mixed-initiative conversations. The VoiceXML’s architecture uses client–server
paradigm to integrate voice services with data services. A voice service is a sequence
20.7 Summary 667

of interaction dialogues between a user and an implementation platform, and a docu-

ment server, which can be external to the implementation platform, and can provide
the dialogues.
The modern speech recognition algorithms are based on statistical modeling tech-
niques trained from hours of speech. The process of speech recognition starts with a
sampled speech signal, which has a good deal of redundancy due to the physical con-
straints on the articulators that produce speech. Consequently, the ASR system can
compress information by extracting a sequence of acoustic feature vectors from the
signal. A system extracts a single multidimensional feature vector every 10 ms that
consists of 39 parameters. The system seeks to infer the spoken word sequence that
could have produced the observed acoustic sequence, using Bayesian probabilistic
approach, which basically, is a process of hypothesis search. Models used for AR
are: Language model (searching the most likely word sequence), Acoustic models
(compute the probability of feature vector components), and Hidden Markov models
(probabilistic finite-state machines).
There are a number of software tools, most as open source, for speech recogni-
tion, speech synthesis, as well for research in automatic speech recognition. These
software tools take input as sound-wave signal (a file) and split the waveform based
on the utterances by the speaker, sample the speech input intervals of about 10 mil-
liseconds, extract the features of input speech. Then using the various models of
probability theory, estimate the probable text, which most likely would have pro-
duced these utterances. These tools were developed (mostly) as research projects.
Among these are: Julius, Kaldi, CMU-Sphinx, Deep Speech tools, and HTK (Hidden
Markov Model).

Exercises

1. Consider alphabet set  = {a, b, c, d}. Create finite automata (recognizers) for
following strings.

a. All strings which start with letter a.

b. All strings which end with letter d.
c. All strings where every c is followed letter d.
d. All strings which have odd number of c’s.

2. Answer followings in brief, giving suitable examples.

a. What is the difference between phoneme and morpheme?
b. What is the difference between language and dialect?
3. Write an equation to compute trigram probability.
668 20 Automatic Speech Recognition

4. The text processing algorithms are usually written in Python, while the ASR
algorithms, which produce the same text, are written in C/C++. Explain what
could have been the reason behind this?
5. What is the fundamental difference between the language model and acoustic
model? Why are they the same so?

References

1. Hannun A et al (2014) Deep Speech: Scaling up end-to-end speech recognition. https://arxiv.

org/abs/1412.5567. Accessed Dec 19, 2017
2. http://julius.osdn.jp/book/Julius-3.2-book-e.pdf. Accessed Dec 19, 2017
3. http://kaldi.sf.net/. Accessed Dec 19, 2017
4. Padmanabham M, Picheny M (2002) Large-vocabulary speech recognition algorithms. Com-
puter 4:42–50
5. Provey D (2011) The Kaldi speech recognition toolkit. IEEE workshop on automatic speech
recognition and understanding. US IEEE Signal Processing Society, Hawaii
6. Ronald C et al (1997) Survey of the state of art in human language technology. Studies in Natural
Language Processing, Cambridge University Press
7. Savitha S, Eric B (2002) Is speech recognition becoming mainstream? Computer 4:38–41
8. http://www.w3.org/Voice/. Accessed Dec 19, 2017
Chapter 21
Machine Vision

Abstract The goal of computer vision is to extract information from images—a

method that can produce a structure from motion, can recover a three-dimensional
model of an object from a sequence of views, e.g., grasping by robot, medical
imaging, and graphical modeling. This chapter presents the machine vision appli-
cations, the basic principle of vision, cognition and classification, and cognitive
architecture. The cognition is—going from image-to-scene, which can be achieved
through inversion by fixing scene parameters, inversion by restricting the problem
domain, and inversion by acquiring additional images. The machine vision tech-
niques are presented here for low-level, middle-level, and high-level vision. The last
one requires indexing of images which can be searched through geometric hashing.
One of advanced areas of vision—the object tracking, is presented in-depth, which
requires the sequences—subtraction of image from background, segmentation of the
image, and learning, followed with tracking. Finally, the axioms of vision, tools for
computer vision, chapter summary, and practice exercises are presented.

Keywords Computer vision goals · 3D models · Vision principles · Vision

application · Cognition · Object classification · Cognitive architecture · Machine
vision techniques · Low-level vision · Middle-level vision · High-level vision ·
Image segmentation · Computer vision tools · Vision axioms · Computer vision
tools

21.1 Introduction

The field of Machine Vision (also called Computer Vision) is growing rapidly. It
is concerned with analysis, modification, and understanding of images. The objec-
tive of this field is to understand what is happening in front of a camera, and use
that understanding to control a computer or robotic system, or making use of these
images provide people new images that are more informative or aesthetically more
pleasing than the original images. The machine vision has vast applications, some of
the important application areas of machine vision are the following: automotive sys-
tems, photography, video surveillance, biometrics, movie productions, Web search,

© Springer Nature India Private Limited 2020 669

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7_21
670 21 Machine Vision

medicines, medical and health sciences, augmented reality, gaming, new user inter-
faces with computers, and in fact, there cannot be an exhaustive list of all possible
areas, as they are continuously increasing in numbers.
As part of advances in machine vision, there are cameras that can focus automat-
ically on our face, and can trigger the shutter when we smile. On the other hand,
using an optical text-recognition system one can transform scanned documents into
text, which can be analyzed or read aloud by a voice synthesizer.
It is common to have driver-assistance systems in the newly built cars that help
the driver to park the car, or give warning signal to the driver in potentially dangerous
situations. Intelligent video surveillance systems exist, that can monitor the security
of public areas. As new smartphones come progressively equipped with more and
more processing power and better resolution of cameras, these phones are becoming
fertile field for potential computer vision applications. These devices have built-in
capability to merge number of photographs into a high-resolution panorama, they can
also read a quick response code, can recognize it and retrieve information about the
product from Internet. The mobile computer vision technology is moving fast toward
touch interface. However, still the computer vision is computationally expensive to
achieve all these in ideal sense.
The computer vision applications have real use only when they can be executed
in real time, for which, it is necessary that processing of each frame needs to be com-
pleted within 30 − 40 msec duration. This is a challenging task, particularly when
the computation takes place, e.g., in a smartphone or some other embedded comput-
ing environment. Generally, it is possible to trade quality for speed or vice versa. For
example, in the panorama stitching algorithm, if time is provided more liberally, it is
possible to find out more matches in source images, and one can synthesize an image
of higher quality. To meet the constraints of time and computations to be carried out,
one can compromise either on quality or spend longer time optimizing the code for
a given hardware architecture.
Further, some of the goals of computer vision or machine vision can be to extract
information from images. For example, an algorithm that produces a structure from
a motion, can recover a 3D model of an object from a sequence of views. As other
examples, from a robot grasping an object it is possible to construct a 3D view;
construction of a 3D object through medical imaging; and graphical modeling to
produce a 3D view. Model-based recognition methods can determine the best matches
of stored objects for image data, for use in visual inspection and image database
searches. Through visual motion analysis, it is possible to recover image motion
patterns for use in vehicle guidance and processing digital video.
The field of computer vision has a close relation to the field of image processing,
as follows: In image processing, the focus is on transformation of images, but in
computer vision, the focus is on extracting information from image data. To extract
a 3D from 2D images is the case of computer vision, and not of an image processing.
However, image filtering and color enhancement are the tasks of image processing.
Though many other computer science areas (computation), and board game play-
ing (chess, and tic-tac-toe), have progressed enough and have speed far higher than
human, the area of computer vision heavily lags behind compared to human capa-
21.1 Introduction 671

bilities. The successful example of visual information processing is recognizing of

printed text. However, the Optical Character Recognition (OCR) systems make mis-
takes that even school level students do not make. The problem with computer vision1
systems is their properties of brittleness—a small change in input causes a big change
in the output.
Learning Outcomes of this Chapter:
1. Summarize the importance of image and object recognition in AI and indicate
several significant applications of this technology. [Familiarity]
2. List at least three image-segmentation approaches, such as thresholding, edge-
based, and region-based algorithms, along with their defining characteristics,
strengths, and weaknesses. [Familiarity]
3. Implement 2D object recognition based on contour-and/or region-based shape
representations. [Usage]
4. Provide at least two examples of a transformation of a data source from one
sensory domain to another, e.g., data interpreted as single-band 2D image to 3D
object. [Familiarity]
5. Describe an algorithm combining features into higher level percepts, e.g., a con-
tour or polygon from visual primitives. [Usage]
6. Describe a classification algorithm that segments input percepts into output cat-
egories and quantitatively evaluates the resulting classification. [Usage]
7. Evaluate the performance of the underlying feature extraction, relative to at least
one alternative possible approach in its contribution to the classification task.
[Assessment]
8. Describe at least three classification approaches, their prerequisites for applica-
bility, their strengths, and their shortcomings. [Familiarity]
9. Tools for Computer Vision. [Usage]

21.2 Machine Vision Applications

Originally the computer vision systems were designed with the aim to be used in
industrial and military applications. Some of the common military applications of
computer vision are: recognition of far-off targets, visual-based guidance systems for
autonomous vehicles, and interpretation of images. In the recent times, many com-
puter vision applications have emerged in medical imaging and multimedia systems,
like preoperative scans of patients in the operating room, complex surgery using
robots, etc. Computer vision techniques are also used for virtual reality applications,
and image database retrieval systems. The successful applications of computer vision
exists as a consequence to two important effects: 1. A technical force push toward
limiting the domain that have sufficient structure or constraint that the many-to-one

1 In
AI literature, the words: “computer vision” and “machine vision” are interchangingly used,
hence in this chapter also, both the terms mean the same.
672 21 Machine Vision

inversion problem can be made tractable, and 2. Economic force push toward having
a problem of sufficient base to be financially viable. The possible solution is an inter-
section of these two conditions. The following is a brief account of machine vision
applications [4].
Visual Inspection
One of the most successful applications of computer vision is in industrial quality
control, for example, in automation of visual inspection tasks. This is in replace-
ment for human visual inspection, which is crucial to quality control across a broad
spectrum of manufacturing, that ranges from low-volume custom products to high-
volume and bulk products.
Assembly and Material Handling
In automatic assembly of products, the central driver is the uncertainty in the factory.
To solve this problem, robot vision is typically used. The robot vision is based on
computer vision algorithms, which have applications in navigation and path control
of robot vehicles and robot manipulators. This application is governed by economic
drive as well as quality control.
Automatic Target Recognition (ATR)
Due to the progress of sophistication of weapon systems, it has become necessary
to provide rapid and accurate identification and tracking of targets. The main drive
force toward the progress in this direction is accuracy of recognition, with tracking
of multiple targets at the same time.
Photo Interpretation
This work requires the digitization of image data, and interpret them using vision
algorithms.
Extraction of 3D Structure
There are many engineering applications, that require construction of complex 3D
geometric models and terrains, such applications are: engineering simulation, numer-
ically controlled machines, virtual reality, and advanced cartography. The input to
such models are: drawings, intensity images, and 3D sensors like laser triangulation.
The algorithms used in these applications derive automatically a 3D slid model of
an object that drives a finite element or graphic simulation program.

21.3 Basic Principles of Vision

One of the most basic and essential characteristics of living beings is the ability to
recognize and identify objects. All the higher animals depend on this ability for their
survival. Without this they are unable to function even in a static and unchanging
environment. Sight is our most impressive sense. It gives us, without conscious
21.3 Basic Principles of Vision 673

Stickman
Eye
Image

Fig. 21.1 Image of a “Stickman”

efforts, the detailed information about the shape of the world around us. It is also the
challenging area in AI.
The most natural place to begin our study is transduction of physical phenomena
into internal representation. Vision starts with an eye—a device for capturing and
focusing the light that bounces off the objects. Every point on an object (other than a
mirror) scatters incident light in all directions. An eye has a lens that directs the light
from one point on an object to just one point on retina (see Fig. 21.1). The image on
retina is upside down, but human mind corrects it. We note that image is 2D, while
the object is 3D [1].
Definition 21.1 (Vision Problem) A vision problem may be stated as, given a 2D
image, infer the objects that produced it, as well as infer their shapes, position, sizes,
and colors.
However, one imprecision exists in the above definition, it is—how all these
characteristics counts for object. For monochrome images, an image may be thought
as a function, giving a gray-level at every point on the image plane. The gray-level
varies from 0 (maximum black) to 1 (maximum bright). The latter is corresponding
to the maximum response the eye can make. We assume that surface on which image
is constructed is flat in x and y coordinates, which is appropriate for computer vision
also. We approximate the gray-level function by a 2D array image, which breaks the
image into squares, and records average gray-level over each square, with a pixel at
a point, say (i, j), is cell image(i, j). The value of gray-level function at each pixel
is in the range [0, 1].
It is challenge to reason back from image to the scene that created that. Figure 21.1
suggests that each spot in image corresponds to one spot in the real world, namely, the
piece of surface the light bounced off in order to make the image spot. The recognition
of images is a process of pattern recognition. The recognition requires learning,
which means mapping the images patterns to some internal representation—this is
scene transformation into images. However, recognition of objects is inversion of this
process, i.e., reconstruction of scene features from the images, which is a complex
task.
Though it is not yet fully established, but it is hypothesized that human follow the
following process for identification or classification of objects:
674 21 Machine Vision

New objects are introduced to a human through activation of sensor stimuli.

The sensors, depending on their physical properties, are sensitive in varying
degrees to certain attributes, serve to characterize the objects, and the sensor
output tends to be proportional to the more prominent attributes. Having per-
ceived a new object, a cognitive model is formed from the stimuli patterns and
stored in the memory. Recurrent experiences in perceiving the same object or
similar objects strengthen and refine the similarity patterns. Repeated percep-
tions result in the generalized models of objects classes which become later
useful in matching, and hence recognition of similar objects.

The recognition is, in fact, a process of establishing a close match between some
new stimulus and the previously stored stimulus pattern. Object recognition is the
task of finding and labeling parts of a 2D image of scene that corresponds to objects
in the scene. For example, from an aerial photo, to identify the various objects, like
building, roads, forests, etc. It is a task as it might be carried out by a human observer
with marking pen, to mark and label the objects.
Object classification is closely related to recognition. The ability to classify var-
ious objects or group of objects accordingly to some commonly shared features is a
form of class recognition. Classification is essential for decision-making, learning,
and many other cognitive acts. Like recognition, classification depends on the ability
to discover common patterns among objects. This ability must be acquired through
some learning process. For the learning to take place, the prominent feature patterns
that characterize classes of objects must be discovered, generalized, and stored for
subsequent recall and comparison.
To recognize an object, it is required to first establish a general description of
each object to be recognized. Most often, a model includes texture, shape, and the
context knowledge about the occurrence of such objects in a scene. For instance,
a mathematical description of a set of shaded rectangles may be used to generate
buildings as objects. A 3D building object could be modeled as a set of rectangular
solids. Texture information might include colors or knowledge about the layout of a
building’s window.
Corresponding to each occurrence or instance of a model in the image, a label
is attached, called model label, that can be thought of as a tag pinned to an area in
the image that we consider showing an instance of the corresponding object model.
The “roads” and “buildings/houses” in Fig. 21.2, shown by arrows, are examples of
model labels. Note that a model may be 2D or 3D, but the labels always show 2D
model instances.
There are many important distinctions about the kind of information we are inter-
ested within a digital image, corresponding the scene of that image. The most ele-
mentary type of information is called syntactic information, which is concerned only
with pixel values, and their meanings. The other information, called semantic infor-
mation, deals with knowledge and meaning of the information. Hence, a syntactic
image operator’s role is to blindly apply an algorithm to the pixel values, without
21.3 Basic Principles of Vision 675

Fig. 21.2 An aerial image of a Hilly Town with labels attached to buildings and roads

concerned to the meaning associated with those pixels. An example is, a procedure
that assembles group of adjacent pixels which have, say, a high contrast with respect
to their neighbors. On the contrary, a semantic operator uses models of the scene and
the image production process. They incorporate symbolic knowledge about how the
image is organized, such as “a part may be lying on top of one another.”

21.4 Cognition and Classification

For a process of mechanized recognition requires learning new objects carried out
through the following steps:
1. The stimuli produced by objects is perceived by the sensory devices. The more
prominent attributes, like size, shape, color, and texture produce the strong stim-
uli. The values of these attributes and their relations are used to characterize
an object in the form of a pattern vector as a string. This string is represented
as a classification tree, a description graph, or some other form. The range of
characteristic attribute value is known as the measurement space.
2. A subset of attributes whose values provide cohesive object grouping or clustering,
consistent with some goals associated with the object classifications, are selected.
The range of the subset of attribute values is feature space.
3. Using the selected attribute values, the object or class characterization models
are learned by forming generalized prototype description, classification rule, or
decision functions. These models are stored for subsequent recognition. The range
of decision function values or classification rules is decision space.
Recognition of familiar objects is achieved through application of the rules learned
by step 3 above, by comparison and matching of object features with the stored
676 21 Machine Vision

Fig. 21.3 Pattern

Sensor Feature Matching Classification
recognition process selection

Learning
Classification
rules

models. Refinements and adjustments can be performed continually thereafter to

improve the quality and speed of recognition. These process steps are shown in
Fig. 21.3.
Conceptual View of Cognitive Architecture
For an artificial system it is a primary requirement that it should build a rich internal
representation of the external environment. This internal representation should be
such that it is supportive to draw the inferences, make the decisions, and allows to
perform the reasoning process in general related to its own tasks. One approach for
representation is classical logic, where symbols are given the meaning by relating
them to abstract entities as per semantics of model-theoretic approach. However,
this approach turns out to be incomplete for the requirement of machine vision, as
it requires to find out two things, 1. The meanings of its symbols within the internal
representation, and 2. Its interaction with the external world.
A machine vision system requires a cognitive architecture with effective internal
representation of the environment. This environment is built through the processes
that are defined over suitable intermediate level. The processes act as intermedi-
ary between sensory data input and the internal symbolic level representation. One
approach to model the visual perception through a process, such that the represen-
tation of information and knowledge, as well as the processing of both, take place
at different levels of abstractions. The two levels of abstraction are: 1. The low-
est level, which is directly related to features of stimuli and 2. The highest level,
where knowledge is in symbolic form, that is concerned with the perceived object.
A general assumption of computer vision is that a vision process concludes with 3D
reconstruction of shapes using some geometric primitives.
For the designs related to cognitive architectures, there are three cognitive repre-
sentation levels, are as follows [2]:
• Sub-symbolic level. The information at this level is related only to sensory data
and to nothing else.
• Linguistic level. This level uses the information represented using symbolic forms.
• Intermediate level. It is also called prelinguistic conceptual level. The characteristic
of information at this level (in terms of metric spaces) is defined by a number of
cognitive dimensions, that is language independent.
The intermediate level representation is useful for generating the essential repre-
sentation of external environment of a cognitive agent. In addition, the intermediate
level provides a precise interpretation of the linguistic level, where interpretation of
21.4 Cognition and Classification 677

conceptual categories is concerned with some standard problems. One such prob-
lem is that, perceptual commonsense concepts never correspond to any clear classic
category which can be described in terms of necessary and sufficient conditions.
For example, the membership in perceptive concepts categories is never in 0 and 1
classes, but usually there is need to consider a prototype of the category.
The information available in a cognitive system strictly depend on data acquired
through measurement process. Hence, the knowledge derived through this informa-
tion at conceptual level will also be affected due to measurement errors. To solve this
problem, a model mapping is carried out between conceptual and linguistic levels
using a connectionist 2 device. The use of Neural Networks (NN) eliminate the need
for exhaustive description of conceptual categories at the symbolic level. The latter
becomes possible because, a prototype is built based on associative mapping taking
place during the training phase of NNs. A measure of similarity between the proto-
type and the given object is implicit in the behavior of the NN, which is determined
during the learning phase of this network.

21.5 From Image-to-Scene

For a creature, of type either biological or mechanical (in case of latter, e.g., robot),
for effective interaction with its environment, it needs to know what are the objects,
and where they exists? The computer vision provides the basic methods to understand
as how to make intelligent decisions about the environment of interactions, on the
basis of sensory inputs it receives. To support the intelligent interaction with the
environment, the vision system must know the information about objects in the
world (with which it interacts). This knowledge about the objects and their positions
become known to the vision system only on the basis of the measurements of inputs
received in the form of reflected brightness. To convert this brightness into the world
of measurements, first it is necessary to know how the objects are mapped into the
image brightness. The magnitude of light recorded by an individual sensor is the
result complex interaction of the following parameters:
• Orientation and position of the object, as well as the position and orientation of
the light sources with respect to the sensor,
• 3D shape of the object,
• The reflective properties of the object’s surface (i.e., the rules of physics that govern
the reflection of light rays from surface of the object),
• Spectral sensitivity and quantum efficiency of the sensor, and
• Spectral properties of the light sources.
For generating an image from a given set of scene parameters (i.e., mapping from
scene to image) is a well-defined process. However, to invert an image to compute
the scene parameters that gave rise to the image is an ill-posed and challenging

2 Labeling of unsegmented sequence of data with recurrent neural networks.

678 21 Machine Vision

problem. This is because, it requires an inversion of a single number to deduce many

parameters. If an image is treated as a set of independent brightness values, there can
be an infinite number of scenes that could have produced this image ! Hence, deciding
the precise scene that might have given rise to so and so image is a challenging task.
However, as human beings (and other living creatures), we generally do not have
any trouble in concluding the true original scene by interpreting any given image.
Thus, the image brightness is generally not independent, and goal of computer vision
is to find out the sufficient additional constraints to invert the brightness into scene
parameters. In the following discussions, we present major approaches to object
recognition through the inversions.

21.5.1 Inversion by Fixing Scene Parameters

One approach to perform inversion from image-to-scene that created the image is
through fixing some scene parameters. For example, if we have information about
the surface reflectivity3 of an isolated object, and about the position of light source,
there is a nonlinear equation available that directly relates image brightness at a point
to the object’s local surface orientation at that point. Since there are two unknown
variables for each measured brightness, we need more constraints, to determine the
scene.
One method, called shape-from-shading, assumes that the surface that is locally
smooth is sufficient to provide solution for the object’s shape. The other methods
are called photometric stereo methods. Their working is based on taking several
images with a single position of camera, but with different light sources, and a set of
brightnesses are used corresponding to a single point, to determine the local surface
shape.

21.5.2 Inversion by Restricting the Problem Domain

There is a class of methods for going from image-to-scene, i.e., image inversion, that
is based on restricting the problem domain. Consider the case of printed character
recognition, like Optical Character Recognition (OCR), where domain consists of
2D objects, with known shapes and restricted range of orientation and positions.
The inversion problem, in this case, is basically separating individual objects (let-
ters, digits, and punctuation marks) in the presence of noisy brightnesses, and then
matching the shapes of those objects against the canonical models. There are in fact
many successful commercial tools available for recognizing the printed characters,
but the problem becomes more complex if we allow in the range of characters, all
the cursive scripts with all their possible variants!

3 The fraction of radiant energy that is reflected from a surface, also called coefficient of reflection.
21.5 From Image-to-Scene 679

Consider the problem of identification and locating planar objects in cluttered

scenes. The image inversion in this case typically requires finding the invariants in
the image forming process. That is, to infer a sudden change in a scene parameter,
like edge of an object, based on the observation of a sudden change in a recorded
brightness in the image. The features like this have the advantage of being insensitive
to variations in light sources and camera positions, hence they are more reliable
indicators of scene features.
After having extracted the hypothesized instances of the object features, the next
step to recover the scene is to match the geometric shapes possessing such features
against the known models. However, in the following, we present an inversion tech-
nique that is based on acquiring additional images.

21.5.3 Inversion by Acquiring Additional Images

Yet another method for image inversion states that, it is possible to achieve image
inversion by acquiring additional images. For example, using stereo vision, which
obtains two views of a scene, it is possible to extract 3D shapes of objects. The
principle of stereo vision is as follows: if one can determine a point in two images that
are projections of a single point in the original scene, and if relative orientation of two
cameras is known, then the method of triangulation can produce a 3D reconstruction
of the shapes of the objects that are existing in the real world. A similar principle
holds for motion sequence of images, where either the camera or the objects in
the real world move between the images. This process again involves a matching
problem, this time there is matching between features from two images, instead of
an image and a model. Most of the machine vision algorithms depend on matching
schemes that measure the similarity between features in the two images, and make
use of multi-resolution methods to control the complexity of such matchings.
Once the 3D shape of the scene is extracted, recognition of the scene is made
possible by matching 3D shape descriptions, while the navigation in the 3D world is
supported by identifying the collision-free paths through the reconstruction world.
A large variety of matching schemes exists, that also includes the methods that
directly search the space, of all possible pairings of models and image features, called,
correspondence space. There are other methods, that directly search the space of all
possible poses of the objects, and there are hybrid methods that can combine both
the type of spaces. Most of the methods rely on geometric constraints to reduce the
complexity of search, the constraints that typically encode information about objects’
shape. The constraints are also used to measure the feasibility of matching the data
features to model features, such that they are consistent with the legal transforma-
tions of the object. The following are typically the difficult parts of this problem:
1. efficiently deciding about the image features that belong to a single object in the
presence of scene clutter and occlusion, and 2. effectively match the partial descrip-
tion of the object in the presence of uncertainty in sensor input.
680 21 Machine Vision

From the above discussions, we find it common that, vision methods attempt
to extract scene parameters, like surface material type and object shape given the
observed brightnesses, and then to use these extracted parameters to match against
known-object models to support the recognition.

21.6 Machine Vision Techniques

The machine vision systems operates on digital images, i.e., quantized collection of
discrete values in 2D space with varying intensity. The methods used in machine
vision are also classified as low-level vision, middle-level vision, and high-level
visions. Though, this is not the only classification for vision systems, but it provides
a useful way of classifying the computer vision problems [9].
The Low-level vision techniques operate directly on images and produce the output
images in the same coordinate system as that of the input. For example, an edge
detection algorithm may take an image with different intensity values as input, and
produce a binary output that indicates the edges in the picture.
The Middle-level vision techniques take an image or the output produced by low-
level vision algorithm as input, and may produce the output something other than
pixels of the image coordinate system. For example, the stereo-vision system may
produce an output shape in 3D using input as two 2D images. As another example,
a structure-from-motion algorithm takes as input a set of image features and produce
at output 3D coordinate of these features.
The high-level vision techniques take as input the results of low- or middle-
level vision algorithms and produce in the output some abstract data structures. For
example, a model-based recognition system may take a set of image features as
input, and provide at output the geometric transformations that map the models in
its database to the respective locations in the image.

21.6.1 Low-Level Vision

The low-level vision computations operate directly on images and produce pixel-
based outputs, that remains within the original image coordinates only. The examples
of these computations are: finding the intensity edges in an image, smoothing images
using filters, computing visual motion fields, and analyzing color information in
images [9].
In the following, we discuss edge detection and smoothing of image, in more
detail.
21.6 Machine Vision Techniques 681

21.6.2 Local Edge Detection

The edge detection of a real-world object is carried out to find out the geometry of the
image and to be used in higher level image processing. There are some physical events
that cause intensity changes, or appearance of an edge in the images. For example, the
object boundaries produce intensity changes which are caused due to discontinuity
in the depth or difference in surface color, or difference in texture. The surface
boundaries produce intensity changes due to a difference in the surface orientation.
However, there are some phenomena of intensity changes that do not reflect directly
on the geometry of the object, the examples are shadows and interreflections [9].
As an example, we shall represent gray-level image by A(x, y), as an intensity of
a function of the image coordinate system (x, y). The intensity of edges in the image
correspond to sudden changes in the value of function A(x, y). For this image, we
compute the gradient magnitude using the expression,
 2  2
∂A ∂A
∇ A =2
+ . (21.1)
∂x ∂y

In simple words, where the squared gradient magnitude (i.e., ∇ A2 ) is large, it
is an indication of an edge. The Laplacian operator (∇ 2 ) is another local differential
operator that is used for edge detection, expressed by

∂2 A ∂2 A
∇2 A = + (21.2)
∂x2 ∂ y2

This second-derivative operator retains the information as which side of the edge
is brighter. The zero crossing of ∇ 2 A indicates that the intensity edges of the image,
and the sign on each side of a zero crossing indicates which side is brighter.
The Machine Vision Systems
The images input to machine vision systems are digitized with respect to both the
space and intensity, and are available as an array A[x, y] of discrete intensity values.
The approximations based on finite difference are used to estimate the derivatives
for computing the local differential operators. A discrete 1D sampled function is
represented as a vector f[i]. The first-order derivative of this vector is df/d x, which
can be approximated as f[i + 1] − f[i]. And, the second-order derivative is computed
as an approximation f[i − 1] − 2f[i] + f[i + 1]. Hence, the gradient magnitude can
be approximated as
 2  2
∂A ∂A
+ ≈ (A[ j + 1, k + 1] − A[ j, k])2 + (A[ j, k + 1] − A[ j + 1, k])2 .
∂x ∂y
(21.3)
682 21 Machine Vision

Image Smoothing and Filtering

Filtering of images in computer vision is carried out using convolution operation.
Consider a function h(x, y) defined in terms of f (x, y), and g(x, y), as
 ∞  ∞
h(x, y) = f (x − ξ, y − η) g(ξ, η)dξ dη (21.4)
−∞ −∞

In the above, h is called the convolution of f and g, and expressed as h = f ⊗ g,

and the value of h at any given point (x, y) depend on the values of f and g at all
points. However, this complex looking computation can be simplified, as convolution
is commutative and associative, hence the computations can be rearranged in any
order it is convenient, to compute them efficiently.
To carry out the discrete approximation for a convolution, the sum of products
can be represented as four nested loops over two arrays representing the sampled
functions f and g. Let the array g[i, j] be m × m, and the array f [x, y] be n × n,
for n > m. The initial values of indexes are 0. Algorithm 21.1 computes the discrete
convolution of sampled functions f and g.
The role of convolution is to smooth an image (i.e., to act as low-pass filter), to
handle the problem of high-frequency variations. The latter is indicated by sudden
change of intensity from a pixel to next pixel. Gaussian method is used for represen-
tation here. The Gaussian function G σ in 1D is expressed by

Algorithm 21.1 Compute discrete convolution of f and g

1: for x ← xmin to xmax do
2: for y ← ymin to ymax do
3: Let s = 0
4: for i ← 0 to m − 1 do
5: for j ← 0 to m − 1 do
6: s ← s + g[i, j] f [x − m/2 + i, y − m/2 + j]
7: end for
8: end for
9: h[x, y] = s
10: end for
11: end for

1 −x 2
G σ (x) = √ e 2σ 2 , (21.5)
2π σ

which is a canonical bell-shaped distribution, also called normal distribution. The

maximum value of this function is obtained at x = 0, and the area of the function
under the curve is 1. The parameter σ controls the width of function G σ , the larger
the σ , the wider the bell. The Gaussian function in 2D (x, y) is given by the following
function.
1 −(x 2 +y 2)
G σ (x, y) = e 2σ 2 . (21.6)
2π σ
21.6 Machine Vision Techniques 683

In discrete-based system, the values of integral steps over some range are used as
approximations, which are generally at ±4σ steps. These values are normalized so
that they sum to 1, just like in the case of continuous values, with integral as 1.

21.6.3 Middle-Level Vision

The middle-level vision techniques receives the images or results from the output of
low-level vision algorithms, and produce some output other than pixels, within the
coordinate systems of the original image. One goal of the middle-level vision is to
extract the 3D geometric information from the images, which are in fact of 2D—
usually referred to as shape-from x, i.e., recovering 3D structure from 2D images.
In fact, presence of a shading in an image reveals information about 3D objects. An
example is, perceiving the shape of a sphere from its image, which we conclude as a
solid rather than a circular disk. This perception is due to the uniform change taking
place in brightness away from the light source [9].
Information about 3D shapes of objects is perceived due to the mirror-like reflec-
tions from the objects. Yet, one more source of 3D shape information is due to change
in location of an object in an image, from one image to another image of the same
object’s views. The commonly used techniques for extracting image shapes from
multiple images are: 1. stereopsis and 2. structure-from-motion. Apart from extrac-
tion of shape-from images, another goal of middle-level vision is to extract structural
description of images. What relations exist between these structures can be found
out using grouping methods, called perceptual grouping. The latter is responsible
for recovering non-accidental alignments of image primitives, like colinear line seg-
ments or co-circular arcs.
Stereopsis
Computer stereo vision is the extraction of 3D information from digital images.
By comparing the information about same object about a scene from two panels.
The human beings compare two images, based on their being superimposed in a
stereoscopic device, such that the image from the right camera being shown to the
observer’s right eye and that from the left camera is shown to the left eye. In a
basic stereo-vision system, two cameras are maintained for observing a scene. The
idea behind this system is that objects closer to camera will appear more displaced
between two views than those which are farther away. By comparing the two images,
it is possible to find out the relative depth information in the form of a map. This map
encodes the difference in horizontal coordinates of corresponding image points.
For any given point in a scene, the amount of this point’s motion between two
views is called disparity. If the camera system is calibrated in world coordinates, then
the actual distance to a point can be found out using the magnitude of its disparity.
Given any two images of a scene, the disparity between two is often the intensity
image having bright points corresponding to larger intensities, i.e., the things that
are closer to the cameras.
684 21 Machine Vision

Example 21.1 Finding out the dept using stereo vision.

Let a simple pinhole camera model has a focal point (optical center) o, such that all
the rays of light pass through this, and are projected onto an image plane A(x, y)
(see Fig. 21.4).
Note that in case of human eyes the image plane is retina of eye. In case of camera,
it is medium (i.e., the 2D surface) inside the camera where image is formed. The
optical axis (parallel to axis z in this figure) of cameras is perpendicular to the image
plane A(x, y) and passes through the focal point o. The focal length f (shown as
(ol , ol ) and (or , or )) is the distance from the optical center o to the image plane
A(x, y).4 We will consider a simple stereo camera geometry where the optical axes
of the two cameras are parallel to one another, and perpendicular to the baseline
b, that connects the two cameras’ centers ol and or . We assume that the both the
cameras have equal focal lengths, and the origin of the world coordinate frame is
placed along the baseline, at the point which is in equal distance between the two
camera centers (at distance b/2 from each ol and or ).
Let the origin of the coordinate system for the left image plane L be the projection
of its optic axis, i.e., ol . Similarly, the origin of the right image plane R is at or . Each
of the L and R planes are (x, y) planes.
We consider a point p = (x, y, z) (see Fig. 21.4) in the world coordinate, that
is projected onto plane L at location pl = (xl , yl ) and onto plane R at location
pr = (xr , yr ). From the geometry of two cameras, we have

xl x + b/2
= , (21.7)
f z

xr x − b/2
= , (21.8)
f z

p
x

ol b or
Origin of world
coordinate frame
f
f

pl ol or pr

Fig. 21.4 A simple stereo camera geometry in image plane A(x, y)

4 The plane (x, y) is formed by lines (ol , or ), and axis y, the latter is perpendicular to plane of paper.
21.6 Machine Vision Techniques 685

yl y y
= r = . (21.9)
f f z

Note that in this simple camera geometry, only the x location (xl and xr ) of a
projected point differs between the left and right images, and the y location (yl and
yl ) of a given point in space is the same for both images. The disparity is distance
between pl and pr , is just xl − xr . We have from Eqs. 21.7 and 21.8,

xl − xr b
= . (21.10)
f z

Hence, if b and f are known, depth z of the point p can be computed from the
disparity, giving the third dimension, hence creating a 3D view using 2D image. If
b and f are unknown, the relative depths of points can be determined, and we can
grasp the relative 3D views of two images. But, in this case, their absolute distance
from the camera cannot be determined. 

21.6.4 High-Level Vision

The High-level vision systems are used to make abstract decisions, or based on the
visual data they perform classification of objects. These methods usually make use
of the outputs produced by low-level or middle-level vision algorithms discussed
above. The high-level vision systems have main applications as object recognition
systems, or as object tracking systems.
• Object recognition systems. These systems are used to find out whether or not a
particular objects is present in the scene, and also determine the locations of other
objects with respect to camera;
• Object tracking systems. As the name implies, the object tracking systems follow
a moving object in a video sequence, and hence can guide a mobile robot or an
autonomous vehicle.
In the following, we will discuss the principles used in the first application.
Object Recognition Systems
The Object recognition systems compare new image to stored object models to
determine whether any of the models is present in the image. Many object recognition
systems perform both the recognition and localization tasks, i.e., identifying what
is the object present in the image, and also recovering its location in the image or
world coordinate systems. The location of an object is called its pose; it is generally
specified using a transformation that maps the model coordinate system to the image
or world coordinate system.
The object recognition systems can be classified based on the type of problems
they solve. A simple recognition task requires identification of 2D objects that are
686 21 Machine Vision

fully un-occluded (all are fully visible), they appear in a uniform background, and the
lighting conditions are controlled, i.e., there are no reflections or shadows in the view.
The large majority of industrial inspections fall in this category of object recognition,
and can be handled accurately using the commercially available systems.
The object recognition problems are difficult, in the following conditions:
Background of the objects is highly textured,
Lighting conditions are unknown and uncontrolled,
Scene comprises too many objects,
Number of objects models are very large in the storage, and
Some objects may be touching and occluding other objects.
In our discussions, we will consider methods to handle images with multiple
objects, the objects are partially occluded, and there is some amount of cutter in the
background. The methods that can solve the object recognition with the presence
of these challenges, first extract geometric information, like intensity edges from an
image. Then, the next step is to compare 2D geometry with 3D geometry.
An alternative method is computing invariant representations of the image geom-
etry, i.e., it remains unchanged in spite of changes in the viewpoints. Such represen-
tations can be used to create an index of object models in a library of collections of
object models. One of the challenges in geometric recognition is to find out which
portion of the image corresponds to a given object from the library collections. The
recognition problem is often defined as recovering a correspondence between local
features of an image and an object model. There are three classes of approaches as
how to search possible match between a model and image feature.
1. Transformation space methods. These methods use the space of possible trans-
formations that map the model to the image.
2. Correspondence methods. These methods use the space of possible corresponding
features to match with the image.
3. Hypothesize and test. The hypothesize and test methods consider k-tuples of
model and data features for matching with the image.
We will be discussing the correspondence-based methods in more detail.

Correspondence Search
It is a tree search to find out which model’s features match with the image features.
Considering a set of image features as A = {a1 , . . . , an }, and a set of model features
as M = {m 1 , ..., m r }, an interpretation of the image can be expressed as a set of pairs,
N = {(m i , a j ), ...}. The interpretations states, which model features correspond to
the image features.
If the model features are unclear (hidden or occluded), and some features are
outside the image features, in that case the set N is any subset of all “pairs of model
and image features.” A search method used, called interpretation tree approach
makes use of pruned search in otherwise exponential space of possible interpretations.
In these interpretations, the method makes use of all possible ways of pairings of
21.6 Machine Vision Techniques 687

models and image features. In simple terms, the search method makes use of pairwise
relations between features to prune a tree of possible model and image feature pairs.
A traversal path in this tree corresponds to a sequence of interpretations.
To limit the number of pairs (m i , a j ) and (m p , aq ) in the interpretations, the
distances m p − m i  and aq − a j  must be equal, i.e., are within the allowable
error limits.
The search for interpretations of an image, a pruned-tree search algorithm is used,
which works as follows:
• Each level of the tree, other than root node, corresponds to an image.
• Each branch at a node corresponds to a model feature or there is special branch
called null face.
• Accordingly, each node of the tree specifies a pair: image feature at that level,
model feature taken from the previous level . This model feature may be null also.
• The three search carried out is Depth-First search (DFS). Any given node is
expanded only when it is pairwise consistent with all the nodes along the path
from the current node back to the root of the tree. That is, a given node is paired
with each node along the path back to the root, and for each such pair, if the
distance constraint is satisfied, the node is expanded otherwise not.
In the above, the null face branch is considered as always consistent.
A path from root to leaf node of the search tree indicates that zero or more model
features exists. However, a null branch path does not account for a model feature.
A path that accounts for k model features is called k-interpretation. Consider that a
reference level to filter out any hypothesis that does not account for enough model
features, is set at a threshold, say k0 . In this case, a matcher algorithm reports only
those k interpretations whose threshold is k > k0 . All these k-interpretations are
guaranteed to be pairwise consistent. Hence, an additional step, to verify the model
is carried out. This step estimates the best transformation for each k-interpretation
and checks that this transformation brings each mode feature within some error range
of each corresponding image feature.

21.7 Indexing and Geometric Hashing

The significance of indexing in vision is to match the images efficiently with the
models stored in the repository. Having this facility, it is possible, in principle, to
search an object in a library of stored models, with time complexity, that is indepen-
dent of number of models stored. To carry out the indexing, it requires finding the
geometric invariants in model recognition—those attributes of images that do not
change due to changes of viewpoints of objects. One key difference of this model
indexing with hashing is that, various instances of the same object in different images
will not generate exactly the same key due to sensing uncertainty.
There are number of methods to use invariants in object recognition. These are
categorized based on: geometric differential, photometric, and thermal properties
688 21 Machine Vision

of images. The indexing methods for structures make use of combinations of sim-
ple features such as points and segments. Other geometric techniques use invariant
properties of curves and co-planar conics. The advantage of using curve features
is lower combinatorial complexity, but it has disadvantage of requiring to extract
features from noisy and cluttered images. However, the intensity information from
images provides their richer description than the geometric features, like points and
line segments.
The geometric hashing is used to explain the invariants in recognition. For this2D
affine transformation is used, with geometric hashing using three points to define a
coordinate system, and with respect to these points, the other points are encoded in
invariant manner. The geometric hashing comprises the following two basic steps:
a. Construction of a model hash table, and
b. Matching of model to image.
The hash table stores redundant and the representation of each object that are
transformation-invariant. For any given model, each ordered triple, m 1 , m 2 , m 3 forms
an affine basis having origin as o = m 1 , and axes x = m 2 − m 1 , y = m 3 − m 1 . For
each such basis every additional model point m i is expressed as (αi , βi ), such that
m i − o = αi x + βi y. The basis triple (o, x, y) and the point m i are then stored in
a hash table using affine invariant indices (αi , βi ). For r number of model points, a
table results with O(r 4 ) entries, i.e., space complexity is combinatorial. The table
is formed using buckets instead of purely a hashing scheme. This is because, due to
the uncertainty in sensing in real-life data makes it difficult to use exact values for
retrieval purpose.
At the time of recognition, the hash table is searched to find out what models are
present in the image. The algorithm steps for recognition of an object are as follows.
i First, the image points are rewritten in terms of an image basis.
ii As the next step, corresponding to an instance of a model that model basis will
be retrieved from the table.
iii An image basis is formed using each ordered triple of image points s1 , s2 , s3 ,
such that origin is O = s1 , and axes are X = s2 − s1 , Y = s3 − s1 .
iv For this image basis, each additional image point si is written as (αi , βi ), such
that si − O = αi X + βi Y .
v The indices (αi , βi ) are used for retrieving matching model basis from the hash
table. Along with retrieving each model basis, corresponding counter is incre-
mented in a histogram.
vi When all the retrieved points for the model basis have been considered for an
image basis, the histogram contains the votes for those model bases which could
match to the current image basis (O, X, Y ).
vii If peak in the histogram for a given model basis (o, x, y) is above a threshold, the
corresponding basis is selected as a potential match. When a next image basis is
chosen, the histogram counts reset.
21.8 Object Representation and Tracking 689

21.8 Object Representation and Tracking

The availability of inexpensive high-quality video cameras and the increasing

demand for automated video analysis, has resulted to lot of interest for object track-
ing. The video analysis comprises three key steps: 1. detection of moving objects, 2.
tracking of these objects from one image frame to another frame, and 3. analysis of
the object tracks to recognize the object behavior, and sometimes its geometry also.
The goal of object tracking is to generate trajectory of an object by locating
its positions in every frame of the video, over a time. Object tracking also provides
information about the region occupied by the object at every time instant in the image.
The two tasks, i.e., detecting the object, and establishing correspondence between the
object instances from frame to frame, can either be performed separately or jointly.
When only one job is required to be performed, i.e., to detect an object, the possible
object regions in every frame are obtained using an algorithm for object detection.
Then, as next step, the tracker can find the correspondence in frame sequences.
In the second approach, when detection and tracking are done together, the object
regions and the correspondence is jointly estimated through iteratively updating the
object location and region information. The information about location and region
are obtained from the previous frames. What type of model is selected to represent
an object’s shape can decide the type of motion or deformation it can undergo. For
instance, if an object is represented as a point, then, only a translational model can
be used for this object [10].
In situations where a geometric shape representation, e.g., an ellipse is used for
the object, parametric motion models like affine5 or transformations like projective
are preferred. The motion of any rigid object in a scene can be approximated by
these representations. For nonrigid objects, silhouette or contour-based approach is
preferred, being the most descriptive form of representation. Both the parametric
as well as nonparametric models can be used to specify their motion using these
approaches.
It is possible to simplify the object tracking job by restricting the motion of
the object or its appearance, or both of these; for example, assuming that object
motion is uniform, i.e., constant velocity, or constant acceleration. The objects can be
represented in many ways, but some representations are considered as more suitable
than others. For instance, fish in an aquarium, boats in the river, vehicles on a road,
and birds in the air, are the sets of objects that may be important to track in a
specific domain. Depending on the type of objects, they may be represented by their
appearances and shapes.
• Points. An object is represented by a point as its centroid (see Fig. 21.5a), or can
be represented by a set of points, as shown in Fig. 21.5b.

5 Affine transformation is a linear mapping method in geometry, such that it preserves points, straight

lines, and planes. For example, after an affine transformation, a set of parallel lines remain parallel,
say, a transformation from parallelogram to rectangle and vice versa. Affine transformation is
typically used to correct for geometric distortions or deformations that occur with nonideal camera
angles.
690 21 Machine Vision

Fig. 21.5 Object representations-I: a Centroid b Multiple points c Rectangular patch

(a) (b) (c)

Fig. 21.6 Object representations-II: a Elliptical patch b Part-based multiple patches c Object
skeleton

• Primitive geometric shapes. An Object can be represented using a shape, e.g., a

rectangle, or an ellipse (see Fig. 21.5c, 21.6a).
• Articulated shape models. The articulated objects comprise body parts, which are
held together with the help of joints. For example, a pet’s body is an articulated
object with hands, legs, head, feet, and the torso connected together by joints (see
Fig. 21.6b).
• Skeletal models. In these models, the object skeleton can be extracted by applying
medial axis transform to the object silhouette (see Fig. 21.6c).
• Object contour. The contour representation defines the boundary of an object (see
Fig. 21.7a, b).
• Object silhouette. The region inside the contour is called the silhouette of the object
(see Fig. 21.7c).
The object tracking is carried out to estimate the trajectory of an object in the
plane of the image, as the object moves in the scene. During the tracking, the tracker
assigns labels to the tracked object to mark its positions in different frames of a video.
Apart from this, depending on what is the domain used, a tracker can also provide
object-centric information, such as area, shape of the object, and about orientation
of the object. However, the tracking process sometimes turn out to be complex, due
to reasons mentioned below.
21.8 Object Representation and Tracking 691

(a) (b) (c)

Fig. 21.7 Object representations-III: a complete object contour (dotted lines) b control points on
object contour c object silhouette

(a) (b) (c)

Fig. 21.8 Object tracking. a (i) Multipoint correspondence (ii) Kernel Tracking b Evolution of
contours c Silhouette Tracking and Evolution of contours

Complex object motions,

Articulated or nonrigid nature of objects,
Occlusion of objects in full or partially,
Complex shapes of objects,
Illumination changes in the scene,
Need of real-time processing of the scenes,
Loss of information due to projection of 3D world on a 2D images, and
Image noise.
The following is brief introduction to the commonly used tracking categories:
Point Tracking
In point tracking, objects detected in consecutive frames are represented by points.
The association between consecutive points is established based on the previous state
of the object, which may comprise object position and its motion. This approach needs
a separate external mechanism to detect the objects in every frame. An example of
correspondence between the objects is shown in Fig. 21.8a, with mapping through
broken lines.
Kernel Tracking
A kernel is concerned with the object shape and its appearance. For example, a kernel
can be a rectangular template or an elliptical shape with an associated histogram (see
Fig. 21.8a, with rectangles). The objects are tracked by computing the motion of
692 21 Machine Vision

the kernel in consecutive frames. The motion is usually in the form of a parametric
transformation, like rotation, translation, or affine.
Silhouette Tracking
Object tracking can also be performed by estimating the object’s region in each frame.
Information encoded inside an object’s region is used by the Silhouette tracking,
which is either in the form of appearance density or shape models (i.e., in the form of
edge maps). Further, given an object model, the silhouettes are tracked by matching
the shapes (Fig. 21.8b), or they can be matched by contour evolution (Fig. 21.8c).
Both of these methods can be used for object segmentation, which is applied in the
temporal domains.
The following are major applications of object tracking.
• Video indexing. It comprises automatic annotation and retrieval of the videos in
multimedia databases.
• Motion-based recognition. Motion-based recognition has applications in identifi-
cation of humans based on gait and in automatic object detection.
• Automated surveillance. It is concerned with monitoring a scene to detect suspi-
cious activities or unlikely events.
• Human–computer Interaction. The HCI is concerned with applications, like ges-
ture recognition and eye-gaze tracking, for data input to computers.
• Vehicle navigation. It is used in video-based path planning and obstacle avoidance
capabilities.
• Traffic monitoring. It is concerned with real-time gathering of traffic statistics to
direct the traffic flow.

21.9 Feature Selection and Object Detection

The feature selection is a process of deciding those attributes of a picture that help
in identifying the object in the picture frame, while object detection or object iden-
tification is a process to identify the object in the frame, based on these features.
Therefore, it is necessary that the features decided should give a strong evidence of
the corresponding object.
Feature Selection
To efficiently distinguish the objects in the feature space, it is desired that the visual
features be unique. The selection of feature is related closely with the objects repre-
sentation. For example, for histogram-based representations, the color is a required
feature, and for contour-based representations, object edges are important features.
Many times, the features with various combinations are useful. Some examples of
features as given below [10].
21.9 Feature Selection and Object Detection 693

Color
The color, in appearance of an object, is primarily influenced by two physical features
of the object: (1) light source’s spectral power distribution, and (2) object’s surface
reflection properties. The color space used for an object is RGB (red, green, blue).
Since, the RGB does not have uniform perceptions, HSV (hue, saturation, value)—
more closer to uniform color space—is preferred as a representation to obtain uniform
perception.
Edges
Since the object boundaries are the cause of strong changes in image intensities,
the edge detection is used to identify these changes. The edges intensities are also
less sensitive to the level of illumination. Hence, the algorithms used for tracking
boundaries of the objects, make use of edges as the important representative features
of the image.
Texture
Texture of an object surface causes an intensity variation in the reflected light from
its surface, hence quantifies properties of the object surface such as smoothness
and regularity. The texture requires an additional processing step to generate the
descriptors.
An object’s representative features usually depend on the application domain;
however, the automatic feature selections is in common practice. The features are
either filter based or they are wrapper based. The methods that are filter based, make
use of general criteria, e.g., the features need to be correlated. On the other hand,
the wrapper methods select the features based on their usefulness in the concerned
problem domain, for example, classifying the performance using the subset of fea-
tures. An example of filter method, called Principle Component Analysis (PCA), is
useful in reduction of number of features. The PCA is based on transformation of
possibly a number of correlated variables into a smaller number of variables that are
not correlated, called principal components. The first principal component chosen
should account for maximum possible variability of the data, and each succeeding
component should account for most of the remaining variability of the components
possible. Adaboost algorithm is an example of wrapper-based method for selection
of discriminatory features for tracking a class of objects. This method identifies a
strong classifier using a combination of moderately less accurate and weak classi-
fiers. Having given a large set of features, it is possible to train one such classifier
for each feature [3].
Among all the features used for object tracking, color is the one most widely used
feature. The colors’ histogram can be used to represent the object’s appearance. In
spite of the popularity of color as an important feature, large majority of color bands
are sensitive to variation in illumination. Thus, in situations where this effect cannot
be stopped, other features are considered to model the object’s appearance.
694 21 Machine Vision

21.9.1 Object Detection

A tracking method always requires a mechanism for object detection, either in every
frame or when the object appears in the video for the first time. The commonly used
approach for object detection is to use the information from a single frame. However,
some method makes use of temporal information computed from a sequence to reduce
the false detection of objects. The temporal information is in the form of frame
differencing, i.e., it highlights the changes in regions from frame to frame. Based on
the object regions available in an image, the trackers establish correspondence from
one frame to next frame of same object, and generate the tracks [10].
Some commonly used methods for object detection as discussed below.
Point Detectors
The point detectors have application in finding the interest points in an image that
have expressive texture in their localities. An important characteristic of interest point
is its invariance with respect to changes in the illumination and camera viewpoint. A
procedure, called Moravec’s operator-based method, is used for finding the interest
points, which computes the variation in image intensity in a 4 × 4 patch in horizontal,
vertical, diagonal, and anti-diagonal directions, and selects the minimum of the four
variations as representative values for the window. A point in this case is declared as
interest point if the intensity variation in a local maxima is 12 × 12 patch.
Image Background Subtraction
It is possible to build a representation of a scene, called background model, and then
to find the deviation from the model for the incoming frame, to detect the object. If
there is a significant change in the image region from the background model, then
it indicates that the object is moving. This approach can also be used to detect if a
new object has been introduced into the scene, or some object has escaped from the
scene. In this method, the pixels representing the regions that undergone the change
are marked for further processing. In such cases, it is common to use connected
component algorithm to obtain the required objects in the image. The process used,
as it appears, is called background subtraction.
Figure 21.9 shows background subtraction results: The part (a) of figure shows
an input image of highway with moving vehicles, pedestrians, and nearby houses
as objects, the part (b) is reconstructed image after projecting input image onto the
eigenspace, and part (c) is an image obtained as difference of two images.
Image Segmentation
Aim of an image-segmentation algorithm is to partition an image into perceptually
similar regions. A segmentation comprises solution of two problems: 1. What is cri-
teria for good partition, and 2. What is procedure for achieving efficient partitioning?
In the following we discuss some of commonly used segmentation techniques that
are useful for object tracking.
21.9 Feature Selection and Object Detection 695

Fig. 21.9 Background subtraction. a input image with objects b reconstructed image c difference
image

Image segmentation using mean-shift clustering

This approach for image segmentation finds the clusters in the joint color+spatial
space, [u, v, w, x, y], where [u, v, w] represents color, and [x, y] represents the spa-
tial location in 2D system. The color may be either RGB or HSV. The algorithm
steps are as given below.
i. For any given image, large number of random hypothetical cluster centers are
initialized, which are chosen from a given data set;
ii. Each cluster center is moved to the mean of the data lying inside a multidimen-
sional ellipsoid whose center is the cluster center;
iii. Using the old and new cluster centers a vector, called as mean-shift vector, is
defined;
iv. This mean-shift vector is iteratively computed until the cluster centers are stabi-
lized (i.e., they do not change their positions).
Image segmentation using graph cuts
The image-segmentation problem can also be solved like partitioning of a graph.
Consider that V = {v1 , v2 , ..., vn } be vertices (pixels) in a graph (image), and the
graph is represented by G = (V, E). Let us assume that this graph is partitioned
into N disjoint subgraphs, with regions A1 , ..., A N , by its weighted edges E. The
partitioning is subject to the condition, that


N
Ai = V, (21.11)
i=1
696 21 Machine Vision

and
Ai ∩ A j = φ, for i = j. (21.12)

In any graph, the total weight of the pruned edges between its two subgraphs is
called a cut. The weights are typically computed based on color of the image, its
brightness, and texture similarity between the vertices. In one approach, minimum
cut (mincut) criteria is used, with the aim to find out the partition that minimizes a
cut. The weights in this approach are defined based on color similarity. However,
this approach has disadvantage of bias toward over-segmenting the image. This over-
segmenting is caused due to increase in cost of a cut with number of edges going
across the partitioned segments.
The problem of over-segmentation can be solved by using a normalized cut, where
the cut depends on the sum of edge weights in the cut, like before. In addition, the cut
also depends on the ratio of, total connection weight of the vertices in each partition
to weight of all vertices of the graph taken together. For the segmentation based
on images, weights between the vertices are defined by the product of the spatial
proximity and color similarity. Once the weights are computed between each pair of
vertices, a weight matrix W and a diagonal matrix D are computed, such that


N
Di,i = Wi, j . (21.13)
i=1

The segmentation is performed first by computing the eigenvectors and the eigen-
values of the generalized eigensystem, as per the expression,

(D − W)y = λDy. (21.14)

Next, the second-smallest eigenvector is used for dividing the image into two seg-
ments. For each of the new segment, this process is performed recursively until a
threshold is reached.
In a segmentation that is based on normalized cut, solution to the generalized
eigensystem for large images can be expensive in terms of memory requirements as
well as time requirements. However, this method requires far less number of manually
selected parameters, compared to that of mean-shift clustering-based method.

21.10 Supervised Learning for Object Detection

The supervised learning can be used to automatically learn different object views
from a set of examples to perform the object detection. Learning of different object
views exempts the need of storing the complete set of templates. The supervised
learning makes use of a set of examples to generate a function which maps any
input to desired output. A standard formulation of supervised learning is in the form
21.10 Supervised Learning for Object Detection 697

of a classification problem. A learner in the supervised learning approximates the

behavior of a function by generating an output either as continuous value or class
labels. When output is continuous value, it is called regression model, and when the
output is class labels, it is called classification model. For carrying out the object
detection, the learning examples comprise pairs: object feature, object class , with
both quantities manually defined.
Selection of proper features is important for effectiveness of the classification,
only those features need to be used that are able to discriminate one class from the
other class. Apart from the features, like, color and texture, other features are also
possible to use, such as area of the object, its orientation, and appearance. These
features can be used in the form of density function, for example, in histogram. Once
the features are finalized, all required appearances of an object can be learned through
supervised learning only. Some of the learning approaches to supervised learning are
decision trees, neural networks, and support vector machines. The learning methods
compute a hypersurface, which separates one object class from the other in a high
dimensional space.
For supervised learning to be useful, it requires a large collection of samples from
each object class, where each collection has been manually labeled. Along with the
supervised learning, co-training is also carried out to reduce the size of manually
labeled data. The co-training approach trains two classifiers using a small set of
labeled data, such that features used for each classifier are independent. Once train-
ing is complete, each classifier assigns unlabeled data to the training set of other
classifiers. Using this method, we start with a small set of labeled data with two sets
of statistically independent features, and the co-training can provide accurate clas-
sification rule. In the following, we discuss two approaches, i.e., adaptive boosting
and support vector machines for object detection.
Adaptive Boosting
This method combines many base and moderately accurate classifiers to produce a
more accurate classifier. The steps (of algorithm) of this classifier are as follows.
1. The training phase of the algorithm constructs an initial distribution of weights
over a training set;
2. A base classifier having least error is selected by a boosting mechanism. This
error is proportional to the weights of the misclassified data;
3. The weights associated with data, that are misclassified by the selected base
classifier, are increased;
4. The iteration step is repeated after selecting a new classifier that performs better
on the misclassified data, and the process is repeated, until there is no misclassified
data.
The simple operators can act as weak classifiers for object detection, e.g., a set
of thresholds. These classifiers are applied to the object features extracted from the
image. In an approach where adaptive-boosting framework is used, say, for detecting
the pedestrians in the presence of other traffic, perceptrons can be chosen to act as the
weak classifiers. These perceptrons can be trained on the image features extracted
698 21 Machine Vision

through a combination of temporal and spatial operators. Operators in the temporal

domains will be in the form of frame differencing, and they encode some form of
motion information (a required feature of pedestrians). When temporal domain makes
use of the frame differencing operator, there is less likelihood of false detection,
because it enforces the object detection in the region where the motion occurs.
Support Vector Machines
When used as a classifier, a Support Vector Machine (SVM)6 is used for classifying
the data into two classes by finding a maximum marginal hyperplane which separates
one class from the other. The margin of this hyperplane is maximized, as defined
by the distance between the hyperplane and its closest data points. The data points
that lie on the boundary of the margin of this hyperplane are called support vectors,
hence the name of this technique.
For detecting an object, these classes, separated by the hyperplane, correspond
to the object class (positive samples) and the non-object class (negative samples).
Using manually generated training examples, labeled as object and non-object, one
hyperplane is computed from among an infinite number of possible hyperplanes
using a method called quadratic programming.7
In spite of being a linear classifier, the SVM can also be used as a nonlinear
classifier by application of a technique, called kernel-trick,8 to the feature vector
extracted from the input. The kernel-trick is applied to a set of data that is not
linearly separable, to transform them to a higher dimensional space so that it becomes
separable.

21.11 Axioms of Vision

The term axiom (also called primitives) is considered from the mathematical usage,
using these as the basic building blocks it is possible to derive new constructs. This
approach has the advantage of abstracting the representation, the type, and the algo-
rithm details, within those components so that it is possible to combine them together.
In the vision axioms, each axiom performs a subset of the tasks such that these tasks
can be grouped together to accomplish a more complex task. The axioms are low-
level enough so that the majority of the subsets of vision problems can be represented
by them, and after combining them, it results in high-level solutions [5].
The axioms share common elements defined globally, such as images and ele-
mentary data types. The time, images, and colors are treated as continuous signals
in the axioms of vision. For example, color is RGB, with each channel represented

6 For more about SVM, refer p. 515 in Chap. 17.

7 Quadratic Programming is process of solving a linearly constrained quadratic optimization prob-
lem (minimizing or maximizing) a quadratic function of several variables subject to linear constraints
on these variables.
8 The technique of kernel-trick does not require explicit mapping used for linear learning algorithms

to learn a nonlinear function.

21.11 Axioms of Vision 699

in the interval of [0, 1]. The width and height of an image are represented in similar
way, with additional value for the aspect ratio. Since time is treated as a continuous
signal, it allows different inputs to the vision system.
The classes of common axioms for vision are: mathematical axioms, source
axioms, model axioms, and construct axioms. These are based in a problem-centric
taxonomy. Each axiom is further divided into three parts: a. description of its tasks,
b. its representations, and c. the processing it would perform when implemented.

21.11.1 Mathematical Axioms

The functionality is encapsulated in these axioms, such as to perform transformations,

optimizations, and other necessary formulations in machine vision.
Optimize
This axiom comprises optimization functions of dynamic programming, linear pro-
gramming, graph cuts, and greedy methods.
Transform
The necessary mathematical descriptions are contained in this axiom to carry out
the transformation of vision objects. It includes operations of linear algebra, geom-
etry, and other formulations necessary for computer vision. The transformation may
comprise the operations, like matrix multiplications and projections.
Similarity
The similarity axiom is used for recognition, correspondence, and tracking. It com-
prises various matrices for evaluating similarity.

21.11.2 Source Axioms

The source axioms are used for encapsulating the source information. Thus, they
also describe the acquired data or the source of the acquired data, such as cameras,
images, and properties of images like noise and illuminations in the images.
Noise
This axiom provides the description of noise in the data.
Light
The lighting models are required to approximate the lighting in a scene. It may have
simple systems like in computer graphics, or variation of intensity across images, to
add the effect of flash, or combining of the images, etc.
700 21 Machine Vision

Camera
The camera axioms provide the general description of the camera, and also include
other models and specific parameters for calibration, like focal length, principal point,
etc.
Image
The image comprises descriptions for low-level image processing techniques, and
representation of image, like resize and filter operations.
Blur
The blur is due to motion of camera or the object. The focus is deliberately used for
creative effect or measurement of depth from focus.

21.11.3 Model Axioms

The model axioms are used for representing abstract concepts used within computer
vision, such that they are accessible.
Model
This axiom contains descriptions of models used in the vision, like geometric model,
probabilistic model, or example-based model. A model unit can provide conversion
from one format to another format.
Object
This axiom provides types and the descriptions of objects.
Shape
It finds the similarly shaped regions within an image, which is carried out in con-
junction with matching and similarity axioms.
Texture
It performs synthesis and analysis with the help of texture descriptor. It also finds
regions of similar texture, with the help of matching and similarity axioms.

21.11.4 Construct Axioms

The construct axioms represent constructs used within vision for the purpose of
modeling and used as output to other modeling packages.
21.11 Axioms of Vision 701

Mesh
The mesh axiom is for centralized description of a mesh, using triangles, quads, etc.
A mesh unit provides an input–output system, for construction of meshes from an
image base or from an arbitrary basis in 3D, for reconstruction of output.
Grid
The grid axiom is used for providing N -dimensional grids’ description, for use in
vision. For example, a 2D grid would be a discretized image, and a 3D grid would
be a set of voxels,9 etc. A grid can be indexed at intersections of grid lines. A grid
unit also provides methods of grid construction, conversion, and for IO methods.
Algorithm Composition
The vision axioms can be combined together through loose coupling to form more
sophisticated algorithms. The basic axioms already encapsulate the sophisticated
concepts and algorithms; and combining these we can create higher level systems
such as correspondence systems, tracking, 3D reconstructions, etc. As an example,
for creating a simple 3D reconstruction such as visual hull, a camera can provide cal-
ibrations, grid can provide volumetric constructs, transformation will give projected
grid points, and mesh can create a model for output from volumetric grid.
It is possible to create a simple tracking system with direct use of axioms: ini-
tialization of this system is done using object, matching, and similarity. And then,
an object can be tracked through a sequence of images through use of move and
optimization axioms.
However, the algorithm composition is not suited to be used by developers,
because it requires expert knowledge of vision. Instead of that use, we look at top
most level, which is problem itself. This is because the problem is decomposed into
smaller level parts, and we use this as a basis for solving the problem. This simplifies
the implementation of computer vision using the axioms.

21.12 Computer Vision Tools

The open source computer vision library, the OpenCV is available under a Berkley
Software Distribution (BSD) license and it is free for use both by academic and
commercial users. The OpenCV has interfaces with C++, C, Python, and Java, and
it is supported by all major Operating Systems. OpenCV was designesd keeping
the computational efficiency as one of the goals, and with a strong focus on real-
time applications. The OpenCV is written in optimized C/C++ code, the library
can take advantage of the availability multicore processing in the computer being
used for running these library programs. The OpenCV is enabled with OpenCL, and
it can take advantage of the hardware acceleration of the underlying heterogeneous

9 Voxels: Each array of elements of 3D volumetric space.

702 21 Machine Vision

computer systems. Usage of OpenCV range from interactive art, to mines inspection,
to stitching maps on the web through advanced robotics [6].
The OpenCV is designed with other goals of building tools for solving computer-
vision problems. It comprises a mixture of low-level image processing functions
and high-level algorithms for the applications of, face recognition, face detection,
pedestrian detection, feature matching, and object tracking.
The following is an example of a simple program in OpenCV to load and display
a given image file, whose name is provided as input at Linux command line [7].
Example 21.2 Load and display an image file.

/* loaddisp.cpp, OpenCV ver 3.2.0 */

#include <opencv2/highgui/highgui.hpp>
#include <opencv2/core/core.hpp>
#include <iostream>
using namespace std;
using namespace cv;
int main(int argc, char** argv)
{
if(argc != 2)
{
cout << "error" << std::endl;
return -1;
}

Mat imagemat;
// Read image file
imagemat = imread(argv[1], CV_LOAD_IMAGE_COLOR);
// Check for invalid input
if(! imagemat.data)
{
cout << "error" << std::endl;
return -1;
}
//Create a display window
namedWindow("Display Window", WINDOW_AUTOSIZE);
//Show the image inside window
imshow("Display Window", imagemat);
//press any key to terminate the program
waitKey(0);
return 0;
}
21.12 Computer Vision Tools 703

The above program is compiled at command level in Ubuntu 18.04.1 by

g++ loaddisp.cpp -ggdb Jpkg-config --cflags --libs opnecvJ

and run by command

./a.out krc.jpg

On run, the screen will display the image in file krc.jpg. 

All modern smartphones and most tablets also contain one or more cameras, and
OpenCV is also available on both Android and iOS operating systems. With all these
components, it is possible to create vision applications for mobile platform also.

21.13 Summary

The vision system has significance due to the fact that, it is one of the essential
characteristics of living beings, and specifically, all the higher animals depend on
this ability for their survival. The vision gives us detailed information about and
around us, without much conscious efforts.
The goal of computer vision or machine vision is to extract the information from
images, so that one can produce a structure from image, e.g., a 3D view from source
images that are 2D. Having this available, it can recover the 3D model of an object,
which in turn can be useful in medical imaging to recover 3D objects, say tumor.
The methods can find best matches for stored objects of image data, for example,
we can search some individual person in a collection of library of photographs, or in
collection of video.
Initially, the vision systems were designed for military applications, like visual
guidance for autonomous vehicles, target recognition, etc. However, in recent many
applications have emerged in medical, for example, for preoperative scan for patients,
virtual reality, image-based data retrieval, assembly and material handling, visual
inspection, and photo interpretation.
Every object, other than mirror, scatters the light incident on it in all directions,
this reflected light is projected on the retina of our eye through a lens, and we are
able to recognize the object, through the interpretation by brain. The vision problem,
which is related to this statement is, given the 2D image, how to infer the shape of
the object that produced this image. However, it is a challenge to reason back from
image to the scene that created it. The recognition requires learning, i.e., mapping
image patterns to some internal representation—transformation of scene to image.
This requires construction of models, i.e., general description of each object that is
to be recognized. A model includes shape, texture, and context knowledge.
Going from image-to-scene is a challenging task. A vision system receives only the
reflected brightness as input in different intensities. To convert these brightnesses into
world measurements, we first need to understand how objects are mapped into image
704 21 Machine Vision

brightnesses. The light received by individual sensor call is a complex interaction

of: position and object’s orientation, its 3D shape, reflection properties of its surface,
properties of light source as well as its spectral sensitivity.
For vision, what is required is inversion of brightness into scene parameters, using
any of the following methods: inversion by fixing scene parameters, by restricting
problem domain, or by acquiring additional images.
A general approach to computer vision is that a vision process ends with 3D recon-
struction of shapes by means of geometric primitives at symbolic level, linguistic
level, and intermediate. However, the interpretation of the conceptual categories at
linguistic levels involves problems as the concepts hardly correspond to classic cat-
egories that can be described in terms of necessary and sufficient conditions. In fact,
the membership of the categories is not 0 and 1. Hence, knowledge in conceptual
categories is affected by measurement errors. The neural networks make it possible
to avoid the exhaustive description of conceptual categories, accordingly, they are
better fit as models.
Depending on the complexities involved, the vision is classified as low-level vision
(image and output are in same coordinate system), medium level vision (input is low-
level vision, and output is something other than pixels of the image, say 3D image),
and high-level vision (input is low and middle-level vision algorithms, and output is
abstract data structures).
One of the applications of vision system is indexing of images, or scenes in a video.
Having this facility, it is possible to lookup an object in a library of stored models,
and that too in time that is independent of number of models in storage. To carry out
the indexing, it requires finding the geometric invariants in model recognition, i.e.,
attributes of images that do not change due to change of viewpoints of the object. For
indexing, we make use of geometric hashing approach, which has two basic steps:
1. construction of a hash table for models, and 2. matching of model to the current
image.
One important area of vision systems is object tracking. Due to lot of demand for
automated video analysis, interest in video analysis has increased. The video anal-
ysis has three steps: a. detecting the moving objects, b. object tracking from frame
to frame, and c. objects’ analysis to recognize their behavior. The major applica-
tions of object tracking are: automated surveillance, video indexing, motion-based
recognition, traffic monitoring, Human–Computer Interaction (HCI), and vehicle
navigation.
The tracking process is a complex task due to loss of information caused by
projections, noise in images, partial and full object occlusions, complex shapes of
the objects, and the requirements of real-time processing.
To detect an object, point detectors are used to: find points of interest, subtraction
of image background, and segmentation of the image. Object detection is performed
by learning different views from a set of examples using supervised learning. A
standard formulation of supervised learning is: classification problem, such that the
learner approximates the behavior of a function by generating an output in the form of
either continuous value, called regression, or labeling of classes, called, classification.
21.13 Summary 705

The support vector machines and adaptive boosting are the examples of supervised
learning.
The complexity of vision systems, where basic primitives are identical, for exam-
ple, in 3D objects, and in dynamics of objects, can be simplified by defining and
constructing axioms of vision. These are the basic building blocks from which we
derive new constructs. The axioms are ways for representations, as well as corre-
spond to certain algorithms within these, and are useful for common subtasks in
vision systems.
The common axioms in vision are: mathematical axioms (transformation, opti-
mization, linearity), source axioms (camera, image, light), model axioms (object,
shape, texture), and construct axioms (mesh, grid, algorithm).
One of the important tool for Vision system research is the open source computer
vision library, OpenCV released under a BSD license, is free for academics as well
as commercial use. It provides C++, C, Python, and Java interfaces, and has been
designed for efficiency of computation with a focus mainly on real-time applications.

Exercises

1. Suggest some experimental observations about human vision that support the
idea that vision is graphics — what we see is explicable only partly by the
optical image itself, but is more strongly determined by top-down knowledge,
model-building, and inference processes.
2. Explain in your own words, with supporting geometry, as how the human vision
system inverts the image formed on the retina of eye(s) to visualize the object.
3. How the vision is a learning and cognition process? Justify.
4. Why it is deterministically not possible to recognize an object given an image?
5. We can always map the coordinates from an object to its image, but mapping
the image coordinates to object is challenging. Why?
6. Suggest any five new applications of machine vision, which are not discussed in
this chapter.
7. Most smartphones available now have feature of face detection. Why this feature
is important? Explain the process of face detection in your own words.
8. What are the syntax and semantics information in a vision? Explain in brief.
9. How the axioms of vision are helpful in machine vision?
10. What is image inversion? Explain some of the techniques for this, and the sig-
nificance of each of them.
11. Give proper examples of low-level, middle-level, and high-level visions.
12. Like, two stereo cameras, our eyes are able to find out the depth of vision, e.g.,
how far roughly it is from the eyes. Write an algorithm, to compute this depth
of vision.
13. Does the width between the centers of human eyes, anyway make difference in
computing accuracy of depth of vision? Give a proof for this, based on some
computations.
706 21 Machine Vision

14. Given any image having number of human faces, write an algorithm, to count
the number of faces in the given image.
15. Write an algorithm for panorama-stitching for given n number of images.
16. Suggest various approaches to speedup the processing of image transformations.

References

1. Charniak E, McDermott D (1998) Introduction to artificial intelligence, Addison-Wesley pub-

lication
2. Chella A et al (1997) A cognitive architecture for artificial vision. Artif Intell 89:73–111
3. Felzenszwalb P et al (2013) Visual object detection with deformable part models. Commun
ACM 56(9):97–105. https://doi.org/10.1145/2494532
4. Grimson WEL, Mundy JL (1994) Computer vision applications. Commun ACM 37(3):44–51
5. Miller G et al (2011) A conceptual structure for computer vision. In: Canadian conference on
computer and robot vision. https://doi.org/10.1109/CRV.2011.29
6. https://opencv.org/. Cited on Dec 2017
7. Pulli K et al (2012) Real-time computer vision with OpenCV. ACMQUEUE 10(4):1–17
8. Suetens P et al (1992) Computational strategies for object recognition. ACM Comput Surv
24(1):5–61
9. Tucker AB Jr (1997) The computer science and engineering handbook, CRC Press
10. Yilmaz A et al (2006) Object tracking: a survey. ACM Comput Surv 38(4):1–45
Further Readings

1. Auger A, Doerr B (2011) Theory of randomized search heuristics—foundations and develop-

ments. Series vol 1, Theoretical Computer Science, World Scientific
2. Briggs R (1985) Knowledge representation in Sanskrit and artificial intelligence. AI Mag AAAI
6(1):32–39
3. Chakbarti S (1999) Mining Web’s link structure. Computer 4:60–67
4. Golding AR et al (1996) Improving accuracy by combining rule-based and case-based reason-
ing. Artif Intell 87:215–254
5. Jansen BJ, Spink A (2007) Sponsored search: is money a motivator for providing relevant
results? Computer 5:52–57
6. Jansen BJ et al (2009) The components and impact of sponsored search. Computer 98–101
7. Joseph M et al (1999) Interactive data analysis: the control project. Computer 4:51–58
8. Koehler J (1996) Planning from second principles. Artif Intell 87:145–186
9. McKinly R (2013) Proof nets for Herbrand’s theorem. ACM Trans Comput Logic 14(1)
10. Minsky M (1970) Form and content in computer science. J ACM 17(2):197–215
11. Natalia D, Iguez iR et al (2014) A survey on ontologies for human behavior recognition. ACM
Comput Surv 46(4)
12. Ritchie GD, Hanna FK (1984) AM: a case study in AI methodology. Artif Intell 23:249–268
13. Ruspini EH (1991) On the semantics of fuzzy logic. Int J Approx Reason 5:45–88
14. Shankar N (2009) Automated deduction for verification. ACM Comput Surv 41(4). https://doi.
org/10.1145/1592434.1592437
15. Sowa JF (1997) Knowledge representation: logical, philosophical, and computational founda-
tions. PWS Publishing, Boston, Mass
16. Vishnupriya R, Devi T (2014) Speech recognition tools for mobile phone—a comparative study.
In: IEEE International Conference on Intelligent Computing Applications (ICICA). https://doi.
org/10.1109/ICICA.2014.93

© Springer Nature India Private Limited 2020 707

K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7
Index

A AQE
Abduction, 35 working, 583
Acquiring additional images, 679 Argumentation theory, 27
Action, 14 Aristotle, 6
Agent Communication Language (ACL), Artificial consciousness, 9
493 Artificial Intelligence (AI), 1
Agent program Goals, 4
semantics, 495 Roots, 5
Agents Sub-field, 10
adaptive agents, 473 Artificial Neural Networks (ANN), 430
agent-based software engineering, 486 Boltzmann learning, 434
architecture, 479 competitive learning rules, 435
buy and sell, 487 error-correction rules, 433
classification, 472 Hebbian rule, 435
collaborative agents, 474 Assertion Box (ABOX), 201
cooperative agents, 474, 481 Atomic formula, 56
coordination, 480 Attribute language-AL , 198
decision maker, 488 Augmented network, 284
description languages, 497 Automated planning, 447
dynamic coalition formation, 482 Automatic Speech Recognition (ASR), 651
mental level modeling, 489 acoustic model, 659
mobile agents, 474, 499 algorithms, 656
personal agents, 474 hypothesis search, 658
planning agents, 450 Julius, 663
prisoner’s dilemma, 483 language model, 658
proactive agents, 474 lexicon, 658
reactive agents, 450 resources, 653
single-agent systems, 476 tools, 662, 664
smart agents, 474 Automatic Speech Recognition (ASR)
static coalition formation, 482 toolkit
taxonomy, 473 CMU-Sphinx, 665
Agent types, 450 Deep Speech, 665
Alan M. Turing, 7 HTK, 666
Algorithm Julius, 663
unification, 79 Kaldi, 664
Algorithm=logic+program, 112 Axiom, 35
Analogical reasoning, 37 construct, 700
© Springer Nature India Private Limited 2020 709
K. R. Chowdhary, Fundamentals of Artificial Intelligence,
https://doi.org/10.1007/978-81-322-3972-7
710 Index

mathematical, 699 squared error algorithms, 530

model, 700 Cognition, 675
of extensibility, 45 Cognitive architecture, 676
of infinity, 46 Cognitive modeling, 148
of power set, 46 Cognitive science, 7
of the null set, 46 Coherent semantic groups, 180
of unordered pairs, 46 Combinatorial game theory, 304
source, 699 Commonsense knowledge, 152
vision, 698 Complementary pair, 33
Axiomatic approach, 43 Completed tableau, 34
Axiomatic system Complex concepts, 197
consistent, 44 Computation, 7
inconsistent, 44 Computer vision, 670
model, 44 tools, 701
Axiomatics, 45 Concept-forming operators, 197
Concept language, 197
Concept learning
B mutual online, 389
Backtracking, 288 Concepts, 196
algorithm, 288 Concept types, 181
Backups-Naur Form, 29 Conceptual Dependency (CD), 204
Backward chaining, 38, 122 inferences, 209
algorithm, 99 primitives, 205
Backward reasoning, 120 Conceptual graphs, 185
Bag-of-Words (BOW), 574 Conceptualization, 149
Bayesian networks, 344 Constraint
Bayes theorem, 340 synthesizing, 281
Berkley Software Distribution (BSD), 701 extended theory, 283
Blocks world, 54 Constraint satisfaction, 273
Brain models, 13 Constraints Satisfaction Problems (CSP),
Branch-and-bound 273
algorithm, 255 algorithms, 287
constraints, 277
constraints propagating, 293
C forward checking, 294
Case Based Reasoning (CBR), 104 heuristics
CD degree, 294
content theory, 211 problem-solving, 280
parser, 207 representation, 276
CD vs Semantic Nets, 211 solution approaches, 285
Classification, 534 theoretical aspects, 298
Classifier, 383 variables, 279
Clausal form, 57 Cost function, 263
Clause, 56 Cryptarithmetics, 295
Clustering, 518 CYC, 149
fuzzy, 534
hard, 534
steps, 523 D
techniques, 531 Daemons, 212
traditional, 523 Data clustering, 519
Clustering algorithms Data mining, 507
nearest neighbor algorithm, 528 algorithms, 515
partitional algorithms, 529 apriori-based algo., 544
Index 711

association rule, 519, 537 Formal system, 15

decision tree, 536 Formula
domains, 509 interpretation, 31, 32
goals, 511 satisfied, 32
scientific applications, 549 unsatisfiable, 32
sequential pattern min. algo., 541 Forward chaining, 38, 123
Data mining algorithms Algorithm, 93
evolution, 512 Forward vs backward chaining, 100
Data streams, 514 Four queens problem, 291
Data structures, 117 Frame
Decision tree, 517 case study, 192
Deduction system, 92 inheritance hierarchies, 189
Default slots, 190
notion, 168 Frame language, 191
Default logic role, 194
syntax, 169 Frames, 188
Default reasoning Fuzzy
algorithm, 170 composition relation, 363
Dempster–Shafer Theory (DST), 356 composition rule, 364
evidence, 356 fuzzy subsets, 365
Depth-First Search (DFS) Fuzzy graph operations, 367
edge branching factor, 226 Fuzzy graphs, 365
Derivation-tree, 30 Fuzzy hybrid systems, 369
Description logic, 195 Fuzzy logic, 361
value restrictions, 202 inferencing, 364
Disagreement set, 79 Fuzzy relation, 366
Discourse analysis, 632 Fuzzy rules, 365
DL, 195 Fuzzy sets, 361
inferences, 203 Fuzzy system, 13
reasoning, 203
DL knowledge representation
architecture, 201 G
Document processing, 630 GA
applications, 261
mutation operator, 261
E Game playing
Earl Stanhope, 6 strategies, 306
Emile Borel, 304 Games
Engineering, 12 n-person, 316
Euclidean geometry, 45 equilibrium Points, 316
Evolution, 8, 13 classification, 305
Exhaustive search, 232 complexities, 318
Expert Configurator (XCON), 102 dynamic games, 312
Expert systems, 12 Grundy’s game, 320
historical events, 329
non-zero-sum, 307
F of imperfect information, 312
Feature extraction, 525 of perfect information, 312
First Order Predicate Logic (FOPL), 195 static games, 312
First-order logic vs axioms, 154 Tic-tac-toe, 321
Fixing scene parameters, 678 two-player
Fluents, 159 representation, 317
Formal logic, 16 strategies, 310
712 Index

zero-sum, 307 maintenance, 568

Game theory, 303 term distributed Arch., 596
Generalization and abstraction, 37 Indexing, 565
Generalized Best-First Search (GBFS), 245 images, 704
Generate and test, 288 Inductive programming, 386
Genetic Algorithm (GA), 13, 259 Inference rules, 41
Geometric hashing, 687 Inferences, 66
George Boole, 6 Information extraction, 630
Gödel-Bernays (GB), 45 output template generation, 633
Gottlob Frege, 6 Information Retrieval (IR), 557
Grammar automatic query expansion (AQE), 579
ambiguous, 619 Bayes inference algorithm, 589
Chomsky hierarchy, 616 Bayes networks
classification, 616 dependent topics, 592
structured descriptive, 635 Bayes probabilistic inference, 588
transformational, 617 Bayesian networks, 587
Graph coloring, 275 query & document representation,
Ground clause, 56 587
Ground literals, 56 Boolean model, 561
concept-based, 574
concept-based algorithms, 578
H cross language IR, 582
Herbrand instantiation, 70 distributed IR, 595
Herbrand’s Universe, 57, 68 fuzzy logic-based IR, 570
Herbrand theorem, 71 information filtering, 582
Heuristic multimedia IR, 581
approaches, 255 probabilistic model, 569
Heuristic search, 239 query refinement, 585
Hidden Markov Model (HMM), 353, 660 question answering, 581
relevant feedback, 585
High-level machine vision, 685
semantic IR, 594
Hilbert, 43
semantic IR on Web, 592
HMM
strategies, 560
parameters, 661
Vector space model, 563
Horn clauses, 114
word sense disambiguation, 586
Horn rule, 70
Instantiations, 60
Human knowledge creation, 184
Intelligent agents, 471
Hypernyms, 153
Interpretation, 66
Hyponyms, 153
procedural, 72
Isomorphic, 44
I
Image filtering, 682 J
Image inversion, 678 John von Neumann, 7
Image smoothing, 682 Joint feature map, 425
Image-to-scene, 677
Incompleteness theorem, 6
Inconsistency K
arc, 281 Kinship relations, 53
path, 282 K-Nearest Neighbor (K-NN), 426
Index K-nearest neighbor algorithm, 425, 426
concept-based, 575 Knowledge engineering, 158
construction, 565 Knowledge representation, 2, 16, 143
document distributed Arch., 595 Kurt Gödel, 6
Index 713

L IBL, 438
Language and reasoning, 151 instance-based learning, 437, 438
Language, logic, ontology, 151 Machine vision, 669
Learning application, 671
argument based, 387 classification, 675
by analogy, 401 hashing, 687
by concept hierarchies, 396 local edge detection, 681
discovery learning, 396 restricting problem domain, 678
explanation based, 406 techniques, 680
inductive, 384 Machine vision systems, 681
k-nearest neighbor, 425 Maxterms, 40
model, 382 Memory
propositional, 392 frontier search, 229
reinforcement, 400 Middle-level machine vision, 683
reinforcement learning, 398 Minimax search, 318
relational, 392 Minterms, 40
single-agent, 391 Model, 32, 57
strategies, 377 Model based reasoning, 103
supervised, 383 Modus ponens, 35, 42
symbol-based, 405 Modus tollens, 42
unsupervised, 384 Morphological, 652
Learning by analogy, 378 Most General Unifier (MGU), 76
Learning by deduction, 378 Multiagent
Learning by induction, 378 coalition algorithm, 485
Learning by instruction, 378 decision criteria, 493
Learning by reinforcement, 379 interactions, 477
Learning task, 437 model structure, 489
Lifting, 78 preferences, 492
Linear classifiers, 515 social view, 500
Literals, 56 taxonomy, 476
Local search Multiagent systems, 475, 476
greedy, 242
Logic, 25
predicate, 52 N
semantics, 55 Naive Bayes classifier, 428
syntax, 55 Nash arbitration, 314
propositional, 28 Natural Language Processing (NLP), 12
semantics, 33 ambiguous grammar, 619
syntax, 33 applications, 608
Logical consequence, 32 commonsense interfaces, 636
Logic and mathematics, 6 commonsense reasoning
Logic program components, 638
control information, 122 commonsense thinking, 638
path finding, 121 components, 609
Logic programming, 111, 112 discourse analysis, 611
Logic v/s control, 116 grammars, 612
Low-level machine vision, 680 NL parsing, 621
NL-question answering, 633
parsing
M probabilistic parsing, 627
Machine learning, 375, 415 top-down, 625
decision-trees, 393 parsing with CFG, 622
deep learning, 436 phrase structure, 613
714 Index

phrase structure grammar, 613 coordination, 467

prepositions, 620 decentralized, 466
semantic analysis, 611 forward planning, 453
syntax analysis, 609 goal allocation, 466
tools, 642 hierarchical, 462
transformational grammar, 617 incremental, 447
Natural Language Toolkit (NLTK), 642 interactive, 447
Neuroscience, 8 language, 456
Newell and Simons, 26 multi-agent, 464
NLTK partial order, 454
examples, 643 propositional Logic, 458
Normal forms, 40 search strategy, 458
NP-complete, 240 Planning graphs, 461
Planning language
STRIPS, 455
O Plausible inference, 148
Object detection, 685 Pragmatics, 652
background subtraction, 694 Predicate logic, 51
image segmentation, 694 Prisoner’s dilemma, 308
point detectors, 694 Probability theory, 339
supervised learning, 696 Prolog, 111
Object representation, 689 arithmetic expressions, 135
Object tracking, 689 backtracking, 135, 136
feature selection, 692 built-in predicates, 129
object detection, 694 cut, 135, 136
Ontolingua, 158 fail, 135
Ontological engineering, 158 list manipulation, 132
Ontologies vs semantic networks, 180 matching, 130
Ontology, 148 procedure call, 119
definition, 593 program efficiency, 137
levels, 152 Proof methods, 39
measurements, 157 Psychology, 7
Sowa’s ontology, 154
Ontology structures, 150
Ontology Web Language (OWL), 146, 196
OpenCV, 701 Q
Oskar Morgenstern, 304 Question Answering (QA)
redundancy based approach, 634
Quick-sort, 120
P
Parse-tree, 30
Parsing R
syntactic, 631 Reasoning
Pattern representation, 525 analogical, 403
Perception, 14 commonsense reasoning, 147
Philosophy, 5, 6 default, 166
Phonological, 652 formal, 37
Physical decomposition, 157 inductive, 36
Physical symbol system, 14 in uncertain environments, 337
Physical Symbol System Hypothesis meta-level, 37
(PSSH), 26 model-based, 38
Planning, 12, 445 nonmonotonic, 42, 165
basic problem, 448 ontologies, 156
classical problem, 449 procedural and numeric reasoning, 37
Index 715

qualitative, 148 greedy, 248

rule-based, 38 optimization, 247
taxonomic reasoning, 144 special cases, 248
temporal reasoning, 146 best-first search, 244
Reasoning architecture, 148 bidirectional, 228
Reasoning modes, 148 blind, 222
Reasoning patterns, 25, 35 blind-search, 217
Recursion rule, 130 breadth-first, 222, 230
Recursive programming, 130 complete, 219, 224
Refutation, 57 complexities, 220, 225
Relative probability, 343 depth-first, 224, 229
Rene Descartes, 6 edge branching factor, 232
Resolution, 40 exhaustive, 217
algorithm, 64 genetic algorithms, 264
refutation, 63 global, 244
Resolution principle, 51, 62 heuristic, 239, 241
Resolution proof, 64 comparison, 254
complexity, 65 informedness, 254
Roger Penrose, 7 hill-climbing, 242
Rote learning, 377 iterative deepening, 227
Rule based reasoning, 89 local, 244
complexity issues, 95 memory requirements, 229
goal determination, 100 minimum cost path heuristics, 249
Rule chaining, 114 monotonicity, 254
Rule selection node branching, 226
efficiency, 97 optimal, 219
Rule-based System (RBS), 90, 91, 102, 105 order-preserving, 264
backward chaining, 98 problem formulation, 230
conflict resolution, 95 representation, 218
forward chaining, 93 space-complexity, 219
forward chaining Algorithm, 93 sponsored search, 328
overview, 91 state-space, 217
preconditions time-complexity, 219
complexities, 98 uniform cost, 223
uninformed, 222
Semantic networks, 144, 179
S benefits, 181
Satisfiability, 57 structure theory, 211
Script, 210 Semantic tableau, 33
roles, 210 Semantic web, 185
scene, 210 Semantics Networks
Search Syntax and Semantics, 182
A∗ Sentence, 56
analysis, 253 Similarity measures, 527
A∗ Search, 249 Simulated annealing, 256
admissibility, 264 Situation
admissible, 254 context, 163
adversarial search, 303 goal, 174
alpha-beta, 324 initial, 174
complexity analysis, 326 Situation calculus, 159
backtracking, 224 action, 159
best-first formalism, 160
analysis, 247 objects, 159
716 Index

situation, 159 Transaction data, 513

Skolem functions, 58 Transitivity, 145
Skolemization, 58 Travelling Salesman Problem (TSP), 221
Socrates, 6 Turing Test, 3
Speech
applications, 651
multimedia-indexing, 652 U
Speech processing, 11 Understanding, 14
Speech Recognition Engine, 663 Unfounded sets, 81
STAR-PLAN system, 193 Unification, 59, 61
State space, 217, 232 algorithm, 61
Stereopsis, 683 Unifier, 61
Stored program concept, 16 most general, 76
Structural SVMs, 423
Structured objects
prediction, 422 V
Substitution components, 59, 60 Valid, 32
Substitutions, 59, 60 Variables
Supervised learning bound, 55
propositional learning, 383 free, 55
Support Vector Machines (SVM), 418, 517 Vision
learning pattern recognition, 419 basic principles, 672
structured SVM, 424 models, 674
training algorithm, 421 Voice Web, 654
Swi-prolog, 112
Synset, 153
Syntactic, 652 W
Synthesis algorithm, 283, 284 Well-formed expressions, 56
WordNet, 153, 594
Working-memory, 92
T World, 652
Tautology, 32 World Ontology, 150
Taxonomy, 144, 149
Term, 56
Terminology Box (TBOX), 201 Z
Theorem Proving, 64 Zermelo-Fraenkel (ZF), 45