Lecture Notes in Electrical Engineering

Volume 68
.
Sio-Iong Ao Burghard Rieger
l l

Mahyar A. Amouzegar

Machine Learning and

Systems Engineering
Editors
Dr. Sio-Iong Ao Prof. Dr. Burghard Rieger
International Association of Engineers Universität Trier
Hung To Road 37-39 FB II Linguistische
Hong Kong Datenverarbeitung
Unit 1, 1/F Computerlinguistik
Hong Kong SAR Universitätsring 15
[email protected] 54286 Trier
Germany

Prof. Mahyar A. Amouzegar

College of Engineering
California State University
Long Beach
CA 90840
USA

ISSN 1876-1100 e-ISSN 1876-1119

ISBN 978-90-481-9418-6 e-ISBN 978-90-481-9419-3
DOI 10.1007/978-90-481-9419-3
Springer Dordrecht Heidelberg London New York
Library of Congress Control Number: 2010936819

# Springer Science+Business Media B.V. 2010

No part of this work may be reproduced, stored in a retrieval system, or transmitted in any form or by any
means, electronic, mechanical, photocopying, microfilming, recording or otherwise, without written
permission from the Publisher, with the exception of any material supplied specifically for the purpose
of being entered and executed on a computer system, for exclusive use by the purchaser of the work.

Cover design: SPi Publisher Services

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Preface

A large international conference on Advances in Machine Learning and Systems

Engineering was held in UC Berkeley, California, USA, October 20–22, 2009,
under the auspices of the World Congress on Engineering and Computer Science
(WCECS 2009). The WCECS is organized by the International Association of
Engineers (IAENG). IAENG is a non-profit international association for the engi-
neers and the computer scientists, which was founded in 1968 and has been under-
going rapid expansions in recent years. The WCECS conferences have served as
excellent venues for the engineering community to meet with each other and to
exchange ideas. Moreover, WCECS continues to strike a balance between theoreti-
cal and application development. The conference committees have been formed
with over two hundred members who are mainly research center heads, deans,
department heads (chairs), professors, and research scientists from over thirty
countries with the full committee list available at our congress web site (http://
www.iaeng.org/WCECS2009/committee.html). The conference participants are
truly international representing high level research and development from many
countries. The responses for the congress have been excellent. In 2009, we received
more than six hundred manuscripts, and after a thorough peer review process
54.69% of the papers were accepted.
This volume contains 46 revised and extended research articles written by
prominent researchers participating in the conference. Topics covered include
Expert system, Intelligent decision making, Knowledge-based systems, Knowledge
extraction, Data analysis tools, Computational biology, Optimization algorithms,
Experiment designs, Complex system identification, Computational modeling,
and industrial applications. The book offers the state of the art of tremendous
advances in machine learning and systems engineering and also serves as an
excellent reference text for researchers and graduate students, working on machine
learning and systems engineering.

Sio-Iong Ao
Burghard B. Rieger
Mahyar A. Amouzegar

v
.
Contents

1 Multimodal Human Spacecraft Interaction in Remote

Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 The MIT SPHERES Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.1 General Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Human-SPHERES Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 SPHERES Goggles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3 Multimodal Telepresence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1 Areas of Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 The Development of a Test Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
4.1 Control via ARTEMIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
4.2 The Servicing Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
5 Results of the Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
5.1 Round Trip Delays due to the Relay Satellite . . . . . . . . . . . . . . . . . . . . . . 11
5.2 Operator Force Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2 A Framework for Collaborative Aspects of Intelligent

Service Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1 Context-Awareness Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2 Robot Grouping and Collaboration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3 Design of the System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.1 Context-Awareness Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Grouping Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Collaboration Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

vii
viii Contents

4 Simulated Experimentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.1 Robot Grouping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.2 Robot Collaboration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3 Piecewise Bezier Curves Path Planning with Continuous

Curvature Constraint for Autonomous Driving . . . . . . . . . . . . . . . . . . . . . . . . 31
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2 Bezier Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.1 The de Casteljau Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2 Derivatives, Continuity and Curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4 Path Planning Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.1 Path Planning Placing Bezier Curves within Segments (BS) . . . . . . 37
4.2 Path Planning Placing Bezier Curves on Corners (BC) . . . . . . . . . . . . 38
5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

4 Combined Heuristic Approach to Resource-Constrained Project

Scheduling Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2 Basic Notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4 Generalisation for Multiproject Schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5 KNapsack-Based Heuristic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6 Stochastic Heuristic Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7 Experimentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

5 A Development of Data-Logger for Indoor Environment . . . . . . . . . . . . . 59

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2 Sensors Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.1 Temperature Sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.2 Humidity Sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.3 CO and CO2 Sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3 LCD Interface to the Microcontroller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4 Real Time Clock Interface to the Microcontroller . . . . . . . . . . . . . . . . . . . . . . 62
5 EEPROM Interface to the Microcontroller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6 PC Interface Using RS-232 Serial Communication . . . . . . . . . . . . . . . . . . . . . 63
7 Graphical User Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
8 Schematic of the Data Logger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Contents ix

9 Software Design of Data Logger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

9.1 Programming Steps for I2C Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
9.2 Programming Steps for LCD Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
9.3 Programming Steps for Sensor Data Collection . . . . . . . . . . . . . . . . . . . 67
10 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
11 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

6 Multiobjective Evolutionary Optimization and Machine Learning:

Application to Renewable Energy Predictions . . . . . . . . . . . . . . . . . . . . . . . . . . 71
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2 Material and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.1 Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.2 Multiobjective Evolutionary Optimization . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.3 SVM-MOPSO Trainings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

7 Hybriding Intelligent Host-Based and Network-Based

Stepping Stone Detections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
2 Research Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4 Proposed Approach: Hybrid Intelligence Stepping Stone
Detection (HI-SSD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6 Result and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.1 Intelligence Network Stepping Stone Detection (I-NSSD) . . . . . . . . 88
6.2 Intelligence Host-Based Stepping Stone Detection (I-HSSD) . . . . . 89
6.3 Hybrid Intelligence Stepping Stone Detection (HI-SSD) . . . . . . . . . . 92
7 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

8 Open Source Software Use in City Government . . . . . . . . . . . . . . . . . . . . . . . 97

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
2 Related Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
3 Research Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5 Survey Execution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6 Survey Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7 Analysis: Interesting Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
7.1 Few Cities Have All Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
x Contents

7.2 Possible Aversion to OSS If Not Currently Using OSS . . . . . . . . 105

7.3 Current OSS Support by Leadership, Management,
and IT Staff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.4 Discrepancy of OSS Awareness: Self, Others . . . . . . . . . . . . . . . . . . . 108
8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

9 Pheromone-Balance Driven Ant Colony Optimization

with Greedy Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
2.1 Ant Colony Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
2.2 Related Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3 Hybrid ACO with Modified Pheromone Update Rules . . . . . . . . . . . . . . 115
4 Experiments and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

10 Study of Pitchfork Bifurcation in Discrete Hopfield

Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
2 Determination of Fixed Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
3 Local Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4 Pitchfork Bifurcation Direction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

11 Grammatical Evolution and STE Criterion . . . . . . . . . . . . . . . . . . . . . . . . . 131

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
2 STE – Sum Epsilon Tube Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
3 STE – Empirical Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
3.1 SSE (Advantages, Disadvantages) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
3.2 STE (Advantages, Disadvantages) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
4 Probabilistic Mapping of SSE to STE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
5 Goodness-of-Fit Tests of Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.1 Uncensored Data – ET10x50 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6 Probabilistic Relationship Between STE and SSE . . . . . . . . . . . . . . . . . . . 141
7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

12 Data Quality in ANFIS Based Soft Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
2 ANFIS Based Inferential Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Contents xi

2.1 Training and Testing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

3 Impact of Data Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
3.1 Experimental Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
3.2 Experimental Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
3.3 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
3.4 Experimental Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
4 Tane Algorithm for Noisy Data Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

13 The Meccano Method for Automatic Volume Parametrization

of Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
2 The Meccano Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
3 Application of the Meccano Method to Complex
Genus-ZeroSolids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
3.1 Example 1: Bust . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
3.2 Example 2: Bunny . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
3.3 Example 3: Bone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
4 Conclusions and Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

14 A Buck Converter Model for Multi-Domain Simulations . . . . . . . . . . . 169

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
2 The Model for Calculating Switching Events . . . . . . . . . . . . . . . . . . . . . . . . 170
3 The Averaged Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
4 Consideration of Switching Losses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
5 Implementation of the Simulation Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
6 Simulation and Laboratory Test Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

15 The Computer Simulation of Shaping in Rotating

Electrical Discharge Machining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
2 Mathematical Modelling of Redm Shaping by End Tool
Electrode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
3 Mathematical Modelling of Redm Shaping by Lateral
Surface of Tool Electrode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
4 Software for Computer Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
5 Experimental Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
xii Contents

16 Parameter Identification of a Nonlinear Two Mass System

Using Prior Knowledge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
2 General Dynamic Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
2.1 Administration Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
2.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
3 Parameter Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
3.1 Levenberg–Marquardt Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
3.2 Jacobian Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
4 Two-Mass-System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
5 Structured Dynamic Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
6 Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
6.1 Excitation Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
6.2 Engine Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
6.3 TMS Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210

17 Adaptive and Neural Learning for Biped Robot

Actuator Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
2.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
2.2 Biped Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
2.3 Uncertain Actuator Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
2.4 Desired Moments Md . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
2.5 Adaptive Control Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
3 Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
3.1 Reference Model for Actuator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
3.2 Inverse Model Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
3.3 MRAC Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
4 MRAC for Walking Biped Actuators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
4.1 Dynamics of Walking Biped . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
4.2 Computation of Desired Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
4.3 Dynamics of Actuators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
4.4 Configuration of MRAC Actuator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
4.5 Convergence Analysis of MRAC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
4.6 Neural Network Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
5.1 First Simulation (Without Disturbance) . . . . . . . . . . . . . . . . . . . . . . . . . . 222
5.2 Second Simulation (Disturbance) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
5.3 Third Simulation (Neural Network Estimation) . . . . . . . . . . . . . . . . . . 223
6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Contents xiii

18 Modeling, Simulation, and Analysis for Battery

Electric Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
2 Steady State Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
2.1 Projected Gravity Force . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
2.2 Aerodynamic Drag . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
2.3 The Rolling Resistance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
2.4 Power Required . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
2.5 Energy Required . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
2.6 Battery Specific Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
2.7 Maximum Cruise Speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
3 Dynamic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
3.1 Power Limited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
3.2 Traction Limited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
3.3 0–60 mph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
3.4 Maximum Gradeability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241

19 Modeling Confined Jets with Particles and Swril . . . . . . . . . . . . . . . . . . . . 243

1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
2 Gas Phase and Turbulence Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
2.1 Standard k e Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
2.2 Renormalization Group (RNG) k e Model . . . . . . . . . . . . . . . . . . . . . . 247
2.3 Realizable k e Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
3 Dispersed Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
4 Simulation Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255

20 Robust Tracking and Control of Mimo Processes

with Input Saturation and Unknown Disturbance . . . . . . . . . . . . . . . . . . . 257
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
2 MRAGPC Design Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
2.1 MRAGPC Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
2.2 Controllers Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
3 Additional Design Schemes for MRAGPC . . . . . . . . . . . . . . . . . . . . . . . . . . 262
3.1 Robust Parallel Compensator (RPC) Scheme
for MIMO Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
3.2 Unknown Disturbance Estimation Scheme
for MIMO Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
4 Simulation Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
4.1 Control of MIMO System Without Disturbance . . . . . . . . . . . . . . . . . 265
xiv Contents

4.2 Control of MIMO System with Disturbance . . . . . . . . . . . . . . . . . . . . . 265

5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268

21 Analysis of Priority Rule-Based Scheduling in

Dual-Resource-Constrained Shop-Floor Scenarios . . . . . . . . . . . . . . . . . . 269
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
2.1 Shop Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
2.2 Priority Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
2.3 Multi/dual-Resource Constrained Scheduling . . . . . . . . . . . . . . . . . . . . 271
3 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
4 Experiments with Static Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
4.1 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
4.2 Analyses of Static Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
5 Long-Term Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
5.1 Long-Term Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
5.2 Analysis of Long-Term Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
6 Conclusion and Further Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280

22 A Hybrid Framework for Servo-Actuated Systems

Fault Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
2 System Under Consideration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
3 Role of Fuzzy Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
4 Design of Fuzzy Logic Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
4.1 Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
4.2 Membership Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
4.3 Rule-Based Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
4.4 Defuzzification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
4.5 Rule Viewer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
5 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295

23 Multigrid Finite Volume Method for FGF-2 Transport

and Binding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
2.1 Mathematical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
2.2 Collocated Finite Volume Discretization . . . . . . . . . . . . . . . . . . . . . . . . 299
2.3 Multigrid Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
Contents xv

3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
4 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309

24 Integrated Mining Fuzzy Association Rules For Mineral

Processing State Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
2 Grinding Process Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
3 The Controller Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
3.1 Fuzzy Logic Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
3.2 Association Rules Miming Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324

25 A Combined Cycle Power Plant Simulator: A Powerful,

Competitive, and Useful Tool for Operator’s Training . . . . . . . . . . . . . . 327
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
2 Antecedent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
3 Architecture Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
3.1 Software Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
3.2 Software Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
3.3 Hardware Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
4 Modeled Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
4.1 Control System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
4.2 DCS Model for Real-Time Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
4.3 The Graphic Visualization Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
4.4 Processes System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
5 Project Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
8 Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339

26 Texture Features Extraction in Mammograms

Using Non-Shannon Entropies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
2 Gray Level Histogram Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
3 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
4 Conclusions and Future . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
xvi Contents

27 A Wideband DOA Estimation Method Based on Arbitrary

Group Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
2 Method of Digital Group Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
3 DOA Estimation Based on Digital Group Delay . . . . . . . . . . . . . . . . . . . . . 356
4 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358

28 Spatial Speaker Spatial Positioning of Synthesized

Speech in Java . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
2.1 Our Research Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
3 System Design and Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
3.1 FreeTTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
3.2 MIT Media Lab HRTF Library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
3.3 Signal Processing Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
3.4 JOAL Library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
3.5 Soundcard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
4 Prototype Applications and Preliminary User Studies . . . . . . . . . . . . . . . 366
4.1 Spatial Audio Representation of a Text File . . . . . . . . . . . . . . . . . . . . . 367
4.2 Spatial Story Reader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
4.3 Multiple Simultaneous Files Reader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
5 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370

29 Commercial Break Detection and Content Based

Video Retrieval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
2 Preprocessing and Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
2.1 Audio Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
2.2 Video Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
3 Commercial Detection Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
3.1 Audio Feature Based Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
3.2 Video Feature Based Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
4 Mechanism for Automatic Annotation and Retrieval . . . . . . . . . . . . . . . . 379
4.1 Automatic Annotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
4.2 Content Based Video Retrieval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
5 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
6 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
Contents xvii

30 ClusterDAM: Clustering Mechanism for Delivery of Adaptive

Multimedia Content in Two-Hop Wireless Networks . . . . . . . . . . . . . . . . 385
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
2 Cluster-Dam Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
2.1 Cluster-based Two-Hop Design for WiMAX Networks . . . . . . . . . 387
2.2 QOAS - Quality Oriented Adaptive Scheme . . . . . . . . . . . . . . . . . . . . . 388
2.3 Other Adaptive Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
3 Simulation Model and Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
3.1 Dumbbell and Double Dumbbell Topology . . . . . . . . . . . . . . . . . . . . . . 389
3.2 Simulation Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391
4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395

31 Ranking Intervals in Complex Stochastic Boolean Systems

Using Intrinsic Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
2 The Intrinsic Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399
2.1 Intrinsic Order Relation on {0,1}n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399
2.2 The Intrinsic Order Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
2.3 Three Sets of Bitstrings Related to a Binary n-tuple . . . . . . . . . . . . . 402
3 Generating and Counting the Elements of Cu and Cu . . . . . . . . . . . . . . . . 404
4 Ranking Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 410

32 Predicting Memory Phases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
2 Phase Classification Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
2.1 Wavelet Based Phase Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
2.2 Activity Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
2.3 Stack Reuse Distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
2.4 Other Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
3 Setvector Based Phase Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
4 Metrics to Compare Phase Classification Techniques . . . . . . . . . . . . . . . 415
5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
5.1 Classification Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
5.2 Computational Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
xviii Contents

33 Information Security Enhancement to Public–Key

Cryptosystem Through Magic Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
2.1 Magic Squares and Their Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
2.2 Construction of Doubly Even Magic Square Based
on Different Views of Fundamental Magic Square . . . . . . . . . . . . . . 427
2.3 Construction of Doubly Even Magic Square of
Order 16 Based on the Properties of 4  4 Magic Square . . . . . . . 428
3 Encryption/Decryption of Plain Text Using RSA Cryptosystem
with Magic Square . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
3.1 Wrapper Implementation-Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
4 Parallel Cryptography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
5 Experimental Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437

34 Resource Allocation for Grid Applications: An Economy Model . . . 439

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
2 Grid Economy Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
3 Resource Management Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
4 Resource Allocation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
5 Design of Economy Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
6 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
7 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448

35 A Free and Didactic Implementation of the Send Protocol

for Ipv6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
2 Neighbor Discovery Protocol Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
3 Vulnerabilities of the Neighbor Discovery Protocol . . . . . . . . . . . . . . . . . 454
4 Secure Neighbor Discovery Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
4.1 Cryptographically Generated Address . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
4.2 Authorization Delegation Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
5 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
6 A Didactic Implementation of the Send Protocol . . . . . . . . . . . . . . . . . . . . 457
7 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463

36 A Survey of Network Benchmark Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
2 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
3 Network Benchmark Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
Contents xix

3.1 Netperf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467

3.2 D-itg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
3.3 NetStress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
3.4 MGEN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
3.5 LANforge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
3.6 WLAN Traffic Visualizer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474
3.7 TTCP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475
4 Comparative Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
5 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 478
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 480

37 Hybrid Stock Investment Strategy Decision Support System . . . . . . . 481

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481
1.1 High Risk Investment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482
2 Finance Theories and Analysis in Stock Price Prediction . . . . . . . . . . . 482
3 Data Mining (DM) and Artificial Intelligence (AI) . . . . . . . . . . . . . . . . . . 483
4 DSS Model for Stock Investment Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . 484
5 Architecture of Stock Investment Strategy Decision
Support System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484
5.1 DM Component . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486
5.2 TA Component . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492

38 Towards Performance Analysis of Ad hoc Multimedia Network . . . 495

1 In-Vehicle Multimedia Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496
1.1 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 498
1.2 Application Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
2 Performance Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
2.1 Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
2.2 Packet Delay Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
2.3 Throughput Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
3 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
3.1 Simulation Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
3.2 Delay Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504
3.3 Throughput Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504
4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506

39 Towards the Performance Optimization of Public-key Algorithms

Using Fuzzy Modular Arithematic and Addition Chain . . . . . . . . . . . . . 507
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
2 Concept of Sum of Squares, Addition Chain, Elliptic Curve,
and Fermat Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
2.1 Sum of Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
xx Contents

2.2 Addition Chain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511

2.3 Elliptic Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512
3 Fuzzy Modular Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
4 Applications of Sum of Squares, and Addition Chain in Reducing
the Number of Multiplication in Modular Exponentiation . . . . . . . . . . . 514
4.1 Pseudocode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514
4.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
5 Implementation of ECC Using Fuzzy Modular Arithmetic . . . . . . . . . . 516
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518

40 RBDT-1 Method: Combining Rules and Decision Tree

Capabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523
3 Rule Generation and Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524
3.1 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524
3.2 Rule Generation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524
4 RBDT-1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
4.1 Attribute Selection Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
4.2 Building the Decision Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528
5 Illustration of the RBDT-1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528
5.1 Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528
6 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529
7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531

41 Computational and Theoretical Concepts for Regulating

Stem Cells Using Viral and Physical Methods . . . . . . . . . . . . . . . . . . . . . . . . 533
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533
2 Methods Used in Gene Therapy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
3 Proposed Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 545

42 DFA, a Biomedical Checking Tool for the Heart Control System . . 547
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 547
2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548
2.1 Finger Blood-Pressure Pulse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548
2.2 DFA Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548
2.3 Volunteers and Ethics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549
3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549
3.1 Extra-Systole . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549
3.2 Alternans with Low Exponent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552
Contents xxi

3.3 Extraordinary High Exponent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552

3.4 Normal Exponent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
3.5 DFA Is Beneficial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554
4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 556

43 Generalizations in Mathematical Epidemiology . . . . . . . . . . . . . . . . . . . . . . 557

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
2 CA And MR Applied to the SNIR Epidemic Model . . . . . . . . . . . . . . . . . 558
2.1 The Standard SIR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559
2.2 The S2IR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 560
2.3 The S3IR, The S4IR and S5IR Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 561
2.4 The SnIR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 562
3 CA and MR Applied to the SNIMR Epidemic Model . . . . . . . . . . . . . . . . 563
3.1 The SI2R Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563
3.2 The S2I2R Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564
3.3 The SnImR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
4 CA and MR Applied to the Staged Progressive SIMR
Epidemic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
4.1 The Staged Progressive SI2R Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566
4.2 The Staged Progressive SI3R Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
4.3 Staged Progressive SImR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 568
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 568

44 Review of Daily Physical Activity Monitoring System Based

on Single Triaxial Accelerometer and Portable Data
Measurement Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
1.1 Measurement of Physical Activity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 570
1.2 Behavioral Observation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 571
1.3 Pedometers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 571
1.4 Accelerometers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 571
2 Material and Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574
2.1 Portable Data Measurement Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574
2.2 Physical Activity Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575
2.3 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 576
3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 576
4 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578

45 A Study of the Protein Folding Problem by a Simulation Model . . . 581

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
2 The Protein Folding Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
xxii Contents

2.1 The Levinthal Paradox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582

2.2 Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 583
3 Approaches to Study the Protein Folding Problem . . . . . . . . . . . . . . . . . . 583
3.1 Latest Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
3.2 The Amino Acid Interaction Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
4 Folding a Protein in a Topological Space by
Bio-Inspired Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 586
4.1 Genetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 586
4.2 Motif Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 588
4.3 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
4.4 Overall Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
4.5 Genetic Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
4.6 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 590
5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592

46 Analysing Multiobjective Fitness Function with Finite

State Automata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595
2 Evolutionary Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 598
2.1 Input-Output Specification (IOS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 598
2.2 Syntax Term (S) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
2.3 Primitive Function (F) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
2.4 Learning Parameter (a1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
2.5 Complexity Parameters (Tmax, b) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
2.6 System Proof Plan (u) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
3 Evolutionary Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
3.1 Single Objective Evolutionary Process . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
3.2 Multi Objective Evolutionary Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
4 Result and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603
4.1 Input-Output Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603
4.2 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604
5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 605
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 605

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607
Chapter 1
Multimodal Human Spacecraft Interaction
in Remote Environments
A New Concept for Free Flyer Control

Enrico Stoll, Alvar Saenz-Otero, and Brent Tweddle

Abstract Most malfunctioning spacecraft require only a minor maintenance oper-

ation, but have to be retired due to the lack of so-called On-Orbit Servicing (OOS)
opportunities. There is no maintenance and repair infrastructure for space systems.
Occasionally, space shuttle based servicing missions are launched, but there are no
routine procedures foreseen for the individual spacecraft.
The unmanned approach is to utilize the explorative possibilities of robots to
dock a servicer spacecraft onto a malfunctioning target spacecraft and execute
complex OOS operations, controlled from ground. Most OOS demonstration mis-
sions aim at equipping the servicing spacecraft with a high degree of autonomy.
However, not all spacecraft can be serviced autonomously. Equipping the human
operator on ground with the possibility of instantaneous interaction with the
servicer satellite is a very beneficial capability that complements autonomous
operations.
This work focuses on such teleoperated space systems with a strong emphasis
on multimodal feedback, i.e. human spacecraft interaction is considered, which
utilizes multiple human senses through which the operator can receive output from
a technical device. This work proposes a new concept for free flyer control and
shows the development of an according test environment.

1 Introduction

On-Orbit Servicing (OOS) has been an active research area in recent times. Two
approaches have been studied: teleoperation by humans and autonomous systems.
Autonomous systems use machine pattern recognition, object tracking, and

E. Stoll (*)
Space Systems Laboratory, Massachusetts Institute of Technology, 77 Massachusetts Avenue,
Cambridge, MA 02139-4307, USA
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 1

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_1, # Springer ScienceþBusiness Media B.V. 2010
2 E. Stoll et al.

acquisition algorithms, as for example DART [1] or Orbital Express [2]. The
research is still in early stages and the algorithms have to be realized in complex
systems.
In contrast, the human eye-brain combination is already very evolved and
trainable. Procedures can be executed by the trained user from the ground. Unfore-
seen incidents can be solved with greater flexibility and robustness. Arbitrary
spacecraft could be approached, i.e. spacecraft which were not explicitly designed
for rendezvous and docking maneuvers. Analogously, inspections and fly-arounds
can be controlled by the human operator. Based on the acquired information the
human operator on ground can decide how to proceed and which servicing mea-
sures to take. Another element in the decision queue is the path planning approach
for the target satellite to the capture object.
Multimodal telepresence, which combines autonomous operations with human
oversight of the mission (with the ability to control the satellites), provides the
benefits of autonomous free-flyers with the evolved human experience. In case
autonomous operations cause the work area to exhibit an unknown and unforeseen
state (e.g. when robotically exchanging or upgrading instruments) the human
operator on ground can support the operations by either finishing the procedure or
returning the system into a state which can be processed by autonomous procedures.
The advantage of multimodal telepresence in this connection is the fact that the
operator will not only see the remote site, but also feel it due to haptic displays. A
haptic interface presents feedback to the human operator via the sense of touch by
applying forces, vibrations or motion.
The applicability of the telepresence approach, with a human operator located in
a ground station, controlling a spacecraft, is mostly limited to the Earth orbit. This is
because the round trip delay increases with increasing distance from operator to the
teleoperator. A decrease of the telepresence feeling is the consequence, which has a
large impact on the task performance. Therefore, as the distance increases, the role
of the autonomy must increase to maintain effective operations.
For an overall and significant evaluation of the benefits of multimodal telepre-
sence a representative test environment is being developed at the MIT Space
Systems Laboratory using the SPHERES satellites on ground and aboard the
International Space Station (ISS).

2 The MIT SPHERES Program

The SPHERES laboratory for Distributed Satellite Systems [3] consists of a set of
tools and hardware developed for use aboard the ISS and in ground based tests.
Three micro-satellites, a custom metrology system (based on ultrasound time-of-
flight measurements), communications hardware, consumables (tanks and bat-
teries), and an astronaut interface are aboard the ISS. Figure 1 shows the three
SPHERES satellites being operated aboard the ISS during the summer of 2008.
1 Multimodal Human Spacecraft Interaction in Remote Environments 3

Fig. 1 SPHERES operations aboard the International Space Station (Picture: NASA)

The satellites operate autonomously, after the crew starts the test, within the US
Destiny Laboratory.
The ground-based setup consists of an analog set of hardware: three micro-
satellites, a metrology system with the same geometry as that on the ISS, a research
oriented GUI, and replenishable consumables. A “guest scientist program” [4]
provides documentation and programming interfaces which allow multiple
researchers to use the facility.

2.1 General Information

The SPHERES satellites were designed to provide the best traceability to future
formation flight missions by implementing all the features of a standard thruster-
based satellite bus. The satellites have fully functional propulsion, guidance, com-
munications, and power sub-systems. These enable the satellites to: maneuver in
6-DoF, communicate with each other and with the laptop control station, and
identify their position with respect to each other and to the experiment reference
frame. The computer architecture allows scientists to re-program the satellite with
new algorithms. The laptop control station (an ISS supplied standard laptop) is used
to collect and store data and to upload new algorithms. It uses the ISS network for
all ground data communications (downlink and uplink). Figure 2 shows a picture of
an assembled SPHERES satellite and identifies its main features. Physical proper-
ties of the satellites are listed in Table 1.
4 E. Stoll et al.

Fig. 2 SPHERES satellite Pressure Control Panel

Regulator
Ultrasound
Pressure Sensors
Gauge

Thrusters Battery

Table 1 SPHERES satellite Property Value

properties
Diameter 0.22 m
Mass (with tank and batteries) 4.3 kg
Max linear acceleration 0.17 m/s2
Max angular acceleration 3.5 rad/s2
Power consumption 13 W
Battery lifetime 2h

SPHERES has been in operation aboard the ISS since May 2006. To date, 21 test
sessions have taken place. The test sessions have included research on Formation
Flight, Docking and Rendezvous, Fluid Slosh, Fault Detection, Isolation, and
Recover (FDIR), and general distributed satellite systems autonomy.

2.2 Human-SPHERES Interaction

Most of the previous test sessions matured autonomous algorithms. However,

future servicing missions and the assembly of complex space structures will not
only depend on increased autonomy, but the ability of humans to provide high-level
oversight and task scheduling will always be critical. SPHERES tests were con-
ducted to develop and advance algorithms for adjustable autonomy and human
system interaction. This research began with basic tests during Test Session 11,
where the crew was asked to move a satellite to multiple corners in a pre-defined
volume. The satellite autonomously prevented collisions with the walls of the ISS.
The test demonstrated the ability of the crew to use the ISS laptop to control
SPHERES. It provided baseline results for future tests. An ongoing sequence of
ISS tests is being conducted in the framework of a program called “SPHERES
Interact”. The goal of the program is to conceive new algorithms that utilize both
human interaction and machine autonomy to complete complex tasks in 6 degrees
of freedom (DoF) environments. Tests during Test Session 19 and 20 included
several scenarios where human interaction helps schedule tasks of a complex
mission (e.g. servicing or assembly). The research area comprises human orienta-
tion, navigation, and recognition of motion patterns. Further, high level human
1 Multimodal Human Spacecraft Interaction in Remote Environments 5

abort commands and collision avoidance techniques for part of this ongoing
research aboard the International Space Station.

2.3 SPHERES Goggles

The SPHERES Goggles is a hardware upgrade to the SPHERES satellites that adds
cameras, lights, additional processing power and a high speed wireless commu-
nications system. Even though it was designed for autonomous operations, it can be
used to support the operator with a visual feedback. The main objective of the
SPHERES Goggles is to provide a flight-traceable platform for the development,
testing and maturation of computer vision-based navigation algorithms for space-
craft proximity operations. Although this hardware was not intended to be launched
to orbit, it was designed to be easily extensible to versions that can operate both
inside and ultimately outside the ISS or any other spacecraft.
The Goggles, which are shown in Fig. 3, were designed to be able to image
objects that are within few meters range and to possess the computational capability
to process the captured images. They further provide a flexible software develop-
ment environment and the ability to reconfigure the optics hardware.
The SPHERES Goggles were used in several parts of the telepresence environ-
ment setup at the MIT SSL ground facilities to support the human operator with a
realistic video feedback which is representative for a camera system used on orbit.
Apart from virtual reality animations of the remote environment it serves as the
only source of visual data in the experiments.

Camera Camera

LED Light LED Light

Fig. 3 Front view of Goggles mounted on SPHERES satellite

6 E. Stoll et al.

3 Multimodal Telepresence

Servicing missions can be differentiated by whether or not a robotic manipulator is

connected to a free flying base (the actual satellite). Different levels of autonomy
can be applied to the control of either and the human operator receives the accord-
ing feedback.

3.1 Areas of Application

Unlike robotic manipulators, where haptic feedback plays an important role for
control as e.g. ETS-VII [5] or Rokviss [6], free flyers are commonly only steered
using visual feedback. That means that even though free flying experiments can be
steered with hand controllers, as for example Scamp [7] or the Mini AERCam [8],
usually no haptic information is fed back to the human operator.
The implementation of haptic feedback into the control of free flyers enriches the
telepresence feeling of the operator and helps the operator on ground to navigate. It
paves the way for new concepts of telepresent spacecraft control. Collision avoid-
ance maneuvers for example can be made perceptible for the human operator, by
placing virtual walls around other spacecraft. Equipping these virtual walls with
sufficient high stiffness means that the operator is not able to penetrate them by
means of the haptic device, since it exerts to the operator a high resistance force.
Areas of fuel optimal paths can be displayed to the operator by implementing an
ambient damping force, featuring a magnitude which is proportional to the devia-
tion of the actual path from the fuel optimal trajectory and area, respectively.
Docking maneuvers can be supported by virtual boundaries as a haptic guiding
cone and damping forces which are increasing with decreasing distance to the
target.
Summarizing the benefits it can be seen that the application of telepresence
control will extend the amount of serviceable spacecraft failures by involving a well
trained human operator. In this connection it is proposed that the task performance
of the operator can be enhanced by feeding back high-fidelity information from the
remote work environment. Here the haptic feedback plays an important role in
human perception and will be tested in a representative test environment.

3.2 The Development of a Test Environment

The key element of the test environment is the Novint Falcon [9], which is a 3-DoF
force feedback joystick. All degrees of freedom are of translational nature and servo
motors are used to feed forces in 3-DoF back to the user. This system has high
utility for space applications since it allows the human operator to control the space
1 Multimodal Human Spacecraft Interaction in Remote Environments 7

application in 3D dimensional space. The Falcon is implemented in a Matlab/

SIMULINK environment via the HaptikLibrary [10], which is a component based
architecture for uniform access to haptic devices. It is used as the interface to
Matlab and reads the positions and button states of the haptic device as well as feeds
calculated forces back to it. By displacing the joystick handle, the human operator is
able to interact with two instances of the remote environment - the virtual instance
in SIMULINK and the hardware (SPHERES) instance on the SSL air table.
The joystick displacement is interpreted by the system, as either position,
velocity or force commands. The received commands are communicated to a
SIMULINK block, containing the satellite dynamics and a state estimator. The
simulation returns the estimated state of the satellite in the virtual entity of the
remote environment. This remote workspace is created using SIMULINK’s Virtual
Reality (VR) toolbox (cp. Fig. 4), allowing for satellite states and environmental
properties to be displayed.
In addition to the Matlab environment, algorithms in C are used as the interface
to the actual SPHERES hardware via the “SPHERES Core” API. Commands are
transmitted via wireless communications to the SPHERES satellites. Torques and
forces are calculated and directly commanded to the thrusters, which will cause a
motion of the SPHERES satellite. The satellites measure their position and attitude
and transmit the information in real-time to the laptop.
By transmitting the actual states to the VR, the operator obtains information of
the estimated and the actual motion of the free flyer, which should be identical if
the communication channel is not delayed and the virtual instance is a good

human visual feedback

operator
haptic joystick
feedback displacement

commands SPHERES
thruster goggles
model
UDP connection

SPHERES communication system

Novint Falcon

HaptikLibrary
force/
torques SPHERES on
satellite air table
forces dynamics /
estimator metrology
data distance
estimated states
information
Position &
virtual reality
attitude
remote
actual determination
environment Ultrasound
states system
Matlab / Simulink
environment C environment
beacons

Fig. 4 General block diagram of the test environment

8 E. Stoll et al.

approximation of reality. In the presence of time delay the predictions should give
the user a feeling for the behaviour of the system (cp. ETS-VII) and enhance the
task performance. This is an important point if a human operator on ground will
steer an application in space. That way the interactions between the autonomous
space operations and a telepresence controlled free flyer can be tested.

4 Experimental Setup

If OOS missions are executed in low Earth orbit (LEO) only limited time windows
are available for telecommands. The common approach for increasing those acqui-
sition times is the usage of geostationary relay satellites. While those satellites do
not have a profound impact on autonomous missions, they will influence the task
performance of an operator on ground directly interacting with a satellite. Thus, this
human spacecraft interaction was tested using a representative test scenario at SSL,
which involved a geostationary relay satellite.

4.1 Control via ARTEMIS

Due to the orbit height of geostationary satellites, the relay of the signal increases
the round trip delay between operator action and feedback to the operator to up to
7 s as in the case of ETS-VII. The delay between telecommand and telemetry is
usually not very intuitively manageable for the human operator and thus a special
area of interest if human spacecraft interaction is considered.
The effect on the human has already been shown for OOS missions in which the
operator on ground steers robotic manipulators via geostationary relay satellites
[11]. It has not been tested, yet, for multimodal human free flyer interaction.
Accordingly, for initial tests a geostationary satellite was introduced in the com-
manding chain. The UDP connection (cp. Fig. 4) was utilized to send the commands
of the Novint Falcon at SSL via a terrestrial internet connection to a ground station
at the Institute of Astronautics of Technische Universitaet Muenchen in Germany.
The telecommands were forwarded via the geostationary relay satellite ARTEMIS
(Advanced Relay Technology Mission) of the European Space Agency (ESA) to a
ground station of ESA in Redu, Belgium. The signal was mirrored in Redu and
retransmitted analogously back to MIT, where again the UDP connection was used
to feed the telecommand into the hardware on the air table and change the posi-
tion of SPHERES in the test environment. That way the SPHERES satellites were
controlled by the Novint Falcon via a geostationary satellite. The round trip delay
characteristics were logged and subsequently implemented into the scenario as a
SIMULINK block. That way the test scenarios could be evaluated in the absence of
a satellite link but with round trip delays representative for commanding a space-
craft in orbit.
1 Multimodal Human Spacecraft Interaction in Remote Environments 9

4.2 The Servicing Scenarios

To show the benefit of multimodal feedback to the operator, two scenarios were
developed and tested. Both are based on a servicing operation, in which three
satellites are involved. The target satellite is the satellite to be serviced. Therefore,
the servicer satellite has to execute proximity operations, approach the target, and
eventually dock with it. The inspector satellite is supporting the operator on ground
with additional data of the remote environment. It carries a camera system and
can yield information on the distance between the two other satellites.

4.2.1 The Human-Controlled Inspector Satellite

In this first scenario the control of the inspector satellite is handed over to the human
operator, while the servicer and the target dock autonomously. The task of the
operator is to ensure that the initial states of the other two satellites are appropriate
for the docking maneuver. Thus, the operator commands the inspector satellite
as depicted in Fig. 5 from its position in front of the other two satellites to a position
behind the two satellites, which is indicated by a virtual checkered marker.
For efficiently accomplishing this circumnavigation, virtual obstacles were
created to avoid collisions with the servicer, the target, and the borders of the
experimental volume. As to be seen in Fig. 5 both of the satellites to dock feature a
virtual collision avoidance sphere. Further, on the left and the right side of the
volume, there are virtual (brick) walls introduced. The Novint Falcon generates
forces in case the operator penetrates those objects. These resistance forces are fed
back to the operator and thus prevent from colliding with the actual hardware on
the SSL air table.

Fig. 5 Virtual and hardware instance of the inspection scenario

10 E. Stoll et al.

A further benefit of using the virtual reality techniques is that the environment
can be augmented with additional data of the remote environment. For example,
arrows can be used for indicating the current velocity and rotation rate (double
arrow) of the inspector. Furthermore, there are two entities of the inspector satellite
to be seen in the VR environment. The dark entity shows the commanded state,
whereas the pale entity shows the actual state of the hardware in the remote
environment. This is of great benefit for the human operator in the presence of
time delays as they occur due to the use of relay satellites.

4.2.2 The Human-Controlled Servicer Satellite

Similar to the first scenario, the inspector, target, and servicer satellite are again
involved in the second scenario. The servicer is supposed to dock with the target,
whereas the inspector is transmitting additional data from the remote scene. In this
scenario the target and the inspector (right upper corner in Fig. 6) are operating
autonomously and the servicer satellite (lower right corner) is controlled by the
human operator via the relay satellite.
Again, the virtual environment is enriched by collision avoidance objects (at the
inspector and the borders of the volume). The task of the operator is to accomplish a
successful docking maneuver. Therefore, the human operator is supposed to com-
mand the servicer at first to a position roughly aligned with centre of the docking
cone, which can be seen in Fig. 6 and approx. 50 cm away from the target. In a
second step the operator is commanding the servicer along the virtual cone until the
berthing takes place.

Fig. 6 Virtual and hardware instance of the docking scenario

1 Multimodal Human Spacecraft Interaction in Remote Environments 11

The docking cone is a mean to simplify the proximity operations for the operator.
Once the servicer has crossed the assistance horizon of the cone, a force field is
applied to the Falcon, which drives the servicer into the docking cone. Inside the
docking cone another force field drives the servicer towards the target. Here, the
forces are proportional to the distance to the target. This helps the operator to
concentrate on the precision of the docking point rather than to worry about relative
velocities and collisions.

5 Results of the Experiments

The two scenarios were controlled via the German ground station [12] at the
Institute of Astronautics and the ESA relay satellite. The human operator at MIT
in Cambridge received instantaneous feedback from the haptic-visual workspace.
To have a representative test conditions the operator had only visual feedback from
the SPHERES Goggles and the Matlab Simulink virtual instance of the remote
environment. Further, the haptic device yielded additional forces for an advanced
human spacecraft interaction in the 3D environment.

5.1 Round Trip Delays due to the Relay Satellite

The occurring round trip delays were logged since they are a first indicator for
the quality of the human task performance. Figure 7 shows an example graph of the
delay characteristics over time. The round trip delays are plotted depending on
the respective UDP packet number. They indicate that the delay in a real OOS mission
can be, except for a couple of outliers, well below 1 s. The outliers occurred due to
the use of a terrestrial internet connection and the lack of synchronization between

1400

1300

1200

1100
delay [ms]

1000

900

800

700

600
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
packet number

Fig. 7 Round trip delay characteristic of the free flyer control via ARTEMIS
12 E. Stoll et al.

the sampling rate of the hardware at MIT and the sampling rate of the satellite
modem at LRT. Nonetheless, a mean of 695.5 ms with a sample standard deviation
of 24.1 ms indicate an acceptable round trip delay [13] for telepresence operations.

5.2 Operator Force Feedback

Navigation in a 3D environment with a sparse number of reference points can be

very complicated for a human operator. The motion with 6-DoF is not only very
unintuitive since the motion in free space is no longer superimposed by gravity as it
is on Earth the case. The equations of motions are further coupled in a way that an
introduced torque about a main axis of inertia of the spacecraft will not necessarily
cause the spacecraft to rotate about the respective axis but about all three axes.

2 10

1.8 9

1.6 8

1.4 7

1.2 6
y [m]

1 5

0.8 4

0.6 3

0.4 2

0.2 1

0
0 0.2 0.4 0.6 0.8 1 1.2 force
x [m] [N]

Fig. 8 Force feedback of the inspection scenario

1 Multimodal Human Spacecraft Interaction in Remote Environments 13

Thus, the human operator has to be supported by technical means in order to

solve complex problems in the remote environment. One of those means is, as
shown in this work, to augment the feedback to the operator. Virtual reality can be
used to show commanded/planned states versus actual states of spacecraft and can
additionally visualize potential dangerous areas.
Since the 3D remote environment is usually projected onto 2D screens, it can be
difficult for the operator to realize where exactly such an area, in which collisions
could take place, is located. Consequently, a haptic device was used which utilizes
another human sense and enriches the perception. Forces are fed back to the
operator site, permitting the operator to enter the respective areas.
Figures 8 and 9 show example forces that were fed back to the operator
depending on the position of the spacecraft in the remote environment. The path

2 10

1.8 9

1.6 8

1.4 7

1.2 6
y [m]

1 5

0.8 4

0.6 3

0.4 2

0.2 1

0
0 0.2 0.4 0.6 0.8 1 1.2
force
x [m] [N]

Fig. 9 Force feedback of the docking scenario

14 E. Stoll et al.

of the spacecraft is indicated by a solid line, whereas the force feedback is labeled
by small circles in gray scale. If a collision avoidance sphere was penetrated as e.g.
in Fig. 8 a restraining force was created proportional to the penetration depth and
the velocity (spring-damper system) of the spacecraft. The same held true for
virtual walls as can be seen in Fig. 9. This figure further shows the force feedback
inside the docking cone. As can be seen, the haptic feedback prevented the human
operator form colliding with the other spacecraft or the experimental boundaries. It
gave the operator a feeling for critical areas and helped the operator to accomplish a
very smooth docking/berthing approach.

6 Summary

This work presented the first tests on haptic feedback for free flyer systems. It
proposes that multimodal feedback from servicer satellites enhances the human task
performance. This feedback supports the operator with an intuitive concept for
collision avoidance and relative navigation. That way, complex tasks in micrograv-
ity can be safely operated from ground.

Acknowledgements This work was supported in part by a post-doctoral fellowship program of

the German Academic Exchange Service (DAAD). The authors would like to express their
gratitude to the ESA ARTEMIS team for providing the opportunity to use the ARTEMIS relay
satellite for their experiments.

References

1. T. Rumford, Demonstration of autonomous rendezvous technology (dart) project summary, in

Proceedings of Space Systems Technology and Operations Conference, Orlando, USA, Apr
2003
2. T. Weismuller, M. Leinz, GN&C technology demonstrated by the orbital express auto-
nomous rendezvous and capture sensor system. 29th AAS Guidance and Control Conference,
Breckenridge, USA, 2006
3. A. Saenz-Otero, A. Chen et al., SPHERES: Development of an ISS laboratory for formation
flight and docking research, IEEE Aerospace Conference (paper #81), Big Sky, Montana,
USA, 9–16 Mar 2002
4. J. Enright, M.O. Hilstad et al., The SPHERES guest scientist program: collaborative science
on the ISS, 2004 IEEE Aerospace Conference (paper #1296), Big Sky, Montana, USA, 7–12
Mar 2004
5. T. Imaida, Y. Yokokohji, T. Doi, M. Oda, T. Yoshikawa, Ground-space bilateral teleoperation
ex-periment using ETS-VII robot arm with direct kinesthetic coupling, in Proceedings of
IEEE International Conference on Robotics and Automation, Seoul, Korea, 2001
6. K. Landzettel et al., ROKVISS verification of advanced light weight robotic joints and tele-
presence concepts for future space missions, in Proceedings of 9th ESA Workshop on
Advanced Space Technologies for Robotics and Automation (ASTRA), Noordwijk, The
Netherlands, Nov 2002
7. C. McGhan, R. Besser, R. Sanner, E. Atkins, Semi-autonomous inspection with a neutral
buoyancy free-flyer, in Proceedings of Guidance, Navigation, and Control Conference,
Keystone, USA, Aug 2006
1 Multimodal Human Spacecraft Interaction in Remote Environments 15

8. S. Fredrickson, S. Duran, J. Mitchel, Mini AERCam inspection robot for human space
missions, in Proceedings of AIAA Space, San Diego, USA, Sept 2004
9. Novint Technologies Inc. (February 2010). http://www.novint.com
10. M. de Pascale, D. Prattichizzo, The Haptik Library: a component based architecture for
uniform access to haptic devices, IEEE Robotics Autom. Mag. 14(4), 64–75 (2007)
11. E. Stoll, Ground verification of telepresence for on-orbit servicing. Dissertation, Lehrstuhl f€ ur
Raumfahrttechnik, Technische Universit€at M€ unchen, 2008, ISBN 978-3-89963-919-3
12. J. Letschnik, E. Stoll, U. Walter, Test environment for time delay measurements of space links
via ARTEMIS, in Proceedings of 4th ESA International Workshop on Tracking, Telemetry
and Command Systems for Space Applications TTC 2007, Darmstadt, Germany, 2007
13. E. Stoll et al., Ground verification of the feasibility of telepresent on-orbit servicing. J. Field
Robotics 26(3), 287–307 (2009)
Chapter 2
A Framework for Collaborative Aspects
of Intelligent Service Robot

Joohee Suh and Chong-woo Woo

Abstract Intelligent service robot is becoming one of the most interesting issues
in the recent Robot research. The service robot monitors its surroundings, and
provides a service to meet a user’s goal. The service often becomes too complex
that one single robot may not handle efficiently. In other words, a group of robots
may be needed to accomplish given task(s) by collaborating each other. We can
define this activity as a robot grouping, and we need to study further to make better
group(s) by considering their characteristics of the each robot. But, it is difficult and
no formal methods to make such a specific group from the many heterogeneous
robots that are different in their functions and structures. This paper describes an
intelligent service robot framework that outlines a multi-layer structure, which is
suitable to make a particular group of robots to solve given task by collaborating
with other robots. Simulated experimentation for grouping from the generated
several heterogeneous is done by utilizing Entropy algorithm. And the collabora-
tion among the robots is done by the multi-level task planning mechanism.

1 Introduction

Ubiquitous computing [1] defines that various computing objects are connected
through the network, so that the system automatically provides services in anytime
in any place. An intelligent robot is an autonomous and dynamic object in this
ubiquitous environment, and it becomes one of the most interesting issues in this
area. It interacts with various surrounding computing devices, recognizes context,

J. Suh (*)
Korea School of Computer Science, Kookmin University, 861-1 Jeongneung-Dong, Seongbuk-
Gu, Seoul
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 17

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_2, # Springer ScienceþBusiness Media B.V. 2010
18 J. Suh and C.‐W. Woo

and provides appropriate services to human and the environment [2]. The defini-
tion of the context [3] is any entity that affects interactions among the computing
objects in this environment. For instance, user, physical location, robots could be
such entities. Therefore, information described for the characteristics of such
entities are defined as a context, and recognizing a situation from the context is
the main focus of the context awareness system [4]. Essentially, the robot is
required to carry on two main issues in this study; understanding the context and
carrying out the context.
First, understanding a context can be done in many different ways, but the
general procedure includes that system perceives raw environmental information
through the physical sensors. And then context modeling and reasoning steps
follows right after preprocessing the raw data. The final result from this procedure
is a context. The approaches on this procedure are vary, and it needs to be further
studied in detail to make the system more efficient, and there are large amount of
studies [2–4] are being reported recently. The second issue is that the robot has to
carry out the generated context to meet the user’s goal. The context can be further
divided into some tasks, which often becomes too complex that one single robot
may not handle in this environment. In other words, a group of robots may be
needed to accomplish given task(s) by collaborating each other. We can define this
activity as a robot grouping, and we need to study further to make better group(s) by
considering their characteristics of the robot and the task.
In this paper, we are describing a development of the social robot framework for
providing an intelligent service with the collaboration of the heterogeneous robots,
in the context awareness environment. The framework is designed as multi-layers
suitable for understanding context, grouping robots, and collaborating among the
robots. In this study, we mainly focused and implemented on the grouping and
collaborating parts of the system. The context understanding part of the system is
designed, but will be discussed in the next study with comprehensive ontological
knowledge representations along with context modeling and context reasoning
mechanisms.

2 Related Works

2.1 Context-Awareness Systems

With the increase of mobile computing devices, the ubiquitous and pervasive
computing is getting popular recently. One part of the pervasive system is the
context awareness system, which is being studied explosively in various directions.
Among the many research results on this issue, the most prominent systems are the
CoBrA [5] (Context Broker Architecture), SOCAM [6] (Service Oriented Context-
Aware Middleware), and Context-Toolkit [7]. The CoBrA and SOCAM used
ontological model aiming for the benefits of easy sharing, reusing, and reasoning
2 A Framework for Collaborative Aspects of Intelligent Service Robot 19

of context information. But, first, these systems were not flexible enough to extend
to other domains when the system needs to expand with other devices or service
with a new domain. And second, they were also limited in formalized and shared
expression of the context, which is needed when the system interoperate or trans-
plant with other systems. Therefore, the ontology becomes one of the most popular
solutions to represent the data for the recent context awareness systems. Short
reviews of the previous systems as follows.
Context-Toolkit: This early context awareness middleware system gains infor-
mation from the connected devices. But, since it does not use ontology, it lacks of
the standardized representation for the context, and also the interoperability
between heterogeneous systems.
CoBrA: This system is developed based on the ontology, so that the standar-
dized representation for the context is possible. But, since the use of ontology is
limited only to a special domain, so called ‘Intelligent Meeting Room’, it does not
guarantee any extensibility to the other diverse domains.
SOCAM: This system is based on Service-oriented structure, which is efficient
middleware system for finding, acquiring, analyzing context information. But, since
it depends on OWL (Web Ontology Language) for reasoning, its reasoning capa-
bility is limited to its own learning module and inference engine.
In our study, we adapted merits of the previously studied systems, and designed a
framework that can overcome the above limitations, such as, the limited standardized
representation or extensibility. For instance, we have adapted context-awareness layer
by adapting the CONCON model [8] to provide extensibility of the ontological
representation.

2.2 Robot Grouping and Collaboration

First of all, our research issue focused on the collaboration among the robots, which
needs to form a group. Therefore, we first need to develop a method of grouping
robots for a given task. Study on the robot grouping is just beginning, but some
related researches are being reported as follows.
For instance, Rodic and Engelbrecht [9] studied initial investigation into feasi-
bility of using ‘social network’ as a coordination tool for multi-robot teams. Under
the assumption of multi-robot teams can accomplish certain task faster than a single
robot, they proposed multi-robot coordination techniques. Inspired by the concept
from the animal colony, Labella et al. [10], showed simple adaptation of an
individual can lead task allocation. They developed several small and independent
modules, called ‘s-bots’, and the collaboration in this system is achieved by means
of communication among them. They claimed that individuals that are mechani-
cally better for retrieval are more likely to be selected. Another point of view for the
collaboration is task allocation among the multi-robot colony. Mataric et al. [11]
had an experimentation comparing between a simulated data with physical mobile
robot experiment. The result showed that there is no single strategy that produces
20 J. Suh and C.‐W. Woo

best performance in all cases. And other approaches are the multi-robot task
allocation by planning algorithm [12–14].
All of these research efforts are being done in many different fields, and the
selection of individual is rather random or simple, which may often result in
inadequacy of performing a given task. Using ‘entropy’ of information theory
[15] could be a good alternative compare to the other informal approaches. Good-
rich argues that the behavioral entropy can predict human workload or measure of
human performance in human robot interaction [16] (HRI) domain. Balch [17]
demonstrated successfully in his experimental evaluation of multi-robot soccer and
multi-robot foraging teams. In our study, we will use the ‘entropy’ metric for
selecting an appropriate robot from the robot colonies by generating decision tree
first, to minimize the complexities of adapting the entropy.

3 Design of the System

As in the Fig. 1, the overall architecture of our system is divided into three main
layers, context-awareness layer, grouping layer, and collaboration layer. Also, we
have two more sub-layers, Physical and Network layer, which will not be discussed
here, since they are not the main issue here. The overall process of the three main
layers works as follows, and entire structure of the system can be viewed in the next
Fig. 2.
l The context-awareness layer generates a context from the raw information, and
then does modeling and reasoning in order to aware of the context.
l The grouping layer creates decision classifier based on the entropy mechanism,
and makes a necessary group.

Context- Grouping Collaboration

Awareness

Layer Layer Layer

Network Layer

Physical Layer

Fig. 1 General structure of social robot

2 A Framework for Collaborative Aspects of Intelligent Service Robot 21

CONTEXT-AWARENESS GROUPING COLLABORATION ROBOT

LAYER LAYER LAYER CLIENT (1)

Situation Recognizer Situation Task Action

Acceptor Allocator Planner
Inference Engine Rule
Classifier
Generator
Low-level
Context Integrator Task
Plan Rule
High-context Learner Planner
model
Actuator
Training Data
Context Provider High-level
low-context Grouper
Plan Rule
.
model
Classifier .
Raw Data Collector
.
Searcher Current ROBOT
Group Info. CLIENT (N)
Temporary & Context
Memory Collector
Group Info Action
Planner

Message Protocol

NETWORK LAYER Low-level

Plan Rule

PHYSICAL SENSOR 1 SENSOR 2 RFID ZIGBEE DEVICE Actuator

LAYER

[ SERVER SYSTEM ARCHITECTURE ] [ ROBOT CLIENT ARCHITECTURE ]

Fig. 2 Structure of the service robot in context awareness environment

l The collaboration layer does multi-level task planning; high-level task planning
generates a set of tasks for the context, the low-level task planning generates a
set of actions for the task.

3.1 Context-Awareness Layer

The context awareness layer receives raw data from surrounding computing
devices including RFID, Zigbee, and so on. Then it transforms the raw data into
a meaningful semantic data by going through some preprocessing steps, and
finally it generates a situation. This work can be done from the next sub modules
as follows.
Raw Data Collector: It simply receives raw data from the Physical layer passes
it to the context provider.
Context Provider: It receives raw data from the raw data collector and trans-
forms the data into standardized context as a preprocessing step according to low
context model. The low context model means that the raw data is formalized but it
is not semantic data.
Context Integrator: It receives standardized context from the context provider
and generates inference level context through the high-level context modeling. The
high level model supports converting the formalized context into a semantic
context.
22 J. Suh and C.‐W. Woo

Situation Recognizer: The context awareness process goes through the

above modules in sequence, and generates situation(s) by using rule-based
inference engine in this sub module. This situation is delivered to the grouping
layer.

3.2 Grouping Layer

When the situation generation is done by the situation acceptor, then this situation is
delivered to the grouping layer to make a group. The grouping layer first receives
information about which robot is connected to the server, and stores this informa-
tion into Group info database. Because currently connected robots through the
network can be the candidates for making a group. The grouping layer consists of
three sub modules, Situation Acceptor, Classifier Generator, and Grouper, and the
details of their work are as follows.
Situation Acceptor: It simply receives information regarding a situation, from
the Context awareness layer, and requests to the Classifier Generator to begin
grouping for this given situation.
Classifier Generator: It generates a classifier (i.e. decision tree) to make a
group for a given specific situation. And also we need to have a set of predefined
training data representing some characteristics of various kinds of robots. In this
study, we generated the classifier based on ID3 decision tree algorithm.
Grouper: The Grouper has the next two sub-modules; the searcher requests
instance information from the connected each robot through the network layer. In
other words, after the request for the instance information is acquired, such as
‘cleaning’, the grouper makes a group by using the classifier that generated from the
Classifier Generator module. The generated group information is stored in the
group info repository, and will be used for collaboration in the collaboration layer
later on.

3.2.1 Context Information for the Robot

For this experiment, we can set up several virtual situations, such as ‘cleaning’
situation, ‘delivery’ situation, ‘conversation’ situation, and so on. The Grouping
layer receives one from the situations, and start making a group that is appropriate
for the service.
The Fig. 3 is the user interface for entering robot attributes and a situation.
From this interface, we can enter five attributes interactively, such as ‘power’,
‘location’, ‘speed’, ‘possession’, ‘IL’, and a situation, such as ‘cleaning’. By using
this interface, we can create robot instances as many as we want arbitrarily and
they represent heterogeneous robots. And also we can set up a situation by just
selecting from the window. This means that each robot has different character-
istics and good for a certain work. For instance, we can set up a robot instance
2 A Framework for Collaborative Aspects of Intelligent Service Robot 23

Fig. 3 User Interface for entering robot attributes and situation

Table 1 Training data to create a decision tree

Index Power Location Speed Possession IL Cleaning
1 Low Near Fast Tray Poor Yes
2 Low Far Low Gripper Normal Yes
3 Middle Near Fast Interpreter Smart No
4 High Near Far Tray Normal No
5 Low Far Normal Tray Smart Yes
6 Low Near Normal Interpreter Normal Yes
7 Middle Far Normal Interpreter Normal No
8 High Near Slow Gripper Smart Yes
9 Middle Near Far Interpreter Smart No
10 High Far Slow Gripper Poor Yes

good for a ‘cleaning’ job as follows. If the robot has low power, location is near,
speed is low, possesses a tray, and so on, then we can consider this robot is good
for a cleaning situation. Similarly, we can create several robot instances for our
experimentation.

3.2.2 Training Data

Each robot’s attributes are described as in the Table 1, which shows that each robot
has different characteristics. For instance, ‘power, and speed’ means the current
robot’s basic characteristics, ‘location’ is the robot’s location to perform the given
24 J. Suh and C.‐W. Woo

context, ‘possession’ means a tool that robot can handle, and ‘IL’ means the robot’s
capability of language interpretation. Since we are not using sensed data from
computing devices for this simulation, we can generate robot instances through
the user interface arbitrarily, as many as possible. Table 1 is a set of training data
that shows ten robot instances for ‘cleaning’ situation.

3.2.3 Decision Tree Generation

The following equation is the entropy algorithm of the information theory, which
will be used to generate a tree.

EntropyðSÞ ¼ P ðþÞ log2 P ðþÞ P ðÞ log2 P ðÞ (1)

X j Su j
GainðS; AÞ ¼ EntropyðSÞ  EntropyðSÞ (2)
u2ValuesðAÞ
S u

We can compute the general entropy using the Eq. (1), and compute the gain
entropy to select a single attribute using Eq. (2).

3.3 Collaboration Layer

It consists of following three sub-components on the server side, and action planner
is on the client (robot) side. The distinctive features of them are as follows.
Group Info. and Context Collector: It collects information of selected situa-
tion and information for grouped robots from the Grouping Layer.
Task Planner: It generates a set of tasks using high-level planning rules
(a global plan) on the server side. For instance, the generated tasks for “cleaning”
situation can be the “sweeping” and “mopping”.
Task Allocator: The task allocator sends the generated tasks to appropriate
robots.
Action Planner: Generated tasks by the task planner are delivered to the client
(robot), and further refined intro a set of actions by the action planner.

3.3.1 Multi-level Task Planning

In collaboration layer, the task is carried out by the multi-level task planning
mechanism as follows (see Fig. 4).
l The task planner gets the situation(s), and generates a set of tasks based on the
high-level planning rules.
2 A Framework for Collaborative Aspects of Intelligent Service Robot 25

HIGH-LEVEL Context Entity

PLAN RULE
(GENERAL
TASK RULE)
Context 1 Context 2 … Context n

Task 1 Task 2 Task 3

Robot Type 1 Robot Type 2

LOW-LEVEL
PLAN RULE
(ROBOT-SPECIFIC
ACTION RULE) Action 1 Action 2 Action 3 Action 1

Fig. 4 Multi-level task planning mechanism

l Then the system allocates the tasks to the appropriate robot who can handle the
specific task.
l When the task allocation is done, then each assigned robot activates the action
planner to generate a set of actions.

4 Simulated Experimentation

The overall experimentation is divided into two parts, robot grouping and robot
collaboration. Robot grouping is done by generating classifier using the Entropy
metric, and the collaboration is done by task planning algorithm.

4.1 Robot Grouping

In this study, the robot grouping simulation experiment begins with generating
virtual robot instances and a situation through the user interface as in the Fig. 3. We
can set up characteristics of each robot by selecting five attributes and also can set
up a virtual situation through the interface. When all the selection is done, then we
can send the information to the server using the start/stop button in the interface.
Figure 5 shows the snapshot of the implemented simulation result for grouping.
We designed the implementation result as six sub windows, and the function of
each window is explained as follows.
l ‘Context’ window: It shows the selected virtual context, such as, ‘cleaning’.
l ‘Training Data’ window: It shows ten training data for the selected situation.
26 J. Suh and C.‐W. Woo

Fig. 5 Snapshot of simulated robot grouping

l ‘Robot Instance’ window: It shows activated instances.

l ‘Robot Grouping Result’ window: It shows grouped robots as instance numbers.
l ‘Tree’ windows: It shows the generated decision tree for the selected context.
l ‘Process’ window: It shows the entropy computation process for the generated
decision tree.

The entire grouping process can be summarized as follows:

l Create arbitrary number of robot instances through the user interface (see Fig. 3),
and this information is sent to server.
l Also, information of a ‘virtual situation’ is sent to server.
l Then, the server creates decision tree using the information of the robot
instances, and decide a group of robots for the specific situation (e.g. cleaning).
l As a result of the final process, the grouped robot instances are shown in the
bottom left window as robot instance numbers.
From the bottom left of the windows in the Fig. 5, we can see that the
robot instance number 1, 4, 6 and 7 are grouped together to perform the task,
‘cleaning’.
2 A Framework for Collaborative Aspects of Intelligent Service Robot 27

4.2 Robot Collaboration

Although there are many different approaches of collaborations in robot stud-

ies, we consider the collaboration as sharing a task among a group of robots.
For instance, a generated task can be further refined into a set of subtask, which
will be carried out by each individual robot. The system generates a set of tasks
or subtasks by the task planning mechanism, which will be described in the
below.

4.2.1 Task Planning Rules

The task planning rules for collaboration layer is divided into two levels, high level
planning rules (general task rules), and the low-level planning rules (robot-specific
action rules). The general task rule can generate set of tasks, and the robot specific
action rule can generate a set of actions to perform the task. For example, if
‘cleaning’ is the selected context, then task planner generates a set of tasks for
the ‘cleaning’ as ‘sweeping’, and ‘mopping’. When the task ‘sweeping’ is assigned
to a robot, the generated action plan is ‘move’, ‘lift’, ‘sweep’, and ‘release’. The
task planner works on the server side, and the action planner is located on the client
side. The sample task planning rules are as in the Table 2.

4.2.2 Multi-level Planning

Our system divides the planning mechanism in two levels, a high-level planning
and low-level planning. It is a kind of hierarchical planning mechanism that is
efficient enough to make the planning mechanism as simple as possible. The high-
level planner generates a set of subtasks to accomplish a given context, and saves
them in a stack that will be taken one by one by the low-level planner. When the
high-level planning is done, the low-level planner gets activated with each one of
the subtasks. A single subtask becomes a goal to accomplish in the low-level
planning process, and it generates a set of actions as a result.

Table 2 Task planning rules

High-level planning rules Low-level planning rules
If (Context ¼ ‘Cleaning’) then (NeedTask If (Task ¼ ‘Sweeping’) then (NeedAction
Sweeping, Mopping) (Move, Lift, Sweeping, Release))
If (Context ¼ ‘Serving’) then (NeedTask If (Task ¼ ‘Take_order’) then (NeedAction
(Take_order, Food_service, Check_out)). . .. (Give_Menu, Help_order)). . ..
28 J. Suh and C.‐W. Woo

5 Conclusion

In this paper, we have described a development of intelligent robot framework for

providing intelligent service in the context awareness environment. The signifi-
cances of our research could be as follows.
First, we have designed intelligent service robot framework and successfully
carried out simulated experimentation by generating several robot instances. The
several heterogeneous robot instances can be generated by the user, through
the user interface windows, as many as possible. Second, the robots are grouped
by the decision classifier based on the entropy metric of the information theory.
This approach will provide more reliable and systematic methods for the robot
grouping. Third, in this study, we consider the collaboration as follows. Once a task
is generated, then it gets refined into a set of subtasks, which will be assigned to
each individual robot. If each robot accomplishes its own subtask, then the original
task is being done as a result. We do not have any interaction among the robot at the
moment. Fourth, the generation of task is being done by the multi-level task
allocation planning mechanism. The high-level task planning is done on the server
side, and detailed action planning is done on the client side (robot).
Our approach may provide some useful solutions for the intelligent service
oriented applications, which require multiple robot instances. Our immediate next
work is to complete the development of context-awareness layer using sensed raw
data from the external physical computing devices. Therefore, when we complete
the development of the entire framework, we are going to setup experimentations
with several humanoid robots equipped with sensors and network communications.
Also, more detailed strategy and also interaction for collaboration among each
individual robot is needed.

Acknowledgement This work was supported by the Seoul R&BD program (10848) of Korea.

References

1. M. Weiser. The computer for the twenty-first century. Sci. Am. 265(3), 94–104 (1991)
2. G.K. Mostefaoui, J. Pasquier-Rocha, P. Brezillon, Context-aware computing: a guide for the
pervasive computing community. Proc. IEEE/ACS Int. Conf. Pervasive Serv. (ICPS’04)
19(23), 39–48 (2004)
3. A.K. Dey, G.D. Abowd, Towards a better understanding of context and context-awareness.
GVU Technical Report GIT-GVU-99-22, Georgia Institute of Technology, 1999
4. M. Baldauf, S. Dustdar, A Survey on context-aware systems. Int. J. Adhoc Ubiquitous
Comput. 2, 263–277 (2004)
5. H. Chen, An intelligent broker architecture for pervasive context-aware systems. Ph.D. thesis,
University of Maryland, 2004
6. T. Gu, H.K. Pung, D.Q. Zhang, A Service-oriented middleware for building context-aware
services. J. Netw. Comput. Appl. 28, 1–18 (2005)
7. D. Salber, A. Dey, G. Abowd, The context toolkit: aiding the development of context-enabled
applications, in Proceedings of CHI’99, Pittsburgh, PA, 1999, pp. 434–441
2 A Framework for Collaborative Aspects of Intelligent Service Robot 29

8. X.H. Wang, T. Gu, D.Q. Zhang, H.K. Pung, Ontology based context modeling and reasoning
using owl, in Proceedings of 2nd IEEE Annual Conference on Pervasive Computing and
Communications Workshops, PERCOMW’04, Orlando, Florida, USA, 2004, pp. 18–22
9. D. Rodic, A.P. Engelbrecht, Social network as a coordination techniques for multi-robot
systems. International conference on system design and applications, Springer, Berlin,
2003, pp. 503–513
10. T.H. Labella, M. Dorigo, J-L. Deneubourg, Self-organized task allocation in a group of robots.
Technical Report No. TR/IRIDIA/2004–6, Universite Libre de Bruxelles, 2004
11. M. Mataric, S. Sukhatme, E. Ostergaard, Multi-robot task allocation in Uncertain Environ-
ment. Auton. Robots 14, 255–263 (2003)
12. M. Sterns, N. Windelinckx Combining planning with reinforcement learning for multi-robot
task allocation. in Proceedings of adaptive agents and MASII, LNAI 3394, 2006, pp. 260–274
13. R. Alami, A. Clodic, V Montreuil, E. Akin, R. Chatila, Task planning for human-robot
interaction, in Proceedings of the 2005 joint conference on Smart Object and ambient
intelligence, 2005, pp. 81–85
14. B. Gerkey, and M. Mataric, Principled communication for dynamic multi-robot task alloca-
tion. Exp. Robotics II, 353–362 (2000)
15. C.E. Shannon, The Mathematical Theory of Communication (University of Illinois Press,
Illinois, 1949)
16. M.A. Goodrich, E.R. Boer, J.W. Crandall, R.W. Ricks, M.L. Quigley, Behavioral entropy in
human-robot interaction, Brigham University, Technical Report ADA446467, 2004
17. T. Balch, Hierarchic social entropy: an information theoretic measure of robot group diversity.
Auton. Robots 8, 09–237 (2000)
Chapter 3
Piecewise Bezier Curves Path Planning
with Continuous Curvature Constraint
for Autonomous Driving

Ji-Wung Choi, Renwick Curry, and Gabriel Elkaim

Abstract We present two practical path planning algorithms based on Bezier

curves for autonomous vehicles operating under waypoints and corridor constraints.
Bezier curves have useful properties for the trajectory generation problem. This
paper describes how the algorithms apply these properties to generate the reference
trajectory for vehicles to satisfy the path constraints. Both algorithms generate the
piecewise-Bezier-curves path such that the curves segments are joined smoothly
with C2 constraint which leads to continuous curvature along the path. The degree
of the curves are minimized to prevent them from being numerically unstable.
Additionally, we discuss the constrained optimization problem that optimizes the
resulting path for a user-defined cost function.

1 Introduction

Bezier Curves were invented in 1962 by the French engineer Pierre Bezier for
designing automobile bodies. Today Bezier Curves are widely used in computer
graphics and animation. The Bezier curves have useful properties for the path
generation problem as described in Section 2 of this paper. Hence many path
planning techniques for autonomous vehicles have been discussed based on Bezier
Curves in the literature. Cornell University Team for 2005 DARPA Grand Chal-
lenge used a path planner based on Bezier curves of degree 3 in a sensing/action
feedback loop to generate smooth paths that are consistent with vehicle dynamics
[5]. Skrjanc proposed a new cooperative collision avoidance method for multiple
robots with constraints and known start and goal velocities based on Bezier curves
of degree 4 [6]. In this method, four control points out of five are placed such that

J.-W. Choi (*)

Autonomous Systems Lab, University of California, Santa Cruz, CA 95064, USA
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 31

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_3, # Springer ScienceþBusiness Media B.V. 2010
32 J.‐W. Choi et al.

desired positions and velocities of the start and the goal point are satisfied. The fifth
point is obtained by minimizing penalty functions. Jolly described Bezier curve
based approach for the path planning of a mobile robot in a multi-agent robot soccer
system [3]. The resulting path is planned such that the initial state of the robot and
the ball, and an obstacle avoidance constraints are satisfied. The velocity of the
robot along the path is varied continuously to its maximum allowable levels by
keeping its acceleration within the safe limits. When the robot is approaching a
moving obstacle, it is decelerated and deviated to another Bezier path leading to the
estimated target position.
Our previous works introduced two path planning algorithms based on Bezier
curves for autonomous vehicles with waypoints and corridor constraints [1, 2]. Both
algorithms join cubic Bezier curve segments smoothly to generate the reference
trajectory for vehicles to satisfy the path constraints. Also, both algorithms are
constrained in that the path must cross over a bisector line of corner area such that
the tangent at the crossing point is normal to the bisector. Additionally, the con-
strained optimization problem that optimizes the resulting path for user-defined cost
function was discussed. Although the simulations provided in that paper showed
the generation of smooth routes, discontinuities of the yaw angular rate appeared at
junction nodes between curve segments. This is because the curve segments are
constrained to connect each other by only C1 continuity, so the curvature of the path
is discontinuous at the nodes. (Section 2 describes this more detail.)
To resolve this problem, we propose new path planning algorithms. The algo-
rithms impose constraints such that curve segments are C2 continuous in order to
have curvature continuous for every point on the path. In addition, they give the
reference path more freedom by eliminating redundant constraints used in previous
works, such as the tangent being normal to the bisector and symmetry of curve
segments on corner area. The degree of each Bezier curve segments are determined
by the minimum number of control points to satisfy imposed constraints while
cubic Bezier curves are used for every segments in previous works. The optimized
resulting path is obtained by computing the constrained optimization problem that
controls the tradeoff between shortest and smoothest path generation. Furthermore,
the proposed algorithms satisfy the initial and goal orientation as well as position
constraint while previous works only made the position constraint satisfied. The
numerical simulation results demonstrate the improvement of trajectory generation
in terms of smoother steering control and smaller cross track error.

2 Bezier Curve

A Bezier Curve of degree n can be represented as

X
n
PðtÞ ¼ Bni ðtÞPi ; t 2 ½0; 1
i¼0
3 Piecewise Bezier Curves Path Planning with Continuous Curvature Constraint 33

where Pi are control points and Bni ðtÞ is a Bernstein polynomial given by
 
n
Bni ðtÞ ¼ ð1  tÞni ti ; i 2 f0; 1; . . . ; ng
i

Bezier Curves have useful properties for path planning:

l They always start at P0 and stop at Pn.
l They are always tangent to P0 P1 and Pn Pn1 at P0 and Pn respectively.
l They always lie within the convex hull consisting of their control points.

2.1 The de Casteljau Algorithm

The de Casteljau algorithm describes a recursive process to subdivide a Bezier

curve P(t) into two segments. The subdivided segments are also Bezier curves.
Let fP00 ; P01 ; . . . ; P0n g denote the control points of P(t). The control points of the
segments can be computed by

Pji ¼ ð1  tÞP j1

i þ t P j1
iþ1 ;
(1)
t 2 ð0; 1Þ; j 2 f1; . . . ; ng; i 2 f0; . . . ; n  jg

Then, fP00 ; P10 ; . . . ; Pn0 g are the control points of one segment and fPn0 ; Pn1
1 ; ...,
Pn0} are the another. (See an example in Fig. 1.) Note that, by applying
the properties of Bezier Curves described in the previous subsection, both of the
subdivided curves end at Pn0 and the one is tangent to Pn1 n
0 P0 and the other is to
Pn0 Pn1
1 at the point. Since Pn0 is chosen on Pn1
0 P1
n1
by using (1), the three points
n1 n n1
P0 , P0 , and P1 are collinear.
Remark 1. A Bezier curve P(t) constructed by control points fP00 ; P01 ; . . . ; P0n g
always passes through the point Pn0 and is tangent to Pn1
0 P1
n1 at Pn .
0

P01 P11
P30 P21
P20 P02
P10 P12
Fig. 1 Subdividing a cubic
Bezier curve with t ¼ 0. 4
P00 P03
by the de Casteljau Algorithm
34 J.‐W. Choi et al.

2.2 Derivatives, Continuity and Curvature

The derivatives of a P
Bezier curve can be determined by its control points. For a
Bezier curve PðtÞ ¼ ni¼0 Bni ðtÞPi , the first derivative is given by

X
n1
_
PðtÞ ¼ Bn1 ðtÞDi (2)
i
i¼0

_
where Di, control points of PðtÞ are

Di ¼ nðPiþ1  Pi Þ

Geometrically, (2) provides us with a tangent vector. The higher order derivative
can be obtained by using the relationship of (2), iteratively.
Two Bezier curves P(t) and Q(t) are said to be Ck at t0 continuous if

Pðt0 Þ ¼ Qðt0 Þ; Pðt _ 0 Þ; . . . ; PðkÞ ðt0 Þ ¼ QðkÞ ðt0 Þ

_ 0 Þ ¼ Qðt (3)

The curvature of a Bezier curve P(t) ¼ (x(t), y(t)) at t is given by

_ xðtÞj
_ yðtÞ  yðtÞ€
jxðtÞ€
kðtÞ ¼ 3 (4)
ð_x2 ðtÞ þ y_ 2 ðtÞÞ2

Lemma 1. For the path constructed by two Bezier curve segments PðtÞjt2½t0 ;t1 
and QðtÞjt2½t1 ;t2  , if P(t) and Q(t) are at least C2 continuous at t1 then the path has
continuous curvature for every point on it.
Proof. The curvature is expressed in terms of the first and the second derivative of a
curve in (4). Since the Bezier curves are defined as polynomial functions of t, their
k-th derivative for all k ¼ 1, 2, . . . are continuous. Hence, they always have
continuous curvature for all t. For two different Bezier curves P(t) and Q(t), it is
_
sufficient that k(t1), the curvature at the junction node is continuous if PðtÞ _
¼ QðtÞ
€ €
and PðtÞ ¼ QðtÞ are continuous at t1.

3 Problem Statement

Consider the control problem of a ground vehicle with a mission defined by

waypoints and corridor constraints in a two-dimensional free-space. Our goal is
to develop and implement an algorithm for navigation that satisfies these con-
straints. That is divided into two parts: path planning and path following.
3 Piecewise Bezier Curves Path Planning with Continuous Curvature Constraint 35

To describe each part, we need to introduce some notation. Each waypoint

is denoted as Wi 2 R2 for i 2 {1, 2, . . ., N}, where N 2 R is the total number of
waypoints. Corridor width is denoted as wj for j 2 {1, . . ., N  1}, j-th widths of
each segment between two waypoints, Wj and Wj+1. For the dynamics of the
vehicle, the state and the control vector are denoted q(t) ¼ (xc(t), yc(t), c(t))T and
u(t) ¼ (v(t), o(t))T respectively, where (xc, yc) represents the position of the center of
gravity of the vehicle. The yaw angle c is defined to the angle from the X axis. v is
the longitudinal velocity of the vehicle. o ¼ c_ is the yaw angular rate. State space
is denoted as C ¼ R3 . We assume that the vehicle follows that
0 1
cos cðtÞ 0
_ ¼ @ sin cðtÞ
qðtÞ 0 AuðtÞ
0 1

With the notation defined above, the path planning problem is formulated
as: Given an initial position and orientation and a goal position and orientation of
the vehicle, generate a path l specifying a continuous sequence of positions and
orientations of the vehicle satisfying the path constraints [4]. In other words, we are
to find a continuous map

l : ½0; 1 ! C

with

lð0Þ ¼ qinit and lð1Þ ¼ qgoal

where qinit ¼ (W1, c0) and qgoal ¼ (WN, cf) are the initial and goal states of the
path, respectively.
Given a planned path, we use the path following technique with feedback
corrections as illustrated in Fig. 2. A position and orientation error is computed
every 50 ms. A point z is computed with the current longitudinal velocity and

z
yerr

p
yerr

Fig. 2 The position error is Y (xc, yc)

measured from a point z, q
projected in front of the
vehicle, and onto the desired X
curve to point p
36 J.‐W. Choi et al.

heading of the vehicle from the current position. z is projected onto the reference
trajectory at point p such that zp is normal to the tangent at p. The cross track error
yerr is defined by the distance between z and p. The steering control o uses a PID
controller with respect to cross track error yerr.
ð
dyerr
do ¼ kp yerr þ kd þ ki yerr dt
dt

4 Path Planning Algorithm

In this section, two path planning methods based on Bezier curves are proposed.
To describe the algorithms, let us denote b ^j as the unit vector codirectional with
the outward bisector of ∠Wj1WjWj+1 for j 2 {2, . . ., N  1} as illustrated in Fig. 3.
The planned path must cross over the bisectors under the waypoint and the corridor
constraints. The location of the crossing point is represented as Wj þ dj  b ^j , where
dj 2 R is a scalar value. The course is divided into segments Gi by bisectors. Gi
indicates the permitted area for vehicles under corridor constraint wi, from Wi to Wiþ1.
Bezier curves constructed by large numbers of control points are numerically
unstable. For this reason, it is desirable to join low-degree Bezier curves together in
a smooth way for path planning [6]. Thus both of the algorithms use a set of low-
degree Bezier curves such that the neighboring curves are C2 continuous at their
end nodes. This will lead to continuous curvature on the resulting path by Lemma 1.
The Bezier curves used for the path plannings are denoted as i PðtÞ ¼
Pni
k¼0 Bk ðtÞ Pk for i 2 {1, . . ., M}, t 2 [0, 1], where M is the total number of
ni i

the Bezier curves and ni is the degree of iP. The planned path l is a concatenation of
all iP such that

lðði  1 þ tÞ=MÞ ¼ i PðtÞ; i 2 f1; . . . ; Mg; t 2 ½0; 1

bˆ 3
W3
G3
W4
G2

Fig. 3 An example of the W2

course with four waypoints. G1 bˆ 2
Y
Gray area is the permitted W1
area for vehicles under a
corridor constraint X
3 Piecewise Bezier Curves Path Planning with Continuous Curvature Constraint 37

4.1 Path Planning Placing Bezier Curves within Segments (BS)

In this path planning method (BS), the Bezier curve iP for i 2 {1, . . ., N  1} are
used within each segment Gi. The adjacent curves, j1P and jP are C2 continuous
^j . The control points of iP, iPk for k ¼ {0, . . ., ni}
at the crossing point, Wj þ dj  b
are determined to maintain these conditions.
l The beginning and the end point are W1 and WN.

1
P0 ¼ W 1 ; N1
PnN1 ¼ WN (5)

l The beginning and end orientation of l are c0 and cf.

1
P1 ¼ W1 þ l0 ðcos c0 ; sin c0 Þ; l0 2 Rþ (6a)

N1
PnN1 1 ¼ WN  lf ðcos cf ; sin cf Þ; lf 2 Rþ (6b)

l
j1
P and jP, 8j 2 {2, . . ., N  1} are C2 continuous at the crossing point.

j1 ^j
Pnj1 ¼ j P0 ¼ Wj þ dj  b (7a)

nj1 ðj1 Pnj1  j1

Pnj1 1 Þ ¼ nj ðj P1  j P0 Þ (7b)

nj1 ðnj1  1Þðj1 Pnj1  2  j1 Pnj1 1 þ j1 Pnj1 2 Þ

(7c)
¼ nj ðnj  1Þðj P2  2j P1 þ j P0 Þ

l The crossing points are bounded within the corridor.

1
jdj j < 8 minðwj1 ; wj Þ (8)
2

l Pk, k ¼ {1, . . ., ni  1} always lie within the area of Gi.

i

i
P1 2 Gi ; . . . ; i Pni 1 2 Gi (9)

Equations (6a, b) are derived by using the tangent property of Bezier curves at
their end points. Equations (7a–c) are obtained by applying (2) and (3). Equation (9)
makes the resulting Bezier curve satisfy the corridor constraint by the convex hull
property. At each crossing point, three control points of each adjacent Bezier curve
are dedicated to the C2 continuity constraint by (2), (4), and Lemma 1. So the
minimum number of control points to satisfy the constraints independent of the
others are six for iP, i 2 {2, . . ., N  2}. On other hand, 1P needs five: three to
38 J.‐W. Choi et al.

connect 2P plus two to connect W1 with slope of c0. Likewise, 2N3

P needs five.
Thus, ni is determined as

4; i 2 f1; N  1g
ni ¼
5; i 2 f2; . . . ; N  2g

Note that 1P0 and N1 PnN1 are fixed in (5). 1P1 and N1 PnN1 1 are fixed in
(6a, b). j1 Pnj1 and jP0 rely on dj in (7a–c). j1 Pnj1 1 and j1 Pnj1 2 rely on jP1
and j  1P2. So the free variables are, 8j 2 {2, . . ., N  1}, P1 ¼ {jP1}, P2 ¼ {jP2}, d
¼ {dj}, and L ¼ {l0, lf}. The number of the variables or the degrees of freedom is
5N  8. The variables are computed by minimizing the constrained optimiza-
tion problem:
X
N 1
min J ¼ Ji (10)
P1 ;P2 ;d;L
i¼1

subject to (8) and (9), where Ji is the cost function of iP(t), given by
ð1
Ji ¼ ½ðai ð_x2 ðtÞ þ y_ 2 ðtÞÞ þ bi ji kðtÞj2 dt (11)
0

where ai 2 R and bi 2 R are constants that control the tradeoff between arc lengths
versus curvatures of the resulting path.

4.2 Path Planning Placing Bezier Curves on Corners (BC)

Another path planning method (BC) adds quadratic Bezier curves on the corner area
around Wj, j 2 {2, . . ., N  1}. The quadratic Bezier curves are denoted as
P
j
QðtÞ ¼ 2k¼0 B2k ðtÞ j Q0k intersects the j-th bisector. The first and the last control
point, j Q00 and j Q02 are constrained to lie within Gj1 and Gj, respectively. Within
each segment Gi, another Bezier curve is used to connect the end points of jQ
with C2 continuity and/or W1 with slope of c0 and/or WN with cf. Hence, jQ is the
curve segments of even number index:

2ð j1Þ
PðtÞ ¼ j QðtÞ; j 2 f2; . . . ; N  1g

The degree of iP is determined by the minimum number of control points to satisfy

C2 continuity constraint, independently:
8
< 4; i 2 f1; 2N  3g
ni ¼ 2; i 2 f2; 4; . . . ; 2N  4g
:
5; i 2 f3; 5; . . . ; 2N  5g
3 Piecewise Bezier Curves Path Planning with Continuous Curvature Constraint 39

Let tc denote the Bezier curve parameter corresponding to the crossing point of
j
Q(t) on the bisector, such that

j ^j :
Qðtc Þ ¼ Wj þ dj  b (12)

Let yj denote the angle of the tangent vector at the crossing point from X-axis:

j _ c Þ ¼ ðjj Qðt
Qðt _ c Þj cos yj ; jj Qðt
_ c Þj sin yj Þ: (13)

The notations are illustrated in Fig. 4. Due to the constraint of j Q00 and j Q02 within
Gj  1 and Gj, the feasible scope of yj is limited to the same direction as Wjþ1 is
with respect to b ^j . In other words, if Wjþ1 is to the right of b^j , then yj must point
^
to the right of bj , and vice versa.
Given j Q00 , j Q02 , dj, and yj, the other control point j Q01 is computed such that the
crossing point is located at Wj þ dj  b ^j and the angle of the tangent vector at the
crossing point is yj. Since each control point is two-dimensional, the degrees of
freedom of jQ(t) are six. Since dj and yj are scaler, representing jQ(t) in terms of
Q0 , Q2 , dj, and yj does not affect the degrees of freedom. However, it comes up
j 0 j 0

with an advantage for corridor constraint. If we compute j Q01 such as above, the
points computed by applying the de Casteljau algorithm such that two subdivided
curves are separated by the j-th bisector are represented as j Q00 and j Q02 as described
in the following. The two curves are constructed by fj Q00 ; j Q10 ; j Q20 g and
fj Q20 ; j Q11 ; j Q02 g. We can test if the convex hull of fj Q00 ; j Q10 ; j Q20 g lies within Gj
j 2 j 1 j 0
 1 and if that of f Q0 ; Q1 ; Q2 g lies within Gj in (26), instead of testing that of
j 0 j 0 j 0
f Q0 ; Q1 ; Q2 g. (Note that j Q01 is not constrained to lie within corridor as shown in
Fig. 4.) So, the convex hull property is tested for tighter conditions against the
corridor constraint without increasing the degrees of freedom.
In order to compute j Q01 , the world coordinate frame T is transformed and rotated
into the local frame jT where the origin is at the crossing point, jQ(tc) and X axis
is codirectional with the tangent vector of the curve at the crossing point, j Qðt _ c Þ.

j Q0
bˆ j 1

j
Q̇ (tc )
Wj

–dj qj
Y

j Q(t )
c
j Q0
2

j
Q00
Fig. 4 The geometry of the
j
Q(t) in corner area around Wj X
40 J.‐W. Choi et al.

Fig. 5 The geometry of the Y

control points of jQ(t) with j Q0
1
respect to jT

∗
at ∗ a(1–t )
j Q1 j Q1
0 1
X
j
Q20
j Q0
0

j
Q02

Let us consider the subdivision ratio, t∗ 2 (0, 1) such that the location of j Q20
computed by applying the de Casteljau algorithm with it is the crossing point. In
other words, t∗ has j Q20 be at the origin with respect to jT frame. Fig. 5 illustrates the
control points of jQ(t) with respect to jT frame. Note that j Q20 is at the origin by the
definition of jT and t∗. j Q10 and j Q11 are on the X axis by the definition of jT and
Remark 1. Let the coordinates of the control points be denoted as Q0i ¼ ðxi ; yi Þ,
i 2 {0, 1, 2}, where all coordinates are with respect to jT.
Lemma 2. Given dj and yj, for jQ(t) to intersect j-th bisector with the crossing
point determined by the dj and (12), and the tangent vector at the point determined
by the yj and (13), it is necessary that y0y2  0.
Proof. Let (x(t), y(t)) denote the coordinate of jQ(t) with respect to jT. By the
definition of jT and Remark 1, Q(t) passes through the origin with tangent slope
_ c Þ ¼ 0. Suppose that
of zero with respect to jT. That is, x(tc) ¼ 0, y(tc) ¼ 0 and yðt
_ > 0 and y€ðtÞ < 0 for t 2 [0, tc).
y0 ¼ y(0) < 0. Since y(t) is a quadratic polynomial, yðtÞ
_ < 0 and y€ðtÞ < 0 for t 2 (tc, 1]. Thus, y2 ¼ y(1) < 0 and y0y2 > 0.
Subsequently, yðtÞ
Similarly, if y0 > 0 then y1 > 0. If y0 ¼ 0 then yðtÞ _ ¼ 0 for t 2 [0, 1] and y2 ¼ 0.
Thus, y0y2 ¼ 0. □
We are to calculate j Q01 depending on whether y0y2 is nonzero. For simplicity,
superscript j is dropped from now on. Without loss of generality, suppose that
y0 < 0 and y2 < 0. Q20 is represented as

Q20 ¼ ð1  t ÞQ10 þ t Q11

by applying (1). So the coordinates of Q10 and Q11 can be represented as

Q10 ¼ ðat ; 0Þ; Q11 ¼ ðað1  t Þ; 0Þ; a>0 (14)

where a > 0 is some constant. Applying (1) with i ¼ 0 and j ¼ 1 and arranging the
result with respect to Q01 by using (14) gives
 
1  t 1  t
Q1 ¼  a 
0
x0 ;   y0 (15)
t t
3 Piecewise Bezier Curves Path Planning with Continuous Curvature Constraint 41

Similarly, applying (1) with i ¼ 1 and j ¼ 1 yields

 
t t
Q01 ¼ a x 2 ;  y (16)
1  t 1  t 2

where a and t∗ are obtained by equations (15) and (16):

1 x0 y2  y0 x2
t ¼ pffiffiffiffiffiffiffiffiffiffiffi ; a¼ pffiffiffiffiffiffiffiffiffiffiffi (17)
1 þ y2 =y0 2y0 y2 =y0

t∗ and a have the square root of y2/y0. So, if y0y2 < 0 then t∗ and a are not
determined, hence, Q(t) is infeasible. That is, (17) ends up with Lemma 2.
If y0 ¼ y2 ¼ 0 then all control points of jQ are on X axis (see proof of Lemma 2).
In the geometric relation of control points and the points computed by applying the
de Casteljau algorithm as shown in Fig. 6, we obtain

x0 ¼ ða þ bÞt
x2 ¼ ða þ gÞð1  tÞ (18)
a ¼ bð1  tÞ þ gt

where a > 0, b > 0, g > 0 are some constants. Using (18), Q01 ¼ ðx1 ; 0Þ is
represented in terms of arbitrary t 2 (0, 1):
 
1 1t t
x1 ¼  x0 þ x2 (19)
2 t 1t

The constraints imposed on the planned path are formulated as follows:

l The beginning and end position of l are W1 and WN.

1
P0 ¼ W 1 ; 2N3
P4 ¼ W N (20)

l The beginning and end orientation of l are c0 and cf.

1
P1 ¼ W1 þ l0 ðcos c0 ; sin c0 Þ; l0 2 R þ (21a)

2N3
P3 ¼ WN  lf ðcos cf ; sin cf Þ; lf 2 R þ (21b)

bt b(1 – t ) gt g (1 – t )

Fig. 6 The geometry of jQ(t) Q 00 Q 10 Q 20 Q 01 Q 11 Q 02

with respect to jT when y0 ¼
y2 ¼ 0 at a (1 – t )
42 J.‐W. Choi et al.

l
i1
P and iP, 8i 2 {2, . . ., 2N  3} are C2 continuous at the junctions.

i1
Pni1 ¼ i P0 (22a)

ni1 ði1 Pni1  i1

Pni1 1 Þ ¼ ni ði P1  i P0 Þ (22b)

ni1 ðni1  1Þði1 Pni1  2  i1 Pni1 1 þ i1 Pni1 2 Þ

(22c)
¼ ni ðni  1Þði P2  2  i P1 þi P0 Þ

l The crossing points are bounded within the corridor.

1
jd j j < minðwj1 ; wj Þ; 8j 2 f2; . . . ; N  1g (23)
2
l ^j .
yj has the same direction as Wjþ1 is with respect to b

mod ðffðWjþ1  Wj Þ  ffb^j ; 2pÞ > p

(24a)
^j ; 2pÞ > p;
) mod ðyj  ffb

mod ðffðWjþ1  Wj Þ  ffb^j ; 2pÞ < p

(24b)
^j ; 2pÞ < p
) mod ðyj  ffb

l j
Q00 and j Q02 with respect to jT satisfies Lemma 2.

y0 y2  0 (25)

where y0 and y2 are with respect to jT.

j 0
l Q0 and j Q10 lie within Gj1. j Q02 and j Q11 lie within Gj.

j
Q00 2 Gj1 ; j Q10 2 Gj1 ; j Q02 2 Gj ; j Q11 2 Gj (26)

l {iP1, . . ., iPni1} always lie within the area of Gi.

i
P1 2 Gi ; . . . ; i Pni 1 2 Gi ; i 2 f1; 3; . . . ; 2N  3g (27)

The free variables are, for j 2 {2, . . ., N  1}, Q ¼ {jQ0, jQ2}, d ¼ {dj},
u ¼ {yj}, and L ¼ {l0, lf}. The degrees of freedom is 6N  10. The variables are
computed by minimizing the constrained optimization problem:

X
N 1
min J ¼ Ji (28)
Q;d;y;L
i¼1
3 Piecewise Bezier Curves Path Planning with Continuous Curvature Constraint 43

subject to (23), (24,b), (25), (26), and (27), where Ji is the cost function of iP(t),
defined in (11). Notice that the convex hull property is tested for j Q10 and j Q11 of the
divided curves instead of j Q01 in (26). Thus, it comes up with tighter conditions for
curves against the corridor constraint.

5 Simulation Results

Simulations were run for the course with four waypoints and identical corridor
width of 8 as illustrated in Fig. 7. The initial and goal orientation are given by
c0 ¼ cf ¼ 0. For path following, the constant longitudinal velocity v(t) ¼ 10 m/s is
used. The magnitude of o is bounded within |o|max ¼ 25 rpm. The PID gains are
given by: kp ¼ 2, kd ¼ 1, and ki ¼ 0.1.

λlBS λcBS
70 70
Planned Path Planned Path
60 Actual Trajectory 60 Actual Trajectory

Vf Vf
50 50
North [m]

North [m]

40 40

30 30

20 20

10 10

V0 V0
0 0
0 20 40 60 80 0 20 40 60 80
East [m] East [m]

λlBC λcBC
70 70
Planned Path Planned Path
60 Actual Trajectory 60 Actual Trajectory

Vf Vf
50 50
North [m]

North [m]

40 40

30 30

20 20

10 10
V0 V0
0 0
0 20 40 60 80 0 20 40 60 80
East [m] East [m]

Fig. 7 The planned paths by BS method (top) and by BC method (bottom). Each pair of results are
obtained by minimizing arc lengths (left) and by minimizing curvatures (right)
44 J.‐W. Choi et al.

In Fig. 7, the reference paths (dashed curves) planned by applying the BS and BC
method are placed in top and bottom row, respectively. The actual trajectories
(solid curves) are generated by using the proposed tracking method. In the figures,
stars indicate the location of control points of Bezier curve segments at an even
number index, and empty circles indicate those of the others. The paths in the left
column are planned by minimizing the summation of (11) with ai ¼ 1 and bi ¼ 0.
Since only arc length is penalized, the cost function leads to paths with minimum
arc length, which we denote llBS and llBC for the BS and BC method, respectively.
On other hand, the paths in the right column are planned by minimizing the cost
function with ai ¼ 0 and bi ¼ 1 so that resulting paths with larger radii of curvature
are provided. We denote them lcBS and lcBC for the BS and BC method, respectively.
lcBS and lcBC generate longer but smoother trajectory guidance than llBS and llBC .
Taking a look at the tracking results on llBS and llBC , the vehicle overshoots in sharp
turns around W3 resulting in a large position error (see Fig. 8) due to the limit on
steering angle rate. The commanded steering angle rate on llBC , marked by ‘*’ in
Fig. 9 undergoes rapid changes and is constrained by the rate limit. However, on lcBS
and lcBC , the vehicle tracks the part of the planned paths accurately thanks to larger
radii of curvature. We can verify that more tangibly in cross track errors plot
provided in Fig. 8. Also, the steering control signal on lcBC , marked by ‘o’ in
Fig. 9 is smoother than that on llBC .

0.2
l
λBC
Cross Track Error [m]

c
λ BC
0.1

– 0.1

– 0.2
Fig. 8 The cross track errors 0 5 10
by BC Time [s]

2
l
λBC
1 λcBC
Yaw Rate [rad / s]

–1
– |w| max
–2

–3
Fig. 9 The steering controls 0 2 4 6 8 10 12
by BC Time [s]
3 Piecewise Bezier Curves Path Planning with Continuous Curvature Constraint 45

The main difference of the proposed algorithm from the previous ones of [1] is
the degree of continuity at the junctions: C2. Assuming that the vehicle tracks a
reference path perfectly and v is continuous, if k is continuous then o is continuous
because o ¼ kv. When v is constant as this simulation, since o is proportional to k,
the continuity characteristic of o tends to follow that of k. Since the previous
algorithms imposed only C1 continuity on junction nodes of the path, k is discon-
tinuous at the nodes. Hence the path underwent the discontinuity of angular rate,
that is, large angular acceleration, which leads to large torque on the vehicle. On
other hand, the proposed algorithm has k continuous along the resulting paths. If the
path has curvature small enough so that the vehicle is able to track it accurately
given a maximum steering angle rate, then the steering control signal will be
smooth, as that of lcBC in Fig. 9.

6 Conclusions

This paper presents two path planning algorithms based on Bezier curves for
autonomous vehicles with waypoints and corridor constraints. Bezier curves pro-
vide an efficient way to generate the optimized path and satisfy the constraints at
the same time. The simulation results also show that the trajectory of the vehicle
follows the planned path within the constraints.

References

1. J. Choi, R.E. Curry, G.H. Elkaim, in Path Planning Based on Bezier Curve for Autonomous
Ground Vehicles (Chapter 19, IEEE Computer Society, 2009), pp. 158–166
2. J. Choi, G.H. Elkaim, Bezier curves for trajectory guidance. Proceedings of the World
Congress on Engineering and Computer Science, WCECS 2008, San Francisco, CA (2008)
3. K.G. Jolly, K.R. Sreerama, R. Vijayakumar, A bezier curve based path planning in a multi-
agent robot soccer system without violating the acceleration limits. Robot. Autonom. Syst. 57,
23–33 (2009)
4. J.C. Latombe, in Robot Motion Planning (Kluwer, Norwell, MA, 1991)
5. I. Miller, S. Lupashin, N. Zych, P. Moran, B. Schimpf, A. Nathan, E. Garcia, in Cornell
University’s 2005 DARPA Grand Challenge Entry, vol. 36, chapter 12 (Springer, Heidelberg,
2007), pp. 363–405
6. I. Skrjanc, G. Klancar, Cooperative collision avoidance between multiple robots based
on bezier curves. Proceedings of the Information Technology Interfaces, 2007 (ITI 2007), pp.
451–456
Chapter 4
Combined Heuristic Approach
to Resource-Constrained Project
Scheduling Problem

Miloš Šeda, Radomil Matoušek, Pavel Ošmera, Čeněk Šandera,

and Roman Weisser

Abstract This chapter deals with the resource-constrained project scheduling

problem that belongs to NP-hard optimisation problems. There are many different
heuristic strategies how to shift activities in time when resource requirements
exceed their available amounts. We propose a transformation of the problem to a
sequence of simpler instances of (multi)knapsack problems that do not use tradi-
tionally predefined activity priorities and enable to maximise limited resources
in all time intervals given by start or end of an activity and therefore to reduce
the total time.

1 Introduction

A classical task of project management is to create the network graph of a

project and, on the basis of knowledge or an estimation of time of activities,
to determine critical activities that would have influenced a project delay. Each
activity draws on certain resources, e.g. financial resources, natural resources
(energy, water, material, etc.), labour, managerial skills. The solution of this
task is well known in the case when we do not take limited resources into
consideration.
The scheduling problem is a frequent task in the control of various systems
such as manufacturing processes [1], project management [2], and service system
control (reservation systems, timetabling).

M. Šeda (*)
Institute of Automation and Computer Science, Faculty of Mechanical Engineering,
Brno University of Technology, Technická 2, Brno 616 69, Czech Republic
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 47

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_4, # Springer ScienceþBusiness Media B.V. 2010
48 M. Šeda et al.

However, in real situations, available capacities of resources are constrained to

certain limits. Project planning under limited resources [1] is difficult because of
l Interdependence among activities due to sharing the same resources
l Dependence of resource consumption on the manner in which the activity is
subdivided
l Dependence on the limits of availability of the various resources
The character of all the three dependencies is basically non-linear. In general,
scheduling problems are NP-hard, consequently no polynomial-time algorithms are
known to guarantee an optimal solution.
Classical methods of network analysis offer some approaches which depend on
whether activities that have a lack of resource(s) may be suspended or not and
whether activities will be put off until the moment when it is possible to perform
them. A question is: Which concurrent activities should be suspended or put off
because of a lack of resource(s)?
An excellent survey [3] with 203 references presents classification, models and
methods for solving the resource-constrained project scheduling problem (RCPSP)
and, in [4], it is critically reviewed and completed. There are single-mode and
multi-mode variants of the problem, resources can be non-renewable or renewable,
activities of project can be interrupted or not and so on.
With respect to NP-hardness of the RCPSP, mainly heuristic methods are used
for its solving. Papers frequently present applications of stochastic heuristics such
as simulated annealing [5, 6], genetic algorithms [6–9], tabu-search [10, 11], ant
systems [12] and swarm optimization method [13].
In this paper we assume that, when resources are not sufficient, a suitable method
for shifting one or more activities has been selected and we touch another problem.
In real situations, the durations of activities can only be estimated and are subject to
change. If we included this new information in our considerations, this could result
in quite a different optimal schedule. Since, obviously, the dates of activities in the
past cannot be changed, it is necessary for the calculation of the entire project with
new data to modify the dates of only those activities that have not yet been finished.

2 Basic Notions

In this paragraph we introduce the notation used in this paper and the basic concepts
of CPM. We consider a network graph G with n topologically ordered vertices [14],
which means that, for each edge (i, j), i appears before j (i < j) and the starting
vertex has number n0 ¼ 1 and ending vertex number n. This ordering can be gained
as follows:
1. Start from the origin and assign n0 to it.
2. Leave all edges outgoing from n0 and assign numbers n0 + 1, . . . , n0 + k1 to k1
vertices that have no input edge.
4 Combined Heuristic Approach to Resource-Constrained Project Scheduling Problem 49

3. Leave all edges outgoing from the vertices numbered in the previous step and
assign numbers n0 + k1 + 1, . . . , n0 + k1 + k2 to k2 vertices that have no input edge.
4. Continue this way until all vertices are numbered.
Assume that edges represent activities of a project and vertices correspond to
beginnings or ends of activities. Denote E(G) the set of edges of a graph G, and V
(G) its set of vertices. If e ¼ (i, j) is an edge in E(G), then we denote its duration
by tij or t(i, j), or te in short. Similarly, requirements of activity (i, j) for a resource r
will be denoted by rij etc.
The following notions refer to start and end vertices of the network graph: Ti(0)
represents the earliest possible start time of vertex i, Tj(1) the latest allowable finish
time of vertex j and TS(i, j) ¼ Tj(1)  Ti(0)  ti j is the total slack of activity (i, j)
or total (activity) float (the amount of time by which the start of a given activity can
be delayed without delaying the completion of the project).
Finally, assume that Vi(e) denotes the starting vertex of an edge e ¼ (i, j) and
Vj(e) is its ending vertex.
Further notation will be introduced when needed.

3 Algorithm

The algorithm is based on time shifting of activities when their total requirements
are higher than the resource limit. This is implemented by prolonging their duration
but distinguishing, for each activity, its starting duration and current duration which
equals the length of shift and starting duration. The greatest advantage of this access
is that whenever we need to compute new earliest possible start times and latest
allowable finish times for activities after shifts or update the actual time duration of
some activities, we can compute the whole project using a simple CPM method and,
in spite of this, the dates of finished activities remain unchanged in the result of the
new calculation. In other words, any change in the present has no effect on results in
the past.
Let us denote
tsij starting duration of activity (i,j)
tcij current duration of activity (i,j)
dij ¼ tcij  tsij interval when activity has no requirements (“sleeps”)
Now we will formulate an algorithm. The symbol:¼ stands for the assignment
operator.
1. Initialization
Using CPM method, we determine for each edge the earliest possible start time
and the latest allowable finish time. Let us assign

t1 :¼ 0 (1)
50 M. Šeda et al.

dij :¼ 0; tcij :¼ tsij for every ði; jÞ 2 EðGÞ (2)

2. Test for finishing

ð0Þ
If t 1 ¼ Tn then algorithm finishes else we continue to step 3.
3. Determination of interval ½ t 1 ; t 2 
The left bound is given and the right bound we determine from the following
formula
n o n o
ð0Þ ð0Þ ð0Þ
t 2 ¼ min Ti j Ti > t 1 [ Ti þ tcij (3)
ði;jÞ2EðGÞ

4. Determination of activities requiring resource in ½t1 ; t2 

Let
n o
ð0Þ ð0Þ
A ¼ ði; jÞ 2 EðGÞj½t1 ; t2   ½Ti þ dij ; Ti þ tcij  (4)

Let us determine the total requirements q of activities from A.

X
q ¼ rij (5)
ði;jÞ 2A

If q > resource limit, then we continue to step 5 else to step 7.

5. Ordering activities from A by their priorities
Let A ¼ fe1 ; :::; ek g. We order all m activities in A into a non-ascending
sequence B as follows. If bk, bl are any elements from B, then the order relation is
defined as follows:
ð0Þ
bk  bl , TVi ðbk Þ þ dbk < t1
else if TS ðbk Þ < TS ðbl Þ
(6)
then if TS ðbk Þ ¼ TS ðbl Þ
else if rbk > rbl

The first condition on the right-hand side means that bk has begun in the previous
interval.
6. Right-shifting
Because of step 4, there exists j< m such that
X
j X
j þ1
rb i  limit and rb i > limit (7)
i ¼1 i ¼1
4 Combined Heuristic Approach to Resource-Constrained Project Scheduling Problem 51

We shift activity bj+1 so that new values of its parameters are obtained from
the old values by the following assignments:
 
ð0Þ
tcb jþ1 :¼ tcb jþ1 þ t2  TPV b þ dbjþ1
ð jþ1 Þ (8)
dbjþ1 :¼ tcbjþ1  tsbjþ1

for the other activities in B (according to their priorities), we either add their
resource requirements and place them into the schedule or shift them if limit has
been exceeded.
Finally, we apply the CPM method again.
7. Next interval

t 1 :¼ t 2 (9)

and return to step 2.

4 Generalisation for Multiproject Schedule

The algorithm described in the previous section can be very simply adapted for
multiproject scheduling with limited resources. The steps of this generalised algo-
rithm must take into account that resources are shared by all the projects. If the
number of projects is N, then, e.g., Step 3 can be modified as follows
n o n o
ð0Þ ð0Þ ð0Þ
t2 ¼ min Ti jTi >t 1 [ Ti þ tcij (10)
ði;jÞ2EðGÞk2½1;N

The other steps may be adapted in a similar way.

5 KNapsack-Based Heuristic

If we consider a schedule as a sequence of time intervals where neighbouring

intervals based on the time an activity starts or finishes, the scheduling problem
may be reduced to a set of the optimal choices of activities in intervals by the order
relation (6).
In the network analysis literature, we may find many other strategies describing
how the ranking of competing activities is accomplished, for instance:
l Greatest remaining resource demand, which schedules as first those activities
with the greatest amount of work remaining to be completed
52 M. Šeda et al.

l Least total float, which schedules as first those activities possessing the least
total float
l Shortest imminent activity, which schedules as first those activities requiring the
least time to complete
l Greatest resource demand, which schedules as first those activities requiring the
greatest quantity of resources from the outset
Using defined priorities as prices, the task investigated in one interval corre-
sponds to the well-known knapsack problem. If the number of the activities sharing
a source in an interval is low, e.g. up to 20, and this is satisfied in most of real
situations, then this problem can be solved exactly by a branch and bound method
or by dynamic programming.
Assuming only one resource with a limited capacity, we can deal with the 0–1
knapsack problem (0–1 KP), which is defined as follows: A set of n items is
available to be packed into a knapsack with capacity of C units. Item i has value
vi and uses up to wi of capacity. We try to maximise the total value of packed items
subject to capacity constraint.
( )
X
n X
n
max vi xi j wi xi  C; xi 2 f0; 1g; i ¼ 1; :::; n (11)
i¼1 i¼1

Binary decision variables xi specify whether or not item i is included in the

knapsack.
The simplest way is to generate all possible choices of items and to determine the
optimal solution among them. This strategy is of course not effective because its
time complexity is O(2n).
Finding the solution may be faster using the branch and bound method [15],
which restricts the growth of the search tree. Avoiding much enumeration depends
on the precise upper bounds (the lower the upper bounds, the faster the finding of
the solution is). Let items be numbered so that

v1 v2 vn
    (12)
w1 w2 wn

We place items in the knapsack by this non-increasing sequence. Let x1, x2, . . . , xp
be fixed values of 0 or 1 and
 
Mk ¼ x j x 2 M; xj ¼ xj ; xj 2 f0; 1g; j ¼ 1; . . . ; p (13)

where M is a set of feasible solutions. if

X
q1 X
p X
q
ð9qÞ ðp<q  ng : wj  C  wj xj < wj (14)
j¼pþ1 j¼1 j¼pþ1
4 Combined Heuristic Approach to Resource-Constrained Project Scheduling Problem 53

then the upper bound for Mk can be determined as follows:

!
X
p X
q1
vq Xp X
q1
UB ðMk Þ ¼ v j xj þ vj þ C w j xj þ wj (15)
j¼1 j¼pþ1
wq j¼1 j¼pþ1

In more complex and more frequent situations when we have more than one
limited resource, we transform the resource-constrained scheduling into a sequence
of multi-knapsack problem (MKP) solutions. MKP is defined as follows:

X
n
maximise vi xi
i¼1
Xn
(16)
subject to wri xi  Cr ; r ¼ 1; ::: ; m;
i¼1
xi 2 f0;1g; i ¼ 1; ::: ; n

For the solution of the task with multiple constraints, we must generalise the
approaches mentioned above.
The combinatorial approach can be applied without any changes, but using the
branch and bound method, we must redefine the upper bound. For its evaluation we
use the following formula
 
UB ðMk Þ ¼ min UB1 ðMk Þ; ::: ; UBm ðMk Þ (17)

where auxiliary bounds UBi ðMk Þ; i ¼ 1; . . . ; m correspond to the given constraints

and are determined as in the 0–1 knapsack problem. Before evaluation of these
auxiliary bounds, the other variables must be sorted again by decreasing values vj/wij.
Evidently, the run time will increase substantially.

6 Stochastic Heuristic Methods

The branch and bound method discussed above is deterministic. Now we will pay
attention to stochastic heuristic methods. These methods are used in situations
where the exact methods would fail or calculations would require a great amount
of time.
Heuristic [16] is a technique which seeks goal (i.e. near optimal) solutions at a
reasonable computational cost without being able to guarantee either feasibility or
optimality, or even in many cases, to state how close to optimality a particular
feasible solution is. The most popular heuristics – genetic algorithms, simulated
annealing, tabu search and neural networks are reviewed in [16]. Examples of their
possible use are described in [13].
54 M. Šeda et al.

Let us briefly deal with the genetic algorithms now. The skeleton for GA is
shown as follows:
Generate an initial population.
Evaluate fitness of individuals in the population.
Repeat select parents from the population.
Recombine parents to produce children.
Evaluate fitness of the children.
Replace some or all of the population by the children.
Until a satisfactory solution has been found.
In the following paragraphs, we briefly summarize GA settings to our scheduling
problem.
Individuals in the population (chromosomes) are represented as binary strings of
length n, where a value of 0 or 1 at the i-th bit (gene) implies that xi ¼ 0 or 1 in the
solution respectively.
The population size N is usually chosen between n and 2n.
An initial population consists of N feasible solutions and it is obtained by
generating random strings of 0s and 1s in the following way: First, all bits in all
strings are set to 0, and then, for each of the strings, randomly selected bits are set to 1
until the solutions (represented by strings) are feasible.
The fitness function corresponds to the objective function to be maximised:
X
n
f ðxÞ ¼ vi xi (18)
i¼1

Pairs of chromosomes (parents) are selected for recombination by the binary

tournament selection method, which selects a parent by randomly choosing two
individuals from the population and selecting the most fit one.
The recombination is provided by the uniform crossover operator. That means
each gene in the child solution is created by copying the corresponding gene from
one or the other parent, chosen according to a binary random number generator. If a
random number is 0, the gene is copied from the first parent; if it is a 1, the gene is
copied from the second parent. After crossover, the mutation operation is applied to
each child. It works by inverting each bit in the solution with a small probability.
We use a mutation rate of 5/n as a lower bound on the optimal mutation rate. It is
equivalent to mutating five randomly chosen bits per string.
If we perform crossover or mutation operations as described above, then the
generated children can violate certain capacity constraints. We can assign penalties
to these children that prevent infeasible individuals from entering the population. A
more constructive approach uses a repair operator that modifies the structure of an
infeasible individual, so that the solution becomes feasible. Its pseudo-Pascal code
is shown below for a more general case of multiple constraints. Once a new feasible
child solution has been generated, the child will replace a randomly chosen
solution. We use a steady-state replacement technique based on eliminating the
individual with the lowest fitness value. Since the optimal solution values for most
4 Combined Heuristic Approach to Resource-Constrained Project Scheduling Problem 55

problems are not known, the termination of a GA is usually controlled by specifying

a maximum number of generations tmax. Usually we choose tmax  5,000.

7 Experimentation

The approach discussed in the previous paragraphs has been implemented in Bor-
land Delphi. Its interface is similar to Microsoft Project. Although Microsoft
Project often is not able to solve a constrained scheduling problem even for very
simple projects (it reports some over-allocation and underallocation may be
unavoidable), we have never come across this situation when using our programme.
Besides the GA approach, the programme implements another popular heuristic
method – simulated annealing – for a comparison, and a user may choose between
these two possibilities. It enables a verification of the quality of computational
results, as the exact results for large projects are not known.
We have tested the proposed approach in many situations. Table 1 shows results
for one real project with 183 activities, 54 vertices in the network graph and 12
limited resources. This project deals with the innovation action of a Benson boiler
in a local power station in the surroundings of Brno. This table contains durations of
schedule for this project designed in 30 tests by the genetic algorithm (GA) and by
the simulated annealing (SA). Parameters for both methods were set in a way such
that the computational time for design of the project schedule should be approxi-
mately 2.4 s. The branch and bound (BB) method found a schedule with a duration
of 291 days in a several times longer CPU time. The results of BB achieved are
favourable because the parallelism of activities was low. The dynamic program-
ming approach could not be used because of insufficient memory.
While the results appear comparable, statistically, the results gained by GA are
better than the SA results. The well-known Kruskal-Wallis test for balanced one-
way design was used to show that samples were taken from the same population. It
has yielded the following results: Average rank for SA ¼ 39.75, for GA ¼ 21.25,
the value of test statistic ¼ 17.5472 and computed significance level ¼ 2.8 105.

Table 1 Durations for schedule (days) designed by GA and SA

GA GA GA SA SA SA
291 294 293 294 293 295
292 292 291 292 294 296
293 291 292 293 296 292
293 293 293 294 292 297
292 292 291 297 294 294
294 291 293 292 291 294
292 291 292 293 298 295
292 292 292 294 294 291
291 293 291 294 292 295
293 291 293 295 293 293
56 M. Šeda et al.

Thus the hypothesis of sampling from the same population is rejected at all
traditionally used significance levels (0.05; 0.01; 0.001).

8 Conclusion

In this chapter, a new technique for computing of the resource-constrained project

scheduling was proposed. The strategy of activity-shifting was replaced by prolong-
ing their duration and dividing them into active and sleeping parts. It makes it
possible to apply a simple CPM algorithm. The proposed algorithm was designed in
a mathematical form and verified for a single version of RCPSP. We also sketched
how to adapt the proposed algorithm for multiproject scheduling with limited
resources.
However, the deterministic approaches are not effective or may not be used in
complex projects with multiple constraints because of the exponentially growing
time in branch and bound method calculations. A genetic algorithm approach is
proposed and compared with simulated annealing.
Further investigation will include fuzzy versions of the problem.

References

1. J. Blazewicz, K.H. Ecker, G. Schmidt, J. Weglarz, Scheduling Computer and Manufacturing

Processes (Springer-Verlag, Berlin, 1996)
2. S.E. Elmaghraby, Activity Networks: Project Planning and Control by Network Models
(Wiley, New York, 1977)
3. P. Brucker, A. Drexl, R. M€ ohring, K. Neumann, E. Pesch, Resource-constrained project
scheduling: notation, classification, models, and methods. Eur. J. Oper. Res. 112: 3–41 (1999)
4. W. Herroelen, E. Demeulemeester, B.D. Reyck, A note on the paper “resource-constrained
project scheduling: notation, classification, models, and methods” by Brucker et al. Eur. J.
Oper. Res. 128, 679–688 (2001)
5. K. Bouleimen, H. Lecocq, A new efficient simulated annealing algorithm for the resource-
constrained project scheduling problem and its multiple mode version. Eur. J. Oper. Res. 149,
268–281 (2003)
6. M. Mika, G. Waligóra, J. Weglarz, Simulated annealing and tabu search for multi-mode
resource-constrained project scheduling with positive discounted cash flows and different
payment models. Eur. J. Oper. Res. 164, 639–668 (2005)
7. A. Azaron, C. Perkgoz, M. Sakawa, A genetic algorithm for the time-cost trade-off in PERT
networks. Appl. Math. Comput. 168, 1317–1339 (2005)
8. K.W. Kim, M. Gen, G. Yamazaki, Hybrid genetic algorithm with fuzzy logic for resource-
constrained project scheduling. Appl. Soft Comput. 2/3F, 174–188 (2003)
9. K.W. Kim, Y.S. Yun, J.M. Yoon, M. Gen, G. Yamazaki, Hybrid genetic algorithm with
adaptive abilities for resource-constrained multiple project scheduling. Comput. Indus. 56,
143–160 (2005)
10. C. Artigues, P. Michelon, S. Reusser, Insertion techniques for static and dynamic resource-
constrained project scheduling. Eur. J. Oper. Res. 149, 249–267 (2003)
4 Combined Heuristic Approach to Resource-Constrained Project Scheduling Problem 57

11. V. Valls, S. Quintanilla, F. Ballestin, Resource-constrained project scheduling: a critical

activity reordering heuristic. Eur. J. Oper. Res. 149, 282–301 (2003)
12 L.-Y. Tseng, S.-C. Chen, A hybrid metaheuristic for the resource-constrained project schedul-
ing problem. Eur. J. Oper. Res. 175(2), 707–721 (2006)
13. H. Zhang, X. Li, H. Li, F. Huang, Particle swarm optimization-based schemes for resource-
constrained project scheduling. Autom. Constr. 14, 393–404 (2005)
14. T.H. Cormen, C.E. Leiserson, R.L. Rivest, C. Stein, Introduction to Algorithms (MIT Press,
Cambridge, MA, 2001)
15. J. Klapka, J. Dvořák, P. Popela, Methods of Operational Research (in Czech) (VUTIUM,
Brno, 2001)
16. C.R. Reeves, Modern Heuristic Techniques for Combinatorial Problems (Blackwell Scientific
Publications, Oxford, 1993)
Chapter 5
A Development of Data-Logger for Indoor
Environment

Anuj Kumar, I.P. Singh, and S.K. Sud

Abstract This chapter describes a development of data logger for indoor environ-
ment. Present work concentrates to environmental parameter (temperature and
humidity) and more polluted contaminants (concentration level of CO and CO2).
In this work four channels have been used for data logger and other four channels is
open to external sensor module. The data collected will be stored in the EEPROM
and output can be taken in note-pad in tabular corresponding to month/date/year
using graphical user interface.

1 Introduction

Environment monitoring system is a complete data logging system. It automatically

measures and records temperature, humidity and other parameters and provides
warnings when readings go out of range [1]. Indoor environment monitoring is
required to protect the building occupant’s health by providing thermally comfort-
able and toxicant free environment [2]. We need a system that monitors as well as
records the indoor environment data easily. In this chapter, we are discussing about
the developed indoor environment monitoring system that monitors and records
indoor temperature, humidity, CO and CO2.
Available technologies for sensing these environmental parameters are developed
by Onseat HOBO, Spectrum Technologies (Watch Dog weather conditions), TandD,
Telaire (Wireless Monitoring Systems), Testo, Log Tag, Measurement Computing
Corporation, Monarch Instruments, MSR Electronics GmbH, P3 International, Quality
Thermistor, S3 Crop, Sensaphone, Sansatronics, Lascar, ICP, Graphtech, Extech
Instruments, Dickson, Dent Instruments, Davis, ACR System Inc, 3M International,

A. Kumar (*)
Instrument Design Development Centre, Indian Institute of Technology Delhi, Electronics Lab,
New Delhi, India 110016
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 59

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_5, # Springer ScienceþBusiness Media B.V. 2010
60 A. Kumar et al.

and Acumen [3]. The drawback of these available systems is that, both air quality
and thermal comfort can not be measured simultaneously. So there is a need to
develop an economical system (prototype data logger) which can help in collecting
data to analyze the necessary environmental parameters simultaneously.
A prototype data logger has been developed to monitor the environmental para-
meters. The developed data logger consist (i) sensors module (ii) LCD (iii) Real
Time Clock (RTC) (iv) EEPROM and (v) PC serial communication. This data
logger is operated through PC using graphical user interface (GUI) in visual basic.

2 Sensors Module

A sensor is a device that measures a physical quantity and converts it into an

equivalent analog or digital signal which can be read by an observer or by an
instrument [4]. We have used temperature, relative humidity, CO, and CO2 sensors
in the developed system.
A gas sensor detects particular gas molecules and produces an electrical signal
whose magnitude is proportional to the concentration of the gas [5]. Till date, no gas
sensor exists that is 100% selective to only a single gas. A good sensor is sensitive to
the measured quantity but less sensitive to other quantities. Available gas sensors are
based on five basic principles. These can be electrochemical, infrared, catalytic bead,
photo ionization and solid-state [6, 7]. We have selected these sensors because they
produce a strong signal for the selected variable especially at high gas concentrations
with adequate sensitivity. They have a fast response time, high stability, long life,
low cost, low dependency on humidity, low power consumption, and compact size
[5]. Four sensors along with their signal conditioning circuit are used to sense the
desired parameter such as temperature, humidity, CO, and CO2. Signal conditioning
circuit for that sensor needs to be connected externally. In software we can select
any of the analog channels. The interface of temperature, humidity, CO, and CO2
sensors with microcontroller PIC 18F4458 is described as follows.

2.1 Temperature Sensor

National semiconductor’s LM 35 IC has been used for sensing the temperature. It is

an integrated circuit sensor that can be used to measure temperature with an
electrical output proportional to the temperature (in  C). The temperature can be
measured more accurately with it than thermistor. The operating and amplifica-
tion circuit is shown in Fig. 1. The output voltage of IC LM 35 is converted to
temperature in  C is given by the following expression [7]

Temp:ð CÞ ¼ ðVout  100Þ=7 C

The measuring temperature range of the instrument is between 15 C and 70 C.
5 A Development of Data-Logger for Indoor Environment 61

Fig. 1 Operating and

amplification circuitof U1
LM35CZ

7
1
temperature sensor 3 +
+4.97 V 2 6 1.722V

0.246V
1 2 – OP07
3

4
8
0

–4.97 V
R2
12 K
R1 2K

Fig. 2 Operating circuit of Vcc

humidity sensor

R 12.5K

1 3
Vin
2 GND
0

Vout

2.2 Humidity Sensor

The sensor circuit develops a linear voltage vs. RH (relative humidity) output,
which is ratio metric to the supply voltage. This means when the supply voltage
varies, the sensor output voltage follows the same proportion. It can operate over
a range of 4–5.8 V supply. At 5 V supply voltage (at room temperature), corres-
ponding to relative humidity variation from 0% to 100% (noncondensing), the
output voltage varies from 0.8 to 3.9 V. The humidity sensor functions with a
resolution of up to 0.5% of relative humidity (RH), with a typical current draw of
only 200 mA, the HIH4000 series is ideally suited for low drain, battery operated
systems.
The operating circuit is shown in Fig. 2. The change in the RH of the surround-
ings causes an equivalent change in the voltage output. The output is an analog
voltage proportional to the supply voltage. Consequently, converting it to relative
humidity (RH) requires both the supply and the sensor output voltages (At 25 C)
and is given by the following expression [7].



RH ¼ Vout =Vsup ply  0:16 =0:0062

The output of the humidity sensor is 2.548 V i.e. the relative humidity is 56%
at 25 C.
62 A. Kumar et al.

TGS 4161

VRL
INPUT OUTPUT
VOLTAGE VC RL VOLTAGE
HEATER
V
VOLTAGE H

Fig. 3 Circuit for CO and CO2 sensor [7]

2.3 CO and CO2 Sensor

The operating circuit of CO and CO2 sensors is shown in Fig. 3. The relationship
between output voltage and gas concentration is given by the following expression
  2
ðVC RL =VOUT Þ  RL 1
c¼ 1
R0 K

where, VOUT ¼ output voltage; VC ¼ input voltage, R0 ¼ electrical resistance of

sensor at zero ppm, K ¼ a constant for particular, RL ¼ sensor load resistance [5, 7].

3 LCD Interface to the Microcontroller

In this work, we are using on-chip analog to digital converter which is on the
microcontroller. This analog to digital converter is having the 12 bit resolution with
programmable acquisition time. It is sensing the analog signal from the sensor at the
variable sampling rate (1 s to 1 h). The sensed value is converted to its digital
equivalent. This digital value is displayed on the LCD (liquid crystal display) and is
interfaced to the microcontroller [8, 9–12].

4 Real Time Clock Interface to the Microcontroller

The IC DS1307 operates as a slave device on the I2C bus. Access is obtained by
implementing a START condition and providing a device identification code fol-
lowed by a register address. Subsequent registers can be accessed sequentially until a
STOP condition is executed. When VCC falls below 1.25 VBAT, the device terminates
an access in progress and resets the device address counter. Inputs to the device will
5 A Development of Data-Logger for Indoor Environment 63

not be recognized at this time to prevent erroneous data from being written to the
device from an out of tolerance system. When VCC falls below VBAT, the device
switches into a low-current battery backup mode. Upon power up, the device switches
from battery to VCC when VCC is greater than VBAT þ0.2 V and recognizes inputs
when VCC is greater than 1.25 VBAT. We are using IC DS1307 as real time clock which
have features such as real-time clock counts in seconds, minutes, hours, day of month,
day of week, month, and year with leap year compensation valid up to 2100, 56 Byte,
Nonvolatile (NV) RAM for data storage, I2C serial interface, programmable square
wave output signal, automatic power fail detect and switch circuitry (consumes less
than 500 nA in battery backup mode oscillator running), and temperature range 40 C
to 85 C [9, 10]. We are using I2C to interface RTC and EEPROM to the micro-
controller. The I2C bus is the most popular of three serial EEPROM protocols.
The I2C chips include address pins as an easy way to have multiple chips on a single
bus while only using two connections to the microcontroller [9].

5 EEPROM Interface to the Microcontroller

The EEPROM will store the digital value which is coming from analog to digital
converter. We will require 52.73 MB of EEPROM if we are sampling all analog
channels at the rate of 1 sample/s. We are using the EEPROM AT24C256
(ATMEL). This will store the sample data at different instants [10–14].

6 PC Interface Using RS-232 Serial Communication

PIC 18F4458 using MAX-232 is interfaced with PC. IC (MAX-232) used to convert
TTL logic level to RS-232 logic level. RS-232 is the serial communication protocol
that does not require the clock along with data lines. Two data lines are there one is
TX and another is RX for serial communication. MAX-432 has two receivers (con-
verts RS-232 logic level to TTL logic) and two drivers. Separate power supply has
been provided because minimum power supply needed is 5 V and MAX-232 con-
sumes a lot of current for operation. External capacitors are required for internal
voltage pump to convert TTL logic level to RS-232 level. For battery operated
application MAX-232 can be used as level converter instead of MAX-232. It is low
supply low power consumption logic converter IC for RS-232 [9, 10, 13].

7 Graphical User Interface

The GUI is one of the important parts for this device as it displays the data from
microcontroller for data monitoring and analysis. The design template has to be user
friendly for best usage. For this chapter, the main objective is to display data received
in graphical form. As transducer detects and translate an analog signal, the data will
64 A. Kumar et al.

go through a conversion at the ADC. This digital data will be stored in EEPROM chip
with the help of Visual Basic 6.0 software. Since the data is using serial RS232
communication, an initialization needs to be done having baud rate, data bits, parity,
stop bit, and the COM port at PC. The baud rate is the number of signal changes per
second or transition speed between Mark (negative) and Space (positive) which
ranges from 110 to 19,200, data bits is the length of data in bit which has one Least
Significant Bit and one Most Significant Bit, the parity bit is an optional bit mainly for
bit error checking. It can be odd, even, none Mark, and Space. Stop bit is used to
frame up the data bits and usually combined with the start bit. These bits are always
represented by a negative voltage and can be 1, 1.5 and 2 stop bits, and COM port is
the selection of the available COM port at PC. The commonly used setting to
establish a serial RS232 communication is 9600 baud rate, none parity, 8 data bits,
1 stop bit, and COM port 1. This can be done by using the GUI monitoring system
where it automatically saves the data received in a notepad. The data saved is the date
and time at which the data collected and the data value it self. Figures 4 and 5,
represents the graphical user interface and logged data in file respectively [12, 13].

8 Schematic of the Data Logger

Figure 6 shows, the full schematic diagram of the data logger for indoor environ-
ment. This data logger has four embedded sensor module and other four channels
are open to be used for the measurement of other environmental parameters.

9 Software Design of Data Logger

This section includes the discussion on software design for all the modules inter-
faced with PIC 18F4458. It also explains the functions of software designed for data
logger [11].

Fig. 4 GUI for the data logger

5 A Development of Data-Logger for Indoor Environment 65

Fig. 5 Representations of the logged data in file

9.1 Programming Steps for I2C Interface

I2C interface is bi-directional. This is implemented by an “Acknowledge” or “ACK”

system allows data to be sent in one direction to one item on the I2C bus, than, that
item will “ACK” to indicate the data received. Normally, the master device controls
the clock line, SCL. This line dictates the timing of all transfers on the I2C bus. Other
devices can manipulate this line, but they can only force the line low. This action
means that item on the bus cannot deal with more data in to any device.

9.1.1 Writing to an I2C Chip

The function of writing to the EEPROM is shown here as “Control IN”, which
represents putting the EEPROM in an “input” mode. Since we are only sending data
to the EEPROM (as shown in Fig. 7), we use “Control IN” byte and later used
“Control OUT”. Next, the EEPROM acknowledges this byte. This is shown by the “A”
after the byte. It is put on the next line to indicate that this is transmitted by the
EEPROM. The Address Byte contains the address of the location of the EEPROM
where we want to write data. Since the address is valid, the data is acknowledged by
the EEPROM. Finally, we send the data we want to write. The data is then acknow-
ledged by the EEPROM. When that finishes, we send a stop condition to complete
the transfer. Remember the “STOP” is represented as the “T” block on the end. Once
the EEPROM gets the “STOP” condition it will begin writing to its memory.
66 A. Kumar et al.

VDD
36
LM35

32.768 KHz
CRYSTAL
1 3 2 AN0 35 R4
R6 10K 1 8
2

DSC130
10 K 2

RTC
3 AN1 34 NC
3V 3 7
+5v 33 6
4 V 5

EEPROM
1 CO 3 DD 32 1
2 0
2 V 31 3 6
SS
+5v 5
30
4 AN2
1 3 4 29
CO PIC18F4458
2 5 28 15 16
2 AN3 27 13 14
9 12

Connector
HIH4000 10
11 26 7 8
1 2 3 5 6

LCD
3 4
12 25 1
24 2
13 D+
0 14 23 0
8MHZ

D–
D+
22
D–
21
USB
Rx Tx Connector

16 15 14 13 12 11 10 9
VDD

+ MAX232
J2
–
1 3 4 5 2 6 7 8

3 C2
2 4 5 C3
7 C1
16 8 9
DB9 C4
CONNECTOR 0

Fig. 6 Full Schematic of the data logger

9.1.2 Reading from an I2C Chip

The transfer will use the “Control IN” byte to load the address into the EEPROM
(as shown in Fig. 8). This sends data to the EEPROM which is why we use the
control in byte. Once the address is loaded, we want to retrieve the data. So, we send
a “Control OUT” byte to indicate to the EEPROM that we want data FROM it.
5 A Development of Data-Logger for Indoor Environment 67

Fig. 7 Writing the data in

I2C chip (Controlling
Window)

Fig. 8 Reading the data from

an I2C chip

The EEPROM will acknowledge this and then send the data we requested. When
we are done getting data, we send a “NACK” to tell the EEPROM that we do not
want more data. If we were to send an ACK at this point, we could get the next byte
of data from the EEPROM. Since we only want to read one byte, we send a
“NACK”.

9.2 Programming Steps for LCD Interface

Set RS ¼ 0 to send command; Send 0b0010 to data lines three times with a delay of
2 ms; to send a byte on four data lines, send higher nibble first and give a RE pulse
of 100 ms at RE; send a set of instruction one after another with a delay of 2 ms
between each command to configure various setting as given in instruction set of
LCD datasheet [9, 13]; send instruction set again.
Set RS ¼ 1; Send higher nibble at four data lines. Send 100 ms RE pulse; Send
lower nibble at data lines. Send RE pulse; Keep track of number of character already
displayed on display panel using LCD_count. Go to line 2 or line 1 according to that.

9.3 Programming Steps for Sensor Data Collection

There are four sensor module connected such as temperature, humidity, CO, and
CO2. Data is collected by the ADC inbuilt in PIC. ADC provides 12 bit of data
after the conversion is completed.
68 A. Kumar et al.

9.3.1 Temperature Sensor Data Collection

Data collection from the temperature sensor needs following actions to be carried
out (a) Selecting the analog channel AN0, sampling frequency, and alignment of
bits for ADRESH and ADRESL, (b) Vref and power on the ADC module by setting
ADCON0, ADCON1 and ADCON2 registers, (c) starting analog to digital conver-
sion by setting ADGO bit high (wait till ADIF flag will not indicate the completion
of conversion), and (d) copy of results from ADRESH and ADRESL to variables.

9.3.2 Humidity Sensor Data Collection

Select the AN3 and set other features of ADC as temperature sensor; after comple-
tion of conversion copy the result in variable.

9.3.3 CO and CO2 Sensor Data Collection

Data collection from the CO sensor needs following actions to be carried out (a)
Selecting the analog channel AN1, sampling frequency, and alignment of bits for
ADRESH and ADRESL, (b) Vref and power on the ADC module by setting
ADCON0, ADCON1 and ADCON2 registers, (c) starting analog to digital conver-
sion by setting ADGO bit high (wait till ADIF flag will not indicate the completion
of conversion), and (d) copy of results from ADRESH and ADRESL to variables.
Now repeat the same process to collect the CO2 data on the channel number
AN2.

10 Results and Discussion

Sensors module, EEPROM, RTC, and LCD have been successfully interfaced to
the microcontroller. EEPROM is successfully storing the logged data with time and
date tag. The sensors data is being displayed on LCD module. A simple GUI has
been designed to store a logged data to a text file, so that it can be analyzed further.

11 Conclusions

We have developed a low cost, 12 bit resolution data logger and successfully
measured temperature, humidity, and concentration of CO and CO2 gases. The
GUI designed gives a lucratively look to the functioning of data logger. Initial
results of the data logger are encouraging and we are working on to improve the
GUI model as well as the accuracy of data logger.
5 A Development of Data-Logger for Indoor Environment 69

References

1. G.L. Tang, Lecture notes on Health and the built environment: Indoor air quality (University
of Calgary, Alberta, Canada)
2. J.D. Richard, G.S. Brager, Thermal comfort in naturally ventilated buildings: revisions to
ASHRAE standard 55. Energy Buildings 34, 549–561 (2002)
3. Microdaq (March 3, 2009); http://www.microdaq.com/data-logger/
4. D.D. Lee, D.S. Lee, Environment gas sensors. IEEE Sensors J. 1(3), 214–215 (2001)
5. N. Kularatna, B.H. Sudantha, An environmental air pollution monitoring system based on the
IEEE 1451 standard for low cost requirements. IEEE Sensors J. 8(4), 415–422 (2008)
6. Sensor industry development and trends, Sensors express (Nov 2002)
7. RS India (April 4, 2009); http://www.rsonlineindia.com
8. R. Luharuka, R.X. Gao, A microcontroller-based data logger for physiological sensing. IEEE
Proc. on Instrument and Measurement Technology Conference, Anchorage, AK, USA, 21–23
May, 175–180 (2002)
9. Data acquisition logging circuits (20 Mar 2009); http://www.hobbyprojects.com/A/
acquistions_data_circuits.html
10. Data sheet of real time clock DS 1307 (8 May 2008). http://www.maxim-ic.com/products/rtc/
real-time-clocks.cfm
11. Microchip (10 Jan 2009). http://www.microchip.com
12. Introduction to data acquisition (10 May 2008). http://zone.ni.com/devzone/concepted.nsf/
webmain/
13. G. Mason, A handheld data acquisition system for use in an undergraduate data acquisition
course. Data Acquisit. 45, 338–393 (2002)
14. D. Malone, Build a versatile data logger. Popular Electronics, 35–42, July 1994
Chapter 6
Multiobjective Evolutionary Optimization
and Machine Learning: Application
to Renewable Energy Predictions

Kashif Gill, Abedalrazq Khalil, Yasir Kaheil, and Dennis Moon

Abstract The inherent variability in climate processes results in significant

impacts on renewable energy production. While a number of advancements have
been made over the years, the accurate energy production estimates and the cor-
responding long-term variability at the full wind farm remains a big challenge. At
the same time, long-term energy estimates and the variability are important for
financial assessment of the wind farm projects. In this chapter, a machine learning
approach to model wind energy output from the wind farm is presented. A multi-
objective evolutionary optimization (MOEO) method has been applied for the
optimization of an Artificial Intelligence learning methodology the “Support Vector
Machines” (SVM). The optimum parameter search is conducted in an intelligent
manner by narrowing the desired regions of interest that avoids getting struck in
local optima. The National Center for Environmental Prediction (NCEP)’s global
reanalysis gridded dataset has been employed in this study. The gridded dataset
for this particular application consists of four points each consisting of five vari-
ables. A 40-years, 6-hourly energy prediction time series is built using the 40-years
of reanalysis data (1968-present) after training against short-term observed farm
data. This is useful in understanding the long-term energy production at the farm
site. The results of MOEO-SVM for the prediction of wind energy are reported
along with the multiobjective trade-off curves.

1 Introduction

It is well-known that the weather characteristics in a given time frame are deter-
mined by a relatively small number of discrete weather systems, each of which may
exhibit very different influences and patterns of development. These fundamental

K. Gill (*)
WindLogics, Inc., 1021 Bandana Blvd., E. # 111, St Paul, MN 55108, USA
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 71

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_6, # Springer ScienceþBusiness Media B.V. 2010
72 K. Gill et al.

causes of variations in on-site wind speeds are inherent in the atmosphere and must
be understood, to the extent possible, for accurate wind resource assessment. In
order to make useful predictions of wind speed/energy, it is therefore important to
develop statistically sound relationships between those wind regimes and the
atmospheric conditions. This is the core of the current methodology described in
this chapter. Machine learning tools have gained immense popularity in the geos-
ciences community due to their success against physically-based modeling
approaches. In brief, machine learning tools are used to determine the relationship
between inputs and output in an empirical framework. These models do not employ
traditional form of equations common in physically-based models or as in regres-
sions, instead have flexible and adaptive model structures that can abstract relation-
ships from data.
The chapter describes a method for training Support Vector Machine (SVM)
in applications for wind energy predictions at a wind-farm level. The SVM is a
powerful learning algorithm developed by Vapnik and is known for its robust
formulation in solving predictive learning problems employing finite data [1].
The method is well-suited for the operational predictions and forecasting of wind
power, which is an important variable for power utility companies. The proposed
methodology employs a Multiobjective Evolutionary Optimization approach for
training the SVM. The goal of an optimization method is to efficiently converge to a
global optimum, in the case of a single objective function, and to define a trade-off
surface in the case of multiobjective problems. Overall, global optimization (GO)
methods have two main categories: deterministic and probabilistic. Deterministic
methods use well-defined mathematical search algorithms (e.g., linear program-
ming, gradient search techniques) to reach the global optimum, and sometimes use
penalties to escape from a local optimum. Probabilistic methods employ probabi-
listic inference to reach the global optimum. Evolutionary optimization algorithms
generally fall in probabilistic category of optimization methods.

2 Material and Methods

The details on support vector machine, the multiobjective evolutionary optimiza-

tion, and the SVM trainings are presented in the current section.

2.1 Support Vector Machines

Support Vector Machines (SVM) has emerged as an alternative data-driven tool

in many traditionally Artificial Neural Network (ANN) dominated fields. SVM
was developed by Vapnik in the early 1990s mainly for applications in classifi-
cation. Vapnik later extended his work by developing SVM for regression. The
SVM can “learn” the underlying knowledge from a finite training data and deve-
lops generalization capability which makes it a robust estimation method.
6 Multiobjective Evolutionary Optimization and Machine Learning 73

The mathematical basis for SVM is derived from statistical learning theory. SVM
consists of a quadratic programming problem that can be efficiently solved and for
which a global extremum is guaranteed [2]. It is sparse (robust) algorithm when
compared against ANN, but what makes it sparse is the use of Structural Risk
Minimization (SRM) instead of Empirical Risk Minimization (ERM) as is the case
in ANN. In SRM, instead of minimizing the total error, one is only minimizing an
upper bound on the error. The mathematical details on SVM can be found in [3–5].
The function f(x) that relates inputs to output has the following form:

X
N
f ðxÞ ¼ wi fðxi Þ þ b (1)
i¼1

Where N (usually N  L) is the number of support vectors which are selected

from the entire dataset of ‘L’ samples. The support vectors are subsets of the training
data points that support the “decision surface” or “hyper-plane” that best fits the data.
The function f(x) is approximated using the SRM induction principle by mini-
mizing the following objective function:

1X N
Minimize R½ f  ¼ C jyi  f ðxi Þje þ kwk2 (2)
K i¼1

The first term (jyi  f ðxi Þje ) in the above expression is what is called e-insensitive
loss function. The samples outside the e-tube are penalized by a penalty termx. The
above formulation makes the solution sparse in the sense that the errors less than e
are ignored. The sparse solution is the one that gives the same error with minimum
number of coefficients describing f(x). The second term in the objective function is
the regularization term added to avoid the consequences of the ill-posedness of the
inverse problem [6].
The Eq. (2) is solved in dual form employing Lagrange multipliers as [3]:

X
N
f ðxÞ ¼ ðai  ai ÞKðx; xi Þ þ b (3)
i¼1

The ai and ai are the Lagrange multipliers and have to greater than zero for the
support vectors i ¼ 1, . . . , N, and Kðxi ; xÞis a kernel function.
There are three main parameters to be determined as part of SVM trainings; the
tradeoff ‘C’, the degree of error tolerance ‘e’, and parameter related to the kernel,
the kernel width, ‘g’. In practice, these may be determined either using a trial-and-
error procedure or an automatic optimization method [7]. The optimization scheme
can be single objective or multiobjective depending upon the nature of the problem.
The initial formulation for training SVM employed a single objective approach
to optimization. It was noticed that single objective methods to train model can
result in optimal parameter sets for the particular single-objective value; but fail to
74 K. Gill et al.

provide reasonable estimates for the other objective. It is therefore preferred to

employ multiobjective search procedures. In the current implementation, Multi-
objective Particle Swarm Optimization (MOPSO) algorithm is used to determine
the SVM parameters.

2.2 Multiobjective Evolutionary Optimization

The goal in an optimization algorithm is to find the minimum (or maximum) of a

real valued function f : S ! < i.e., finding optimum parameter set x 2 S such that:

f ðx Þ  f ðxÞ; 8x 2 S (4)

Where S  <D is the feasible range for x (the parameter set, having D-dimen-
sions) and <D represents the D-dimensional space of real number.
The current multiobjective methodology employs Swarm Intelligence based evo-
lutionary computing multiobjective strategy called Multiobjective Particle Swarm
Optimization (MOPSO) [8]. The PSO method has been developed for single objec-
tive optimization by R. C. Eberhart and J. Kennedy [9]. It has been later extended to
solve multiobjective problems by various researchers including the method by [8].
The method originates from the swarm paradigm, called Particle Swarm Optimi-
zation (PSO), and is expected to provide the so-called global or near-global optimum.
PSO is characterized by an adaptive algorithm based on a social-psychological
metaphor [9] involving individuals who are interacting with one another in a social
world. This sociocognitive view can be effectively applied to computationally
intelligent systems [10]. The governing factor in PSO is that the individuals, or
“particles,” keep track of their best positions in the search space thus far obtained,
and also the best positions obtained by their neighboring particles. The best position
of an individual particle is called “local best,” and the best of the positions obtained
by all the particles is called the “global best.” Hence the global best is what all the
particles tend to follow. The algorithmic details on PSO can be found in [8, 9,
11, 12]. The approach in [8] presents a multiobjective framework for SVM optimi-
zation using MOPSO.
A multiobjective approach differs from a single objective method in that the
objective function to be minimized (or maximized) is now a vector containing more
than one objective function. The task, therefore, of the optimization method is to
map out a trade-off surface (otherwise known as Pareto front), unlike finding a
single scalar-valued optimum in case of single objective problems. The multi-
objective approach to the PSO algorithm is implemented by using the concept
of Pareto ranks and defining the Pareto front in the objective function space.
Mathematically, a Pareto optimal front is defined as follows: A decision vector
~
x1 2 S is called Pareto optimal if there does not exist another ~ x2 2 S that dominates
it. Let P  <m be a set of vectors. The Pareto optimal front P  P contains all
vectors ~x1 2 P, which are not dominated by any vector ~ x2 2 P:
6 Multiobjective Evolutionary Optimization and Machine Learning 75

P ¼ f~ x2 2 P : ~
x1 2 Pj 6 9~ x2 ~
x1 g (5)

The idea of Pareto ranking is to rank the population in the objective space and
separate the points with rank 1 in a set P* from the remaining points. This
establishes a Pareto front defined by a set P*. All the points in the set P* are the
“behavioral” points (or non-dominated solutions), and the remaining points in set
P become the “non-behavioral” points (or inferior solution or dominated solutions).
The reason behind using the Pareto optimality concept is that there are solutions
for which the performance of one objective function cannot be improved without
sacrificing the performance of at least one other.
In the MOPSO algorithm, as devised in [8], the particles will follow the nearest
neighboring member of the Pareto front based on the proximity in the objective
function (solution) space. At the same time, the particles in the front will follow the
best individual in the front, which is the median of the Pareto front. The term
follow means assignments done for each particle in the population set to decide the
direction and offset (velocity) in the subsequent iteration. These assignments are
done based on the proximity in the objective function or solution space. The best
individual is defined in a relative sense and may change from iteration to iteration
depending upon the value of objective function.
An example test problem is shown in Fig. 1. The test function presents a
maximization problem (as in [13]):

Maximize F ¼ ðf1 ðx; yÞ; f2 ðx; yÞÞ (6)

8.6
MOPSO Front

8.4 True Front

8.2
f 2(x,y)

7.8

7.6

7.4
–4 –2 0 2 4 6 8
f 1(x,y)

Fig. 1 Test function results

76 K. Gill et al.

where

x
f1 ðx; yÞ ¼ x2 þ y; f2 ðx; yÞ ¼ þyþ1
2

subject to

x x
0 þ y  6:5; 0 þ y  7:5; 0 5x þ y  30; x; y 0
6 2

in the range 0  x; y  7. The true front and the front from the MOPSO algorithm
are shown in Fig. 1 after 5,000 function evaluations. It can be noticed that MOPSO
was able to reproduce the true front for this test case.

2.3 SVM-MOPSO Trainings

The unique formulation of MOPSO helps it to avoid getting struck in local optima,
when making a search in the multi-dimensional parameter domain. In the current
research, the MOPSO is used to parameterize the three parameters of SVM namely;
the trade-off or cost parameter ‘C’, the epsilon ‘e’, and the kernel width ‘g’. The
MOPSO method uses a population of parameter sets to compete against each other
through a number of iterations in order to improve values of specified multi-
objective criteria (objective functions) e.g., root mean square error (RMSE), bias,
histogram error (BinRMSE), correlation, etc. The optimum parameter search is
conducted in an intelligent manner by narrowing the desired regions of interest and
avoids getting struck in local optima.
In the earlier efforts, a single objective optimization methodology has been
employed for optimization of three SVM parameters. The approach was tested on
a number of sites and results were encouraging. However, it has been noticed that
using a single objective optimization method can result in sub-optimal predictions
when looking at multiple objectives. The single objective formulation (using PSO)
employed coefficient of determination (COD), as the only objective function, but
it was noticed that the resulting distributions were highly distorted when compared
to the observed distributions. The coefficient of determination (COD) is linearly
related to RMSE and can range between 1 and 1; the value of 1 being a perfect
fit. It is shown in Fig. 2 where a trade-off curve is presented between BinRMSE vs.
COD. It can be noticed that COD value increases with the increase in BinRMSE
value. The corresponding histograms are also shown in Fig. 3 for each of the
extreme ends (maximum COD and minimum BinRMSE) and the “compromise”
solution from the curve. The histograms shown in Fig. 3 make it clear that the best
COD (or RMSE) is the one with highest BinRMSE and indeed misses the extreme
ends of the distribution. Thus no matter how tempting it is to achieve best COD
6 Multiobjective Evolutionary Optimization and Machine Learning 77

0.55
BinRMS vs COD
Best BinRMS
0.54 Best COD
Compromise
0.53
COD

0.52

0.51

0.5

0.49
300 350 400 450 500 550 600 650 700
BinRMS

Fig. 2 Trade-off curve between BinRMSE and COD

2500
Observed
Best BinRMS
2000 Compromise
Best COD

1500
Counts

1000

500

0
0 2 4 6 8 10
× 104
Power (kw)

Fig. 3 Histogram comparison between Observed, BinRMS best, COD best, and the best compro-
mise solution

value it does not cover the extreme ends of the distribution. On the other hand the
best BinRMSE comes at the cost of lowest COD (or highest RMSE) and is not
desired either. Thus it is required to have a multiobjective scheme that simulta-
neously minimize these objectives and provide a trade-off surface and therefore a
compromise solution can be chosen between the two objectives. Figure 3 also
shows the histogram for the “compromise” solution which provides a decent
histogram when compared to observed data.
78 K. Gill et al.

3 Application

The current procedures primarily employ SVM for building regression models for
assessing and forecasting wind resources. The primary inputs to the SVM come
from the National Center for Environmental Prediction (NCEP)’s reanalysis gridded
data [14] centered on the wind farm location. The target is the measurements of wind
farm aggregate power. The current training uses a k-fold cross validation scheme
referred to as “Round-Robin strategy”. The idea within the “Round-Robin” is to
divide the available training data into two sets; use one for training and hold the
other for testing the model. In this particular “Round-Robin strategy” data is divided
into months. The training is done on all the months except one, and the testing is
done on the hold-out month. The previous operational methods employ manual
calibration for the SVM parameters in assessment projects and a simple grid-based
parameter search in forecasting applications.
The goal in using MOPSO is to explore the regions of interest with respect to
the specific multiobjective criteria in an efficient way. Another attractive feature of
MOPSO is that it results in a so-called Pareto parameter space, which accounts for
parameter uncertainty between the two objective functions. Thus the result is an
ensemble of parameter sets cluttered around the so called global optimum with respect
to the multiobjective space. This ensemble of parameter sets also gives tradeoffs on
different objective criteria. The MOPSO-SVM method is tested on the data from an
operational assessment site in North America. The results are compared with observed
data using a number of evaluation criteria on the validation sets.
As stated above, the data from four NCEP’s grid points each consisting of five
variables (total 20 variables) is used. It has been noticed that various normalization
and pre-processing techniques on the data may help to improve SVM’s prediction
capabilities. In the current study, input data has also been tried by pre-processing it
using Principal Component Analysis (PCA). In that case, the PC’s explaining 95%
of the variance are included as inputs to SVM. The comparison is made with SVM
that does not use PCA as pre-processing step.
MOPSO require a population consisting of parameter sets to be evolved through
a number of iterations competing against each other to obtain an optimum (mini-
mum in this case) value for the BinRMSE and RMSE. In the current formulation, a
50 member population is evolved for 100 iterations within MOPSO for wind power
predictions at the wind farm.

4 Results and Discussion

The wind farm site is located in Canada and has 29 months of energy data available.
The MOPSO is used to train SVM over the available 17 months of training data
(at 6-hourly time resolution) using a “Round-Robin” cross-validation strategy. This
gives an opportunity to train on 16 months and test the results on a ‘hold-out’ 1
month test set. By repeating the process for all the 17 months, gives a full 17 months
6 Multiobjective Evolutionary Optimization and Machine Learning 79

of test data to compare against the observed. Since there is 29 months of data
available for this site, a full 1 year of data is used in validation (completely unseen
data). The results that follow are the predictions on the validation set. The results
are shown on the normalized (between 1 and 1) dataset. The MOPSO-SVM
results are shown with and without PCA pre-processing.
The trade-off curve for MOPSO SVM optimization for the two objectives is
shown in Fig. 4. The trade-off between BinRMS vs. RMSE is shown for the SVM
using original input compared against SVM using Principal Components (PCs) as
inputs. It can be noticed that there is little difference between the two approaches.
The PCA-SVM produced a better objective function result for BinRMS, where as
simple SVM provided a better objective function result for RMSE.
Figure 5 shows the monthly mean wind power for the 12 months compared
against the observed data. The results are shown for SVM prediction with and
without the pre-processing using PCA. As stated above, there is a very little dif-
ference between the two approaches and a good fit has been found. It can be noticed
that predictions are in reasonable agreement with the observed data. The results in
Fig. 6 show histogram of observed vs. the predicted wind power data at the 6-hourly
time resolution (the prediction time step). The results are shown for SVM prediction
with and without the pre-processing using PCA. It can be noticed that the distribu-
tions are well-maintained using MOPSO methodology and a reasonable agreement
between observed and predicted power is evident from Fig. 6. A number of good-
ness-of-fit measures are evaluated in Table 1, which are monthly root mean square
error (RMSE), monthly coefficient of determination (COD), instantaneous RMSE,
instantaneous COD, and BinRMSE (histogram bin RMSE). The results in Table 1
are presented for SVM prediction with and without the pre-processing using PCA.
Both monthly and instantaneous wind power are of significant interest, and thus are

0.56

0.54 SVM
PCA-SVM
0.52 Poly. (PCA-SVM)
Poly. (SVM)
0.5
RMSE

0.48

0.46

0.44

0.42

0.4
5 7 9 11 13 15 17
BinRMS

Fig. 4 Trade-off curve between the two objectives BinRMS vs. RMSE for SVM and PCA-SVM
80 K. Gill et al.

0.3
0.2
0.1
0
–0.1 SVM
Predicted

PCA-SVM
–0.2
–0.3
–0.4
–0.5
–0.6
–0.7
–0.7 –0.6 –0.5 –0.4 –0.3 –0.2 –0.1 0 0.1 0.2 0.3
Observed

Fig. 5 Scatter plot for the mean monthly wind energy data for SVM and PCA-SVM

400
Observed
350 PCA SVM
SVM
300

250

200

150

100

50

0
–1 – 0.8 – 0.6 – 0.4 – 0.2 0 0.2 0.4 0.6 0.8 1
Power (kw)

Fig. 6 Histogram for SVM and PCA-SVM along with observed

included in the current analysis. It can be noticed that not only monthly but also
instantaneous power predictions are in close agreement with the observed.

5 Conclusions

Machine learning methods have gained in popularity in solving prediction problems in

the area of geosciences. Due to their ability to model complex processes, the machine
learning algorithms are preferred over computationally rigorous process-based
6 Multiobjective Evolutionary Optimization and Machine Learning 81

Table 1 Wind energy Goodness measure SVM PCA-SVM

goodness-of-fit
Monthly RMSE 0.053 0.064
Monthly COD 0.954 0.934
Instantaneous RMSE 0.381 0.381
Instantaneous COD 0.734 0.735
BinRMSE 37.758 36.491

models which are usually over-parameterized and require immense amounts of data
to calibrate. Support Vector Machines are well-suited for the problems exhibiting
high degrees of spatial and temporal variability, issues of nonlinearity, conflicting
scales, and hierarchical uncertainty.
In the current chapter, a multiobjective evolutionary computing method MOPSO
is used to optimize the three parameters of SVM for wind energy predictions. The
approach has been tested on data from a wind farm using NCEP’s re-analysis grid
data. The prediction strategy employs SVM which is parameterized for the two
objective functions. The approach is also tested by pre-processing the input data
using PCA. A number of graphical and tabular results in the form of goodness-of-fit
measures are presented for wind energy predictions. The results also show a trade-
off curve for the two objectives employed in the MOPSO. The trade-off curve is
helpful in identifying the appropriate parameter set for SVM in order to achieve
the desired accuracy for the two objective problem. The SVM predictions at the
farm level produced excellent agreement with the observed data for the validation
set. Overall, the results have been encouraging and it is recommended to use
MOPSO-SVM approach for other operational projects in the area of renewable
energy predictions and forecasting. While further modifications and advancements
are underway, the current procedure is sound enough to be applied in operational
settings.

References

1. V. Cherkassky, F. Mulier, Learning from Data - Concepts, Theory, and Methods (Wiley, New
York, 1998), p. 441
2. B. Sch€olkopf, K. Sung, J.C. Chris, C. Burges, F. Girosi, T. Poggio, V. Vapnik, Comparing
support vector machines with Gaussian kernels to radial basis function classifiers. IEEE Trans.
Signal Process. 45(11), 2758–2765 (1997)
3. V. Vapnik, The Nature of Statistical Learning Theory (Springer, New York, 1995), p. 188
4. V. Vapnik, Statistical Learning Theory (Wiley, Hoboken, NJ, 1998), p. 736
5. N. Cristianini, J. Shaw-Taylor (eds), An Introduction to Support Vector Machines and Other
Kernel-Based Learning Methods (Cambridge University Press, Cambridge, 2000), p. 189
6. A. Tikhonov, V. Arsenin, Solution of Ill-posed Problems (W.H. Winston & Sons, Washington,
DC, 1977), p. 258
7. M.K. Gill, T. Asefa, Y. Kaheil, M. McKee, Effect of missing data on performance of learning
algorithms for hydrologic predictions: Implications to an imputation technique. Water Resour.
Res. 43, W07416 (2007). doi:10.1029/2006WR005298.
82 K. Gill et al.

8. M.K. Gill, Y.H. Kaheil, A. Khalil, M. McKee, L. Bastidas, Multiobjective particle swarm
optimization for parameter estimation in hydrology. Water Resour. Res. 42, W07417 (2006).
doi:10.1029/2005WR004528 (2006)
9. R.C. Eberhart, J. Kennedy, A new optimizer using particle swarm theory, in Proceedings of
the Sixth International Symposium on Micro Machine and Human Science, 1995, MHS’95.
doi:10.1109/MHS.1995.494215 (IEEE Press, Piscataway, NJ (1995), pp. 39–43
10. J. Kennedy, R.C. Eberhart, Particle swarm optimization, in Proceedings of IEEE International
Conference on Neural Networks, IV, vol. 4. doi:10.1109/ICNN.1995.488968 (IEEE Press,
Piscataway, NJ (1995), pp. 1942–1948
11. J. Kennedy, R.C. Eberhart, Y. Shi, Swarm Intelligence (Morgan Kaufmann, San Francisco,
CA, 2001)
12. R. C. Eberhart, R. W. Dobbins, P. Simpson, Computational Intelligence PC Tools (Elsevier,
New York, 1996)
13. H. Kita, Y. Yabumoto, N. Mori, Y. Nishikawa, Multi-objective optimization by means of the
thermodynamical genetic algorithm, in Parallel Problem Solving From Nature—PPSN IV,
eds. by H.-M. Voigt, W. Ebeling, I. Rechenberg, H.-P. Schwefel (Springer-Verlag Berlin,
Germany, 1996), Sept. Lecture Notes in Computer Science, pp. 504–512
14. Kalnay et al. The NCEP/NCAR 40-year reanalysis project. Bull. Am. Meteor. Soc. 77,
437–470 (1996)
Chapter 7
Hybriding Intelligent Host-Based
and Network-Based Stepping Stone Detections

Mohd Nizam Omar and Rahmat Budiarto

Abstract This paper discusses the idea of hybriding intelligent host-based and
network-based stepping stone detections (SSD) in order to increase detection
accuracy. Experiments to measure the True Positive Rate (TPR) and False Positive
Rate (FPR) for both Intelligent-Network SSD (I-NSSD) and Intelligent-Host SSD
(I-HSSD) are conducted. In order to overcome the weaknesses observed from each
approach, a Hybrid Intelligent SSD (HI-SSD) is proposed. The advantages of
applying both approaches are preserved. The experiment results show that HI-
SSD not only increases the TPR but at the same time also decreases the FPR.
High TPR means that accuracy of the SSD approach increases and this is the main
objective of the creation of HI-SSD.

1 Introduction

When the Internet was created, security was not a priority. The TCP/IP protocol
security mechanism was thought to be sufficient at the beginning. However, as the
Internet usage increases, its security mechanism has become more and more
problematic [1]. In fact, as the internet is used widely, the number of attacks also
continues to increase. Therefore, attacks or intrusions have always occurred from
time to time. There are many techniques that can be used by an attacker or intruder
to execute network attacks or network intrusions. A persistent attacker usually
employs stepping stones as a way to prevent from being detected [2]. By using
Stepping Stone Detection (SSD), the attacker can be detected.
However, due to the complicated patterns of the stepping stones used by the
attackers, detection of these stepping stones becomes a challenging task. More

M.N. Omar (*)

College of Arts and Sciences, Universiti Utara Malaysia, 06010 UUM Sintok, Kedah, Malaysia
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 83

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_7, # Springer ScienceþBusiness Media B.V. 2010
84 M.N. Omar and R. Budiarto

intelligent techniques are required in order to detect accurately the existence of

stepping stones by analyzing the traffic pattern from one node to another. Beginning
with a research conducted by Standiford-Chen and Herberlein [3] to the latest
research by Wu and Huang [4], problems such as accuracy [5] and Active Pertur-
bation Attacks (APA) [6] still remain the top issues among the researchers. An
Intelligent-Host-Based SSD (I-HSSD) [7] and an Intelligent Network-based SSD
(I-NSSD) [8] have been introduced in our previous research. Nevertheless, they
have yet to provide a high rate of accuracy in detecting the stepping stones. We
propose a Hybrid Intelligent SSD (HI-SSD) so as to tackle this accuracy problem.
The proposed HI-SSD will be compared with other approaches: the I-HSSD and
I-NSSD to examine the hybrid approach that has been applied in this research.
Experiments will be conducted using a well-planned dataset, and the performance
in terms of True Positive Rate (TPR) and False Positive Rate (FPR) [9] will become
our main benchmark.
The rest of this article is structured as follows. Section 2 gives research terms
used in this research. Section 3 discusses related works and Section 4 describes the
proposed approach. In Section 5, we discuss further the experiment and then, a
discussion of the results is in Section 6. Finally, we summarize the overall research
and present possible future works in Section 7.

2 Research Terms

Before we start in more detail on the experiment, there are several research terms or
terminologies used in this work which need to be clarified. In Fig. 1, there are five
hosts involved in an SSD environment. Host A is a source of attack and Host E is a
victim.
From Fig. 1, Stepping Stones (SS) are hosts A, B, C, D and E. SS ¼ {A, B, C, D, E}
where hosts A, B, C, D and E contain the same packet that flows through each host.
A Connection Chain (CC), on the other hand, is the connection between hosts A, B,
C, D and E. Therefore CC is a, b, c and d. CC ¼ {a, b, c, d}. In Stepping Stone
Detection (SSD) research using the Network-based approach (N-SSD), either SS ¼
{A, B, C, D, E} or CC ¼ {a, b, c, d}, can be used to denote the existence of stepping
stones.

A B C D E

a b i c d
inbound outbound
packet of packet of
Host C Host C

Fig. 1 Detecting a stepping stone chain

7 Hybriding Intelligent Host-Based and Network-Based Stepping Stone Detections 85

Host-based SSD (HSSD), in contrast, works by determining whether or not

inbound and outbound connections contain the same packet flow. By referring to
Fig. 1, if Host C is referred to, the inbound connection is CCc ¼ b and the outbound
connection is CCc ¼ c. In this case, CC2 ¼ CC3 or b ¼ c.
The challenge to SSD is to find the right solution by determining SS or CC.
Overall, SSD research can be divided into statistical- and AI-based SSD, with most
of the AI approach comprising the most recent research on SSD. As explained
before, SSD research begins with the introduction of statistical-based research such
as ON/OFF [10], Thumbprint [3], Deviation [11], and so forth.

3 Related Works

Research by Thames et al. [12] applied the hybrid concept by combining Bayesian
Learning Network (BLN) and Self-Organizing Map (SOM) for classifying network-
based and host-based data collected within LAN for network security purposes. The
experiment was conducted by using four types of analyses (i) BLN with network
and host-based data, (ii) BLN with network data, (iii) hybrid BLN-SOM analysis
with host and network-based data and (iv) hybrid BLN-SOM analysis with net-
work-based data. The four different types of analyses were required to compare one
result to another.
Meanwhile, Bashah et al. [13] proposed a system that combines anomaly, misuse
and host-based detection for Intrusion Detection System (IDS). This research only
proposed an architecture that combines fuzzy logic and the SOM approach without
any implementations/experiments.
Inspired by the above two research works, we hybrid the I-HSSD and I-NSSD
approaches and compare the hybrid approach (HI-SSD) with the non-hybrid
approaches (I-HSSD and I-NSSD) in terms of accuracy.

4 Proposed Approach: Hybrid Intelligence Stepping Stone

Detection (HI-SSD)

The main components of HI-SSD are intelligence and hybrid. The intelligence
component comes from the use of the Self-Organization Map (SOM) approach and
the hybrid component is created from the combination of Host-based SSD and
Network-based SSD. Both components of intelligence and the hybrid are discussed
in detail our previous research [7, 8, 14]. Figure 2 shows the HI-SSD architecture.
Figure 2 shows the overall HI-SSD architecture that involves I-HSSD and
I-NSSD. In this architecture an intrusion detection system (IDS) is used as a trigger
to detect any network intrusion. When an intrusion occurs, I-NSSD starts to capture
the network packet in a defined range. At the same time, each host also captures the
86 M.N. Omar and R. Budiarto

Host
for each
host I-HSSD

IDS

SSD
attack is
detected SSD list

similar?

I-NSSD

Fig. 2 HI-SSD architecture

network packet as well. When I-NSSD finishes its process to detect stepping stones,
information about related hosts involved in the chain of stepping stones is pro-
duced. Each host listed in I-NSSD as a stepping stone node then executes a self-
examination to check whether it is being used as a stepping stone or not. Results on
both I-NSSD’s list and I-HSSD’s list then are compared. Similarity between these
lists shows the real stepping stone host.
For testing purposes, only the functions of HI-SSD, I-NSSD and I-HSSD are
involved in the experiment. The development of a fully-functional HI-SSD will
become our future work.
The HI-SSD will contain a stepping stone list each from I-NSSD and I-HSSD,
while the I-NSSD and I-HSSD will contain a stepping stone list for every network
and host respectively. Comparisons will be measured on the TPR and the FPR on
each component.

5 Experiment

For the dataset arrangement, Telnet Scripting Tool v.1.0 [15] is used. This is to
guarantee a uniform pattern of telnet operations during the execution of the
experiment. Here, Telnet represents the interactive connection most frequently
used by SSD-related research. Moreover, there is no other dataset that is suitable
in this research. Jianhua and Shou-Hsuan [16] used their own dataset. Staniford-
Chen and Herberlein [3] also agreed with that.
The experiment is run in a controlled environment so as to avoid any interfer-
ence with outside networks. Wireshark [17], on the other hand, is used to capture
network packets that flow in each host. After the Telnet Scripting tool has ended its
run, information pertaining to the packets is converted into text-based form. This is
7 Hybriding Intelligent Host-Based and Network-Based Stepping Stone Detections 87

done in order to proceed to the consequent processes to acquire the appropriate

information needed later. In this research, only time information is needed. This
time information from the experiment is transferred into m-type file to be used
later with Matlab 6.1 [18] software. In Matlab 6.1, the time information is used
as the input to create, train, and lastly, to plot the SOM graph. The result from
the visualization is taken as the result of this research and will be discussed in the
next section.
Although the result for HI-SSD is obtained from I-NSSD and I-HSSD, this work
will show the benefit of stepping stone detection when information obtained from a
combination of the two approaches is compared to two separate sets of information
when using I-NSSD or I-HSSD alone.
Figure 3 shows the layout of the overall experiment. From the figure, it explicitly
shows that four hosts are involved in the stepping stone.
From these four stepping stone hosts, there are three connection chains (C1, C2,
and C3) involved. That means that only four stepping stones or three connections
should be detected by the SSD approach. Increased number of detections causes
lower TPR and decreased number of detections causes higher FPR. Table 1 shows
the relationships for each host and its connection chains.
Based on the location of each host in Fig. 1, Table 1 shows the number of
possible connection chains for each host and its list. The number of connection
chains has been made based on the assumption that Host 1 and Host 4 are the
starting and ending points respectively. Host 2 and Host 3 contain two different
types of connection chains, one and two. However, connection chains that contain

Host 4 Host 2

C3 C2 C1

Host 3 Host 1

Fig. 3 Experiment layout

Table 1 Host and its Host No. of connection chain Connection chain list
relationship
Host 1 3 c1, c2, c3
Host 2 1 c1
2 c2, c3
Host 3 2 c1, c2
1 c3
Host 4 3 c1, c2, c3
88 M.N. Omar and R. Budiarto

just one connection can be eliminated because one connection does not mean that it
is a stepping stone connection. In fact, research by RTT-based SSD [19–21] agreed
that only connection chains with more than three connections can be considered as
stepping stone connections. Based on our previous research [8], the existence of
two connections onwards is enough for a chain to be identified as a possible
stepping stone chain.
To calculate the effectiveness of the tested approach, TPR and FPR are used.

number of false positive

FPR ¼ (1)
number of possible negative instances

False Positive Rate (FPR) refers to the fraction of negative instances that are
falsely reported by the algorithm as being positive. In this situation, the algorithm
has detected the connection chains which exist even though it is not true.

number of true positive

TPR ¼ (2)
number of possible true instances

True Positive Rate (TPR) refers to the fraction of true instances detected by the
algorithm versus all possible true instances. The discussion on the result and its
analysis will be presented.

6 Result and Analysis

Result on each type of SSD will be discussed in each of the following sub-section.
As described before, we have chosen to use SOM as the intelligent approach. As a
result, the SOM graph will represent each compared solution.

6.1 Intelligence Network Stepping Stone Detection (I-NSSD)

In I-NSSD, the results are obtained from the execution of SOM approach through
the arrival time for the overall captured data in the network. In this case, only one
graph has been produced. It is different from I-HSSD that needs one graph for each
host involved. Figure 4 shows the result.
In contrast to the I-HSSD approach that depends on the number of possible
straight lines which can be created to determine the number of connection chains,
the I-NSSD approach is based on the number of possible groups generated. In
Fig. 4, four possible groups of SOM nodes can be counted (labeled as a, b, c and d).
Thus, there are four connection chains involved in the experiment. However, the
true number of connection chains involved in this research is only three. Therefore,
7 Hybriding Intelligent Host-Based and Network-Based Stepping Stone Detections 89

Fig. 4 Node of SOM in I-NSSD

there exist false positive reports in the I-HSSD experiment. By using formula (1),
FPR for I-NSSD is 33.3%. For TPR, I-NSSD shows 100% achievement.

6.2 Intelligence Host-Based Stepping Stone Detection (I-HSSD)

As described previously, I-HSSD involves every host that has been listed in
I-NSSD. In this case, arrival time for each host is run by using the SOM approach.
Each result is an output from the execution.
Figure 5 shows that three possible directions can be traced (e, f and g). That
means there are three possible connections which exist in the originating host,
Host 1. Based on the value from Table 1, Host 1 obtains 100% TPR and 0% FPR.
Figure 6 on the other hand shows that there are two connection chains (h and i)
that could possibly exist in Host 2 as the monitored host. Based on the number of
connection chains from Table 1, Host 2 got 100% TPR and 0% FPR.
In Fig. 7, similar to Fig. 6, there are two connection chains (j and k) that
could possibly exist in this graph. Based on the number of connection chains
from Table 1, Host 3 also got 100% TPR and 0% FPR.
Figure 8 shows the last node of SOM that needs to be observed. From the graph,
there are two possible obtainable directions. That means that two connection chains
90 M.N. Omar and R. Budiarto

Fig. 5 Node of SOM on Host 1

Fig. 6 Node of SOM on Host 2

7 Hybriding Intelligent Host-Based and Network-Based Stepping Stone Detections 91

Fig. 7 Node of SOM on Host 3

Fig. 8 Node of SOM on Host 4

92 M.N. Omar and R. Budiarto

Table 2 I-HSSD experiment Host No. of connection No. of TPR FPR

result chain connection (%) (%)
chain detected
1 3 3 100 0
2 2 2 100 0
3 2 2 100 0
4 3 2 66.7 0

(l and m) have been detected. However, based on the number of connection chains
in Table 1, three connection chains should have been detected in Host 4. For that
reason, I-HSSD in Host 4 miss-detects another connection. In this case, the TPR
becomes 66.7%. The FPR on the other hand also gives 0%.
From the result on each host, it shows that only the result of Host 4 does
not achieve 100% TPR. This can be considered as a weakness of the I-HSSD.
Table 2 shows the result of the I-HSSD experiment.
Based on the overall result of the I-HSSD experiment, connection chains have
been successfully detected in Host 1, Host 2 and Host 3. In Host 4, although the
connection chains have also been successfully detected, the number of connection
chains detected is less than the actual number involved. This makes the TPR at just
66.7%. On the other hand, the FPR is the same for all hosts.

6.3 Hybrid Intelligence Stepping Stone Detection (HI-SSD)

As described earlier, HI-SSD is constructed from the combination of I-HSSD and

I-NSSD. Therefore, the information from I-HSSD and I-NSSD is still used in the
HI-SSD. As shown in Fig. 2, the information obtained from I-HSSD and I-NSSD
are rearranged so as to generate a more accurate stepping stone detection result
compared to the use of I-HSSD or I-NSSD alone.
For this purpose, the first information acquired from I-NSSD is observed. From
the result, four connection chains are detected. At this level, false detection still not
has been discovered. All information needed is distributed to the related hosts. The
host which is related to the list then runs the I-HSSD function by itself. This is
to check whether or not the host is used as a stepping stone. In this case, Host 1,
Host 2, and Host 3 have successfully detected the number of existing connection
chains. However, Host 4 has just detected two connection chains compared to three
connection chains that should have been detected.
Although miss-detection has occurred here, the number of maximum possible
hosts from Host 1 helps a lot to balance the Host 4 result. From the I-HSSD result, it
is shown that there are only three connection chains involved compared to four
proposed by I-NSSD. Combining I-HSSD and I-NSSD avoids false and miss-
detections. By using both I-HSSD and I-NSSD, it is shown that 100% TPR and
7 Hybriding Intelligent Host-Based and Network-Based Stepping Stone Detections 93

Table 3 Experiment result Type of stepping stone TPR (%) FPR (%)
detection approach
I-NSSD 100 33.3
I-HSSD 91.67 0
HI-SSD 100 0

0% FPR has been achieved. For a clear picture on the overall HI-SSD, the related
algorithm is given as follows.
Begin
I-NSSD:
capture network packet
collect arrival time
execute SOM
count the group of node, n
list involved host, l
sent n & l to I-HSSD
I-HSSD:
activate I-HSSD on selected host based on l
for 1 to n
execute SOM
count the possible straight line,
identify the existence of connection chain for the host, e
end for
End
Result of the experiment according to the percentage of TPR and FPR obtained
from I-NSSD, I-HSSD and HI-SSD respectively is tabulated in Table 3.
Table 3 shows the experiment result of I-NSSD, I-HSSD and HI-SSD. As
described previously, HI-SSD is a combination of I-NSSD and I-HSSD. Therefore,
the TPR and FPR for HI-SSD is the combination of I-NSSD’s TPR and FPR and
I-HSSD’s TPR and FPR. In the other words, the average TPR and FPR of I-NSSD
and I-HSSD become the result of HI-SSD. In this case, 100% TPR and 0% FPR for
HI-SSD is better than the use of I-NSSD or I-HSSD alone. Although I-NSSD
shows 100% TPR, but the 33.3% FPR can adversely affect the overall function of
stepping stone detection. In the I-HSSD, 100% FPR makes this approach free of
false detection. However, 91.67% TPR is not enough to render this approach to be
fully-functional. As a result, combining both I-NSSD and I-HSSD to form the
HI-SSD will balance the TPR and the FPR at the same time.

7 Conclusion and Future Work

The goal of this research is to prove the effectiveness of the proposed HI-SSD
which is actually a hybrid or combination of I-NSSD and I-HSSD. From the
experiment, it is shown that the HI-SSD is more accurate compared to both
94 M.N. Omar and R. Budiarto

I-NSSD and I-HSSD. Therefore, it is proven that HI-SSD is more accurate com-
pared to I-NSSD or I-HSSD.
In the future, we will improve our approach to be more robust towards active
perturbation attacks such as delay, packet drop and chaffing. Testing on the active
perturbation problem would not only be executed onto our proposed solution, but it
would also involve our closest research that used the Data Mining approach. By
doing this, we not only can measure the proposed approach’s capabilities but at the
same time can also compare it with other similar approaches.
Only one dataset is used in this experiment. With the intention of verifying the
capability of the proposed HI-SSD, we need different types of dataset. Our observa-
tion on the datasets used in SSD-based research has shown that there are two kinds
of datasets which can be used; datasets generated by ourselves and public datasets.
By using various kinds of datasets instead of just one type of dataset, the true
capabilities of the approach can be examined more. As such, testing the proposed
HI-SSD on top of different datasets is one of our future works.
Another future plan is to compare the proposed HI-SSD approach with the statistical-
based approaches. Therefore, to ascertain that the proposed approach is better than
statistical-based approaches, it is good to involve the different approaches together
in the experiment.
Lastly, for a true experience of a fully-functional stepping stone detection, the
overall approach should be translated into a full system. If a complete system is
created, comparisons with other approaches could be made more easily.

References

1. C. Chamber, J. Dolske, J. Iyer, “TCP/IP Security”, Department of Computer and Information

Science (Ohio State University, Ohio 43210). Available at: http://www.linuxsecurity.com/
resource_files/documentation/tcpip-security.html
2. K.D. Mitnick, W.L. Simon, The art of intrusion: the real stories behind the exploits of hackers,
intruders & deceivers” (Wiley, 10475 Crosspoint Boulevard, Indianapolis, IN 46256, 2005)
3. S. Staniford-Chen, L.T. Herberlein, Holding intruders accountable on the Internet, in
Proceedings of the 1995 IEEE Symposium on Security and Privacy, Oakland, CA, 1995,
pp. 39–49
4. H. Wu, S.S. Huang, Neural network-based detection of stepping-stone intrusion. Exp. Sys.
Appl. 37(2), 1431–1437 (March 2010)
5. A. Blum, D. Song, S. Benkataraman, Detection of Interactive Stepping Stone: Algorith-
m0020and Confidence Bounds, vol 3224/2004. Lecture Notes in Computer Science (Springer
Berlin/Heidelberg, 1 Oct 2004), pp. 258–277
6. X. Wang, Tracing Intruder BStepping Stone. Ph.D. Thesis, North Carolina State University,
Raleigh, 2004.
7. M.N. Omar, R. Budiarto, Intelligent host-based stepping stone detection approach, in Pro-
ceeding of The World Congress on Engineering and Computer Science 2009, vol II, San
Francisco, USA, 20–22 Oct 2009, pp. 815–820
8. M.N. Omar, R. Budiarto, Intelligent network-based stepping stone detection approach. Int. J.
Appl. Sci. Eng. Technol. 53–135, 834–841 (2009)
9. A. Almulhem, I. Traore, Detecting connection-chains: a data mining approach. Int. J. Netw.
Security. 11(1), 62–74 (Jan 2010)
7 Hybriding Intelligent Host-Based and Network-Based Stepping Stone Detections 95

10. Y. Zhang, V. Paxson, Detecting stepping stones, in Proceedings of the 9th USENIX Security
Symposium, Denver, CO, 2000, pp. 67–81.
11. K. Yoda, H. Etoh, Finding connection chain for tracing intruders, in Proceedings of the 6th
European Symposium on Research in Computer Security (LNCS 1985), Toulouse, France,
2000, pp. 31–42
12. J.L. Themes, R. Abler, A. Saad, Hybrid intelligent system for network security. ACM
southeast regional conference 2006, 2006.
13. N. Bashah, I.B. Shanmugam, A.M. Ahmed, Hybrid intelligent intrusion detection system,
Proc. World Acad. Sci. Eng. Technol. 6 (2005)
14. M.N. Omar, L. Serigar, R. Budiarto, Hybrid stepping stone detection method. in Proceeding of
1st International Conference on Distrubuted Framework and Application (DFmA 2008),
Universiti Sains Malaysia, Penang, 21–22 Oct 2008, pp. 134–138
15. Wareseeker (2008) [Online] Available: http://wareseeker.com/freeware/telnet-scripting-tool-
1.0/19344/TST10.zip 8 Feb 2008
16. Y. Jianhua, S.H. Shou-Hsuan, A real-time algorithm to detect long connection chains of
interactive terminal session”, in Proceedings of the 3rd International Conference on Informa-
tion Security, China, 2004, pp. 198–203
17. Wireshark (2009). [Online] Available: http://www.wireshark.org 8 Feb 2009
18. H. Duane, L. Bruce, Mastering MATLAB A Comprehensive Tutorial and Reference (Prentice-
Hall, New Jersey, 1996)
19. K.H. Yung, Detecting long connection chains of interactive terminal sessions, in Proceedings
of the International Symposium on Recent Advance in Intrusion Detection (RAID 2002),
Zurich, Switzerland, 2002, pp. 1–16
20. J. Yang, S.S. Huang, A real-time algorithm to detect long connection chains of interactive
terminal sessions, in Proceedings of the 3rd International Conference on Information Security
(INFOSECU 2004), Shanghai, China, 2004, pp. 198–203.
21. J. Yang, S.S. Huang, Matching TCP packets and its application to the detection of long
connection chains on the internet, in Proceedings of the 19th International Conference on
Advance Information Networking and Applications (AINA 2005), Tamkang University,
Taiwan, 2005, pp. 1005–1010
Chapter 8
Open Source Software Use in City Government
Is Full Immersion Possible?

David J. Ward and Eric Y. Tao

Abstract The adoption of open source software (OSS) by government has been a
topic of interest in recent years. National, regional, and local government are using
OSS in increasing numbers, yet the adoption rate is still very low. This study
considers if it is possible from an organizational perspective for small to
medium-sized cities to provide services and conduct business using only OSS.
We examine characteristics of municipal government that may influence the
adoption of OSS for the delivery of services and to conduct city business. Three
characteristics are considered to develop an understanding of city behavior with
respect to OSS: capability, discipline, and cultural affinity. Each of these general
characteristics contributes to the successful adoption and deployment of OSS by
cities. Our goal was to determine the organizational characteristics that promote the
adoption of OSS. We conducted a survey to support this study resulting in 3,316
responses representing 1,286 cities in the Unites States and Canada. We found
most cities do not have the requisite characteristics to successfully adopt OSS on a
comprehensive scale and most cities not currently using OSS have no future plans
for OSS.

1 Introduction

All city governments seek to deliver services in the most efficient manner possible.
Whether in direct support of service delivery or support of conducting the business
of government, Information Technology (IT) has become and integral component
of operations at all levels of government.
In the past 5 years there has been a trend by some national, regional, and local
governments toward use of open source software (OSS) and open standards as a

D.J. Ward (*)

CSU Monterey Bay, 100 Campus Center, Seaside, CA 93955, USA
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 97

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_8, # Springer ScienceþBusiness Media B.V. 2010
98 D.J. Ward and E.Y. Tao

first choice rather than a curiosity. Recently, The Netherlands has mandated that all
national government agencies will use open standards formatted documents by
April 2009 [1]. The U.S. Navy has clarified the category and status of OSS to
promote a wider use of OSS to provide seamless access to critical information [2].
The U.S. Congress is recognizing the potential value of OSS in the National
Defense Authorization Act for fiscal year 2009. OSS is identified as an objective
in the procurement strategy for common ground stations and payloads for manned
and unmanned aerial vehicles, and in the development of technology-neutral
information technology guidelines for the Department of Defense and Department
of Veteran Affairs.
Small to medium size cities, populations less than 500,000 [3], may have serious
limitations in funding IT efforts. Escalating costs of service delivery coupled with
reduce revenue will force governments to seek novel ways to reduce operating
costs. With limited revenue their budgets seem to force cities to under fund IT infra-
structure in favor of applying resources to increasing labor requirements to deliver
services. Careful and deliberate selection of IT solutions can reduce the labor required
for service delivery freeing that labor to be harvested for other purposes.
Considerable research has been conducted on the topic of e-government and the
development of models to explain e-government maturity. There have also been
ample studies of the trends in adoption of OSS by government at various levels.
However, little research has been done in the characteristics of regional and local
government that would promote adoption and successful deployment of OSS.
The support of city leadership and management as well as the IT staff are
required for successful adoption of any technology, not just OSS. Vision, strategy
and government support are important for success of IT projects, while insufficient
funding and poor infrastructure are major factors for failure [4].
This research focused on the perspectives of city leadership, management, and
IT staff with respect to OSS and its adoption by city government. While OSS was
the focus, the approach used in this study can be applied to investigating the
adoption of most technologies or practices.

2 Related Research

The literature review revealed little research with similar characteristics of this
study. While much work can be found on OSS in a wide variety of topics, the body
of research covering municipal OSS adoption is relatively limited. Of the literature
found relating to OSS adoption by government, a significant portion of that work
examines the adoption of OSS by developing countries.
A fair number of studies have been conducted to examine the extent to which
government entities are using OSS. The studies tend to collect, categorize, and
report the current state of open source adoption [5, 6]. A study of Finnish munici-
palities [7] had a methodology similar to the methodology used in this study.
8 Open Source Software Use in City Government 99

While the Finnish study considered only IT managers as survey subjects, our survey
considered city leaders, managers, and IT staff for subjects, as our focus is examin-
ing organizational behavioral influences on OSS adoption rather than technical and
budgetary influences.
In the current economic climate government at all levels are facing funding
crises as costs of operations increase and revenue decreases. The question of what
can be done to reduce IT operating costs is now a very important one. There are
more benefits to using OSS than just reduced acquisition costs [8]. Restrictive
licensing, vendor lock-in, and high switching costs can be eliminated, which, in
the long term, also may reduce costs.
The level of knowledge of a user with respect to OSS and Closed Source
Software (CSS) will influence their decision to use OSS.
There are two topologies of consumers: a) “informed” users, i.e. those who know about the
existence of both CSS and OSS and make their adoption decision by comparing the utility
given by each alternative, and b) “uniformed” users, i.e. those who ignore the existence of
OSS and therefore when making their adoption decision consider only the closed source
software [9].

We can apply this observation to municipal organizations. The informed municipal

organization, which knows about the existence of open-source alternatives to
commercial products, may make adoption decisions based on the value provided
by each. The uninformed organization either ignores the existence of open-source
alternatives or is unaware of OSS alternatives to commercial products. The unin-
formed organization may have misperceptions of OSS. These misperceptions may
include OSS usability, deployment, and support. A common misperception is that
an organization must have a programmer on staff in order to deploy and maintain
OSS. While in the distant past within the open source era, this may have been true,
the current maturity of most open source applications may require only a competent
IT technician.
The adoption of OSS by municipal government is a technology transformation
problem. Technology transformation is not about the technology, but the organiza-
tion adopting the technology, in this case city governments. Implementation of
technologies to support government efforts in its self does not guarantee success
[10]. Organizations must change to embrace the new technologies in order to use
them effectively [11].
e-Government cannot be achieved by just simply implementing available soft-
ware [12]. Fundamental differences between the public and private sectors influ-
ence the rate at which ICT is employed [13]. The real opportunity is to use IT to
help create fundamental improvement in the efficiency, convenience and quality of
service [14]. Adopting new or different technology requires an organization to
change, with varying degree, at all levels. An organization with a culture that
embraces change will be more successful at adopting new technologies than an
organization with a strong status quo bias.
100 D.J. Ward and E.Y. Tao

3 Research Goals

Our approached in this study was to focus on organizational characteristics required

to promote OSS adoption instead of OSS as technology needing to be implemented.
One fundamental question ultimately coalesced; is it possible for small to medium
cities to use only OSS to deliver services and conduct city business? This is not a
question of technical feasibility or product availability, but one of organizational
characteristics. If we assume OSS products exist in the application domains city
government requires for IT operations, then successful adoption and deployment
becomes an organizational behavior issue.
Three general domains of interest evolved from the core question; capability,
discipline, and cultural affinity. Having the capability to deploy and manage infor-
mation technology was deemed very important for a city to successfully deploy
OSS technologies. A city that has a mature management structure in place that
demonstrates forethought in planning and a deliberate approach to budgeting and
acquisition will have the level of discipline to deploy OSS on a comprehensive
scale. Cultural affinity we define as the predisposition toward or against the use of
OSS. The awareness, support, and understanding of OSS by key city personnel
(leadership, management, and IT Staff) will influence a city’s affinity toward the
use of OSS. From these domains we developed additional questions to guide this
research effort: Do cities have the necessary organizational characteristics to adopt
and use only OSS to deliver services and conduct city business? What are the basic
IT capabilities of cities? Do these capabilities support the adoption and deployment
of open source technologies? Do cities plan and budget for IT in a deliberate
manner that would support the adoption and deployment of open source technolo-
gies? Does the organizational culture of cities promote the adoption of open source
technologies?

4 Methodology

The following sections describe the methodology used for this study. We begin by
describing in general the process, followed by the survey design, survey execution,
and subject selection.
A survey was conducted to collect data from municipal IT managers, IT staff,
city leadership, city management, and city employees. The survey was adminis-
tered online using SurveyMonkey.com. The collection period was initially sched-
uled for 30 days from June 1, 2008 through June 30, 2008.
The survey required soliciting responses from subjects to provide insight into the
characteristics of their cities with respect to IT capability, organizational discipline,
and cultural affinity to OSS. Presenting direct questions would not produce useful
data, as subjects may not have the requisite knowledge in the subject areas. One
goal in the design of the survey was to reduce the number of aborted attempts by
subjects. An aborted attempt is the failure to complete the survey once started.
8 Open Source Software Use in City Government 101

We identified six subject classifications; IT Manager, IT Staff, City Leadership,

City Management, City Employee, Other. These classifications were used to tailor
the set of questions presented to the subject to keep the set within the knowledge
domain of the subject. All survey questions were framed in the context of the subject
and their perception of their city. This was an important aspect of the design as the
survey questions solicit information that indirectly relates to the research questions.
The survey was divided into four sections. The first three sections address in
three interest domains; capability, disciple, and cultural affinity. The fourth section
solicited demographic information.
The first section of the survey collected data regarding city IT capability addres-
sing the capability dimension characteristic of cities. The second section solicited
responses related to the city’s IT strategy addressing the discipline dimension
characteristic of cities. The third section solicited responses related to the subject’s
perspectives and opinions about IT and OSS and the subject’s impression of the city
leadership, management, and IT staff’s perspectives of IT and OSS. These questions
were intended to reveal the city’s cultural affinity to the adoption of OSS.

5 Survey Execution

For the announcement strategy we used three channels to contact potential subjects;
magazines related to city management, municipal associations, and direct e-mail.
Announcing the survey through a magazine was deemed to have potential for
generating significant level of exposure for those subjects who are more likely to
read city government related magazines. Several magazines were contacted for
assistance to announce the survey. Two magazines responded, the Next American
City Magazine and American City and County Magazine. The Next American City
magazine provided a half page ad space to announce this survey. American City and
County Magazine announced the survey in an article posted on the front page of
its website. Although the potential exposure was thought to be high, the magazine
announcement channel only produced 20 responses of which only one response was
valid for analysis.
Municipal associations were thought to be the best vehicle for reaching the
largest number of subjects. The rationale behind this was the assumption that
individuals affiliated with municipal associations might be more inclined to respond
to a survey announcement received from their association. The expectation was
that the greatest number of responses would result from municipal associations.
Individuals affiliated with municipal associations may also have greater interest
in supporting this research as they may see a potential benefit for their city.
A total of 116 municipal associations were contacted to assist with announcing
the survey to their members, 28 associations approved the request for assistance and
forwarded the announcement to their members.
The municipal associations were identified via a search of the Internet. Most of
the associations found were regional providing representation within a county,
102 D.J. Ward and E.Y. Tao

multi-county, state, or multi-state area. Each municipal association was sent an

initial survey announcement with a reminder sent within 7 days of the initial
announcement.
We received 207 responses from subjects indicating they learned of the survey
through a municipal association.
To reach the greatest number of potential subjects a direct e-mail approach was
also used. Direct e-mail has proven to be a very effective and economical means of
reaching the greatest number of survey subjects [15].
Individuals were contacted via email addresses harvested from municipal web-
sites. A commercial email-harvesting program, Email Spider (www.gsa-online.de),
was used to collect the municipal email addresses.
The collection process harvested over 80,000 email addresses from the munici-
pal websites. Invalid email addresses, those not associated with a city government
domain name (i.e. yahoo.com, aol.com), were excluded to produce a set of 60,000
addresses to which the announcements were sent.
Survey announcements were emailed to the potential subjects over a 2-week
period. Reminder emails were sent within 7 days of the initial announcements. Survey
responses can be increased with a reminder email that includes more than just a link
to the survey15. With this in mind the reminder email included the original announce-
ment text with an introductory paragraph explaining the email was a reminder.
The collection period defined in the research design was for 1 month running
from June 1, 2008 through June 30, 2008. Toward the end of the collection period,
between 18 and 25 June, the response activity maintained a significant level (an
average of 143 per day) prompting this researcher to extend the collection period 15
days ending on July 15, 2008. Additionally, many automated email responses
indicated the recipients were on vacation during the month of June. In the days
following distribution of the initial survey announcement email and reminder
emails increased response activity was observed. We anticipated response activity
would increase during the first part of July as potential subjects returned from
vacation. During the 15 day extended collection period 1443 responses were
collected, 43.5% of the total responses.

6 Survey Results

The total estimated exposures to the survey announcement were in excess of

60,000. An exposure for the purpose of this study is defined as the delivery of a
survey announcement to a potential subject. Several factors prevent an accurate
tally of total survey announcement exposures. We did not have access to the
membership numbers for the municipal associations that forwarded the survey
announcement or access to the web site page hit counts for the article on the
magazine website.
Of the 60,000 e-mails sent directly to city leaders, managers, and employees,
53,900 may have reached the addressee. 6,100 of the original 60,000 email
8 Open Source Software Use in City Government 103

announcements were returned as undeliverable, a non-existent address, or reported

by an email server as potential “spam”.
A total of 3,316 individuals responded to the survey announcement resulting in a
response rate of 6%. There are 1,286 distinct cities represented in this response set.
A sample set was created to include cases from cities with populations less than
300,000 and an indicated primary duty of IT Manager, IT staff, City Leader, or City
Manager. While the survey data included responses from cities with populations
greater than 300,000, those responses were too few in number to permit valid analysis
(Fig. 1).
Reponses from city employees (not city leaders, managers, or IT staff) and
individuals not affiliated with a city were excluded from the sample set used for
analysis. The responses from city employees had very limited or no value as these
subjects had little knowledge of city IT capability, strategy, or the views of the city
leadership, management, and IT staff regarding OSS. These criteria produced a
sample set of 1,404 cases representing 1,206 cities.
Cities use a wide variety of desktop operating systems. Within the IT staff
sample sub-set, 15 different operating systems were identified. Virtually all cities
(99.7%) deploy one or more versions of Microsoft Windows on desktop computers.
Twenty percent of the respondents indicate Linux is used on desktop computers.
The survey instrument did not collect the degree to which Linux is deployed on the
desktop in the respondents’ city.
An interesting observation was 13% of the IT staff indicating that Mac OS X was
deployed on desktop computers in their city. Since Mac OS X can only be installed
on Apple Inc. hardware [16], we can conclude these cities are using Apple com-
puters to support service delivery or to conduct city business (Fig. 2).
Server side operating systems showed similar deployment rates with Microsoft
Windows most widely used at 96.5% and Linux following in second at 40%. A
small number of cities continue to use Novell Netware and DecVMS/OpenVMS
indicating some cities may be maintaining legacy software to support service
delivery or city operations (Fig. 3).

Fig. 1 Response by primary

duty
104 D.J. Ward and E.Y. Tao

Fig. 2 Desktop OS deployment

Fig. 3 Server OS deployment

7 Analysis: Interesting Findings

7.1 Few Cities Have All Characteristics

Analysis of the survey data indicates few cities have all the characteristics that
would enable successful comprehensive adoption and deployment of OSS. Of the
1,206 distinct cities in the sample set, just ten cities satisfied all characteristics
within the three dimensions.
Ten cities, listed in Table 1, satisfied the following criteria; has an IT Depart-
ment, IT support is handled in-house, currently uses OSS, has well defined IT
strategy, has IT line item in budget, IT is sufficiently funded, total cost of ownership
acquisition strategy, uses budget for software acquisition.
8 Open Source Software Use in City Government 105

Table 1 Cities satifying all City State Population

selection criteria
Balwin Missouri 30,000
Northglen Colorado 31,000
Houma Louisiana 32,400
Ipswitch Massachusetts 12,000
Largo Florida 73,000
Layton Utah 64,300
Redding California 80,800
Santa Monica California 87,200
Tomball Texas 10,200
Ulysses Kansas 5,600

Largo, Florida is of particular interest. Largo, has embraced the use of OSS. It
has deployed Linux as the operating system for its 400 desktop clients saving the
city an estimated $1 million per year in hardware, software, licensing, maintenance,
and staff costs [17].

7.2 Possible Aversion to OSS If Not Currently Using OSS

Of the 460 Municipal IT managers and staff in the sample set, 56% indicated their
city was not currently using OSS while 39% indicated their city was using OSS.
Considering the widespread use of OSS in the commercial sector, the relatively
high percentage of cities in this survey not currently using OSS required further
investigation (Fig. 4).
Of the Cities currently using OSS, 76% are planning to use OSS in the future and
10% have no plans to use OSS in the future. The high percentage of cities planning
to use OSS in the future that currently use OSS can be expected. It is more likely an
organization will continue to use a software product once deployed and established
than to abandon the product (Fig. 5).
The cities currently not using OSS provides a more interesting observation. Of
the 259 IT managers and staff indicating their city is currently not using OSS, 82%
indicated their city has no plans to use OSS in the future, 9% indicated their city did
plan to use OSS in the future, and 9.7% (25) did not know.
The number of dedicated IT staff at the respondent cities may not be an
influencing factor in decisions to use OSS in the future. While 74% of the cities
not planning to use OSS in the future have IT staff numbering ten or less, 71% of
cities currently using OSS also have ten or less IT staff.
The organizational support for using OSS appears to be a significant influencing
factor for a city’s future plans for OSS use. The survey design included questions
regarding the respondent’s perception of the Leadership, Management, and IT staff
views of OSS. The subjects were asked if the city leadership, management, and IT
staff support the use of OSS. For the cities not planning to use OSS in the future
only 6% of the respondents indicated the city leadership supports the use of OSS,
8% of respondents indicated city management supports the use of OSS, and 33%
106 D.J. Ward and E.Y. Tao

Fig. 4 Cities currently using

OSS

Fig. 5 OSS plans for cities

currently not using OSS

indicated city IT staff supports use of OSS. For cities currently using OSS the
responses were 22% leadership, 33% management, and 71% IT staff.

7.3 Current OSS Support by Leadership, Management,

and IT Staff

IT managers and staff report a significant difference in the perceived current support
of OSS and the support of OSS if using OSS would reduce IT operating costs. 11%
of IT managers and staff indicate they agree their city leadership currently supports
the use of OSS. The IT managers’ and staff’s perception of city management’s
current support of OSS is similar, if somewhat higher, to their perception of city
leadership. Sixteen percent agree their city management currently supports OSS.
The IT managers’ and staff’s perception of city IT staff’s current support of OSS,
that is their perception of themselves, was significantly higher than their perception
of city leadership and management support of OSS with 26% agreeing the city IT
staff supports the use of OSS (Fig. 6).
8 Open Source Software Use in City Government 107

When asked about support of OSS if it meant saving money, 36% of the IT staff
agrees their city leadership would support OSS if it would save money, a threefold
increase. 41% agree their city management would support OSS if it would save
money, a 150% increase. However, only 36% agreed the city IT staff would support
OSS to save money, just a 50% increase. While these results indicate city leadership
and management may be motivated to support OSS given the potential cost savings,
IT staff may not share those same motivations (Fig. 7).
Of note is the drop in frequency (from 70% to 50%) of respondents indi-
cating a neutral position or those who did not know. The possibility of reducing
the costs of information technology is a significant influence on IT strategy and
technology adoption.

Fig. 6 Current support of OSS

Fig. 7 Support if OSS would save money

108 D.J. Ward and E.Y. Tao

7.4 Discrepancy of OSS Awareness: Self, Others

The survey data suggests a discrepancy between the subject’s own awareness of OSS
and their perception of city leadership, management, and IT staff’s awareness of OSS.
Within the sample set 69% of the respondents indicated they are aware of OSS.
However, their responses regarding their city leadership, management, and IT staff’s
awareness of OSS show that most respondents perceive the leadership, management
and IT staff as generally unaware of OSS. The high frequency of those individually
aware of OSS could be attributed to the survey attracting individuals interested in OSS.

8 Conclusion

The results indicate cities in general do not have the necessary characteristics to
successfully adopt OSS to deliver services and conduct city business on a compre-
hensive scale. The key indicators point to significant deficiencies in the three
domains: capability, discipline, and cultural affinity.
While a majority of cities in the study show some characteristics that indicate the
adoption of OSS is possible, and indeed on a trivial level (with a few notable
exceptions) some cities are using OSS, still most cities lack key characteristics in
the three domains to enable a successful comprehensive adoption of OSS.
The data suggest many cities may have an adequate level of discipline to support
open source adoption with IT line items in the city budget and sufficient IT funding.
However, a significant number of cities make software purchases on an ad hoc
basis, indicating potential lack of organizational planning capability.
A city’s Culture, with respect to IT decision making, appears to be a significant
barrier to open source adoption. City leadership and management of cities that do not
support the use of OSS are generally unaware of OSS as an alternative to commercial
software. Cities currently using OSS are highly likely to continue to use OSS in the
future while cities not presently using OSS have no future plans to use OSS.
Because the cities represented in this study in general do not exhibit the
indicators in the three domains examined we conclude most cities do not have the
capability, discipline, and cultural affinity to successfully adopt OSS on more than a
trivia level.

References

1. Associated Press (14 Dec 2007) Dutch Government Ditches Microsoft, Moves to Open-
Source Software. Foxnews.com. Retrieved 7 Feb 2008, from Fox News Web site: http://
www.foxnews.com/story/0,2933,316841,00.html
2. R.J. Carey, Department of the Navy Open Source Software Guidance (-) Washington, DC, U.
S. Navy. Retrieved 8 Feb 2008, from Open Source Software Institute Web site: http://oss-
institute.org/Navy/DONCIO_OSS_User_Guidance.pdf
3. V. Henderson, Medium size cities. Reg. Sci. Urban Econ. 27(6), 583–612 (1997)
8 Open Source Software Use in City Government 109

4. D. Gichoya, Factors affecting the successful implementation of ICT projects in government.

Electron. J. E Govern. 3(4), 175–184 (2005)
5. J. Mtsweni, E. Biermann, An investigation into the implementation of open source software
within the SA government: an emerging expansion model. In Proceedings of the 2008 Annual
Research Conference of the South African institute of Computer Scientists and information
Technologists on IT Research in Developing Countries: Riding the Wave of Technology
(Wilderness, South Africa, October 06–08, 2008). SAICSIT ‘08, vol. 338, ACM, New York,
pp. 148–158
6. W. Castelnovo, M. Simonetta, The evaluation of e-government projects for small local
government organisations. Electron. J. e-Government 5(1), 21–28 (2007)
7. M. V€alim€aki, V. Oksanen, J. Laine, An empirical look at the problems of open source
adoption in Finnish municipalities, in Proceedings of the 7th International Conference
on Electronic Commerce (Xi’an, China, August 15–17, 2005). ICEC ‘05, vol. 113, ACM,
New York, pp. 514–520
8. H. Thorbergsso, T. Bj€ orgvinsson, Á. Valfells, Economic benefits of free and open source
software in electronic governance, in Proceedings of the 1st International Conference on
Theory and Practice of Electronic Governance (Macao, China, December 10–13, 2007).
ICEGOV ‘07, vol. 232, ACM, New York, pp. 183–186
9. S. Comino, F. Manenti, Free/open software: public policies in the software market (July 2004)
10. Z. Ebrahim, Z Irani, E-Government adoption: architecture and barriers. Business Process
Manage. J. 11(5), 580–611 (2005)
11. M. Murphy, Organisational change and firm performance. OECD Directorate for Science,
Technology, and Industry Working Papers (2002), http:////www.oecd.org//sti
12. P. Alpar, Legal requirements and modeling or processes in government. Electron. J.
E-Govern. 3(3), 107–116 (2005)
13. D. Swedberg, Transformation by design: an innovative approach to implementation of
e-government. Electron. J. E-Govern. 1(1), 51–56 (2003)
14. J. Borras, International technical standards for e-government. Electron. J. E-Govern. 2(2),
75–80 (2004)
15. D. Kaplowitz, T. Hadlock, R. Levine, A comparison of web and mail survey response rates.
Public Opin. Quart. 68(1), 94–101 (2004)
16. Apple Inc. (2008) Software License Agreement for MAC OS X, http://manuals.info.apple.
com/en/Welcome_to_Leopard.pdf
17. L. Haber, ZDnet. (2002) City saves with Linux, thin clients, http://techupdate.zdnet.com/
techupdate/stories/main/0,14179,2860180,00.html
Chapter 9
Pheromone-Balance Driven Ant Colony
Optimization with Greedy Mechanism

Masaya Yoshikawa

Abstract Ant colony optimization (ACO), which has been based on the feeding
behavior of ants, has a powerful solution searching ability. However, since proces-
sing must be repeated many times, the computation process also requires a very
long time. In this chapter, we discuss a new ACO algorithm that incorporates
adaptive greedy mechanism to shorten the processing time. The proposed algorithm
switches two selection techniques adaptively according to generation. In addition,
the new pheromone update rules are introduced in order to control the balance of
the intensification and diversification. Experiments using benchmark data prove the
validity of the proposed algorithm.

1 Introduction

Combinatorial optimization problems can be used for a variety of fields, and

various solutions for these types of problems have been studied. Among these
solutions, a heuristic solution, referred to as meta-heuristics, has attracted much
attention in recent years. Meta-heuristics is a general term for an algorithm that is
obtained by technological modeling of biological behaviors [1–10] and physical
phenomena [11]. Using this approach, an excellent solution can be obtained within
a relatively short period of time. One form of meta-heuristics is ant colony optimi-
zation (ACO), which has been based on the feeding behavior of ants. The fundamen-
tal elements of ACO are the secretion of pheromone by the ants and its evaporation.

M. Yoshikawa
Department of information engineering, Meijo university, 1-501 Shiogamaguchi,
Tenpaku Nagoya 468-8502, Japan
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 111

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_9, # Springer ScienceþBusiness Media B.V. 2010
112 M. Yoshikawa

Using these two elements, ants can search out the shortest route from their nest to a
feeding spot. Fig. 1 shows the actual method for searching the shortest route.
Figure 1a shows the case where two ants A and B have returned from a feeding
spot to the nest via different routes. In this case, ants A and B have secreted
pheromone on their routes from the nest to the feeding spot and from the feeding
spot to the nest. At this point, a third ant C moves from the nest to the feeding spot,
relying on the pheromone trail that remains on the route.
ACO is performed based on the following three preconditions: (1) the amount of
pheromone secreted from all ants is the same, (2) the moving speed of all ants is the
same, and (3) the secreted pheromone evaporates at the same rate. In the above
case, since the moving distance from the nest to the feeding spot is longer for route
A than for route B, a larger amount of pheromone will have evaporated along route
A than along route B, as shown in Fig. 1b. Therefore, ant C will select route B
because a larger amount of pheromone remains on this route. However, ant C does
not go to the feeding spot by a route that is the same as route B; rather, ant C goes to
the feeding spot by route C and then returns to the nest, as shown in Fig. 1c. Another
ant D then goes to the feeding spot either by route B or by route C (on which a larger
amount of pheromone remains). The fundamental optimization mechanism of ACO
is to find a shorter route by repeating this process.
For this reason, ACO has a powerful solution searching ability. However, since
processing must be repeated many times, the computation process also requires a
very long time. In order to shorten the processing time, this study proposes a new
ACO algorithm that incorporates adaptive greedy selection. The validity of the
proposed algorithm is verified by performing evaluation experiments using bench-
mark data.
This chapter is organized as follows: Section 2 indicates the search mechanism
of ACO, and describes related studies. Section 3 explains the proposed algorithm
with modified pheromone update rules. Section 4 reports the results of computer
simulations applied to the travelling salesman problem (TSP) benchmark data.
We summarize and conclude this study in Section 5.

a b c

Fig. 1 Example of pheromone communication: (a) two routes on initial state, (b) positive
feedback reinforcement using pheromone information, and (c) example of another route
9 Pheromone-Balance Driven Ant Colony Optimization with Greedy Mechanism 113

2 Preliminaries

2.1 Ant Colony Optimization

ACO is a general term for the algorithm that can be obtained by technological
modeling of the feeding behavior of ants. The basic model of ACO is the ant system
(AS) [1] designed by M. Dorigo. The ant colony system (ACS) [2] is one of the
modified algorithms of the AS. The ACS provided a better solution when the TSP
was to be solved, compared to a genetic algorithm (GA) [9, 10] and a simulated
annealing (SA) [11] method. Therefore, this study adopts ACS as its basic algo-
rithm. Henceforth, ACO also denotes the ACS in this paper.
The processing procedure of ACO is explained using an example in which ACO
is applied to the TSP. In ACO, each of several ants independently creates a
travelling route (visiting every city just one time). At this time, each ant determines
the next destination according to the probability pk calculated from formula (1).

½tði; jÞ½ði; jÞb

pk ði; jÞ ¼ P (1)
½tðl; jÞ½ðl; jÞb
l2nk

Where, the value  (i,j) in formula (1) is called the static evaluation value. It is
the reciprocal of the distance between city i and city j, and is a fixed value. On the
other hand, t (i,j) is referred to as the dynamic evaluation value. It expresses the
amount of pheromone on the route between city i and city j, and this changes during
the process of optimization. The term b represents a parameter and nk represents a
set of unvisited cities.
In ACO, the probability is high for selecting a route whose distance is short and
whose pheromone amount is large. That is, selection using a roulette wheel is
performed, as shown in Fig. 2.

Fig. 2 Example of the selection probability

114 M. Yoshikawa

Here, the amount of pheromone on each partial route is determined by

two types of rules: the local update rule and the global update rule. The local
update rule is expressed by formula (2), and it is applied whenever each ant
moves.

tði; jÞ ð1  cÞtði; jÞ þ ct0 (2)

In formula (2), c is a decay parameter in local update rule, t0 is the initial value
of pheromone. The global update rule is expressed by formula (3), and it is applied
after each ant completes a travelling route.

tði; jÞ ð1  rÞtði; jÞ þ rDtði; jÞ (3)


1=Lþ if ði; jÞ 2 T þ
Dtði; jÞ ¼ (4)
0 otherwise

In formula (3), r represents the evaporation rate of pheromone in the global

update rule, Tþ represents the shortest travelling route among the completed
travelling routes, and Lþ represents the length of the partial route that composes
the shortest travelling route.
In ACO, the amount of pheromone is larger on a route that has been more
often selected, and is smaller, due to evaporation, on a route that has been less
often selected. In addition, a rule called pseudo-random-proportional rule is also
introduced. For this rule, an ant is forced to move to a city, with the selection
probability being the highest at a certain probability. In other words, a random
number q is generated between 0 and 1. If q is smaller than threshold value qt, an
ant will move to a city whose value of numerator in formula (1) is the largest. If
q is larger than the threshold value, the usual selection by formula (1) will be
performed.

2.2 Related Studies

Many studies [1–8] have been reported for the use of ACO, such as those that have
applied ACO to network routing problems [3] and to scheduling problems [4].
A number of studies [5, 6] have also been reported on pheromone, which is an
important factor for controlling the intensification and diversification of a search.
Hybridization with other algorithms, such as a GA [7] or SA [8] method, has also
been studied. However, no study has yet been reported that incorporates adaptive
greedy selection into an ACO, as is proposed in this study.
9 Pheromone-Balance Driven Ant Colony Optimization with Greedy Mechanism 115

3 Hybrid ACO with Modified Pheromone Update Rules

In ACO, when x represents the number of cities, y represents the total number
of ants, and z represents the number of processing repetitions, the number of
calculations for formulae (1), (2), and (3) are expressed by (x  1)  y  z, a
x  y  z, and z, respectively. Since the number of processing repetitions in
ACO is large, a processing time problem is inherent in ACO. In this study,
incorporating a greedy selection mechanism into the ACO reduced the number
of calculations in formulae (1) and (2). Consequently, the processing time was
shortened.
A greedy selection approach only employs static evaluation values, with no use
of dynamic evaluation values, when selecting the next destination city. That is, an
ant moves from the present city to the nearest city. Since static evaluation values are
constant in the optimization process, once the calculation is performed in the early
stage, it is not necessary to re-calculate. Therefore, the processing time can be
shortened by performing a greedy selection. However, since greedy selection favors
the local optimal solution, the balance between formula (1) and the greedy selection
becomes important.
In order to control this trade-off relationship, the greedy selection is adaptively
used in this study. The adaptive greedy selection is explained using the TSP of six
cities, as shown in Fig. 3.
In Fig. 3, when city A is set as the start city, the following three selections are
performed using formula (1): (1) movement from city A to city B; (2) movement
from city B to city C; and (3) movement from city C to city D. The greedy selection
is also applied to the other movements; namely, from city D to city E and from
city E to city F.
In contrast, when city D is set as the start city, the following three selections
are performed using the greedy selection: (1) movement from city D to city E,

Fig. 3 Example of adaptive greedy selection

116 M. Yoshikawa

(2) movement from city E to city F, and (3) movement from city F to city A.
Formula (1) is applied to the other movements; namely, from city A to city B and
from city B to city C. In this study, the proportion of the greedy selection and the
selection by formula (1) is changed according to generation, as shown in Fig. 4.
Figure 4 shows an example obtained when the total number of processing
repetitions was set at 1,000, and the proportion of the greedy selection and the
selection by formula (1) was changed every 200 repetitions. In this example, the
selection by formula (1) was performed until the 200th repetition. From the 201st
repetition to the 400th repetition, the selection by formula (1) was applied to 80%
of cities, and the greedy selection was applied to 20% of cities. In other words,
for the problem of 100 cities, the selection by formula (1) was performed on 80
cities and the greedy selection was performed on 20 cities. Similarly, from the
401st repetition to the 600th repetition, the selection by formula (1) was applied
to 60% of cities, and the greedy selection was applied to 40% of cities. Thus, by
changing the proportion of two selection methods according to generation, the
trade-off relationship between the accuracy of the solution and the processing
time can be controlled.
In optimization algorithms, controlling the trade-off relationship between the
intensification and diversification of search is also important. In the adaptive greedy
selection proposed in this study, the pressure of intensification is strong. Therefore,
in order to strengthen the pressure of diversification, the local update rule and the
global update rule are partially changed.
First, the local update rule is generally applied whenever an ant moves. How-
ever, it is not applied when the greedy selection is performed. Next, the global
update rule is generally applied to the best travelling route; i.e., the shortest
travelling route. It is also applied to the second shortest travelling route. By adding
these two modifications, the proposed algorithm can realize a balance between the
intensification and diversification of the search.

Fig. 4 Example of the

proportion of the greedy
selection and the selection by
formula (1)
9 Pheromone-Balance Driven Ant Colony Optimization with Greedy Mechanism 117

4 Experiments and Discussion

In order to verify the validity of the proposed algorithm, several comparison

experiments were performed. Each experiment was performed 10 times. Table 1
shows the conditions used in the experiments. In the table, the designation ACS/
Greedy indicates that the city selection of the first half was performed using formula
(1), and that of the second half was performed using the greedy selection. Similarly,
the designation 80/20 indicates that 80% of the cities were selected using formula
(1), and selection of 20% of the cities used the greedy selection. In each table,
technique (a) represents the conventional ACO. Techniques (b) and (c) represent
algorithms in which the proportion of the greedy selection and the selection by
formula (1) was not changed according to generation. Techniques (d), (e), (f), and
(g) represent algorithms in which the proportion was changed according to genera-
tion. These four algorithms are proposed in this study. Tables 2–4 show the
experimental results.
As shown in Tables 2–4, the processing time was shortened when the greedy
selection was used, regardless of its proportion. By changing the proportion of the
greedy selection and the selection by formula (1) according to generation, the
ability to search a solution also improved, when compared with the case where
the proportion was fixed. In order to examine the optimal proportion, changes in
solutions obtained using techniques (d), (e), (f), and (g) were plotted, as shown in
Fig. 5. In this figure, the horizontal axis represents the number of processing

Table 1 Experimental conditions

Technique Generation (# repetitions)
1–200 201–400 401–600 601–800 801–1000
(a) ACS (N/A)
(b) Greedy/ACS 50/50
(c) ACS/Greedy 50/50
(d) Greedy/ACS 0/100 20/80 40/60 60/40 80/20
(e) ACS/Greedy 100/0 80/20 60/40 40/60 20/80
(f) Greedy/ACS 80/20 60/40 40/60 20/80 0/100
(g) ACS/Greedy 20/80 40/60 60/40 80/20 100/0

Table 2 Results of qt ¼ 0.25

Technique Best Average Worst Time (s)
(a) ACS 21839 22049 22393 20.00
(b) Greedy/ACS 23196 23482 23690 14.73
(c) ACS/Greedy 22348 22599 22915 15.39
(d) Greedy/ACS 22103 22295 22619 15.64
(e) ACS/Greedy 21997 22264 22709 16.35
(f) Greedy/ACS 21967 22161 22535 15.93
(g) ACS/Greedy 21731 22109 22543 16.42
118 M. Yoshikawa

Table 3 Results of qt ¼ 0.50

Technique Best Average Worst Time (s)
(a) ACS 21700 21876 22056 18.90
(b) Greedy/ACS 23082 23490 23700 14.24
(c) ACS/Greedy 22348 22707 23122 14.85
(d) Greedy/ACS 21665 22114 22330 15.36
(e) ACS/Greedy 21866 22062 22386 15.72
(f) Greedy/ACS 21824 22026 22346 15.31
(g) ACS/Greedy 21817 22149 22721 15.81

Table 4 Results of qt ¼ 0.75

Technique Best Average Worst Time (s)
(a) ACS 21558 21768 22262 18.14
(b) Greedy/ACS 23218 23561 23773 13.80
(c) ACS/Greedy 22336 22978 23429 14.37
(d) Greedy/ACS 21787 22038 22503 14.78
(e) ACS/Greedy 21713 21888 22181 15.21
(f) Greedy/ACS 21826 21901 21977 14.73
(g) ACS/Greedy 21665 21975 22312 15.27

Fig. 5 Comparisons of
selective techniques (d), (e),
(f), and (g)

repetitions and the vertical axis represents the travelling route length of the optimal
solution.
As shown in Fig. 5, when the number of processing repetitions was small as the
termination condition, techniques (d) and (e) showed excellent performance.
When the number of processing repetitions was sufficient, techniques (f) and
(g) also exhibited excellent performance. Thus, the selection of technique accord-
ing to the given experimental condition (termination condition) was clearly
important.
9 Pheromone-Balance Driven Ant Colony Optimization with Greedy Mechanism 119

5 Conclusion

In this chapter, we proposed a new technique that incorporates a greedy mechanism

to shorten the processing time of an ACO. A hybrid technique that uses two types of
selection techniques (i.e., the conventional selection technique and the greedy
selection technique) is introduced to generate one solution. A technique was also
developed for adaptively changing these two selection techniques according to
generation. In order to control the balance of the intensification and diversification,
the update rule of pheromone was improved in each phase of global update and
local update. The validity of the proposed technique was verified by comparative
experiments using benchmark data. As a future task, the proposed technique should
be applied to practical problems other than the TSP.

References

1. M. Dorigo, V. Maniezzo, A. Colorni, Ant system: optimization by a colony of cooperating

agents. IEEE Trans. Syst. Man Cybern. Part B. 26(1), 29–41 (1996)
2. M. Dorigo, L.M. Gambardella, Ant colony system: a cooperative learning approach to the
traveling salesman problem. IEEE Trans. Evol. Comput. 1(1), 53–66 (1997)
3. T.H. Ahmed, Simulation of mobility and routing in ad hoc networks using ant colony
algorithms. Proc. Int. Conf. Inf. Technol. Coding Comput. 2, 698–703 (2005)
4. M. Yoshikawa, H. Terai, A Hybrid Ant Colony Optimization Technique for Job-Shop Scheduling
Problems, Proc. of 4th IEEE/ACIS International Conference on Software Engineering Research,
Management & Applications (SERA 2006), 95–100 (2006)
5. A. Hara, T. Ichimura, N. Fujita, T. Takahama, “Effective Diversification of Ant-Based Search
Using Colony Fission and Extinction”, Proc, of IEEE Congress on Evolutionary Computation
(CEC 2006), 1028–1035 (2006)
6. Z. Ren, Z. Feng, Z. Zhang, GSP-ANT: an efficient ant colony optimization algorithm with
multiple good solutions for pheromone update. Proc. IEEE Int. Conf. Intelligent Comput.
Intelligent Syst. 1, 589–592 (2009)
7. G. Shang, J. Xinzi, T. Kezong, “Hybrid Algorithm Combining Ant Colony Optimization
Algorithm with Genetic Algorithm”, Proc. of Chinese Control Conference (CCC 2007),
701–704 (2007)
8. Y.-H. Wang, P.-Z. Pan, A Novel Bi-Directional Convergence Ant Colony Optimization with SA
for Job-Shop Scheduling, Proc. of International Conference on Computational Intelligence and
Software Engineering (CiSE 2009), 1–4 (2009)
9. J. Holland, Adaptation in Natural Artificial Systems, 2nd edn. (The University of Michigan
Press, MIT Press, 1992)
10. D.E. Goldberg, Genetic Algorithms in Search Optimization, and Machine Learning (Addison
Wesley, Reading MA, 1989)
11. R.A. Rutenbar, Simulated annealing algorithms: an overview. IEEE Circuits Dev. Mag. 5(1),
19–26 (1989)
Chapter 10
Study of Pitchfork Bifurcation in Discrete
Hopfield Neural Network

R. Marichal, J. D. Piñeiro, E. González, and J. Torres

Abstract A simple two-neuron model of a discrete Hopfield neural network is

considered. The local stability is analyzed with the associated characteristic model.
In order to study the dynamic behavior, the Pitchfork bifurcation is examined. In the
case of two neurons, one necessary condition for yielding the Pitchfork bifurcation
is found. In addition, the stability and direction of the Pitchfork bifurcation are
determined by applying the normal form theory and the center manifold theorem.

1 Introduction

The purpose of this paper is to present some results on the analysis of the dynamics
of a discrete recurrent neural network. The particular network in which we are
interested is the Hopfield network, also known as a Discrete Hopfield Neural
Network in [1]. Its state evolution equation is

X
N X
M
xi ðk þ 1Þ ¼ win f ðxi ðkÞÞ þ w0 im um ðkÞ þ w00i (1)
n¼1 m¼1

where
xi(k) is the ith neuron output
um(k) is the mth input of the network
win, w0 im are the weight factors of the neuron outputs, network inputs and
w"i is a bias weight
N is the neuron number

R. Marichal (*)
System Engineering and Control and Computer Architecture Department, University of La
Laguna, Avda. Francisco Sánchez S/N, Edf. Informatica, 38206 Tenerife, Canary Islands, Spain
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 121

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_10, # Springer ScienceþBusiness Media B.V. 2010
122 R. Marichal et al.

M is the input number

f(·) is a continuous, bounded, monotonically increasing function, such as the
hyperbolic tangent
This model has the same dynamic behavior as the Williams–Zipser neural
network. The relationship between the Williams–Zipser states and Hopfield states is

Xh ¼ WXWZ

where
Xh are Hopfield states
XWZ are the Williams-Zipser states
W is the weight matrix without the bias and input weight factor
We will consider the WilliamsZipser model in order to simplify the mathe-
matical calculations.
The neural network presents different classes of equivalent dynamics. A system
will be equivalent to another if its trajectories exhibit the same qualitative behavior.
This is made mathematically precise in the definition of topological equivalence [2].
The simplest trajectories are those that are equilibrium or fixed points that do not
change in time. Their character or stability is given by the local behavior of nearby
trajectories. A fixed point can attract (sink), repel (source) or have directions of
attraction and repulsion (saddle) of close trajectories [3]. Next in complexity are
periodic trajectories, quasi-periodic trajectories or even chaotic sets, each with its
own stability characterization. All of these features are similar in a class of
topologically equivalent systems. When a system parameter is varied, the system
can reach a critical point at which it is no longer equivalent. This is called a
bifurcation, and the system will exhibit new behavior. The study of how these
changes can be carried out will be another powerful tool in the analysis.
With respect to discrete recurrent neural networks as systems, several results on
their dynamics are available in the literature. The most general result is derived
using the Lyapunov stability theorem in [4], and establishes that for a symmetric
weight matrix, there are only fixed points and period two limit cycles, such as
stable equilibrium states. It also gives the conditions under which only fixed-point
attractors exist. More recently Cao [5] proposed other, less restrictive and more
complex, conditions. In [6], chaos is found even in a simple two-neuron network
in a specific weight configuration by demonstrating its equivalence with a one-
dimension chaotic system (the logistic map). In [7], the same author describes
another interesting type of trajectory, the quasi-periodic orbits. These are closed
orbits with irrational periods that appear in complex phenomena, such as frequency-
locking and synchronization, which are typical of biological networks. In the same
paper, conditions for the stability of these orbits are given. These can be simplified,
as we shall show below.
Passeman [8] obtains some experimental results, such as the coexisting of the
periodic, quasi-periodic and chaotic attractors. Additionally, [9] gives the position,
number and stability types of fixed points of a two-neuron discrete recurrent
network with nonzero weights.
10 Study of Pitchfork Bifurcation in Discrete Hopfield Neural Network 123

There are some works that analyze the Hopfield continuous neural networks
[10, 11], like [12–15]. These papers show the stability of Hopf-bifurcation with two
delays.
Firstly, we analyze the number and stability-type characterization of the fixed
points. We then continue with an analysis of the Pitchfork bifurcation. Finally, the
simulations are shown and conclusions are given.

2 Determination of Fixed Points

For the sake of simplicity, we studied the two-neuron network. This allows for an
easy visualization of the problem. In this model, we considered zero inputs so as to
isolate the dynamics from the input action. Secondly, and without loss of generality
with respect to dynamics, we used zero bias weights. The activation function is the
hyperbolic tangent.
With these conditions, the network mapping function is

x1 ðk þ 1Þ ¼ tanhðw11 x1 ðkÞ þ w12 x2 ðkÞÞ

x2 ðk þ 1Þ ¼ tanhðw21 x1 ðkÞ þ w22 x2 ðkÞÞ (2)

where x(k) and y(k) are the neural output of step k.

The fixed points are solutions of the following equations

x1;p ¼ tanhðw11 x1;p þ w12 x2;p Þ

x2;p ¼ tanhðw21 x1;p þ w22 x2;p Þ (3)

The point (0, 0) is always a fixed point for every value of the weights. The
number of fixed points is odd because for every fixed point (x1,p, x2,p), (x1,p, x2,p)
is also a fixed point.
To graphically determine the configuration of fixed points, we redefine the above
equations as

a tanhðx1;p Þ  w11 x1;p

x2;p ¼ ¼ Fðx1;p ; w11 ; w12 Þ
w12
a tanhðx2;p Þ  w22 x2;p
x1;p ¼ ¼ Fðx2;p ; w22 ; w21 Þ (4)
w21

Depending on the diagonal weights, there are two qualitative behavior functions.
We are going to determine the number of fixed points using the graphical represen-
tation of the above Eq. (4). First, we can show that the graph of the F function has a
maximum and a minimum if wii > 1 or, if the opposite condition holds, is like the
hyperbolic arctangent function (Fig. 1).
124 R. Marichal et al.

Fig. 1 The two possible behaviors of the F function. The left figure corresponds to the respective
diagonal weight lower than unity. The right shows the opposite condition

The combination of these two possibilities with another condition on the ratio of
slopes at the origin of the two curves (4) gives the number of fixed points. The latter
condition can be expressed as

jW j ¼ w11 þ w22  1

where |W| is the weight matrix determinant.

If w11 > 1, w22 > 1 and |W| > w11 + w22  1, then there can exist nine, seven or
five fixed points. When this condition fails, there are three fixed points.
When a diagonal weight is less than one, there can be three or one fixed points.

3 Local Stability Analysis

In the process below, a two-neuron neural network is considered. It is usual for the
activation function to be a sigmoid function or a tangent hyperbolic function.
Considering the fixed point Eq. (3), the elements of the Jacobian matrix at the
fixed point (x, y) are
 
w11 f 0 ðx1 Þ w12 f 0 ðx1 Þ
J¼
w21 f 0 ðx2 Þ w22 f 0 ðx2 Þ

The associated characteristic equation of the linearized system evaluated at the

fixed point is

l2  ½w11 f 0 ðx1 Þ þ w22 f 0 ðx2 Þl þ jW j f 0 ðx1 Þ f 0 ðx2 Þ ¼ 0 (5)

where w11, w22 and |W| are the diagonal elements and the determinant of the matrix
weight, respectively.
We can define new variables

w11 f 0 ðx1 Þ þ w22 f 0 ðx2 Þ

s1 ¼
2
10 Study of Pitchfork Bifurcation in Discrete Hopfield Neural Network 125

σ2

SOURCE ZONE

SINK ZONE
σ1

Pitchfork bifurcation
|W| = w11 + w22 –1

SADDLE ZONE SADDLE ZONE

SOURCE ZONE

Fig. 2 The stability regions and the Pitchfork bifurcation line at the fixed point (0, 0)

s2 ¼ jW j f 0 ðx1 Þ f 0 ðx2 Þ

The eigenvalues of the characteristic Eq. (5) are defined as

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
l ¼ s 1  s21  s2

The Pitchfork bifurcation appears when two complex conjugate eigenvalues

reach the unit circle. It is easy to show that this limit condition is.
The boundaries between the regions shown in Fig. 2 are the bifurcations.
At these limit zones the fixed point changes its character. The Pitchfork bifurcation
is represented by the line |W| ¼ w11 + w22  1 in Fig. 2.

4 Pitchfork Bifurcation Direction

In order to determine the direction and stability of the Pitchfork bifurcation, it is

necessary to use the center manifold theory [2]. The center manifold theory
demonstrates that the mapping behavior in the bifurcation is equivalent to the
complex mapping below:
126 R. Marichal et al.

uðk þ 1Þ ¼ uðkÞ þ cð0ÞuðkÞ3 þ oðuðkÞ4 Þ (6)

The parameter c(0) is [2]

1
cð0Þ ¼ hp; Cðq; q; qÞi (7)
6

where C is the third derivative terms of the Taylor development, the notation
< . , . > represents the scalar product, and p, q are the eigenvector Jacobian
matrix and its transpose, respectively. These vectors satisfy the normalization
condition

hp; qi ¼ 1

The above coefficients are evaluated for the critical parameter of the system
where the bifurcation takes place. The c(0) sign determines the bifurcation direc-
tion. When c(0) is negative, a stable fixed point becomes an unstable fixed point and
two additional stable symmetrical fixed points appear. In the opposite case, c(0)
positive, an unstable fixed point becomes a stable fixed point and two additional
unstable symmetrical fixed points appear.
In the neural network mapping, p and q are
 
d e
q¼ ; 1 (8)
e þ d w21 X2;0

 
e
p¼ ; 1 (9)
w12 X1;0

where

d ¼ w11 X1;0  1

e ¼ w22 X2;0  1

X1;0 ¼ 1  x21;0

X2;0 ¼ 1  x22;0

x1,0 and x2.0 are the fixed point coordinates where the bifurcation appears.
10 Study of Pitchfork Bifurcation in Discrete Hopfield Neural Network 127

The Taylor development term is

X
2
@fi
Ci ðq; q; qÞ ¼ qj qk ql
j;k;l¼1
@x j @x k @xl

X
2
¼ f 000 ð0Þ dik dil wij wik wil qj qk ql
j;k;l¼1

X
2
¼ f 000 ð0Þ w3ij q2j qj (10)
j¼1

where dij is the Kronecker delta.

In order to determine the parameter c(0), it is necessary to calculate the third
derivate of the mapping (1) given by Eq. (10)

@fi
¼ 2ð1  x2i Þð3x2i  1Þwij wik wil
@xj @xk @xl

The Taylor development term is then

X
2
Ci ða; b; cÞ ¼ 2 ð1  x2i Þð3x2i  1Þwij wik wil aj bk cl :
j;k;l¼1

Taking into account the previous equations and the q autovector Eq. (8)
2 3
2X1;0 ð3x21:0 1Þw312

Cðq; q; qÞ ¼ 2 4 d3
ð13x22;0 Þ
5 (11)
2
X2;0

It can be shown in Eq. (3) that the zero is always a fixed point. Replacing the
expressions for C(q,q,q), q and p given by Eqs. (8) (9) and (11), respectively, and
evaluating them at the zero fixed point, the previous c(0) coefficient is

1 w2 ðw22  1Þ þ ðw11  1Þ3

cð0Þ ¼ h p; Cðq; q; qÞi ¼ 12
6 3ð1  w11 Þ2 ð2  w11  w22 Þ

The previous expression is not defined for the following cases

a. w11 ¼ 1
b. w11 + w22 ¼ 2.
In this paper we only consider condition (a) since condition (b) shows the
presence of another bifurcation, known as Neimark-Sacker and which is analyzed
in another paper [16].
128 R. Marichal et al.

Taking into account condition a) and the bifurcation parameter equation

jW j ¼ w11 þ w22  1

then

w12 w21 ¼ 0

In this particular case, the eigenvalues match with the diagonal elements of the
weight matrix

l1 ¼ w11
l2 ¼ w22

The new q and p eigenvectors are given by

q ¼f1; 0g
 
w12
p ¼ 1; 
w22  1

The c(0) coefficient is

1 1 1
cð0Þ ¼ hp; Cðq; q; qÞi ¼  w311 ¼  :
6 3 3
Therefore, in this particular case, the coefficient of the normal form c (0) is
negative, a stable fixed point becomes a saddle fixed point and two additional stable
symmetrical fixed points appear.

5 Simulations

In order to examine the results obtained. The simulation shows the Pitchfork
bifurcation (Fig. 3). The Pitchfork bifurcation is produced by the diagonal element
weight matrix w11. Figure 3a shows the dynamic configuration before the bifurcation
is produced, with only one stable fixed point. Subsequently, when the bifurcation is
produced (Fig. 3b), two additional stable fixed points appear and the zero fixed point
changes its stability from stable to unstable (the normal form coefficient c is negative).

6 Conclusion

In this paper we considered the Hopfield discrete two-neuron network model. We

discussed the number of fixed points and the type of stability. We showed the bifurcation
Pitchfork direction and the dynamical behavior associated with the bifurcation.
The two-neuron networks discussed above are quite simple, but they are poten-
tially useful since the complexity found in these simple cases might be carried
10 Study of Pitchfork Bifurcation in Discrete Hopfield Neural Network 129

a
1

0.8

0.6

0.4

0.2

X2 0

–0.2

–0.4

–0.6

–0.8

–1
–1 – 0.5 0 0.5 1
X1

b
1

0.8

0.6

0.4

0.2

0
X2
–0.2

–0.4

–0.6

–0.8

–1
–1 – 0.5 0 0.5 1
X1

Fig. 3 The dynamic behavior when the Pitchfork bifurcation is produced. + and D are the saddle
and source fixed points, respectively. (a) w11 ¼ 0.9, w12 ¼ 0.1, w21 ¼ 1 and w22 ¼ 0.5; (b) w11 ¼
1.1, w12 ¼ 0.1, w21 ¼ 1 and w22 ¼ 0.5

over to larger Hopfield discrete neural networks. There exists the possibility of
generalizing some of these results to higher dimensions and of using them to design
training algorithms that avoid the problems associated with the learning process.
130 R. Marichal et al.

References

1. R. Hush, B.G. Horne, Progress in supervised neural network. IEEE Signal Process. Mag. 8–39
(1993)
2. Y.A. Kuznetsov, Elements of Applied Bifurcation Theory, 3rd edn. (Springer-Verlag,
New York, 2004)
3. K.T. Alligood, T.D. Sauer, J.A. Yorke, Chaos: An Introduction to Dynamical Systems
(Springer Verlag, New York, 1998)
4. C.M. Marcus, R.M. Westervelt, Dynamics of iterated-map neural networks. Phys. Rev. A 40,
501–504 (1989)
5. Cao Jine, On stability of delayed cellular neural networks. Phys. Lett. A 261(5–6), 303–308
(1999)
6. X. Wang, Period-doublings to chaos in a simple neural network: an analytical proof. Complex
Syst. 5, 425–441 (1991)
7. X. Wang, Discrete-time dynamics of coupled quasi-periodic and chaotic neural network
oscillators. International Joint Conference on Neural Networks (1992)
8. F. Pasemann, Complex dynamics and the structure of small neural networks. Netw. Comput.
Neural Syst. 13(2), 195–216 (2002)
9. Tino et al., Attractive periodic sets in discrete-time recurrent networks (with emphasis on
fixed-point stability and bifurcations in two-neuron networks). Neural Comput. 13(6),
1379–1414 (2001)
10. J. Hopfield, Neurons with graded response have collective computational properties like those
of two-state neurons. Procs. Nat. Acad. Sci. USA 81, 3088–3092 (1984)
11. D.W. Tank, J.J. Hopfield, Neural computation by concentrating information in time, Proc.
Nat. Acad. Sci. USA, 84, 1896–1991 (1984)
12. X. Liao, K. Wong, Z. Wu, Bifurcation analysis on a two-neuron system with distributed
delays. Physica D 149, 123–141 (2001)
13. S. Guo, X. Tang, L. Huang, Hopf bifurcating periodic orbits in a ring of neurons with delays,
Physica D, 183, 19–44 (2003)
14. S. Guo, X. Tang, L. Huang, Stability and bifurcation in a discrete system of two neurons with
delays. Nonlinear Anal. Real World. doi:10.1016/j.nonrwa.2007.03.002 (2007)
15. S. Guo, X. Tang, L. Huang, Bifurcation analysis in a discrete-time single-directional network
with delays. Neurocomputing 71, 1422–1435 (2008)
16. R. Marichal et al., Bifurcation analysis on hopfield discrete neural networks. WSEAS Trans.
Syst. 5, 119–124 (2006)
Chapter 11
Grammatical Evolution and STE Criterion
Statistical Properties of STE Objective Function

Radomil Matousek and Josef Bednar

Abstract Grammatical evolution (GE) is one of the newest among computational

methods (Ryan et al. 1998; O’Neill and Ryan 2001). Basically, it is a tool used to
automatically generate Backus-Naur-Form (BNF) computer programmes. The
method’s evolution mechanism may be based on a standard genetic algorithm
(GA). GE is very often used to solve the problem of a symbolic regression,
determining a module’s own parameters (as it is also the case of other optimization
problems) as well as the module structure itself. A Sum Square Error (SSE) method
is usually used as the testing criterion. In this paper, however, we will present the
original method, which uses a Sum epsilon Tube Error (STE) optimizing criterion.
In addition, we will draw a possible parallel between the SSE and STE criteria
describing the statistical properties of this new and promising minimizing method.

1 Introduction

The general problem of optimizing a nonlinear model or a specific problem of

nonlinear regression in statistics may be seen as typical problems that can success-
fully be approached using evolutional optimization techniques. A genetic algorithm
method, for example, may be very efficient in determining the parameters of
models representing a multimodal or non-differentiable behaviour of the goal
function. On the other hand, there are numerous mathematical methods that may,
under certain conditions, be successful in looking for a minimum or maximum. The
classical optimization methods may differ by the type of the problem to be solved
(linear, nonlinear, integer programming, etc.) or by the actual algorithm used.
However, in all such cases, the model structure for which optimal parameters are

R. Matousek (*)
Dept. of Mathematic and Applied Computer Science, Brno University of Technology, FME,
Technicka 2896/2, 61600 Brno, Czech Republic
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 131

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_11, # Springer ScienceþBusiness Media B.V. 2010
132 R. Matousek and J. Bednar

Table 1 Grammar G which was used with respect of the approximation tasks
Approximation task Grammar and example of generating function
Trigonometric (data: ET20x50) G ¼ {þ,,*,/,sin,cos,variables,constants}
notes:
unsigned integer constants ∈ [0,15], real constants ∈ [0,1]
y ¼ sinðxÞ þ sinð2:5xÞ
Polynomial (data: ET10x50) G ¼ {þ,,*,/, variables, constants}
notes:
unsigned integer constants ∈ [0,15], real constants ∈ [0,1]
y ¼ 3x4  3x þ 1

to be found is already known. But it is exactly the process of an adequate model

structure design that may be of key importance in statistical regression and numerical
approximation.
When a problem of regression is dealt with, it is usually split into two parts. First,
a model is designed to fit the character of the empirically determined data as much
as possible. Then its optimum parameters are computed for the whole to be
statistically adequate.
In the event of a pure approximation problem, the situation is the same: a model
is designed (that is, its structure such as a polynomial P(x) ¼ ax2 þ bx þ c) with its
parameters being subsequently optimized (such as a, b, c) using a previously chosen
minimization criterion, which is usually a Sum Square Error (SSE) one.
GE uses such data representation as to deal with the above problem in a
comprehensive way, that is, combining the determination of a model structure
and finding its optimum parameter in a single process. Such a process is usually
called the problem of a symbolic regression. The GE optimisation algorithm usually
uses a GA with the genotype-phenotype interpretation using a context-free Backus-
Naur-Form (BNF) based grammar [1, 2]. The problem solved is then represented by
a BNF structure, which a priori determines its scope – an automatic computer
program generating tool. Further we will focus on issues of symbolic regression of
trigonometric and polynomial data, that is, data that can be described using a
grammar containing trigonometric functions and/or powers.
The advanced GE algorithm implemented was using a binary GA. The GA
implemented the following operators: tournament selection, one- and two-point
structural crossing-over, and structural mutations. Numerous modern methods of
chromosome encryption and decryption were used [3, 4]. The grammar G was used
with respect to the class to be tested (Table 1).

2 STE – Sum Epsilon Tube Error

To keep further description as simple as possible, consider a data approximation

problem ignoring the influence of a random variable, that is, disregarding any errors
of measurement. The least-square method is used to test the quality of a solution in
11 Grammatical Evolution and STE Criterion 133

an absolute majority of cases. Generally, the least-squares method is used to find

a solution with the squared errors r summed over all control points being minimal.
A control point will denote a point at which the approximation error is calculated
using the data to be fitted, that is, the difference between the actual y value and the
approximation y^, (r¼y^ y).
X
SSE ¼ ri2 (1)
i

However, in the event of a GE problem of symbolic regression, this criterion

may not be sufficient. For this reason a new, Sum Epsilon-Tube Error (STE)
evaluation criterion has been devised. For each value e, this evaluation criterion
may be calculated as follows:
(
X 0 ri 2
= ½e; e
STEe ¼ ee ðri Þ; ee ðri Þ ¼ (2)
i 1 ri 2 ½e; e

where e is the radius of the epsilon tube, i is the index set of the control points,
e is an objective function determining whether the approximated value lies within
the given epsilon tube. The workings of this criterion may be seen as a tube of
diameter epsilon that stretches along the entire approximation domain (Fig. 2). The
axis of such a tube is determined by the points corresponding to the approximated
values. Such points may be called control points.
The actual evaluating function then checks whether an approximation point lies
within or out of the tube (2). The epsilon value changes dynamically during the
optimization process being set to the highest value at the beginning and reduced
when the adaptation condition is met.
The variable parameters of the STE method are listed in Table 2. At the
beginning of the evolution process, the algorithm sets the diameter of the epsilon
tube. If the condition (3), which indicates the minimum number of points that need
to be contained in the epsilon cube, is met, the epsilon tube is adapted, that is, the

Table 2 Parameters of the STE minimization method

Parameter Value Description
steCP Unsigned integer The number of control points
steEpsStart Positive real number Epsilon initial value (e-upper bound)
steEpsStop Positive real number Final epsilon value (e-lower bound)
steEpsReduction Unsigned integer Percentage of value decrease on adapting the value
steAdaptation Unsigned integer Percentage indicating the success rate (control points
are in the e tube) at which epsilon is adapted
steModel “Interest-Rate or Model of the computation of a new epsilon value if
Linear Model” adaptation condition is met
steIteration Unsigned integer This may replace the steEpsStop parameter, and then
this parameter indicates the number of adaptations
until the end of the computation
134 R. Matousek and J. Bednar

current epsilon value is reduced. This value can either be reduced using an interest-
rate model with the current epsilon value being reduced by a given percentage of
the previous epsilon value or using a linear model with the current epsilon value
being reduced by a constant value.
The condition for the adaptation of the epsilon tube to be finished may either
be chosen as a limit epsilon value (the steEpsStop parameter) or be given by the
number of adaptive changes (the steIteration parameter).

ðsteCP=STEe Þ  100  steAdaptation (3)

3 STE – Empirical Properties

Practical implementations of the symbolic regression problem use STE as an

evaluating criterion. The residua obtained by applying STE to the approximation
problems implemented have been statistically analyzed. It has been found out that
these residua are governed by the Laplace distribution (4) with m ¼ 0 and b ¼ 1/sqrt
(2) in our particular case. The residua was standardised for requirement of our
experiments (the data are standardized by subtracting the mean and dividing by the
standard deviation).
8  
  > exp  m  x
< if x < m
1 jx  m j 1 b 
f ðxjm; bÞ ¼ exp  ¼  x m (4)
2b b 2b >
: exp  if x  m
b

A random variable has a Laplace (m, b) distribution if its probability density function
is (4). Here, m is a location parameter and b > 0 is a scale parameter. If m ¼ 0 and
b ¼ 1, the positive half-line is exactly an exponential distribution scaled by ½.
An intuitive approach to approximation shows better results for the STE method
(Fig. 1) mainly because there are no big differences between the approximating
functions (Figs. 3 and 4). The reason for this is obvious: very distant points
influence the result of approximation a great deal for SSE while, for STE, the
weight of such points are relatively insignificant.
Next, the main advantages and disadvantages are summarized of the minimiza-
tion methods using the SSE and STE criteria.

3.1 SSE (Advantages, Disadvantages)

þ a classical and well-researched method both in statistical regression and numeri-

cal approximation.
11 Grammatical Evolution and STE Criterion 135

Fig. 1 Comparison of SSE and STE principles. Common principle of residual sum of squares
(above). The principle of the Sum of epsilon-Tube Error (below)

Histogram for ET10 × 50

300 Distribution
Laplace
250

200
frequency

150

100

50

0
–5 –3 –1 1 3 5
ET10 × 50

Fig. 2 Histogram of the standardized residua (from the experiment ET10x50) and the probability
density function of the Laplace distribution
136 R. Matousek and J. Bednar

Fig. 3 The final approximations (symbolic regression) of polynomial data (see Table 1). STE
method was used (left), respectively SEE method was used (right)

Fig. 4 The final approximations (symbolic regression) of trigonometric data (Table 1). STE
method was used (left), respectively SEE method was used (right)

þ a metric is defined.
þ a N(m ¼ 0, s2) residua error distribution may be assumed with descriptive
statistics for this distribution being available.
 more time-consuming.
 being used excessively, this method hardly provides an incentive for users to gain
new insights [5].

3.2 STE (Advantages, Disadvantages)

þ less time consuming.

þ more intuitive and readable.
þ when using a GE-based symbolic regression, the Laplace distribution may be
assumed (which is one of the original results of this paper).
þ better results are generated for the problem class.
 mathematical processing is worse with no metric defined.
 the method being new, it is not so well described.
11 Grammatical Evolution and STE Criterion 137

4 Probabilistic Mapping of SSE to STE

This part shows how the results obtained by a SSE-based minimization method may
be used to derive a result corresponding to the values of a STE-based minimization
criterion. In other words, a procedure will be shown for using SSE to obtain a
corresponding STE. The procedure is based on the following:
l It is assumed that, when applying the least-squares method, the residua are
normally distributed with N(m ¼ 0, s2). Here, the standard deviation may be
arbitrary, but a reduction to the standardized normal distribution N(0, 1) is
possible.
l In order to reach a sufficient statistical significance, the experiment simulated
10,000 instances of solution to an approximation problem (nRUN). This prob-
lem was discretized using 100 control points.
l Different sizes eps ¼ {2, 1, 0.5, 0.25, 0.125} of epsilon tubes were chosen to
simulate possible adaptations of the epsilon tube by (3).
l The frequencies were calculated of the values lying within the given epsilon
tubes.
This simulation and the subsequent statistical analysis were implemented using
the Minitab software package. Table 3 provides partial information on the proce-
dure used.
Where r is residua (X  N(0,1)), i is control point index (i ∈ [1, 100] in our
particular case), nRUN is given implementation (nRUN ∈ [1, 10000] in our
particular case).
Next, for each residuum value, the number of cases will be determined of these
residua lying in the interval given by a particular epsilon tube. This evaluation
criterion denoted by STE (Sum epsilon-Tube Error) is defined by (2) and, for the
particular test values, corresponds to:

X
100
STEe ¼ ee ðri Þ; e ¼ f2; 1; 0:5; 0:25; 0:125g (5)
i¼1

By a statistical analysis, it has been determined that STE behaves as a random

variable with a Binomial distribution

Table 3 A fragment of a table of the simulated residua values with standardized normal
distribution and the corresponding ee ðri Þ values
r i nRUN e¼2 e¼1 e ¼ 0.5 e ¼ 0.25 e ¼ 0.125
0.48593 1 1 1 1 1 0 0
0.07727 2 1 1 1 1 1 1
1.84247 3 1 1 0 0 0 0
1.29301 4 1 1 0 0 0 0
0.34326 5 1 1 1 1 0 0
... ... ... ... ... ... ... ...
0.02298 100 10,000 1 1 1 1 1
138 R. Matousek and J. Bednar

Table 4 Table of probabilities ps

e PðNð0; 1Þ<  eÞ PðNð0; 1Þ<eÞ pe
2 0.022750 0.977250 0.954500
1 0.158655 0.841345 0.682689
0.5 0.308538 0.691462 0.382925
0.25 0.401294 0.598706 0.197413
0.125 0.450262 0.549738 0.099476

Distribution Plot
Normal; Mean = 0; StDev = 1
0.5
0.954
0.4

0.3
Density

0.2

0.1

0.0
–2 0 2
X

Fig. 5 Calculating probability pe with e ¼ 2

KBiðn; pe Þ; (6)

where

pe ¼ PðX  Nð0; 1Þ 2 ½e; eÞ

For the epsilons chosen, the STE characteristic has a binomial distribution given
by (6) where the probabilities pe are calculated from Table 4. For the case of e ¼ 2,
Fig. 5 shows the probability calculation.
The Fig. 6 shows the probability functions for epsilon tubes with value k
indicating, with a given probability, the number of control points for which the
e(r) function takes on a value of one, that is, the number of control points contained
in a given epsilon tube.
The correctness of the calculation was proved empirically. The STE values
simulated (10,000 runs) were compared with distribution (6). The following figure
(Fig. 7) displays an empirical probability density function obtained from Table 3 by
(5) compared with the probability density function as calculated by (6) for e ¼ 1.
It can be seen in Fig. 7 that the empirical distribution is close the calculated one,
which was verified using a goodness-of-fit test.
11 Grammatical Evolution and STE Criterion 139

0.20
Variable
Bi (100, 0.9545)
Bi (100, 0.6827)
0.15
Bi (100, 0.3829)
Bi (100, 0.1974)
Bi (100, 0.0995)
p(k)

0.10

0.05

0.00

0 20 40 60 80 100
k

Fig. 6 STE probability density functions for e-tubes with e ¼ {2, 1, 0.5, 0.25, 0.125}

Bi(100, 0.6826895) vs Empirical Probabilities

0.09
Variable
0.08 Bi (100, 0.6826895)
0.07 Empirical Prob.

0.06
0.05
p(k)

0.04
0.03
0.02
0.01
0.00

50 60 70 80 90
k

Fig. 7 Comparing empirical and calculate probability density functions of occurrences within the
epsilon tube for e ¼ 1

5 Goodness-of-Fit Tests of Data Sets

Residua which we obtained from our symbolic regression experiments were

analyzed for Laplace distribution coincidence. We have tested many residua data
for this distribution. The examples of specific results which confirmed our hypothesis
about Laplace distribution follows. There we denoted data sets ET10x50 (500
values ranging from 5.12124 to 3.27653) for polynomial problem and ET20x50
140 R. Matousek and J. Bednar

(1000 values ranging from 5.12124 to 5.11919) for goniometrical problem which is
described by Table 1.

5.1 Uncensored Data – ET10x50

Fitted distribution is Laplace, where mean is 0.00198 and scale is 1.55075. This
analysis shows the results of fitting a Laplace distribution to the data on ET10x50
(Tables 5 and 6). We can test whether the Laplace distribution fits the data
adequately by selecting Goodness-of-Fit Tests from the list of Tabular Options.
The visual fiiting of Laplace distribution is in the Fig. 2.
This pane shows the results of tests run to determine whether ET10x50 can be
adequately modeled by a Laplace distribution. The chi-squared test divides the
range of ET10x50 into nonoverlapping intervals and compares the number of
observations in each class to the number expected based on the fitted distribution.
The Kolmogorov-Smirnov test computes the maximum distance between the
cumulative distribution of ET10x50 and the CDF of the fitted Laplace distribution.
In this case, the maximum distance is 0.0540167. The other statistics compare the
empirical distribution function to the fitted CDF in different ways.
Since the smallest P-value amongst the tests performed is greater than or equal to
0.05, we can not reject the idea that ET10x50 comes from a Laplace distribution
with 95% confidence. Analogous procedure can be use in case of ET20x50 data sets
and another.

Table 5 Table of goodness-of-fit tests for ET10x50: *chi-squared test

Lower limit Upper limit Observed frequency Expected frequency Chi-squared
At or below 3.0 2 2.39 0.06
3.0 2.33333 1 4.33 2.57
2.33333 1.66667 15 12.19 0.65
1.66667 1.0 39 34.27 0.65
1.0 0.333333 78 96.36 3.50
0.333333 0.333333 217 201.82 1.14
0.333333 1.0 85 95.77 1.21
1.0 1.66667 37 34.06 0.25
1.66667 2.33333 18 12.11 2.86
2.33333 3.0 5 4.31 0.11
above 3.0 3 2.38 0.16
*Chi-squared ¼ 13.1719 with 8 d.f. P-Value ¼ 0.10607

Table 6 Table of goodness-of-fit tests for ET10x50: *Kolmogorov-Smirnov test

Laplace Laplace
DPLUS 0.0397308 DN 0.0540167
DMINUS 0.0540167 *P-Value 0.108113
11 Grammatical Evolution and STE Criterion 141

6 Probabilistic Relationship Between STE and SSE

Fundamental difference between STE and SSE optimal criteria is metric regarding
Eqs. (1) and (2). The SSE contrary to the STE has define the Euclidean metric (in this
case triangle inequality exist). Hence, exact relation form STE to SSE optimal criteria
doesn’t exist. But, if we try to use probabilistic relation, we can obtain relationship
for transformation of Laplace distribution to Normal distribution. In other words, we
can transform STE solution to SSE solution in probabilistic point of view.
We suppose (from caption IV and on the base chi-square goodness test) that
the residua has Laplace distribution in case of STE criteria. As we can see in 0,
Laplace distribution is a typical distribution for some numerical methods. Laplace
distribution is also called double exponential distribution. It is the distribution
of differences between two independent varieties with identical exponential distri-
bution 0. The probability is given by (7)
1 jxmj=b
PðxÞ ¼ e (7)
2b
and the moments about the mean mn are related to the moments about 0 by (8)
! ! (
X X
n floor ðj=2Þ
nj n j 2k n2k n!bn for n even
mn ¼ ð1Þ b m Gð2k þ 1Þ ¼ (8)
j¼0 k¼0 j 2k 0 for n odd

Common uses moments are in the Table 7. Because the Mean Square Error
(MSE) given by MSE ¼ SSE/n is estimation of the variance s2 we can derive from
Table 7 and given parameter b the Sum Square Error by Eq. (9).

SSE ¼ s2 n ¼ 2b2 n (9)

7 Conclusion

l A new STE evaluation criterion for GE was introduced using epsilon tubes.
Based on the experiments conducted, this criterion seems to be more appropriate
for the symbolic regression problem as compared with the SSE method.

Table 7 Common uses moments in normal and Laplacian probabilistic distributions

Relationship Description
m ¼ m Mean
s2 ¼ 2b2 Variance
g1 ¼ 0 Skewness
g2 ¼ 3 Kurtosis
142 R. Matousek and J. Bednar

l Using statistical tools the method of transforming the results obtained by SSE
minimisation into results obtained by STE has been shown.
l Using a particular class implemented using approximations obtained by an STE
minimization, the empirically obtained residua were found to be Laplace
distributed. It was confirmed on the base statistical approach of goodness-of-fit
tests.
l The properties of such residua and transformation of the STE-based minimi-
zation criterion into a SSE-based was shown by means of probabilistic
relationship.

Acknowledgement This work was supported by the Czech Ministry of Education in the frame
of MSM 0021630529 and by the GACR No.: 102/091668.

References

1. M. O’Neill, C. Ryan, Grammatical Evolution: Evolutionary Automatic Programming in

an Arbitrary Language (Kluwer, London, USA, 2003), ISBN 1-4020-7444-1
2. C. Ryan, M. O’Neill, J.J. Collins, Grammatical evolution: solving trigonometric identities, in
Proceedings of MENDEL ‘98 (Brno, Czech Republic, 1998), pp. 111–119, ISBN 80-214-1199-6
3. P. Osmera, R. Matousek, O. Popelka, T. Panacek, Parallel grammatical evolution for circuit
optimization, in Proceedings of MENDEL 2008 (Brno, Czech Republic, 2008), pp. 206–213,
ISBN 978-80-214-3675-6
4. R. Matousek, GAHC: hybrid genetic algorithm, in The Springer Book Series, eds. by S.L. Ao,
B. Rieger, S.S. Chen. Lecture Notes in Electrical Engineering: Advances in Computational
Algorithms and Data Analysis, vol 14 (Springer, The Netherlands 2008), pp. 549–562, ISSN
1876-1100, ISBN 978-1-4020-8918-3
5. P. Popela, Numerical techniques and available software, in Chapter 8 in Part II, Stochastic
Modeling in Economics and Finance, eds J. Dupacova, J. Hurt, J. St_ep_an (Kluwer Academic
Publishers) pp.206–227
6. M. Abramowitz, I.A. Stegun (eds), Handbook of Mathematical Functions with Formulas,
Graphs, and Mathematical Tables, 9th printing (Dover, New York, 1972)
7. E.W. Weisstein, Laplace Distribution. From MathWorld – A Wolfram. http://mathworld.
wolfram.com/LaplaceDistribution.html
8. W.H. Pressm, S.A. Teukolsky, Numerical Recipes, The Art of Scientific Computing, 3rd edn.
(Cambridge University Press, 2007) ISBN 978-0521880688
9. J. Roupec, GA-based parameters tuning in grammatical evolution, in Proceedings of MENDEL
2007 (Prague, Czech Republic, 2007), pp. 66–71, ISBN 978-80-214-3473-8
Chapter 12
Data Quality in ANFIS Based Soft Sensors

S. Jassar, Z. Liao, and L. Zhao

Abstract Soft sensor are used to infer the critical process variables that are
otherwise difficult, if not impossible, to measure in broad range of engineering
fields. Adaptive Neuro-Fuzzy Inference System (ANFIS) has been employed to
develop successful ANFIS based inferential model that represents the dynamics
of the targeted system. In addition to the structure of the model, the quality of the
training as well as of the testing data also plays a crucial role in determining the
performance of the soft sensor. This paper investigates the impact of data quality on
the performance of an ANFIS based inferential model that is designed to estimate
the average air temperature in distributed heating systems. The results of the two
experiments are reported. The results show that the performance of ANFIS based
sensor models is sensitive to the quality of data. The paper also discusses how to
reduce the sensitivity by an improved mathematical algorithm.

1 Introduction

A soft sensor computes the value of the process variables, which are difficult to
directly measure using conventional sensors, based on the measurement of other
relevant variables [1]. The performance of soft sensors is sensitive to the inferential
model that represents the dynamics of the targeted system. Such inferential models
can be based on one of the following modeling techniques:
l Physical model
l Neural network
l Fuzzy logic
l Adaptive Neuro-Fuzzy Inference System

S. Jassar (*)
Ryerson University, 350 Victoria Street, Toronto M5B2K3, Canada
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 143

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_12, # Springer ScienceþBusiness Media B.V. 2010
144 S. Jassar et al.

Recent research demonstrates the use of ANFIS in the development of an

inferential model that estimated the average air temperature in a distributed heating
system [2]. The estimated average air temperature allows for a closed-loop boiler
control scheme, resulting in higher energy efficiency and improved comfort. In the
current practice, due to the absence of economic and technically reliable method for
measuring the overall comfort level in the buildings, the boilers are normally
controlled to maintain the supply water temperature as a predefined level that
normally does not reflect the heating demand of the buildings [3].
For ANFIS based inferential models, when estimation/prediction accuracy is
concerned, it is assumed that both the data used to train the model and the testing
data to make estimations are free of errors [4]. But rarely a dataset is clean before
extraordinary effort having been made to clean the data. For this problem of
average air temperature estimation, with the measurement errors in the input
variables of the model, it is not unusual to have some uneven patterns in the dataset.
The research presented in this chapter aims to analyze the impact of data quality of
both training and testing datasets on the estimation accuracy of the developed
model.

2 ANFIS Based Inferential Model

As an AI technique, “Soft Computing”, integrates the powerful artificial intelli-

gence methodologies such as neural networks and fuzzy inference systems. While
fuzzy logic performs an inference mechanism under cognitive uncertainty, neural
networks posses exciting capabilities such as learning, adaption, fault-tolerance,
parallelism and generalization. Since Jang proposed ANFIS, its applications are
numerous in various fields, including engineering, management, health, biology
and even social sciences [5].
ANFIS is a multi-layer adaptive network-based fuzzy inference system. An
ANFIS consists of a total of five layers to implement different node functions to
learn and tune parameters in a fuzzy inference system (FIS) structure using a hybrid
learning mode. In the forward pass of learning, with fixed premise parameters,
the least squared error estimate approach is employed to update the consequent
parameters and to pass the errors to the backward pass. In the backward pass
of learning, the consequent parameters are fixed and the gradient descent method
is applied to update the premise parameters. One complete cycle of forward and
backward pass is called an epoch. Premise and consequent parameters will
be identified for membership function (MF) and FIS by repeating the forward and
backward passes. ANFIS has been widely used in prediction problems and other
areas.
ANFIS based inferential model represents the dynamic relationship between
the average air temperature, Tavg, and three easily measurable variables: external
temperature, T0, solar radiation, Qsol, and energy consumed by the boilers, Qin [6].
The FIS structure is generated by Grid partitioning method.
12 Data Quality in ANFIS Based Soft Sensors 145

Grid partition divides the data space into rectangular sub-spaces using axis-
paralleled partition based on pre-defined number of MFs and their types in each
dimension. The wider application of grid partition in FIS generation is blocked by
the curse of dimensions. The number of fuzzy rules increases exponentially when
the number of input variables increases. For example, if there are m MFs for each
input variable and a total of n input variables for the problem, the total number of
fuzzy rules is mn. It is obvious that the wide application of grid partition is
threatened by the large number of rules. According to Jang, grid partition is only
suitable for cases with small number of input variables (e.g. less than 6). In this
research, the average air temperature estimation problem has three input variables.
It is reasonable to apply the grid partition to generate FIS structure, ANFIS-GRID.
Gaussian type MF is used for characterizing the premise variables. Each input
has 4 MFs, thus there are 64 rules. The developed structure is trained using hybrid
learning algorithm [5]. The parameters associated with the MFs change through
training process.

2.1 Training and Testing Data

Experimental data obtained from a laboratory heating system is used for training
and testing of the developed model [7]. The laboratory heating system is located in
Milan, Italy. The details of experimental data collection for the four variables, Qin,
Qsol, T0 and Tavg, are given by the authors [2, 6]. The dataset used for the training of
ANFIS-GRID has 1800 input-output data pairs and is shown in Fig. 1.
The experimental data used for checking the performance of the developed
model is shown in Fig. 2. The testing dataset has 7,132 data pairs, which is large
enough as compared to training dataset used for the development of the model.

3 Impact of Data Quality

Data quality is generally recognized as a multidimensional concept [8]. While no

single definition of data quality has been accepted by the researchers working in this
area, there is agreement that data accuracy, currency, completeness, and consis-
tency are important areas of concern [9]. This Chapter is primarily considering the
data accuracy, defined as conformity between a recorded value and the actual data
value.
Several studies have investigated the effect of data errors on the outputs of
computer based models. Bansel et al. studied the effect of errors in test data on
predictions made by neural network and linear regression models [10]. The training
dataset applied in the research was free of errors. The research concluded that the
error size had a statistically significant effect on predictive accuracy of both the
linear regression and neural network models.
146 S. Jassar et al.

15000
Qin (W)

10000
5000
0
0 50 100 150
Time (Hour of the Year)
15000
Qsol (W)

10000
5000
0
0 50 100 150
Time (Hour of the Year)
10
T0 (Deg C)

5
0
–5
0 50 100 150
Time (Hour of the Year)
22
Tavg (Deg C)

20
18
16
0 50 100 150
Time (Hour of the Year)

Fig. 1 Training dataset (February 2000: day 22 to day 27)

O’Leary investigated the effect of data errors in the context of a rule-based

artificial intelligence system [11]. He presented a general methodology for analyz-
ing the impact of data accuracy on the performance of an artificial intelligence
system designed to generate rules from data stored in a database. The methodology
can be applied to artificial intelligence systems that analyze the data and generate a
set of rules of the form “if X then Y”. It is often assumed that a subset of the
generated rules is added to the system’s rule base on the basis of the measure of the
“goodness” of each rule. O’Leary showed that the data errors can affect the subset
of rules that are added to the rule base and that inappropriate rules may be retained
while useful rules are discarded if data accuracy is ignored.
Wei et al. analyzed the effect of data quality on the predictive accuracy of
ANFIS model [12]. The ANFIS model is developed for predicting the injection
profiles in the Daqing Oilfields, China. The research analyzed the data quality using
TANE algorithm. They concluded that the cleaning of data has improved the
accuracy of ANFIS model from 78% to 86.1%.
12 Data Quality in ANFIS Based Soft Sensors 147

× 104
2
Qin (W)

0
0 100 200 300 400 500 600
Time (Hour of the Year)
× 104
2
Qsol (W)

0
0 100 200 300 400 500 600
Time (Hour of the Year)

20
T0 (Deg C)

–20
0 100 200 300 400 500 600
Time (Hour of the Year)

25
Tavg (Deg C)

20

15
0 100 200 300 400 500 600
Time (Hour of the Year)

Fig. 2 Testing dataset (February 2000: day 1 to day 21)

In this chapter, the developed ANFIS-GRID based inferential model is trained

and tested using the experimental data collected from a laboratory heating system
[7]. The data collected has some uneven patterns. In this section we will discuss the
experiments conducted to examine the impact of data quality on the predictive
performance of the developed ANFIS-GRID model.

3.1 Experimental Methodology

Data errors may affect the accuracy of the ANFIS based models in two ways. First,
the data used to build and train the model may contain errors. Second, even if
training data are free of errors, once the developed model is used for estimation
tasks a user may use input data containing errors to the model.
148 S. Jassar et al.

The research in this area has assumed that data used to train the models and data
input to make estimation of the processes are free of errors. In this study we relax
this assumption by asking two questions: (1) What is the effect of errors in the test
data on the estimation accuracy of the ANFIS based models? (2) What is the effect
of errors in the training data on the predictive accuracy of the ANFIS based models?
While many sources of error in a dataset are possible, we assume that the
underlying cause of errors affect data items randomly rather than systematically.
One source of inaccuracy that may affect a dataset in this way is the measurement
errors caused by reading the equipment. This type of error may affect any data item
in the dataset and may understate or overstate the actual data value. This study does
not address the effect of systematic data errors on the estimations made by the
ANFIS based models.
Two experiments are conducted to examine the research targets. Both the
experiments used the same application (estimation of average air temperature)
and the same dataset.
Experiment 1 examines the first question: What is the effect of errors in the test
data on the estimation ability of the ANFIS based models? Experiment 2 examines
the second question: How the errors of the training data affect the accuracy of the
ANFIS based models?

3.2 Experimental Factors

There are two factors in each experiment: (1) fraction-error and (2) amount-error.
Fraction-error is the percent of the data items in the appropriate part of the dataset
(the test data in experiment 1 and the training data in experiment 2) that are
perturbed. Amount-error is the percentage by which the data items identified in
the fraction-error factor are perturbed.

3.2.1 Fraction-Error

Since fraction-error is defined as a percent of the data items in a dataset, the number
of data items that are changed for a given level of fraction-error is determined by
multiplying the fraction-error by the total number of data items in the dataset.
Experiment 1: The test data used in experiment 1, shown in Fig. 2, has 4 data
items (1 value for each of the 4 input and output variables for 1 entry of the total
7,132 data pairs). This experiment examines all of the possible number of data
items that could be perturbed. These four levels for fraction-error factor are: 25%
(one data item perturbed), 50% (two data items perturbed), 75% (three data items
perturbed), and 100% (four data items perturbed)
Experiment 2: The training data used in experiment 2 contains 1,800 data pairs
(1 value for each of the 4 input and output variables for 1,800 entries). Four levels of
the fraction-error factor are tested: 5% (90 data items are perturbed), 10% (180 data
12 Data Quality in ANFIS Based Soft Sensors 149

items are perturbed), 15% (270 data items are perturbed), and 20% (360 data items
are perturbed).

3.2.2 Amount-Error

For both the experiments, the amount-error factor has two levels: (1) plus or minus
5% and (2) plus or minus 10%. The amount-error applied to the dataset can be
represented by the following set of equations:

y0 ¼ y  0:05  y (1)

y0 ¼ y  0:1  y (2)

For Eqs. (1) and (2), y0 is the value of the variable after adding or subtracting the
noise error to the unmodified variable y.

3.3 Experimental Design

The experimental design is shown in Table 1. Both the experiments have four levels
for the fraction-error factor and two levels for the amount-error. For each combina-
tion of fraction-error and amount-error, four runs with random combinations of the
input and output variable are performed.
Although the levels of the fraction-error are different in the two experiments, the
sampling procedure is the same. For each fraction-error level, the variables are
randomly selected to be perturbed. This is repeated a total of four times per level.
Table 2 shows the combinations of the variables for experiment 1.
Second, for each level of the amount-error factor, each variable is randomly
assigned either a positive or negative sign to indicate the appropriate amount-error
to be applied. Table 3 shows the randomly assigned amount-error levels in

Table 1 Experimental design

Experiment 1 (errors in the test data)
Fraction-error levels (25%, 50%, 75%, and 100%) 4
Amount-error levels (5%, and 10%) 2
Number of random combinations of the variables considered within each fraction-error 4
level
Total number of samples considered 7132
Experiment 2 (errors in the training data)
Fraction-error levels (5%, 10%, 15%, and 20%) 4
Amount-error levels (5%, and 10%) 2
Number of random combinations of the variables considered within each fraction-error 4
level
Total number of samples considered 1800
150 S. Jassar et al.

Table 2 Four combinations of the variables for each fraction-error level in Experiment 1
Fraction-error level Input and output variable combination
(%) 1 2 3 4
25 (Qin) (Qsol) (T0) (Tavg)
50 (Qin, T0) (Qin, Tavg) (Qsol, T0) (Tavg, T0)
75 (Qin, T0, Qsol) (Qin, Tavg, T0) (Qin, Tavg, Qsol) (T0, Tavg, Qsol)
100 (Qin, T0, Qsol, (Qin, T0, Qsol, (Qin, T0, Qsol, (Qin, T0, Qsol,
Tavg) Tavg) Tavg) Tavg)

Table 3 Randomly assigned percentage increase (+) or decrease () for a given amount-error
level in Experiment 1
Fraction-error level Input and output variable combination
(%) 1 2 3 4
25 (Qin) (Qsol) (T0) (Tavg)
 + + 
50 (Qin, T0) (Qin, Tavg) (Qsol, T0) (Tavg, T0)
+, , +,+ ,+
75 (Qin, T0, Qsol) (Qin, Tavg, T0) (Qin, Tavg, Qsol) (T0, Tavg, Qsol)
+,, ,+, +,+, +,,+
100 (Qin, T0, Qsol, (Qin, T0, Qsol, (Qin, T0, Qsol, (Qin, T0, Qsol,
Tavg) Tavg) Tavg) Tavg)
,+,,+ +,+,+,+ ,,+,+ ,,,+

experiment 1. The procedure for experiment 2 differs only in the number of

variables that were randomly selected to be perturbed for the four levels of the
fraction-error factor.

3.4 Experimental Result

For both the experiments, the measured average air temperature values and ANFIS-
GRID estimated average air temperature values are compared using Root Mean
Square Error (RMSE) as a measure of estimation accuracy.

3.4.1 Experiment 1 Results: Errors in the Test Data

Estimation accuracy results, using the simulated inaccuracies for amount-error and
fraction-error for the average air temperature estimation are given in Fig. 3. It
shows that as fraction-error increases from 25% to 100%, RMSE increases results
in a decrease in predictive accuracy. As amount-error increases from 5% to 10%,
RMSE increases also indicating a decrease in estimation accuracy. Both fraction-
error and amount-error have an effect on predictive accuracy.
12 Data Quality in ANFIS Based Soft Sensors 151

Fig. 3 RMSE ( C) values as error level in the test data varies

Fig. 4 RMSE ( C) values as error level in the training data varies

3.4.2 Experiment 1 Results: Errors in the Training Data

Predictive accuracy results, using the simulated inaccuracies for amount-error and
fraction-error for the average air temperature estimation are given in Fig. 4. It
shows that as fraction-error increases from 5% to 20%, RMSE increases indicating
a decrease in predictive accuracy.
152 S. Jassar et al.

Table 4 Approximate functional dependencies detected using the TANE algorithm

Index Approximate dependencies Number of rows with conflicting tuples
1 Qin, Qsol, T0 ! Tavg 42
2 Qin, T0, Tavg ! Qsol 47
3 Qin, Qsol, Tavg ! T0 43
4 Qin, T0, Tavg ! Qsol 54

4 Tane Algorithm for Noisy Data Detection

Data quality analysis results show that the errors in the training data as well as in the
testing data affect the predictive accuracy of the ANFIS based soft sensor models.
This section discusses an efficient algorithm, TANE algorithm, to identify the noisy
data pairs in the dataset.
The TANE algorithm, which deals with discovering functional and approximate
dependencies in large data files, is an effective algorithm in practice [13]. The
TANE algorithm partitions attributes into equivalence partitions of the set of tuples.
By checking if the tuples that agree on the right-hand side agree on the left-hand side,
one can determine whether a dependency holds or not. By analyzing the identified
approximate dependencies, one can identify potential erroneous data in the relations.
In this research, relationship of the three input parameters (Qin, Qsol, and T0) and
the average air temperature (Tavg) is analyzed using TANE algorithm. For equiva-
lence partition, all the four parameters are rounded off to zero decimal points.
After data pre-processing, four approximate dependencies are discovered, as
shown in Table 4. Although all these dependencies reflect the relationships among
the parameters, the first dependency is the most important one because it shows that
the selected input parameters have consistent association relationship with the
average air temperature except a few data pairs, which is a very important depen-
dency for average air temperature estimation.
To identify exceptional tuples by analyzing the approximate dependencies, it is
required to investigate the equivalence partitions of both left-hand and right-hand
sides of an approximate dependency. It is non-trivial work that could lead to the
discovery of problematic data. By analyzing the first dependency, conflicting tuples
are identified. The total dataset has 7,132 data pairs. For the first approximate
dependency from Table 4, 42 conflicting data pairs are present which needs were
fixed using data interpolations for better performance of ANFIS-GRID model.

5 Results

The developed ANFIS-GRID model is validated using experimental results [7]. The
model performance is measured using RMSE
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
u1 X N  
RMSE ¼ t Tavg ðiÞ  T^avg ðiÞ (3)
N i¼1
12 Data Quality in ANFIS Based Soft Sensors 153

Measured Temperature
ANFIS-GRID (raw data) Estimated Temperature
22

21
Tavg (Deg C)

20

19

18

17

16
0 100 200 300 400 500 600
Time (Hour of the Year)
Measured Temperature
ANFIS-GRID (clean data) Estimated Temperature
22

21
Tavg (Deg C)

20

19

18

17

16
0 100 200 300 400 500 600
Time (Hour of the Year)

Fig. 5 Comparison of ANFIS-GRID estimated and the measured temperature values

For Eq. (3), N is the total number of data pairs, T^avg is the estimated and Tavg is the
experimental value of average air temperature.
Initially, ANFIS-GRID model uses the raw data for both the training as well as
the testing. Figure 5 compares ANFIS-GRID estimated average air temperature
values with the experimental results. First plot in Fig. 5 shows that ANFIS-GRID
estimated average air temperature values are in agreement with the experimental
results, with RMSE 0.56 C. However there are some points at which estimation is
not following the experimental results. For example, around 152–176 and 408–416
h of the year time, there is a significant difference between estimated and experi-
mental results.
For checking the effect of data quality on ANFIS-GRID performance, the
training and testing datasets are cleaned using TANE algorithm. The conflicting
data pairs are replaced with the required data pairs. Then the cleaned dataset is
applied for the training and the testing of ANFIS-GRID model. A comparison of
the model output with clean data and the experimental results is shown in second
plot of Fig. 5.
154 S. Jassar et al.

Fig. 6 Comparison of results

Figure 6 clearly shows the effect of data quality on predictive accuracy of

ANFIS-GRID model. The RMSE is improved by 37.5–0.35 C. RMSE is
considered as a measure of predictive accuracy. Less RMSE means less difference
between the estimated values and the actual values. Finally, predictive accuracy
is improved with decrease in RMSE. Figure 6 also shows the improvement in
R2 values.

6 Conclusion

ANFIS-GRID based inferential model has been developed to estimate the average
air temperature in the distributed space heating systems. This model is simpler
than the subtractive clustering based ANFIS model [2] and may be used as the air
temperature estimator for the development of an inferential control scheme. Grid
partition based FIS structure is used as there are only three input variables. The
training dataset is also large enough as compared to the modifiable parameters
of the ANFIS. As the experimental data is used for both the training as well as
the testing of the developed model, it is expected that data can have some
discrepancies. TANE algorithm is used to identify the approximate functional
dependencies among the input and the output variables. The most important
approximate dependency is analyzed to identify the data pairs with uneven patterns.
The identified data pairs are fixed and again the developed model is trained and
tested with the cleaned data. Figure 6 shows that the RMSE is improved by 37.5%
and R2 is improved by 12%. Therefore, it is highly recommended that the quality
of datasets should be analyzed before they are applied in ANFIS based modelling.
12 Data Quality in ANFIS Based Soft Sensors 155

References

1. M.T. Tham, G.A. Montague, A.J. Morris, P.A. Lant, Estimation and inferential control.
J. Process Control 1, 3–14 (1993)
2. S. Jassar, Z. Liao, L. Zhao, Adaptive neuro-fuzzy based inferential sensor model for estimat-
ing the average air temperature in space heating systems. Build. Environ. 44, 1609–1616
(2009)
3. Z. Liao, A.L. Dexter, An experimental study on an inferential control scheme for optimising
the control of boilers in multi-zone heating systems. Energy Build. 37, 55–63 (2005)
4. B.D. Klein, D.F. Rossin, Data errors in neural network and linear regression models:
an experimental comparison. Data Quality 5, 33–43 (1999)
5. J.S.R. Jang, ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans. Syst.
Man Cybern. 2, 665–685 (1993)
6. Z. Liao, A.L. Dexter, A simplified physical model for estimating the average air temperature
in multi-zone heating systems. Build. Environ. 39, 1013–1022 (2004)
7. BRE, ICITE, Controller efficiency improvement for commercial and industrial gas and
oil fired boilers, A craft project, Contract JOE-CT98-7010, 1999–2001
8. J. Wang, B. Malakooti, A feed forward neural network for multiple criteria decision making.
Comput. Oper. Res. 19, 151–167 (1992)
9. Y. Huh, F. Keller, T. Redman, A. Watkins, Data quality. Inf. Softw. Technol. 32, 559–565
(1990)
10. Bansal, R. Kauffman, R. Weitz, Comparing the modeling performance of regression and
neural networks as data quality varies. J. Manage. Inf. Syst. 10, 11–32 (1993)
11. D. O’Leary, The impact of data accuracy on system learning. J. Manage. Inf. Syst. 9, 83–98
(1993)
12. M. Wei, et al., Predicting injection profiles using ANFIS. Inf. Sci. 177, 4445–4461 (2007)
13. Y. Huhtala, J. Karkkainen, P. Porkka, H. Toivonen, TANE: an efficient algorithm for
discovering functional and approximate dependencies. Comput. J. 42, 100–111 (1999)
Chapter 13
The Meccano Method for Automatic Volume
Parametrization of Solids

R. Montenegro, J.M. Cascón, J.M. Escobar, E. Rodrı́guez, and G. Montero

Abstract In this paper, we present significant advances of the novel meccano

technique for simultaneously constructing adaptive tetrahedral meshes of 3-D com-
plex solids and their volume parametrization. Specifically, we will consider a solid
whose boundary is a surface of genus zero. In this particular case, the automatic
procedure is defined by a surface triangulation of the solid, a simple meccano
composed by one cube and a tolerance that fixes the desired approximation of the
solid surface. The main idea is based on an automatic mapping from the cube faces to
the solid surface, a 3-D local refinement algorithm and a simultaneous mesh untan-
gling and smoothing procedure. Although the initial surface triangulation can be a
poor quality mesh, the meccano technique constructs high quality surface and volume
adaptive meshes. Several examples show the efficiency of the proposed technique.
Future possibilities of the meccano method for meshing a complex solid, whose
boundary is a surface of genus greater than zero, are commented.

1 Introduction

Many authors have devoted great effort to solving the automatic mesh generation
problem in different ways [4, 15, 16, 28], but the 3-D problem is still open [1].
Along the past, the main objective has been to achieve high quality adaptive meshes
of complex solids with minimal user intervention and low computational cost.
At present, it is well known that most mesh generators are based on Delaunay
triangulation and advancing front technique, but problems, related to mesh quality
or mesh conformity with the solid boundary, can still appear for complex geo-
metries. In addition, an appropriate definition of element sizes is demanded for

R. Montenegro (*)
Institute for Intelligent Systems and Numerical Applications in Engineering (SIANI), University
of Las Palmas de Gran Canaria, 35017 Las Palmas de Gran Canaria, Spain
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 157

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_13, # Springer ScienceþBusiness Media B.V. 2010
158 R. Montenegro et al.

obtaining good quality elements and mesh adaption. Particularly, local adaptive
refinement strategies have been employed to mainly adapt the mesh to singularities
of numerical solution. These adaptive methods usually involve remeshing or nested
refinement.
We introduced the new meccano technique in [2, 3, 23, 24] for constructing
adaptive tetrahedral meshes of solids. We have given this name to the method
because the process starts with the construction of a coarse approximation of the
solid, i.e. a meccano composed by connected polyhedral pieces. The method builds
a 3-D triangulation of the solid as a deformation of an appropriate tetrahedral mesh
of the meccano. A particular case is when meccano is composed by connected
cubes, i.e. a polycube.
The new automatic mesh generation strategy uses no Delaunay triangulation, nor
advancing front technique, and it simplifies the geometrical discretization problem
for 3-D complex domains, whose surfaces can be mapped to the meccano faces.
The main idea of the meccano method is to combine a local refinement/derefine-
ment algorithm for 3-D nested triangulations [20], a parameterization of surface
triangulations [8] and a simultaneous untangling and smoothing procedure [5]. At
present, the meccano technique has been implemented by using the local refine-
ment/derefinement of Kossaczky [20], but the idea could be implemented with
other types of local refinement algorithms [17]. The resulting adaptive tetrahedral
meshes with the meccano method have good quality for finite element applications.
Our approach is based on the combination of several former procedures (refine-
ment, mapping, untangling and smoothing) which are not in themselves new, but
the overall integration is an original contribution. Many authors have used them in
different ways. Triangulations for convex domains can be constructed from a coarse
mesh by using refinement/projection [25]. Adaptive nested meshes have been
constructed with refinement and derefinement algorithms for evolution problems
[7]. Mappings between physical and parametric spaces have been analyzed by
several authors. Significant advances in surface parametrization have been done
in [8, 10, 11, 22, 27, 29], but the volume parametrization is still open. Floater et al.
[12] give a simple counterexample to show that convex combination mappings over
tetrahedral meshes are not necessarily one-to-one. Large domain deformations can
lead to severe mesh distortions, especially in 3-D. Mesh optimization is thus key for
keeping mesh shape regularity and for avoiding a costly remeshing [18, 19]. In
traditional mesh optimization, mesh moving is guided by the minimization of
certain overall functions, but it is usually done in a local fashion. In general, this
procedure involves two steps [13, 14]: the first is for mesh untangling and the
second one for mesh smoothing. Each step leads to a different objective function. In
this paper, we use the improvement proposed by [5, 6], where a simultaneous
untangling and smoothing guided by the same objective function is introduced.
Some advantages of the meccano technique are that: surface triangulation is
automatically constructed, the final 3-D triangulation is conforming with the
object boundary, inner surfaces are automatically preserved (for example, interface
between several materials), node distribution is adapted in accordance with the
object geometry, and parallel computations can easily be developed for meshing
13 The Meccano Method for Automatic Volume Parametrization of Solids 159

the meccano pieces. However, our procedure demands an automatic construction

of the meccano and an admissible mapping between the meccano boundary and the
object surface must be defined.
In this paper, we consider a complex genus-zero solid, i.e. a solid whose boundary
is a surface that is homeomorphic to the surface of a sphere, and we assume that the
solid geometry is defined by a triangulation of its surface. In this case, it is sufficient
to fix a meccano composed by a single cube and a tolerance that fixes the desired
approximation of the solid surface. In order to define an admissible mapping between
the cube faces and patches of the initial surface triangulation of the solid, we intro-
duce a new automatic method to decompose the surface triangulation into six patches
that preserves the same topological connections than the cube faces. Then, a discrete
mapping from each surface patch to the corresponding cube face is constructed by
using the parameterization of surface triangulations proposed by M. Floater in [8–11].
The shape-preserving parametrizations, which are planar triangulations on the cube
faces, are the solutions of linear systems based on convex combinations.
In the near future, more effort should be made in developing an automatic con-
struction of the meccano when the genus of the solid surface is greater than zero.
Currently, several authors are working on this aspect in the context of polycube-maps,
see for example [22, 27, 29]. They are analyzing how to construct a polycube for a
generic solid and, simultaneously, how to define a conformal mapping between the
polycube boundary and the solid surface. Although harmonic maps have been exten-
sively studied in the literature of surface parameterization, only a few works are
related to volume parametrization, for example a procedure is presented in see [21].
In the following Section we present a brief description of the main stages of the
method for a generic meccano composed of polyhedral pieces. In Section 3 we
introduce applications of the algorithm in the case that the meccano is formed by a
simple cube. Finally, conclusions and future research are presented in Section 4.

2 The Meccano Method

The main steps of the general meccano tetrahedral mesh generation algorithm are
summarized in this section. A detailed description of this technique can be analyzed
in [2, 23, 24]. The input data are the definition of the solid boundary (for example by
a given surface triangulation) and a given tolerance (corresponding to the solid
surface approximation). The following algorithm describes the whole mesh gener-
ation approach.
Meccano tetrahedral mesh generation algorithm
1. Construct a meccano approximation of the 3-D solid formed by polyhedral pieces.
2. Define an admissible mapping between the meccano boundary faces and the solid
boundary.
3. Build a coarse tetrahedral mesh of the meccano.
160 R. Montenegro et al.

4. Generate a local refined tetrahedral mesh of the meccano, such that the mapping of the
meccano boundary triangulation approximates the solid boundary for a given precision.
5. Move the boundary nodes of the meccano to the object surface with the mapping defined
in 2.
6. Relocate the inner nodes of the meccano.
7. Optimize the tetrahedral mesh with the simultaneous untangling and smoothing proce-
dure.

The first step of the procedure is to construct a meccano approximation by

connecting different polyhedral pieces. Once the meccano approximation is fixed,
we have to define an admissible one-to-one mapping between the boundary faces of
the meccano and the boundary of the object. In step 3, the meccano is decomposed
into a coarse and valid tetrahedral mesh by an appropriate subdivision of its initial
polyhedral pieces. We continue with a local refinement strategy to obtain an
adapted mesh which can approximate the boundaries of the domain within a
given precision. Then, we construct a mesh of the solid by mapping the boundary
nodes from the meccano faces to the true solid surface and by relocating the inner
nodes at a reasonable position. After those two steps the resulting mesh is tangled,
but it has an admissible topology. Finally, a simultaneous untangling and smooth-
ing procedure is applied and a valid adaptive tetrahedral mesh of the object is
obtained.
We note that the general idea of the meccano technique could be understood
as the connection of different polyhedral pieces. So, the use of cuboid pieces, or a
polycube meccano, are particular cases.

3 Application of the Meccano Method to Complex

Genus-ZeroSolids

In this section, we present the application of the meccano algorithm in the case of
the solid surface being genus-zero and the meccano being formed by a single cube.
We assume as datum a triangulation of the solid surface.
We introduce an automatic parametrization between the surface triangulation of
the solid and the cube boundary. To that end, we automatically divide the surface
triangulation into six patches, with the same topological connection that cube faces,
so that each patch is mapped to a cube face. These parametrizations have been
done with GoTools core and parametrization modules from SINTEF ICT, available
in the website http://www.sintef.no/math_software. This code implements Floater’s
parametrization in C++. Specifically, in the following application we have used the
mean value method for the parametrization of the inner nodes of the patch triangula-
tion, and the boundary nodes are fixed with chord length parametrization [8, 10].
We have implemented the meccano method by using the local refinement of
ALBERTA. This code is an adaptive multilevel finite element toolbox [26] devel-
oped in C. This software can be used to solve several types of 1-D, 2-D or 3-D
problems. ALBERTA uses the Kossaczky refinement algorithm [20] and requires
13 The Meccano Method for Automatic Volume Parametrization of Solids 161

an initial mesh topology [25]. The recursive refinement algorithm could not termi-
nate for general meshes. The meccano technique constructs meshes that verify the
imposed restrictions of ALBERTA in relation to topology and structure. The
minimum quality of refined meshes is function of the initial mesh quality.
The performance of our novel tetrahedral mesh generator is shown in the
following applications. The first corresponds to a Bust, the second to the Stanford
Bunny and the third to a Bone. We have obtained a surface triangulation of these
objects from internet.

3.1 Example 1: Bust

The original surface triangulation of the Bust has been obtained from the website
http://shapes.aimatshape.net, i.e. AIM@SHAPE Shape Repository. It has 64,000
triangles and 32,002 nodes. The bounding box of the solid is defined by the points
(x, y, z)min ¼ (120, 30.5, 44) and (x, y, z)max ¼ (106, 50, 46).
We consider a cube, with an edge length equal to 20, as meccano. Its center is
placed inside the solid at the point (5, 3, 4). We obtain an initial subdivision of
Bust surface in seven maximal connected subtriangulations by using the Voronoi
diagram associated to the centers of the cube faces. In order to get a compatible
decomposition of the surface triangulation, we apply an iterative procedure to
reduce the current seven patches to six.
We map each surface patch SiS to the cube face SiC by using the Floater
parametrization [8]. The definition of the one-to-one mapping between the cube
and Bust boundaries is straightforward once the global parametrization of the Bust
surface triangulation is built.
Fixing a tolerance e2 ¼ 0.1, the meccano method generates a tetrahedral mesh of
the cube with 147, 352 tetrahedra and 34, 524 nodes, see a cross section of the cube
mesh in Fig. 1a. This mesh has 32, 254 triangles and 16, 129 nodes on its boundary
and it has been reached after 42 Kossaczky refinements from the initial subdivision
of the cube into six tetrahedra. The mapping of the cube external nodes to the Bust
surface produces a 3-D tangled mesh with 8, 947 inverted elements, see Fig. 1b. The
location of the cube is shown in this figure. The relocation of inner nodes by using
volume parametrizations reduces the number of inverted tetrahedra to 285. We
apply our mesh optimization procedure [5] and the mesh is untangled in two
iterations. The mesh quality is improved to a minimum value of 0.07 and an average
qk ¼ 0:73 after 10 smoothing iterations.
We note that the meccano technique generates a high quality tetrahedra mesh
(see Figs. 1c, d): only one tetrahedron has a quality lower than 0.1, 13 lower than
0.2 and 405 lower than 0.3.
The CPU time for constructing the final mesh of the Bust is 93.27 s on a Dell
precision 690, 2 Dual Core Xeon processor and 8 Gb RAM memory. More precisely,
the CPU time of each step of the meccano algorithm is: 1.83 s for the subdivision
of the initial surface triangulation into six patches, 3.03 s for the Floater
162 R. Montenegro et al.

a b

c d

Fig. 1 Cross sections of the cube (a) and the Bust tetrahedral mesh before (b) and after (c) the
application of the mesh optimization procedure. (d) Resulting tetrahedral mesh of the Bust
obtained by the meccano method
13 The Meccano Method for Automatic Volume Parametrization of Solids 163

parametrization, 44.50 s for the Kossaczky recursive bisections, 2.31 s for the exter-
nal node mapping and inner node relocation, and 41.60 s for the mesh optimization.

3.2 Example 2: Bunny

The original surface triangulation of the Stanford Bunny has been obtained from the
website http://graphics.stanford.edu/data/3Dscanrep/, i.e. the Stanford Computer
Graphics Laboratory. It has 12, 654 triangles and 7, 502 nodes. The bounding box
of the solid is defined by the points (x, y, z)min ¼ (10, 3.5, 6) and (x, y, z)max ¼
(6, 2, 6).
We consider a unit cube as meccano. Its center is placed inside the solid at the
point (  4.5, 10.5, 0.5). We obtain an initial subdivision of the Bunny surface in
eight maximal connected subtriangulations using Voronoi diagram. We reduce the
surface partition to six patches and we construct the Floater parametrization from
each surface patch SiS to the corresponding cube face SiC . Fixing a tolerance e2 ¼
0.0005, the meccano method generates a cube tetrahedral mesh with 54, 496
tetrahedra and 13, 015 nodes, see Fig. 2a. This mesh has 11, 530 triangles and 6,
329 nodes on its boundary and has been reached after 44 Kossaczky refinements
from the initial subdivision of the cube into six tetrahedra.
The mapping of the cube external nodes to the Bunny surface produces a 3-D
tangled mesh with 2, 384 inverted elements, see Fig. 2b. The relocation of inner
nodes by using volume parametrizations reduces the number of inverted tetrahedra
to 42. We apply eight iterations of the tetrahedral mesh optimization and only
one inverted tetrahedron can not be untangled. To solve this problem, we allow the
movement of the external nodes of this inverted tetrahedron and we apply eight
new optimization iterations. The mesh is then untangled and, finally, we apply
eight smoothing iterations fixing the boundary nodes. The resulting mesh quality is
improved to a minimum value of 0.08 and an average qk ¼ 0:68, see Figs. 2c, d. We
note that the meccano technique generates a high quality tetrahedra mesh: only one
tetrahedron has a quality below 0.1, 41 below 0.2 and 391 below 0.3.
The CPU time for constructing the final mesh of the Bunny is 40.28 s on a Dell
precision 690, 2 Dual Core Xeon processor and 8 Gb RAM memory. More precisely,
the CPU time of each step of the meccano algorithm is: 0.24 s for the subdivision of
the initial surface triangulation into six patches, 0.37 s for the Floater parametrization,
8.62 s for the Kossaczky recursive bisections, 0.70 s for the external node mapping
and inner node relocation, and 30.35 s for the mesh optimization.

3.3 Example 3: Bone

The original surface triangulation of the Bone has been obtained from http://www-c.
inria.fr/gamma/download/..., and it can be found in the CYBERWARE Catalogue.
This surface mesh contains 274, 120 triangles and 137, 062 nodes.
164 R. Montenegro et al.

a b

c d

Fig. 2 Cross sections of the cube (a) and the Bunny tetrahedral mesh before (b) and after (c) the
application of the mesh optimization procedure. (d) Resulting tetrahedral mesh of the Bunny
obtained by the meccano method

Steps of the meccano technique are shown in Fig. 3. The resulting mesh has
47, 824 tetrahedra and 11, 525 nodes. This mesh has 11, 530 triangles and 5, 767
nodes on its boundary and it has been reached after 23 Kossaczky refinements
from the initial subdivision of the cube into six tetrahedra. A tangled tetrahedra
mesh with 1, 307 inverted elements appears after the mapping of the cube external
nodes to the bone surface. The node relocation process reduces the number of
inverted tetrahedra to 16. Finally, our mesh optimization algorithm produces a
high quality tetrahedra mesh: the minimum mesh quality is 0.15 and the average
quality is 0.64.
13 The Meccano Method for Automatic Volume Parametrization of Solids 165

a b

c d

Fig. 3 Cross sections of the cube (a) and the Bone tetrahedral mesh before (b) and after (c) the
application of the mesh optimization procedure. (d) Resulting tetrahedral mesh of the Bone
obtained by the meccano method

4 Conclusions and Future Research

The meccano technique is a very efficient adaptive tetrahedral mesh generator for
solids whose boundary is a surface of genus zero. We remark that the method
requires minimum user intervention and has a low computational cost. The proce-
dure is fully automatic and it is only defined by a surface triangulation of the solid, a
cube and a tolerance that fixes the desired approximation of the solid surface.
A crucial consequence of the new mesh generation technique is the resulting
discrete parametrization of a complex volume (solid) to a simple cube (meccano).
166 R. Montenegro et al.

We have introduced an automatic partition of the given solid surface triangula-

tion for fixing an admissible mapping between the cube faces and the solid surface
patches, such that each cube face is the parametric space of its corresponding patch.
The mesh generation technique is based on sub-processes (subdivision, mapping,
optimization) which are not in themselves new, but the overall integration using a
simple shape as starting point is an original contribution of the method and has
some obvious performance advantages. Another interesting property of the new
mesh generation strategy is that it automatically achieves a good mesh adaption to
the geometrical characteristics of the domain. In addition, the quality of the resulting
meshes is high.
The main ideas presented in this paper can be applied for constructing tetrahe-
dral or hexahedral meshes of complex solids. In future works, the meccano tech-
nique can be extended for meshing a complex solid whose boundary is a surface of
genus greater than zero. In this case, the meccano can be a polycube or constructed
by polyhedral pieces with compatible connections. At present, the user has to define
the meccano associated to the solid, but we are implementing a special CAD
package for more general input solid.

Acknowledgements This work has been partially supported by the Spanish Government, “Secre-
tarı́a de Estado de Universidades e Investigación”, “Ministerio de Ciencia e Innovación”, and
FEDER, grant contract: CGL2008-06003-C03.

References

1. Y. Bazilevs, V.M. Calo, J.A. Cottrell, J. Evans, T.J.R. Hughes, S. Lipton, M.A. Scott, T.W.
Sederberg, Isogeometric analysis: toward unification of computer aided design and finite
element analysis. Trends in Engineering Computational Technology (Saxe-Coburg Publica-
tions, Stirling, 2008), pp. 1–16
2. J.M. Cascón, R. Montenegro, J.M. Escobar, E. Rodrı́guez, G. Montero, A new meccano
technique for adaptive 3-D triangulations. Proceedings of the 16th International Meshing
Roundtable (Springer, New York, 2007), pp. 103–120
3. J.M. Cascón, R. Montenegro, J.M. Escobar, E. Rodrı́guez, G. Montero, The Meccano method
for automatic tetrahedral mesh generation of complex genus-zero solids. Proceedings of the
18th International Meshing Roundtable (Springer, New York, 2009), pp. 463–480
4. G.F. Carey, in Computational Grids: Generation, Adaptation, and Solution Strategies (Taylor
& Francis, Washington, 1997)
5. J.M. Escobar, E. Rodrı́guez, R. Montenegro, G. Montero, J.M. González-Yuste, Simultaneous
untangling and smoothing of tetrahedral meshes. Comput. Meth. Appl. Mech. Eng. 192,
2775–2787 (2003)
6. J.M. Escobar, G. Montero, R. Montenegro, E. Rodrı́guez, An algebraic method for smoothing
surface triangulations on a local parametric space. Int. J. Num. Meth. Eng. 66, 740–760 (2006)
7. L. Ferragut, R. Montenegro, A. Plaza, Efficient refinement/derefinement algorithm of nested
meshes to solve evolution problems. Comm. Num. Meth. Eng. 10, 403–412 (1994)
8. M.S. Floater, Parametrization and smooth approximation of surface triangulations. Comput.
Aid. Geom. Design 14, 231–250 (1997)
9. M.S. Floater, One-to-one piece linear mappings over triangulations. Math. Comput. 72,
85–696 (2002)
13 The Meccano Method for Automatic Volume Parametrization of Solids 167

10. M.S. Floater, Mean value coordinates. Comput. Aid. Geom. Design 20, 19–27 (2003)
11. M.S. Floater, K. Hormann, Surface parameterization: a tutorial and survey. Advances in
Multiresolution for Geometric Modelling, Mathematics and Visualization (Springer, Berlin,
2005), pp. 157–186
12. M.S. Floater, V. Pham-Trong, Convex combination maps over triangulations, tilings, and
tetrahedral meshes. Adv. Computat. Math. 25, 347–356 (2006)
13. L.A. Freitag, P.M. Knupp, Tetrahedral mesh improvement via optimization of the element
condition number. Int. J. Num. Meth. Eng. 53, 1377–1391 (2002)
14. L.A. Freitag, P. Plassmann, Local optimization-based simplicial mesh untangling and
improvement. Int. J. Num. Meth. Eng. 49, 109–125 (2000)
15. P.J. Frey, P.L. George, in Mesh Generation (Hermes Sci. Publishing, Oxford, 2000)
16. P.L. George, H. Borouchaki, in Delaunay Triangulation and Meshing: Application to Finite
Elements (Editions Hermes, Paris, 1998)
17. J.M. González-Yuste, R. Montenegro, J.M. Escobar, G. Montero, E. Rodrı́guez, Local refine-
ment of 3-D triangulations using object-oriented methods. Adv. Eng. Soft. 35, 693–702 (2004)
18. P.M. Knupp, Achieving finite element mesh quality via optimization of the Jacobian matrix
norm and associated quantities. Part II-A frame work for volume mesh optimization and the
condition number of the Jacobian matrix. Int. J. Num. Meth. Eng. 48, 1165–1185 (2000)
19. P.M. Knupp, Algebraic mesh quality metrics. SIAM J. Sci. Comput. 23, 193–218 (2001)
20. I. Kossaczky, A recursive approach to local mesh refinement in two and three dimensions.
J. Comput. Appl. Math. 55, 275–288 (1994)
21. X. Li, X. Guo, H. Wang, Y. He, X. Gu, H. Qin, Harmonic volumetric mapping for solid
modeling applications. Proceedings of the ACM Solid and Physical Modeling Symposium,
Association for Computing Machinery, Inc., 2007, pp. 109–120
22. J. Lin, X. Jin, Z. Fan, C.C.L. Wang, Automatic PolyCube-Maps. Lecture Notes in Computer
Science 4975, 3–16 (2008)
23. R. Montenegro, J.M. Cascón, J.M. Escobar, E. Rodrı́guez, G. Montero, Implementation in
ALBERTA of an automatic tetrahedral mesh generator. Proceedings of the 15th International
Meshing Roundtable (Springer, New York, 2006), pp. 325–338
24. R. Montenegro, J.M. Cascón, J.M. Escobar, E. Rodrı́guez, G. Montero, An automatic strategy
for adaptive tetrahedral mesh generation. Appl. Num. Math. 59, 2203–2217 (2009)
25. A. Schmidt, K.G. Siebert, in Design of Adaptive Finite Element Software: The Finite Element
Toolbox ALBERTA. Lecture Notes in Computer Science and Engineering, vol. 42. (Springer,
Berlin, 2005)
26. A. Schmidt, K.G. Siebert, ALBERTA – an adaptive hierarchical finite element toolbox. http://
www.alberta-fem.de/
27. M. Tarini, K. Hormann, P. Cignoni, C. Montani, Polycube-Maps. ACM Trans. Graph. 23,
853–860 (2004)
28. J.F. Thompson, B. Soni, N. Weatherill, in Handbook of Grid Generation (CRC Press, London,
1999)
29. H. Wang, Y. He, X. Li, X. Gu, H. Qin, Polycube splines. Comput. Aid. Geom. Design 40,
721–733 (2008)
Chapter 14
A Buck Converter Model for Multi-Domain
Simulations

Johannes V. Gragger, Anton Haumer, and Markus Einhorn

Abstract In this work a buck converter model for multi-domain simulations is

proposed and compared with a state-of-the-art buck converter model. In the pro-
posed model no switching events are calculated. By avoiding the computation of
the switching events in power electronic models the processing time of multi-
domain simulations can be decreased significantly. The proposed model calculates
any operation point of the buck converter in continuous inductor current conduction
mode (CICM) while considering the conduction losses and switching losses. It is
possible to utilize the proposed modeling approach also for other dc-to-dc converter
topologies. Laboratory test results for the validation of the proposed model are
included.

1 Introduction

For the efficient utilization of multi-domain simulation software it is of high

importance to have fast simulation models of power electronic components on
hand. Especially in simulations of vast and complex electromechanical systems
(e.g. power trains of hybrid electric vehicles [11] or drive systems in processing
plants [7]) it is crucial to limit the processing effort to a minimum. Many times such
electromechanical systems contain power electronic subsystems such as rectifiers,
inverters, dc-to-dc converters, balancing systems (for energy sources), etc. When
simulating these power electronic devices together with the other electrical and
mechanical components of the application, computing the quantities of the power
electronic models requires a large share of the available processing power if
switching events are calculated in the power electronic models. Simulation models

J.V. Gragger (*)

Electric Drive Technologies, Austrian Institute of Technology, Giefinggasse 2, 1210 Vienna,
Austria
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 169

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_14, # Springer ScienceþBusiness Media B.V. 2010
170 J.V. Gragger et al.

Fig. 1 Topology of a
conventional buck converter

including power electronic devices with switching frequencies around 100 kHz
require at least four calculation points within simulation times of around 10 ms for
calculating the switching events. However, if the energy flow in an electromechan-
ical system has to be investigated by simulation it is not necessary to calculate the
switching events in the power electronic model as long as the relevant losses are
considered.
In this work two different buck converter models are described. The first model,
model A, which is state-of-the-art describes the behavior of a conventional buck
converter, as shown in Fig. 1, including the calculation of switching events. This
means that in model A the switching of the semiconductors in the circuit is imple-
mented with if-clauses. Therefore, model A directly calculates the ripple of the current
through the storage inductor, and the ripple of the voltage across the buffer capacitor.
Due to the if-clauses in model A the duration of the computing time is very high.
The second model in this work, indicated as model B, describes the behavior of
the buck converter without calculating the switching events with if-clauses. Only
the mean and RMS values of the voltages and currents are calculated. Therefore, the
computation times of model B are significantly shorter than the computation times
of model A.
In both models the conduction losses are considered by an ohmic resistance of
the storage inductor, the knee voltage and the on-resistance of the diode, and the on-
resistance of the MOSFET. Linear temperature dependence is implemented for the
ohmic resistances of the storage inductor, the knee voltage and the on-resistance of
the diode and the on-resistance of the MOSFET in both buck converter models.
The switching losses are calculated assuming a linear dependency on the switch-
ing frequency, the blocking voltage and the commutating current between the
MOSFET and the diode. A controlled current source connected to the positive
and the negative pin of the supply side of the buck converter is used to model the
switching losses. This current source assures that the energy balance between the
supply side and the load side of the buck converter is guaranteed.

2 The Model for Calculating Switching Events

If the buck converter circuit in Fig. 1 is operated in continuous inductor current

conduction mode (CICM) the circuit can be in two different states. As long as
the MOSFET S is on and the diode D blocks the current, the buck converter is
14 A Buck Converter Model for Multi-Domain Simulations 171

in state 1. The corresponding equivalent circuit of the buck converter in state 1 is

shown in Fig. 2.
vin is the input voltage and vout is the output voltage of the converter. RS
indicates the on-resistance of the MOSFET and iS denotes the current through the
MOSFET. RL represents the ohmic contribution of the storage inductor L and iL is
the current through L. C stands for the buffer capacitor, iload indicates the output
current of the converter and iD represents the current through the diode (in state 1,
iD ¼ 0).
After S switched from on to off the diode begins to conduct. If S is off and D
conducts, the circuit is in state 2. The corresponding equivalent circuit of the buck
converter in state 2 is shown in Fig. 3 where RD is the on-resistance and VD
represents the knee voltage of the diode.
Discontinuous inductor current conduction mode (DICM) could be considered in
a third state where S and D are open at the same time. The buck converter is in
DICM if S is open and the current passing through the diode becomes zero.
A buck converter model for calculating switching events can be implemented
according to the pseudo code given in Alg. 1 where d stands for the duty cycle,
fs represents the switching frequency, and t indicates the time. scontrol, the
Boolean control signal of the MOSFET, is true during

ton ¼ dT s (1)

and false during

toff ¼ ð1  dÞT s (2)

in a switching period T s ¼ f1 . In Alg. 1 only CICM is considered. However, it is

s
easy to modify the model so that DICM can be simulated as well.
The basic principle of the modeling approach described in Alg. 1 is used in many
state-of-the-art simulation tools. A disadvantage of such a model is the processing

Fig. 2 Equivalent circuit of

the buck converter in state 1.
Switch S is on

Fig. 3 Equivalent circuit of

the buck converter in state 2.
Switch S is off
172 J.V. Gragger et al.

Algorithm 1 Pseudo code of a buck converter model for calculating switching events in CICM
Model:
BuckConverter
Parameter:
L, C, RS,RL,RD,VD,fs
Real variables:
vin,vout,iS,iL,iD,iload,t,d
Boolean variables:
scontrol
Equations:
if(scontrol ¼ true),
consider equations corresponding to the equivalent circuit of state 1
(Fig. 2)
else
consider equations corresponding to the equivalent circuit of state 2
(Fig. 3)

effort that is caused by the if-clauses. Strictly speaking, the whole set of equations
describing the circuit changes whenever the converter switches from state 1 to
state 2 and vice versa. In such a model the relevant conduction losses are considered
inherently. For the consideration of the switching losses a model expansion as
described in Section 4 is necessary.

3 The Averaged Model

If the dynamic behavior of the buck converter is not of interest but the energy flow
needs to be investigated it is possible to model the buck converter without calculat-
ing the switching events. Assuming the buck converter is in steady state the integral
of the inductor voltage vL over one switching period Ts equals zero [9]. Hence,
ð Ts ð ton ð Ts
vL dt ¼ vL dt þ vL dt ¼ 0: (3)
0 0 ton

During the time ton the equivalent circuit of state 1 describes the behavior of the
buck converter. In the circuit in Fig. 2 the inductor voltage is given by

vL;state 1 ¼ vin  vout  vRL  vRS;state 1 ; (4)

where the voltage across RL

vRL ¼ iL RL (5)

and the voltage across RS

vRS;state 1 ¼ iL RS ; (6)
14 A Buck Converter Model for Multi-Domain Simulations 173

with the mean value of the inductor current

i
L ¼ i
load : (7)

The equivalent circuit of state 2 (shown in Fig. 3) describes the behavior of the buck
converter during the time toff ¼ T s  ton . In state 2 the inductor voltage

vL;state 2 ¼ vout  vRL  vRD;state 2  V D ; (8)

where vRL is given by (5) and the voltage across RD

vRD;state 2 ¼ iL RD : (9)

Combining (3) with (4) and (8) one can derive

dT s ½vin  vRL  vout  vRS;state 1  þ ð1  dÞT s ½vout  vRL  vRD;state 2  V D  ¼ 0:

(10)

From (10) it is possible to find the mean value of the output voltage by

dðvin  vRS;state 1 þ vRD;state 2 þ V D Þ  ðvRL þ vRD;state 2 þ V D Þ ¼ vout : (11)

vout is a function of the duty cycle d, the input voltage vin, and the mean value of
the load current iload . Consequently, it is possible to calculate the average output
voltage with considering the conduction losses if there are relations for d, vin, and
iload available in other models, which is usually the case. The result of (11) can be
used as the input of a voltage source that is linked to the connectors of the load side
of the buck converter model. Please note that with (11) only the influence of the
conduction losses on the average output voltage is considered. In order to calculate
the influence of the conduction losses on the supply current the conduction losses
of the individual elements (MOSFET, diode, and inductor) need to be known. From
the equivalent circuits in Figs. 2 and 3 it appears that by approximation (with the
assumption that vout only changes insignificantly in one switching period) in state 1
the inductor current rises with a time constant

L
tstate 1 ¼ (12)
RS þ RL

and in state 2 the inductor current decays exponentially with

L
tstate 2 ¼ : (13)
RD þ RL

Provided that the time constants tstate 1 and tstate 2 are much larger than the
switching period Ts (which applies practically to all buck converters with proper
174 J.V. Gragger et al.

design), the instantaneous current through the inductor can be assumed to have a
triangular waveform such as

iL;state 1 if nT s <t  ðn þ dÞT s
iL ¼ (14)
iL;state 2 if (n + d)Ts <t (n þ 1)Ts

with n ¼ 0, 1, 2, 3, . . . and

DI L DI L
iL;state 1 ¼ iload  þ t (15)
2 dT s

DI L DI L
iL;state 2 ¼ iload þ  t; (16)
2 ð1  dÞT s

where the current ripple

vout þ V D
DI L ¼ ð1  dÞT s : (17)
L

Considering the two states of the buck converter circuit and using (14)–(17) the
waveform of the current through the MOSFET

iL;state 1 if nT s < t  ðn þ dÞT s
iS ¼ (18)
0 if ðn þ dÞT s < t  ðn þ 1ÞT s

and the waveform of the current through the diode


0 if nT s <t  ðn þ dÞT s
iD ¼ (19)
iL;state 2 if ðn þ dÞT s <t  ðn þ 1ÞT s :

For calculating the conduction losses of the individual elements in the converter,
the RMS values of the current through the MOSFET IS,rms, the current through the
diode ID,rms and the inductor current IL,rms have to be available. Applying the
general relation
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð
1 t0 þT 2
I rms ¼ iðtÞ dt (20)
T t0

to (18) and (19) results in

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 
DI 2L
I S;rms ¼ d I L;min þ I L;min DI L þ
2
; (21)
3
14 A Buck Converter Model for Multi-Domain Simulations 175

with

DI L
I L;min ¼ iload  (22)
2

and
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 
DI 2L
I D;rms ¼ ð1  dÞ I L;max  I L;max DI L þ
2
; (23)
3

with

DI L
I L;max ¼ i
load þ : (24)
2

Using (21) – (24) and considering

iL ¼ iS þ iD (25)

the RMS value of the inductor current can be written as

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
I L;rms ¼ I 2S;rms þ I 2D;rms : (26)

The conduction losses of the MOSFET PS,con and the storage inductor PL,con can
be calculated by

PS;con ¼ RS I 2S;rms (27)

and

PL;con ¼ RL I 2L;rms : (28)

When calculating the conduction losses of the diode also the portion of the power
emission contributed by the knee voltage has to be taken into account. Since the
knee voltage is modeled as a constant voltage source the mean value of the current
through the diode

i
D ¼ ð1  dÞiload (29)

has to be used to calculate the respective contribution to the conduction losses. The
total conduction losses in the diode can be written as

PD;con ¼ RD I 2D;rms þ V D i
D : (30)

176 J.V. Gragger et al.

Using (27), (28), and (30) the total amount of conduction losses can be calculated
by

Ptot;con ¼ PS;con þ PD;con þ PL;con : (31)

4 Consideration of Switching Losses

Models on different levels of detail for switching loss calculation have been
published. In many MOSFET models the parasitic capacitances are considered
and in some also the parasitic inductances at the drain and at the source of the
MOSFET are taken into account. In [1] a model considering the voltage depen-
dence of the parasitic capacitances is proposed. A model in which constant parasitic
capacities as well as parasitic inductances are considered is suggested in [2] and in
[4] voltage dependent parasitic capacities together with the parasitic inductances
are used for the calculation.
In data sheets such as [6] an equation combining two terms is used. In the first
term constant slopes of the drain current and the drain source voltage are assumed
and in the second the influence of the output capacitance is taken into account. Also
for this approach the parasitic capacities as well as the switching times (or at least
the gate switch charge and the gate current) have to be known. In [10] is stated that
the approach in [6] leads to an overestimation of the switching losses in the
MOSFET.
A general approach for switching loss calculation in power semiconductors
using measurement results with linearization and polynomial fitting is presented
in [3]. In [12] the switching losses are considered to be linear dependent on the
blocking voltage, the current through the switch, and the switching frequency. This
approach was initially developed for modeling switching losses in IGBTs but it can
also be applied to the calculation of MOSFET switching losses. In [8] a modified
version of the model proposed in [12] is presented. The difference is that in [8] the
switching losses are dependent on the blocking voltage, the current through the
switch, and the switching frequency with higher order.
In the presented work the approach described in [12] is used to model the
switching losses. The two buck converter models in Section 2 and 3 can be
expanded with switching loss models using

fs iload vin
Pswitch ¼ Pref;switch ; (32)
f ref;s iref;load vref;in

where Pswitch represents the sum of the actual switching losses in the MOSFET
and the diode of the buck converter, fs denotes the actual switching frequency, iload
is the actual commutating current between the diode and the MOSFET, and vin is
the actual blocking voltage of the diode and the MOSFET. Pref,switch represents a
14 A Buck Converter Model for Multi-Domain Simulations 177

measured value of the switching losses at a reference operation point defined by

fref,s, iref,load, and vref,in.
With this approach no knowledge of the parasitic capacitances and inductances
is needed. Neither the switching energy nor the switching times need to be known.
For the accuracy of the model given in (32) the precision of the measurement results
at the reference operation point is very important.

5 Implementation of the Simulation Models

The buck converter models described in Sections 2 and 3 got implemented with
Modelica modeling language [5] using the Dymola programming and simulation
environment. Modelica is an open and object oriented modeling language that
allows the user to create models of any kind of physical object or device which
can be described by algebraic equations and ordinary differential equations.
In both models the conduction losses with linear temperature dependence and
the switching losses (calculated according to Section 4) are considered.
In Fig. 4 the scheme of model A, the model calculating the switching events (as
explained in Section 2) is shown.
The conduction losses are inherently considered in Alg. 1 and the switching
losses are considered by means of a controlled current source with

Pswitch
imodel A ¼ (33)
vin

whereas Pswitch is calculated by (32).

Figure 5 illustrates the scheme of the averaged buck converter model (as
explained in Section 3) with consideration of the switching and conduction losses.
The basic components of model B are the current source controlled with i∗model B,
the voltage source controlled with v ¼ vout , and the power meter measuring the
averaged output power Pout ¼ vout i
load .

Fig. 4 Model A (buck

converter model)
178 J.V. Gragger et al.

Fig. 5 Model B (buck

converter model)

In model B the control signal of the voltage source v∗ is computed according to

(11) and the control signal of the current source i∗model B is calculated by

Pswitch þ Ptot;con þ Pout

imodel B ¼ : (34)
vin

In (34) Pswitch is given by (32), Ptot,con is calculated from (31), and Pout is the
output signal of the power meter in Fig. 5.

6 Simulation and Laboratory Test Results

The approach applied in model A is well established. Therefore the results of model
A are used as a reference for the verification of model B. For the comparison of the
two models a buck converter ( fs ¼ 100 kHz) supplied with a constant voltage of
30 V and loaded with a constant load current of 40 A was simulated using model A
and model B. In the two simulations the duty cycle was decreased step by step with
Dd ¼ 0.1 every 0.03 s starting from d ¼ 0.8 to d ¼ 0.2.
The purpose of model B is to calculate the efficiency and the electric quantities
in steady state. The supply current signals and the load voltage signals in Fig. 6
show that after the transients decay both models reach the same operation point.
Please note that in Fig. 6 the instantaneous supply current signal computed with
model A is averaged over a switching period.
Both simulations were computed on a state-of-the-art PC with 3 GHz dual core
and 3 GB RAM. It took only 2.8 s to process the results of the simulation with
model B whereas the CPU time for the simulation with model A was 36 s. The large
difference between the CPU times indicates that it is much more efficient to use
model B if the energy flow through a converter is the focus of the simulation.
For the validation of the two simulation models several laboratory tests have
been conducted. In order to avoid core losses an air-cored coil was implemented as
the storage inductor. As the passive and the active switch two IRFPS3810 power
MOSFETs were chosen whereas the body diode of one of the MOSFETs was used
as the freewheeling diode. The temperatures of the two MOSFETs and the air-cored
coil were measured with type-K thermocouples.
14 A Buck Converter Model for Multi-Domain Simulations 179

Fig. 6 Supply current (top) 40

current (A)
and load voltage (bottom) model A
30 model B
simulated with model A and B
20
10
0
0 0.04 0.08 0.12 0.16 0.2
30

voltage (V)
model A
20 model B

10

0
0 0.04 0.08 0.12 0.16 0.2
time (s)

Fig. 7 Measured and 1

simulated efficiency of
the buck converter with 0.8
vin ¼ 30 V and iload ¼ 40 A
efficiency (1)

0.6
measured
simulated
0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
duty cycle (1)

The buck converter used in the laboratory tests was operated with a similar duty
cycle reference signal as used for the results in Fig. 6. However, the step time of the
duty cycle signal in the laboratory test was significantly longer compared to the
signal used for the results in Fig. 6. Because of this, the temperatures of the
semiconductors increased significantly. Figure 7 shows the measured efficiency
of the circuit under test and the respective (steady state) simulation results of model
A and B. The measured and simulated results show satisfactory coherence.
In Fig. 8 the measured losses of the buck converter operated with d ¼ 0.2, and
d ¼ 0.8 during warm-up tests are compared with the results of a simulation carried
out with model B. In the top diagram it can be seen that the conduction losses
decrease with increasing time and temperature. This is because the knee voltage of
the freewheeling diode has a negative temperature coefficient and at d ¼ 0.2 the
freewheeling diode conducts 80 % of the time in a switching period. In the bottom
diagram of Fig. 8 the conduction losses raise with increasing time and temperature.
The reason for this is the positive linear temperature coefficient of the on-resistance
of the MOSFET and the longer duration in which the MOSFET conducts during a
180 J.V. Gragger et al.

Fig. 8 Warm-up test results d = 0.2

120
at d ¼ 0.2 (top) and d ¼ 0.8

power (W)
measured
(bottom). Measured and simulated
simulated losses of the buck 115
converter with vin ¼ 30 V
and iload ¼ 40 A 110
0 50 100 150 200
d = 0.8
120

power (W)
110
measured
simulated
100
0 50 100 150 200
time (s)

switching period. Please note that the MOSFET dissipates more energy and reaches
higher temperatures if the buck converter is operated with high duty cycles.

7 Conclusion

An analytical approach to calculate the steady state behavior of a buck converter

including the consideration of conduction losses is described. The presented model B
is generated from the derived equations and expanded so that switching losses and
temperature dependence of the conduction losses are considered. For the verification
of the described modeling approach two simulation models (models A and B) are
programmed with Modelica language. In steady state model A, the model calculating
switching events, matches the behavior of model B, the model based on the approach
of system averaging. When comparing the CPU times of models A and B it appears
that model B can be computed more than 10 times faster than model A. Conse-
quently, it is recommended to preferably use model B in simulations whenever only
the steady state values of the electrical quantities in the buck converter are of interest.
This is for instance the case in energy flow analyzes and in simulations for core
component dimensioning of electromechanical systems. The simulation results of
model A and B show satisfying conformity with the laboratory test results.

References

1. L. Aubard, G. Verneau, J.C. Crebier, C. Schaeffer, Y. Avenas, Power MOSFET switching

waveforms: an empirical model based on a physical analysis of charge locations. The 33rd
Annual Power Electronics Specialists Conference, IEEE PESC, 3, 1305–1310 (2002)
14 A Buck Converter Model for Multi-Domain Simulations 181

2. Y. Bai, Y. Meng, A.Q. Huang, F.C. Lee, A novel model for MOSFET switching loss
calculation. The 4th International Power Electronics and Motion Control Conference,
IPEMC, 3, 1669–1672 (2004)
3. U. Drofenik, J.W. Kolar, A general scheme for calculating switching- and conduction-losses
of power semiconductors in numerical circuit simulations of power electronic systems.
Proceedings of the International Power Electronics Conference, IPEC, 2005
4. W. Eberle, Z. Zhang, Y.-F. Liu, P. Sen, A simple switching loss model for buck voltage
regulators with current source drive. The 39th Annual Power Electronics Specialists Confer-
ence, IEEE PESC, 2008, pp. 3780–3786
5. P. Fritzson, Principles of Object-Oriented Modeling and Simulation with Modelica 2.1. (IEEE
Press, Piscataway, NJ, 2004)
6. Datasheet of the IRF6603, N-Channel HEXFET Power MOSFET. International Rectifier,
2005
7. H. Kapeller, A. Haumer, C. Kral, G. Pascoli, F. Pirker, Modeling and simulation of a large
chipper drive. Proceedings of the 6th International Modelica Conference, 2008, pp. 361–367
8. K. Mainka, J. Aurich, M. Hornkamp, Fast and reliable average IGBT simulation model with
heat transfer emphasis. Proceedings of the International Conference for Power Conversion,
Intelligent Motion and Power Quality, PCIM, 2006
9. N. Mohan, T.M. Undeland, W.P. Robbins, Power Electronics – Converters, Applications, and
Design, 2nd edn. (Wiley, New York, 1989)
10. Z.J. Shen, Y. Xiong, X. Cheng, Y. Fu, P. Kumar, Power MOSFET switching loss analysis:
A new insight. Conference Record of the IEEE Industry Applications Conference, 41st IAS
Annual Meeting, 3, 1438–1442 (2006)
11. D. Simic, T. B€auml, F. Pirker, Modeling and simulation of different hybrid electric vehicles in
modelica using Dymola. Proceedings of the International Conference on Advances in Hybrid
Powertrains, IFP, 2008
12. D. Srajber, W. Lukasch, The calculation of the power dissipation of the IGBT and the inverse
diode in circuits with the sinusoidal output voltage. Conference Proceedings of Electronica,
1992, pp. 51–58
Chapter 15
The Computer Simulation of Shaping in Rotating
Electrical Discharge Machining

Jerzy Kozak and Zbigniew Gulbinowicz

Abstract The effect of the tool electrode wear on the accuracy is very important
problem in the Rotating Electrical Discharge Machining (REDM). Two mathematical
models of REDM are presented: the first one considers machining with the face of
the end tool electrode and the second one considers EDM with the lateral side of the
electrode. The software for computer simulation of EDM machining with the side
and face of the electrodes has been developed. This simulation model for NC
contouring EDM using rotating electrode may also be applied for tool electrode
path optimization. The experimental results confirm the validity of the proposed
mathematical models and the simulation software.

1 Introduction

Today’s manufacturing industries are facing challenges from advanced difficult-

to-machine materials (tough super alloys, ceramics, and composites), stringent
design requirements (high precision, complex shapes, and high surface quality),
and machining costs. The greatly-improved thermal, chemical, and mechanical
properties of the material (such as improved strength, heat resistance, wear resis-
tance, and corrosion resistance) are making ordinary machining processes unable to
machine them economically. The technological improvement of manufacturing
attributes can be achieved by high efficiency Rotating Electrical Discharge Machining
(REDM: Electrical Discharge Grinding-EDG or Electrical Discharge Milling),
Abrasive Electrodischarge Grinding (AEDG), Rotary Electrochemical Arc/Discharge
Machining – RECAM/RCDM or grinding using metallic bond diamond wheels.
These machining processes use a rotating tool.

J. Kozak (*)
Warsaw University of Technology, ul. Narbutta 85, 02-524 Warsaw, Poland
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 183

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_15, # Springer ScienceþBusiness Media B.V. 2010
184 J. Kozak and Z. Gulbinowicz

The Electrical Discharge Machining process is widely used to machine compli-

cated shapes with high accuracy on hard and advanced materials including
hardened steels, super-alloys, tungsten carbides, electrically conductive engineer-
ing ceramics, polycrystalline diamonds etc. The material is removed by the thermal
energy generated by a series of electrical discharges between the tool electrode and
workpiece immersed in the dielectric. There is no direct contact between the tool
and workpiece during the machining process. The workpiece can be formed, either
by replication of a shaped tool electrode or by 3D movement of a simple electrode
like in milling.
The fabrication of complex shaped multiple electrodes used in the die-sinking
micro and macro Electrical Discharge Machining (EDM) is expensive and time
consuming [1]. Moving a simple shaped electrode along designed tool paths has
been proposed as a solution to some of these problems [2–5]. A numerical control
system enables the workpiece or tool electrode to move a previously programmed
path. In this way, very complex 2 or 3 dimensional shape according to the design
requirement can be shaped.
However, the tool wear during machining adversely affects the accuracy of
the machined components. The problem of tool wear in 3D micro and macro
EDM using simple shaped electrodes has been addressed by applying the Uniform
Wear Method to maintain the electrode shape unchanged and compensate for the
longitudinal tool wear [5–8].
Tool wear in EDM is characterized by the relative tool-electrode wear (RTW)
which is generally defined as ratio tool wear rate (TWR, which is the volume of tool
material removed per unit time) to material removal rate (MRR, which is the
volume of workpiece material removed per unit time):

TWR
n¼ (1)
MRR

Depending upon the operating parameters of REDM, the relative wear may be
0.01–2. Changes in dimensions of tool due to wear during machining is expected to
reflect in the actual depth of cut and finally in the profile and dimensional accuracy
of machined parts.
Results of investigation of EDM with rotating electrode reported in studies
[9, 10] show a slope curvilinear profile of bottom surface of machined groove
due to the wear of disk electrode. Controlling the path of electrode can reduce shape
error. In the paper [3] preliminary analysis of shape error indicates that one of the
main factors leading to shape errors is wheel wear. More extended study of this
problem based on mathematical modeling and experiments is reported in articles
[11, 12], where the general differential equation described relationship between tool
wear, initial and final shape of machined surface has been derived. Effect of wheel
wear on dimensional accuracy of grinding is known from theory of tolerances.
During machining of parts loaded together in a pocket, and in one pass of the rotary
tool, height of parts achieved is random variable in the machined set. Based on the
15 The Computer Simulation of Shaping in Rotating Electrical Discharge Machining 185

a b
Tool electrode
Workpiece

V
f Z
x, y

Fig. 1 Example of EDM using rotating tool electrode: (a) machining with the face of end
electrode, (b) machining by the side of rotating electrode

Fig. 2 Principal scheme for y = g(x)

mathematical modeling of
REDM by end tool electrode
Vf
y y = f(x)

a(0)

y = F(x)

assumption of constant wear rate, uniform probability density function (PDF) has
been obtained for height of machined parts [11].
Two cases of machining operations are taken for mathematical modeling: the
first one considers machining with the face of the tool electrode and the second one
considers EDM with the side of the electrode (Fig. 1).
In first case, the technique of integrating Uniform Wear Method with CAD/
CAM software has been successful in generating very complex 3D cavities, it
involves a time consuming, empirical approach for selecting tool paths and machin-
ing parameters. Therefore, it is necessary to develop a theoretical model which
accounts for the effect of tool wear on the surface profile generation during each
pass. As the number of passes can be very large, a corresponding computer
simulation software simulation also needs to be developed.

2 Mathematical Modelling of Redm Shaping

by End Tool Electrode

The principal scheme of shaping using the end tool electrode process is presented
in Fig. 2. The initial profile of workpiece is given by function y ¼ f(x). The
electrode is controlled to move along the tool head path y ¼ g(x). However, the
186 J. Kozak and Z. Gulbinowicz

longitudinal tool wear results in the profile of machined surface y ¼ F(x), which is
different from g(x).
The purpose of this mathematical modeling and computer simulation is to
determine surface profile y ¼ F(x), taking into account the change in tool
length which occurs due to the wear of electrode. The final profile depends on
input parameters, such as depth of cut a0, initial profile y ¼ f(x), diameter of tool d0
and tool head path y ¼ g(x).
Let us consider the case of machining presented in Fig. 2, when the initial
surface is y ¼ f(x) and g(x) ¼ constant.
In determining the profile of machined surface, the following assumptions
are made:
– Changes in tool shape are neglected because the Uniform Wear Method
is applying.
– The gap between tool electrode and workpiece is neglected.
– Material removal rate MRR is equal:

MRR ¼ ð f ðxÞ  FðxÞÞ  d0  Vf (2)

– Tool wear rate TWR, defined as the volume of tool material removed per unit
time, for g(x) ¼ const. is:

p  d02 dF
TWR ¼ (3)
4 dt

or

p  d02 dF dx p  d02 dF
TWR ¼ ¼ Vf (4)
4 dx dt 4 dx

where Vf is feed rate.

After substituting (2) and (4) in (1) for n, the equation describing the profile
of machined surface takes the following form:

dF
þ mF ¼ mf ð xÞ (5)
dx

with initial condition F(0) ¼ – a(0) (Fig. 2), where m ¼ 4v/p·d0 is the wear
factor.
In many cases it is possible to assume, that relative wear during machining
is constant (m ¼ const.), i.e. it is not dependent on the actual depth of cut. For
this condition and for f(x) ¼ 0, the solution of (5) becomes:

FðxÞ ¼ a0 expðmxÞ (6)

15 The Computer Simulation of Shaping in Rotating Electrical Discharge Machining 187

While considering accuracy, the profile is conveniently described in coordinates

relative to required allowance, a0 i.e.:
hðxÞ ¼ F  ða0 Þ ¼ a0 ½1  expðmxÞ (7)
For more universal application of (6) a non-dimensional form of notations
is used. The non-dimensional variables are defined as h ¼ h=a0 and x ¼ x=L,
where L is the length of workpiece. In non-dimensional form, (7) can be written as:
h ¼ 1  expðA  xÞ (8)
where A ¼ m · L ¼ 4v L/(p · d0).
Profiles of machined surface obtained for different values of A are presented in
Fig. 3. As result of wear, the depth of cut is decreasing and the tool wear rate (TWR)
is changing during machining. For comparison in Fig. 3 the dashed line (for A ¼ 1)
represents the profile neglecting changes in TWR during machining.
In general case when g ¼ g(x), mathematical model of REDM process is
described as follows:
dF dg
þ mF ¼ mf ð xÞ þ (9)
dx dx
with initial condition F(0) ¼ a0.
The machining accuracy may be improved by control motion of tool electrode.
For example, the tool wear can be compensated by moving tool or workpiece along
y-axis, to obtain flat surface. In the case when initial surface is flat f(x) ¼ 0 and the
depth of cut is equal a0, the required flat shape is F(x) ¼ a0. Therefore, based on
solution (9) the linear path of tool is needed, which can be described as follows:

y ¼ gðxÞ ¼ H0  a0  m  x (10)
where H0 is initial position of tool head.

Fig. 3 Profile of machined

surfaces for different values
of A
188 J. Kozak and Z. Gulbinowicz

For machining with constant feed rate Vf along x-axis, the compensation of
the tool wear can be obtained by adding the relative motion of the tool/workpiece
with constant feed rate Vy along y-axis equal:

4n  a0
Vy ¼ a0  m  Vf ¼  Vf (11)
p  d0

This theoretical conclusion about linear path of tool electrode for compensation
of wear has been confirmed by the experiments [5].

3 Mathematical Modelling of REDM Shaping by Lateral

Surface of Tool Electrode

The principal scheme of shaping process using the lateral side of tool electrode is
presented in Fig. 4. The purpose of this mathematical modeling and computer
simulation is to determine the profile of generated surface y ¼ F(x), taking into
account the change in tool diameter which occurs due to the tool wear of rotary tool
during machining (Fig. 4). The final surface profile depends on the input data, such
as depth of cut a0, initial surface profile y ¼ f(x), diameter of the tool d0 and the
curvilinear path of the center of tool y ¼ g(x).
For the modeling purpose the following assumptions were made:
– The inter electrode gap s is constant and included to the effective radius of tool
i.e. R ¼ R(electrode) þ s.
– Actual depth of cut, a, is determined by the position of points A and B, which are
point tangency of the tool electrode to the generated profile and intersection
point of the tool and the initial profile y ¼ f(x), respectively (Fig. 4).
– Feed rate along axis y is significant lower in compare to the feed rate Vf i.e. the
value of dg/dx << 1.
– Changes in the tool shape along the axis of rotation are neglected.

Fig. 4 Scheme of machining

with curvilinear path of
rotating tool
15 The Computer Simulation of Shaping in Rotating Electrical Discharge Machining 189

After machining for certain time t, the center of the tool has reached the
coordinate xC ¼ x, the feed rate is V(t) ¼ Vf and the effective tool radius due to
the wear is R(t).
The material removal rate MRR at dg/dx << 1 is:

MRR ¼ a  b  Vf ¼ ½ðf ðzÞ  FðxÞ  b  Vf (12)

where b is width of workpiece and x is function of x, x ¼ x(x).

The coordinates of intersection point B(x, f(x)) can be calculated from the
relation AC ¼ BC (Fig. 4):

ðx  xÞ2 þ ½Fð xÞ  gðxÞ2 ¼ ðB  xÞ2 þ ½f ðBÞ  gðxÞ2 (13)

The tool wear rate TWR, is given by:

dR
TWR ¼ 2p  b  RðtÞ  (14)
dt

Since

dR dR dx dR
¼ ¼ Vf (15)
dt dx dt dx

and according to the relation from Fig. 4,

R2 ðtÞ ¼ ½gðxÞ  Fð xÞ2 þ ðx  xÞ2 (16)

the rate of change of tool radius can be expressed as following:

   
dR 1 dg dF dx dx
¼ ½ gð xÞ  Fð x Þ   þ ð x  xÞ 1 (17)
dt R dx dx dx dx

Substituting (12) and (17) into (1) and performing transformation, the profile of
machined surface can be described by (18):

dF n F dx
þ ¼
dx 2p ½gðxÞ  F dx
  (18)
n f ð BÞ dx dg ð x  xÞ dx
¼ þ  1
2p ½gðxÞ  F dx dx ½gðxÞ  F dx

with initial condition F ¼ F(0) ¼ g(0)  R(0).

Generally, the value of relative tool wear is depending on the depth of cut a ¼ f
(x)-F(x) i.e. v ¼ v(a), and this function can be only determined experimentally.
190 J. Kozak and Z. Gulbinowicz

An additional equation to (18) and (13) can be derived from the condition for
tangency of the tool to the profile machined surface at the point A as follows:

dF
½gðxÞ  Fð xÞ ¼xx (19)
dx

The described above mathematical model can be used in following processes:

REDM/EDG (side EDM milling), AREDM, RECAM/RCDM and grinding (in this
case from definition of the G-ratio, v ¼ 1/G).
Based on the presented mathematical model, computer simulation of evolution
of workpiece profile can be carried out by using in-house developed software and
two main tasks can be formulated:
1. The tool path, g(x), and initial shape of surface, f(x), are known but the resulting
shape of the machined surface, F(x), and needs to be predicted.
2. For a required shape of the machined surface, F(x), the tool path, g(x), needs to
be determined in order to compensate for the tool wear.
These tasks have been solved numerically using the Finite Difference Method
and iterative procedure.
In many cases the changes in the radius and the differences between coordinates
x and x (for points A and B, respectively) are usually small when compared to the
initial radius. Therefore dx/dx ffi 1 and (18) will evolve in:

dF n F n f ð xÞ dg
þ ¼ þ (20)
dx 2p ½gð xÞ  F 2p ½gð xÞ  F dx

This simplified mathematical model can be successfully used to plan REDM

operations and in solving the mentioned above two main tasks.

4 Software for Computer Simulation

The developed software supports process design for the REDM. The software is
easy to use, menu driven application allowing evaluating the effect of various
parameters on process performance.
Programming of kinematics of the tool electrode can be achieved by supplying
program with tool path in x direction described. Geometry of the surface at the
beginning of machining can be defined in two ways:
– By supplying software with the surface equation, y ¼ f(x)
– By supplying software with coordinates of the surface (for example, a surface
just obtained a machining process cycle can be used as a starting surface for the
next machining operation)
15 The Computer Simulation of Shaping in Rotating Electrical Discharge Machining 191

The important feature of the software is the capability to define relative tool wear
as function of depth of cut.
Input date is inserted in the overlap Data. The windows for input data are shown
in Fig. 5.
Results of simulation can be viewed in two ways: tables of coordinates of points
of the workpiece surface and 2D graphs.
Fig. 6 shows an example of a graph of REDM machined surface defined
by function y ¼ a/(1 þ bsin(2x/). The tool path was given by g(x) ¼ H0  a0 þ
R0 ¼ constant (in Fig. 7 is shown shifted tool path to axis x).

Fig. 5 Overlap data

Fig. 6 The profiles from software screen (R0 ¼ 10 mm, a0 ¼ 0.6, n ¼ 0.3)
192 J. Kozak and Z. Gulbinowicz

In Fig. 7 is shown simulation of REDM process with compensation of tool wear.

An effect of periodical changes in the depth on final shape is appeared on machined
surface.

5 Experimental Verification

The theoretical model and simulation results were verified in Rotary Electrical
Discharge Machining (REDM) on M35K Mitsubishi, EDM – NC machine tool
using electrode with diameter of 8.2 and 18 mm. The range for workpiece move-
ment was 0–160 mm for x-axis. Experiments were performed on tool steel material
P233330 (HRC 56) with copper electrode using the following process parameters:
Pulse voltage of U: 40 V, Pulse current I: 55 and 120 A, Pulse on time tp ¼ pulse off

Fig. 7 The surface profile and tool path at EDM with compensation of tool wear (R0 ¼ 10 mm,
a0 ¼ 0.6, n ¼ 0.3)

Fig. 8 Comparison of theoretical and experimental results in non-dimensional system coordi-

nates. Setting parameters: tp ¼ t0 ¼ 120 ms, n ¼ 200 rpm
15 The Computer Simulation of Shaping in Rotating Electrical Discharge Machining 193

Table 1 The machining conditions for experiments carried out on EDIOS-16

Setting parameters
Test t (min) a0 (mm) I (A) U (V) n (rev/min) TWR (mm3/min) MRR (mm3/min)
a 25 1 6 90 600 0,033 9,077
b 19 1,5 10 50 2,700 0,254 2780
c 18 0,5 5 100 6,000 1,954 3,262

a
–1,08
0 10 20 30 40 50 60
–1,09

–1,1
Y, mm

–1,11

–1,12
Experimental
–1,13
Simulation
–1,14
X, mm

b
0 10 20 30 40 50 60
–0,84
–0,85
–0,86
–0,87
Y, mm

–0,88
–0,89
– 0,9
–0,91 Experimental
–0,92
Simulation ni(y)
–0,93
X, mm

c
0
0 10 20 30 40 50 60
– 0,1

– 0,2

– 0,3
Y, mm

Experimental
– 0,4 Simulation

– 0,5

– 0,6

– 0,7
X, mm

Fig. 9 Comparison of theoretical and experimental results

194 J. Kozak and Z. Gulbinowicz

time t0: 60 and 120 ms, Depth of cut a0: 1.2 and 1.5 mm, Feed rate Vf: 0.15, 1.0, 2.0,
and 6.0 mm/min, Rotation speed n: 200 rpm. For different combination of
setting parameters for verification used non-dimensional system coordinates X ¼
v·x/(2p·R0) and Y ¼ y=a0 . The results are presented in Fig. 8.
Further experiments were carried out on EDIOS-16 machine tool using copper
electrode with diameter of 4 mm, pulse on time 160 ms pulse off time 10 ms and the
maximal travel distances in x axis were 0–50 mm.
The machining conditions are shown in the Table 1.
Experimental results are compared with simulation prediction in Fig. 9.
Experimental verifications show high accuracy of developed mathematical
model and computer simulation. An overall average of 6% (of the initial depth of
cut) deviation was found between simulation and experimental results.

6 Conclusion

The study showed a good agreement of theoretical and experimental results of

modeling of REDM process. The developed software can be useful for analysis of
the REDM process, parameter optimization and surface prediction. Computer
simulation of REDM has a significant potential to be used in industry.

Acknowledgements This work was supported in part by the European Commission project
“Micro-Technologies for Re-Launching European Machine Manufacturing SMEs (LAUNCH-
MICRO)”.

References

1. M. Kunieda, B. Lauwers, K.P. Rajurkar, B.M. Schumacher, Advancing EDM through funda-
mental insight into the process. Ann. CIRP 54(2), 599–622
2. P. Bleys, J.P. Kruth, B. Lauwers, Milling EDM of 3D shapes with tubular electrodes,
Proceedings of ISEM XIII, 2001, pp. 555–567
3. P. Bleys, J. P. Kruth, Machining complex shapes by numerically controlled EDM. Int.
J. Electric. Mach. 6, 61–69 (2001)
4. J.P. Kruth, B. Lauwers, W. Clappaert, A study of EDM pocketing, Proceedings of ISEM X,
1992, pp. 121–135
5. Z.Y. Yu, T. Masuzawa, M. Fujino, Micro-EDM for three-dimensional cavities, development
of uniform wear method. Ann. CIRP 47(1), 169–172 (1998)
6. K.P. Rajurkar, Z.Y. Yu, 3D micro EDM using CAD/CAM., Ann. CIRP 49(1), 127–130 (2000)
7. Z.Y. Yu, T. Masuzawa, M. Fujino, A basic study on 3D Micro-EDM. Die Mould Technol. 11
(8), 122–123 (1996) (in Japanese)
8. Z.Y. Yu, J. Kozak, K.P. Rajurkar, Modeling and simulation of micro EDM process. Ann.
CIRP 52(1), 143–146 (2003)
9. Y. Uno, A. Okada, M. Itoh, T. Yamaguchi, EDM of groove with rotating disk electrode, Int.
J. Electr.Mach. 1, 13–20 (1996)
15 The Computer Simulation of Shaping in Rotating Electrical Discharge Machining 195

10. J. Quian, H. Ohmori, T. Kato, I. Marinescu, Fabrication of micro shapes of advanced materials
by ELID- Grinding. Trans. NAMRI/SME 27, 269–278 (2000)
11. J. Kozak, Effect of the wear of rotating tool on the accuracy of machined parts, in Proceedings
of the 2nd International Conference on Advances in Production Engineering APE-2, vol. I.
Warsaw, 2001, pp. 253–262
12. J. Kozak, Z. Gulbinowicz, D. Gulbinowicz, Computer simulation of rotating electrical
machining (REDM). Arch. Mech. Eng. XL(1), 111–125 (2004)
Chapter 16
Parameter Identification of a Nonlinear Two
Mass System Using Prior Knowledge

C. Endisch, M. Brache, and R. Kennel

Abstract This article presents a new method for system identification based on
dynamic neural networks using prior knowledge. A discrete chart is derived from a
given signal flow chart. This discrete chart is implemented in a dynamic neural
network model. The weights of the model correspond to physical parameters of the
real system. Nonlinear parts of the signal flow chart are represented by nonlinear
subparts of the neural network. An optimization algorithm trains the weights of the
dynamic neural network model. The proposed identification approach is tested with
a nonlinear two mass system.

1 Introduction

If the system is unknown and no prior knowledge is available the identification

process can start with an oversized neural network. To find a model of the system as
simple as possible a pruning algorithm deletes unnecessary parts of the network.
This approach reduces the complexity of the network. As shown in [2, 3] this
identification approach is able to remove weights from an oversized general
dynamic neural networks (short GDNN, in GDNNs the layers have feedback
connections with time delays, see Fig. 1).
The approach presented in this article is completely different. The assumption is
that prior knowledge is available in form of a continuous signal flow chart. The
model is derived in two steps. First the structure is altered in such a way, that all
backward paths contain integrator blocks. In the second step the integrator blocks
are discretized by Implicit and Explicit Euler approximation. In this article the

C. Endisch (*)
Institute for Electrical Drive Systems, Technical University of Munich, Arcisstraße 21, 80333
M€unchen, Germany
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 197

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_16, # Springer ScienceþBusiness Media B.V. 2010
198 C. Endisch et al.

p1(t) T a1(t) a2(t) a3(t) = ŷ(t)

D IW1, 1 Σ f1 LW2, 1 Σ f2 LW3, 2 Σ f3
R1´1 L S1×1
S2×1 S3×1
S1´R1 S2×S1 S3×S2

1 b1 1 b2 1 b3

S1×1 S2×1 S3×1

T T T
D LW1, 3 D LW2, 3 D LW3, 3
L L L
S1×S3 S2×S3 S3×S3
T T
D LW1, 2 D LW2, 2
L L
S1×S2 S2×S2
T
D LW1, 1
L
S1×S1

Layer 1 Layer 2 Layer 3

Fig. 1 Example of a three-layer GDNN with feedback connections in all layers – the output of a
tapped delay line (TDL) is a vector containing delayed values of the TDL input. Below the matrix-
boxes and below the arrows the dimensions are shown. Rm and Sm respectively indicate the
dimension of the input and the number of neurons in layer m. y^ is the output of the GDNN

parameters are trained with the powerful Levenberg–Marquardt (LM) algorithm

[5]. Real time recurrent learning (RTRL) is used to calculate the necessary Jacobian
matrix.
In [6, 7] a so called structured recurrent neural network is used for identifying
dynamic systems. On a first glance the structured recurrent network seems to be
similar to the approach in this article. However, the approach presented in this
article does neither depend on a state observer nor on system-dependent derivative
calculations. No matter what model is used for the identification process, the
derivative calculations are conducted for the GDNN-model in general [9–11].
The second section presents the general dynamic network (GDNN). For imple-
menting a special model structure administration matrices are introduced. Section 3
explains the optimization method used to train the network parameters throughout
this article. Section 4 describes the mechatronic system to be identified. The
structured GDNN is constructed in Section 5. In Section 6 the identification
approach is tested with a nonlinear dynamic system. Finally, Section 7 summarizes
the results.

2 General Dynamic Neural Network

Figure 1 shows an example of a three-layer GDNN with feedback connections. De

Jesus described the GDNN-model in his doctoral thesis [10]. The sophisticated
formulations and notations of the GDNN-model allow an efficient computation of
16 Parameter Identification of a Nonlinear Two Mass System 199

the Jacobian matrix using real time recurrent learning (RTRL) [4, 9, 11, 16]. In this
article we follow these conventions suggested by De Jesus. The simulation equation
for layer m is calculated by
X X X X
nm ðtÞ ¼ LW m;l ðdÞ  al ðt  dÞ þ LW m;l ðdÞ  pl ðt  dÞ þ bm ;
g l2I m d2DI m;l g
l 2 Lfm d 2 DLm;l

(1)

where nm ðtÞ is the summation output of layer m; pl ðtÞ is the l-th input to the
network, IW m;l is the input weight matrix between input l and layer m, LW m;l is the
g
layer weight matrix between layer l and layer m, bm is the bias vectorg of layer m,
m, l
DL is the set of all delays in the tapped delay line between layer l and layer m,
DIm, l is the set of all input delays in the tapped delay line between input l and layer
m, Im is the set of indices of input vectors that connect to layer m and Lfm is the set
of indices of layers that directly connect forward to layer m. The output of layer m
is

am ðtÞ ¼ f m ðnm ðtÞÞ; (2)

where f m ðÞ are either nonlinear tanh- or linear activation functions. At each point
in time the Eqs. 1 and 2 are iterated forward through the layers. Time is incremented
from t ¼ 1 to t ¼ Q. (See [10] for a full description of the notation used here.) In
order to construct a flexible model-structure, it is necessary that only particular
weights in the weight matrices do exist. This is realized by the introduction of
administration matrices.

2.1 Administration Matrices

For each weight matrix there exists one weight administration matrix to mark which
weights are used in the GDNN-model. The layer weight administration matrices
AL m;l ðdÞ have the same dimensions as the layer weight matrices LW m;l ðdÞ, the
g g
input weight administration matrices AI m;l ðdÞ have the same dimensions as the
g
input weight matrices IW ðdÞ and the bias weight administration vectors Abm
m;l
g
have the same dimensions as the bias weight vectors bm. The elements of the
administration matrices can have the boolean values 0 or 1, indicating if a weight
is valid or not. If e.g. the layer weight lw m;l ðdÞ ¼ ½LW m;l ðdÞk;i from neuron i of
g g
layer l to neuron k of layer m with ak;idth-order time-delay is valid, then
  m;l
AL m;l ðdÞ ¼ alk;i ðdÞ ¼ 1. If the element in the administration matrice equals
g k;i
to zero, the corresponding weight has no influence on the GDNN. With these
definitions the kth output of layer m can be computed by
200 C. Endisch et al.

!
X X X
Sl

k ðtÞ
nm ¼ lwm;l
k;i ðdÞ  m;l
alk;i ðdÞ  ali ðt  dÞ
l2Lfm d2DLm;l i¼1
!
X X X
Rl
þ iwm;l
k;i ðdÞ  aim;l
k;i ðdÞ  pli ðt  dÞ þ bm
k  abk ;
m

l2Im d2DI m;l i¼1

k ðtÞ ¼ f k ðnk ðtÞÞ;

am m m
(3)

where Sl is the number of neurons in layer l and Rl is the dimension of the lth input.
By setting certain entries of the administration matrices to one a certain GDNN-
structure is generated. As this model uses structural knowledge from the system, it
is called Structured Dynamic Neural Network (SDNN).

2.2 Implementation

For the simulations throughout this paper the graphical programming language
Simulink (Matlab) is used. SDNN and the optimization algorithm are implemented
as S-function in Cþþ.

3 Parameter Optimization

First of all a quantitative measure of the network performance has to be defined. In

the following we use the squared error

1 X
Q
Eðwk Þ ¼  ðy  y^q ðwk ÞÞT  ðyq  y^q ðwk ÞÞ
2 q¼1 q

1 X T
Q
¼  e ðw Þ  eq ðwk Þ; (4)
2 q¼1 q k

where q denotes one pattern in the training set, yq and y^qðwk Þare the desired target
and the actual model output of the q-th pattern respectively. The vector wk is
composed of all weights in the SDNN. The cost function Eðwk Þ is small if the
training process performs well and large if it performs poorly. The cost function
forms an error surface in a (N þ 1)-dimensional space, where N is equal to the
number of weights in the SDNN. In the next step this space has to be searched in
order to reduce the cost function.
16 Parameter Identification of a Nonlinear Two Mass System 201

3.1 Levenberg–Marquardt Algorithm

All Newton methods are based on the second-order Taylor series expansion about
the old weight vector wk:

Eðwkþ1 Þ ¼ Eðwk þ Dwk Þ

1 (5)
¼ Eðwk Þ þ gTk  Dwk þ  DwTk  H
k
 Dwk :
2
If a minimum on the error surface is found, the gradient of the expansion Eq. 2 with
respect to Dwk is zero:

rEðwkþ1 Þ ¼ gk þ H
k
 Dwk ¼ 0: (6)

Solving Eq. 6 for Dwk results in the Newton method

1
Dwk ¼ H
k
 gTk ;
1
(7)
wkþ1 ¼ wk  H
k
 gk :

1
The vector  H k
 gTk is known as the Newton direction, which is a descent
direction, if the Hessian matrix H k
is positive definite. The LM approach approx-
imates the Hessian matrix by [5]

H
k

J T ðwk Þ  
J ðwk Þ (8)

and it can be shown that

gk ¼ 
J T ðwk Þ  eðwk Þ; (9)

J ðwk Þ is the Jacobian matrix:

where 
2 3
@e1 ðwk Þ @e1 ðwk Þ @e1 ðwk Þ

6 @w1 @w2 @wN 7
6 7
6 @e2 ðwk Þ @e2 ðwk Þ @e2 ðwk Þ 7
6  7
J ðwk Þ ¼ 6
 6 @w1 @w2 @wN 7 7 (10)
6 .. .. .. .. 7
6 7
4 @e .ðw Þ @e .ðw Þ . .
@eQ ðwk Þ 5
Q Q
k k

@w1 @w2 @wN

The Jacobian matrix includes first derivatives only. N is the number of all weights in
the neural network and Q is the number of evaluated time steps. With Eqs. 4, 5 and
6 the LM method can be expressed with the scaling factor mk
202 C. Endisch et al.

 1
wkþ1 ¼ wk  J T ðwk Þ  
J ðwk Þ þ mk  I 
J T ðwk Þ  eðwk Þ; (11)

where I is the identity matrix. As the LM algorithm is the best optimization

method for small and moderate networks (up to a few hundred weights), this
algorithm is used for all simulations in this paper.
LM optimization is usually carried out offline. In this paper we use a sliding time
window that includes the information of the last Q time steps. With the last Q errors
the Jacobian matrix  J ðwk Þ from Eq. 10 is calculated quasi-online. In every time
step the oldest training pattern drops out of the time window and a new one (from
the current time step) is added – just like a first in first out (FIFO) buffer. If the time
window is large enough, it can be assumed that the information content of the
training data is constant. With this simple method we are able to implement the LM
algorithm quasi-online. For the simulations in this paper the window size is set to
Q ¼ 25000 using a sampling time of 1ms.

3.2 Jacobian Calculations

To create the Jacobian matrix, the derivatives of the errors have to be computed, see
Eq. 10. The GDNN has feedback elements and internal delays, so that the Jacobian
cannot be calculated by the standard backpropagation algorithm. There are two
general approaches to calculate the Jacobian matrix for dynamic systems: By
backpropagation through time (BPTT) [15] or by real time recurrent learning
(RTRL) [16]. For Jacobian calculations the RTRL algorithm is more efficient
than the BPTT algorithm [11]. According to this the RTRL algorithm is used in
this paper. The interested reader is referred to [2, 8, 11] for further details.

4 Two-Mass-System

The considered plant (shown in Fig. 2) is a nonlinear two-mass flexible servo

system (TMS), which is a common example for an electrical drive connected to a
work machine via flexible shaft. Figure 3 displays the signal flow chart of the TMS,

Fig. 2 Laboratory setup of

the TMS
16 Parameter Identification of a Nonlinear Two Mass System 203

d MR2

MW
1 MD 1
1
s·J1 s c s·J2
− '˙ 1 − − '˙ 2
M1 MB 1 − − Δ'˙ Δ' MC MB2

MR1

Fig. 3 Signal flow chart of the TMS with friction

where the spring constant c and the damping d model the shaft between the two
machines [14]. ’_ 1 and ’_ 2 denote the rotation speed of the main engine and the work
machine respectively. The torque of inertia of the machines are depicted by J1 and
J2. The motor torque is M1 and the torque at the work machine is M2. MB1 and MB2
are the acceleration torques of the main engine and the work machine respectively.
The difference of the rotation speeds is denoted by D’_ and D’ is the difference of
the angles. The torque of the spring and the damping torque are depicted by MC and
MD respectively. The friction torques of the engine and the working machine are
MR1 and MR2 respectively. The objective in this paper is to identify the linear TMS-
parameters and the characteristics of the two friction torques.

5 Structured Dynamic Neural Networks

To construct a structured network with the help of GDNNs it is necessary to

redraw the signal flow chart from Fig. 3 because the implementation of algebraic
loops is not feasible [1]. All feedback connections must contain at least one time
delay, otherwise the signals cannot be propagated through the network correctly.
This goal is accomplished by inserting integrator blocks in the feedback loops.
Figure 4 displays the redrawn version of Fig. 3. By using the Euler approximation
it is possible to discretize the redrawn signal flow chart. The Implicit Euler
approximation y(t) ¼ y(t  1) þ x(t)T replaces all integrator blocks in the forward
paths and the Explicit Euler approximation y(t) ¼ y(t  1) þ x(t  1)T replaces all
integrator blocks in the feedback paths. This approach ensures that all feedback
connections contain the necessary time delays. The resulting discrete signal flow
chart, which can be implemented as a SDNN, is displayed in Fig. 5. z 1 denotes a
first order time delay and T is the sampling time. All other denotations are the
same as in Fig. 3. In total the SDNN consists of 16 layers. The summing junctions
depict the neurons of the network. Every summing junction is marked with a
number, which denotes the layer of the neuron and its position within the layer.
204 C. Endisch et al.

d MR2
1
s·J1 1
s
1 MW
1 MD 1
s·J1 J2 1
s c s
− '˙ 1 − − '˙ 2
M1 − 1 − Δ'˙ Δ' MC MB2 1
s·J1 s
MR1

Fig. 4 Redrawn signal flow chart of the TMS from Fig. 3

−1
z
1.1 10.1
T
d J1
8.1
T MD T
J1 c·T J2
M1 1.2 2.1 − '˙ 7.1 8.2 MC 13.1 14.1 16.1 '˙ 2
1
3.1 − 5.1 − Δ'˙ 9.1 11.1 −
z
−1 T z
−1
z
−1
J1
6.1 9.2
4.1 12.1
.. ..
. .
1.3 15.1
−1 −1
z z
1.4

−1
15.2
z

Fig. 5 Structured recurrent network of a nonlinear TMS

For instance, 15. 1 marks the first neuron of the 15th layer. The friction of the
engine and the work machine can be modeled by an optional number of neurons in
the 5th layer and in the 11th layer respectively. These are the only neurons with
tanh-transfer functions. All other neurons have linear transfer functions. The
connections in Fig. 5 which do neither belong to a linear parameter (depicted as
box) nor to a friction-subpart are initialized with 1 or –1. The optimization
algorithm is able to tune the parameters corresponding to the spring constant c,
the damping d and the torque of inertia J2 and the friction weights of the work
machine. As it is not possible to identify the two torques of inertia as well as
the two characteristic curves of the TMS simultaneously, the engine parameters
are determined in a first no-load-identification which is conducted in idle running,
see Section 6.
16 Parameter Identification of a Nonlinear Two Mass System 205

6 Identification

6.1 Excitation Signal

The system is excited by an APRBS-signal (Amplitude Modulated Pseudo Random

Binary Sequence [13]) combined with a bias produced by a relay. The APRBS-
signal has an amplitude range between –7 and 7 Nm and an amplitude interval
between 10 and 250 ms. The relay output switches to –4 Nm if the rotation speed of
the TMS is greater than 10 rads and it switches to 4 Nm if the rotation speed is smaller
than 10 rads . The suggested excitation signal ensures that the TMS, which is
globally integrating, remains in a well defined range for which the SDNN is able
to learn the friction. Moreover, the output of the relay is multiplied by 0.2 for a
rotation speed in the range of 2 to 2 rad
s . Thus, the SDNN receives more informa-
tion about the friction in the region of the very important zero crossing. The
resulting output of the TMS can be regarded in upper panel of Fig. 8. This excitation
signal is used in all the following identification processes.

6.2 Engine Parameters

For a successful TMS-identification it is necessary to identify the engine para-

meters in idle mode first. The obtained values are used as fix parameters in the
identification of the whole TMS in Chapter 6. The upper left drawing of Fig. 6
displays the signal flow chart of the nonlinear engine, where M1 is the torque, ’_ 1
is the rotation speed, and J1 denotes the torque of inertia. In order to be able to
discretize the signal flow chart, we insert the integrator block in the backward
path. The resulting signal flow chart is shown in the upper right drawing of Fig. 6.
As explained above, for discretizing the signal flow chart the Implicit Euler
approximation has to replace the integrator in the forward path, whereas the
Explicit Euler approximation replaces the integrator in the backward path. The
sampling time T is incorporated in the gain corresponding to the torque of inertia J1.
The lower drawing of Fig. 6 displays the resulting discrete signal flow chart, which
can be implemented as a SDNN in which all the summing junctions are regarded as
neurons. The neurons in layer 5 model the friction of the engine.
For the identification process the engine is excited with the signal explained in
Section 6. The quasi-online calculated cost function E(wk) Eq. 4 with Q ¼ 25000 is
minimized by the LM optimization algorithm Eq. 11. The sampling time is set to
T ¼ 1ms and the identification process starts after 5 s, seen in the upper panel of
Fig. 7. The SDNN-model of Fig. 6 has three neurons in the fifth layer with six
weights to model the friction. These weights are initialized randomly between –0.5
and 0.5. Table 1 shows two different initial values for the torque of inertia J1 and the
corresponding results. The calculated results are mean values between the last 25,
000 optimization steps. Figure 7 displays the characteristic curve of the friction at
206 C. Endisch et al.

signal flow chart redrawn signal flow chart

1 1 1
s .J 1 s J1
'˙ 1 '˙ 1
M1 M1 1
− − s

MR1 MR1

resulting SDNN
T
J1
1.1 2.1 3.1 '˙ 1
M1 − 5.1
z−1

6.1
4.1
..
.
1.2

z−1

Fig. 6 Signal flow chart of the engine (upper left side), redrawn signal flow chart of the engine
(upper right side) and resulting SDNN (lower drawing)

t ¼ 100 s identified by the SDNN. We observe that the network is able to model the
jump due to the static friction with just three neurons. The following identification
of the whole TMS uses this friction curve from Fig. 7 and the result J1 ¼ 0. 1912
from Table 1 for the torque of inertia.

6.3 TMS Parameters

To identify the parameters of the whole TMS we excite the plant with the torque
signal described in Section 6 and use the SDNN-model constructed in Chapter 5.
The torque of inertia of the engine J1 and its friction are initialized according to the
results of Section 6. These weights remain unchanged during the whole identifica-
tion process, whereas the weights corresponding to the torque of inertia of the work
machine J2, the spring constant c, the damping d and the work machine friction are
trained. The work machine friction is modeled by three neurons with tanh functions
in the 11th layer, see Fig. 5. The six weights of this nonlinear subpart are initialized
16 Parameter Identification of a Nonlinear Two Mass System 207

1
torque of inertia J1

0.5

0
0 10 20 30 40 50 60 70 80 90 100
Start time [sec]

0.75
friction torque [Nm]

0.5
0.25
0
−0.25
−0.5
−0.75
−10 −8 −6 −4 −2 0 2 4 6 8 10
rad
rotation speed s

Fig. 7 Identification of the engine parameters – Torque of inertia and friction curve identified by
the nonlinear subpart in the 5th layer

Table 1 Initial values and J1 E

final values of the torque of [kg m2] (Mean value)
inertia of the engine J1
Initial value 0. 6
Result (mean value) 0.1912 5.629  10 2
Initial value 0.01
Result (mean value) 0.1912 4.387  10 2

randomly between –0.5 and 0.5. The upper panel of Fig. 8 displays the outputs of
the TMS and the SDNN-model during the identification process for the first set of
initial values of Table 2. The lower panel of this figure shows only 5 s for a detailed
view. The identification process starts after 5 s. The sampling time is set to T ¼ 1
ms. The quasi-online calculated cost function Eðwk Þ Eq. 4 with Q ¼ 25000 is
minimized by the LM optimization algorithm Eq. 11 and is depicted in the middle
panel of Fig. 8. Due to the quasi-online approach the cost function value increases
until t ¼ 25 s, until the training data window is completely filled up. The results in
Table 2 are mean values of the last 25000 optimization steps.
Figure 9 displays the developing of the damping, the spring constant and the
torque of inertia during the identification process for the first initialization of
Table 2. Figure 10 shows the characteristic curve of the work machine friction
208 C. Endisch et al.

rad 10
s SDNN 'ˆ˙2
system '˙ 2
rotation speed

−10
10 20 30 40 50 60 70 80 90 100
time [sec]

100
cost function

10−1

10−2

10−3
10 20 30 40 50 60 70 80 90 100
time [sec]

10
rad
s

5
rotation speed

0
SDNN 'ˆ˙2
−5 system '˙ 2
90 91 92 93 94 95
time [sec]

^_ 2 and the real TMS ’_ 2 with resulting cost function

Fig. 8 Output signals of the SDNN model ’

Table 2 Initial values and final values for the identification of the TMS
   
Nms Nm
J2 Error
[kgm2] rad rad (mean value)
d c
Initial value 0.7 0.4 100
Result (mean value) 0.3838 0.2406 477.2 3.792 102
Initial value 0.1 0.7 800
Result (mean value) 0.3881 0.0919 477.6 6.400 102

identified by the SDNN after 100 s. In addition to that Table 2 shows the results of a
second identification run with another initialization. The resulting torque of inertia
and spring constant are almost equal. Only the damping shows different final
values. The higher mean error (compared to the first identification) implies that
the second optimization process ended in a local minimum.
16 Parameter Identification of a Nonlinear Two Mass System 209

1
damping d

0.5

0
0 10 20 30 40 50 60 70 80 90 100
Start
time [sec]

1000
spring constant c

500

0
0 10 20 30 40 50 60 70 80 90 100
Start
time [sec]

1
torque of inertia J2

0.5

0
0 10 20 30 40 50 60 70 80 90 100
Start
time [sec]

Fig. 9 Linear parameter signals during the identification

1
friction torque [Nm]

0.5

−0.5

−1
−10 −8 −6 −4 −2 02 4 6 8 10
rotation speed rad
s

Fig. 10 Friction curve identified by the nonlinear subpart in the 11th layer
210 C. Endisch et al.

7 Conclusion

The system identification approach in this article is based on a given continuous

signal flow chart. Before discretizing the signal flow chart, integrator blocks are
inserted in all feedback loops. The Implicit Euler approximation replaces all
integrator blocks in the forward paths, whereas the Explicit Euler approximation
is used to replace the integrator blocks in the feedback paths. Since the resulting
discrete flow chart contains time delays in all feedback connections it can be
implemented as a SDNN without algebraic loops. Physical parameters of the
system are represented by certain weights of the SDNN. Nonlinear subparts of
the network model nonlinearities of the system. The suggested identification
approach is tested with a nonlinear TMS. The results verify that the suggested
approach enables us to identify the torques of inertia, the spring constant, the
damping and the two friction curves of the TMS. Problems may occur if the
optimization algorithm gets stuck in a local minimum. In this case the identification
process has to be restarted with a different set of initial values. With the help of the
mean error the success of the identification process can be validated.

References

1. M.J. Brache, Identification of dynamic systems using previous knowledge, Diploma Theses,
Lehrstuhl f€ur Elektrische Antriebssysteme, Technische Universit€at M€ unchen, 2008
2. C. Endisch, C. Hackl, D. Schr€ oder, Optimal brain surgeon for general dynamic neural net-
works, in Lecture Notes in Artificial Intelligence (LNAI) 4874 (Springer, Berlin, 2007), pp.
15–28
3. C. Endisch, C. Hackl, D. Schr€ oder, System identification with general dynamic neural net-
works and network pruning. Int. J. Computat. Intel. 4(3), 187–195 (2008)
4. C. Endisch, P. Stolze, C. Hackl, D. Schr€oder, Comments on backpropagation algorithms for a
broad class of dynamic networks. IEEE Trans. Neural Network 20(3), 540–541 (2009)
5. M. Hagan, B.M. Mohammed, Training feedforward networks with the marquardt algorithm.
IEEE Trans. Neural Networks 5(6), 989–993 (1994)
6. C. Hintz, B. Angerer, D. Schrder, Online identification of mechatronic system with structured
recurrent neural networks. Proceedings of the IEEE-ISIE 2002, L’Aquila, Italy, pp. 288–293
7. C. Hintz, Identifikation nichtlinearer mechatronischer Systeme mit strukturierten rekurrenten
Netzen. Dissertation, Lehrstuhl fr Elektrische Antriebssysteme, Technische Universit€at
M€unchen, 2003
8. O. De Jesus, M. Hagan. Backpropagation Algorithms Through time for a general class of
recurrent network. IEEE International Joint Conference Neural Network, Washington, 2001,
pp. 2638–2643
9. O. De Jesus, M. Hagan, Forward perturbation algorithm for a general class of recurrent
network. IEEE International Joint Conference Neural Network, Washington, 2001, pp.
2626–2631
10. O. De Jesus, Training general dynamic neural networks. Ph.D. dissertation, Oklahoma State
University, Stillwater, OK, 2002
11. O. De Jesus, M. Hagan backpropagation algorithms for a broad class of dynamic networks.
IEEE Trans Neural Networks 18(1), 14–27 (2007)
16 Parameter Identification of a Nonlinear Two Mass System 211

13. O. Nelles, in Nonlinear System Identification (Springer, Berlin, 2001).

14. D. Schr€oder, in Elektrische Antriebe – Regelung von Antriebssystemen, 2nd edn. (Springer,
Berlin, 2001)
15. P.J. Werbos, Backpropagation through time: what it is and how to do it. Proc. IEEE 78(10),
1550–1560 (1990)
16. R.J. Williams, D. Zipser, A learning algorithm for continually running fully recurrent neural
networks. Neural Comput. 1, 270–280 (1989)
Chapter 17
Adaptive and Neural Learning for Biped
Robot Actuator Control

Pavan K. Vempaty, Ka C. Cheok, Robert N.K. Loh,

and Micho Radovnikovich

Abstract Many robotics problems do not take the dynamics of the actuators into
account in the formulation of the control solutions. The fallacy is in assuming that
forces/torques can be instantaneously and accurately generated. In practice, actua-
tor dynamics may be unknown. This paper presents a Model Reference Adaptive
Controller (MRAC) for the actuators of a biped robot that mimics a human walking
motion. The MRAC self-adjusts so that the actuators produce the desired torques.
Lyapunov stability criterion and a rate of convergence analysis is provided. The
control scheme for the biped robot is simulated on a sagittal plane to verify the
MRAC scheme for the actuators. Next, the paper shows how a neural network (NN)
can learn to generate its own walking gaits using successful runs from the adaptive
control scheme. In this case, the NN learns to estimate and anticipate the reference
commands for the gaits.

1 Introduction

Biped walking dynamics is highly non-linear, has many degrees of freedom and
requires developing complicated model to describe its walking behavior. Many
novel approaches have emerged in the field of biped walking to address this
complicated control mechanism. Existing biped walking methods [3, 6, 7, 10]
give precise stability control for walking bipeds. However these methods require
highly precise biped walking dynamics. In recent years, biped walking through
imitation has been a promising approach, since it avoids developing complex
kinematics of the human walking trajectory and gives the biped a human like

P.K. Vempaty (*)

Department of Electrical and Computer Engineering, Oakland University, Rochester, Michigan
48309, USA
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 213

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_17, # Springer ScienceþBusiness Media B.V. 2010
214 P.K. Vempaty et al.

walking behavior. These methods combine the conventional control schemes to

develop the walking mechanism for the bipeds. Examples include, imitation based
on intelligent control methods like genetic algorithm [3, 14], fuzzy logic [15],
neural network approach [8], and other methods such as adaptation of biped
locomotion [10], learning to walk through imitation [8] and reinforcement learning
[9, 12, 15]. But these methods cannot adapt their behavior to the changes in the
dynamics of the process and the character of the disturbances [7]. Therefore,
adaptive control approaches [1, 2, 5, 11] are useful.
To address the problem of poorly known actuator dynamic characteristics and
unpredictable variations of a biped system, we propose a Lyapunov based model
reference adaptive control system (MRAC) method for the biped walking control.
The controlled plant (biped actuator) adapts itself to the reference model (desired
dynamics for the actuators). Lyapunov’s stability criterion and convergence analy-
sis are shown for parameter tuning. Through this scheme, a robot can learn its
behavior through its reference models. The paper further describes a NN scheme for
estimating and anticipating reference commands needed for walking gaits. Data
from successful adaptive control biped runs were used to train the NN.

2 Problem Description

2.1 Objective

Consider the objective of controlling a biped robot so that it imitates the movement
of a person. Figure 1 shows the basic idea where the human movement is repre-
sented by yd and the biped movement by y. The biped motion is determined by the
actuators which are controlled by the inputs ua. The overall objective is to find the
adaptive ua such that y ! yd.
The actuator dynamics have uncertainties including non-linearities, unknown
parameter values and delays, which have not been widely addressed. Figure 2 shows
the adaptive actuator objective where the actuator output moment M is made to
follow a required Md, which will be computed from the desired requirement that
y tracks yd. Figure 2 also shows a NN scheme for generating an estimate ŷd of

ua y
yd

Uncertain M q
ua
Actuators Biped q̇
Fig. 1 Human movements, Dynamics
Dynamics y = Cqq
biped robot, its actuators
17 Adaptive and Neural Learning for Biped Robot Actuator Control 215

Fig. 2 MRAC scheme Md M F

yd
for the biped walker Desired Adaptive
Biped
Torque Actuator
Dynamics
ŷd Generator Scheme
q, q̇
Cyq

Neural
Network

the reference command signal. This paper deals with the formulation and
simulation aspects of the MRAC actuator and neural network learning schemes.

2.2 Biped Dynamics

Equations for describing the dynamics of a biped robot were introduced in [4], and
can be summarized as follows.

Aq ðqÞ€ _ M; FÞ
q ¼ Bðq; q; (1)

where q is the generalized coordinates of the robot, q_ and q

€ are the first and second
derivatives, M the moments/torques applied to the joints in the robot and F the
reaction forces at the contact of the robot’s feet and ground surface.

2.3 Uncertain Actuator Dynamics

The literature often assumes that M can be readily generated without considering
the dynamics of the actuators. For example, if a set of desired torques are calculated
as Md, then it would assume that M ¼ Md and applied directly as inputs to the
robot. However, this is not a valid assumption since in practice the moments M will
be generated by actuators which normally have unknown parameters, time delays
and non-linearities. The moments M can be modelled as the states of

_
x_ a ðtÞ ¼ Aa xa ðtÞ þ Ba ua ðtÞ þ da ðqðtÞ; qðtÞ; text ðtÞÞ
(2)
M ¼ C a xa

where ua are inputs of the actuators, and da(q, q, _ text) represents disturbance
torques to the actuators due to robot movements. t ext is an external disturbance
torque. We assume that the moments/torques M can be measured; for example, by
measuring the currents in motors or pressure in hydraulics. In pre-tuned actuators,
we can assume that M ¼ xa, i.e., Ca ¼ I.
216 P.K. Vempaty et al.

2.4 Desired Moments Md

The desired moments Md can be derived as the output of a controller that operates
on yd and y. For example,


Md ðsÞ ¼ Gc ðsÞ yd ðsÞ  yðsÞ (3)

where s is the Laplace variable, Gc(s) is the controller transfer function. The
controller Gc is designed to generate the desired moments Md, required for the
adaptive actuator scheme by using the information from y and yd.

2.5 Adaptive Control Approach

In the presence of actuator unknown parameters and uncertainties in (2). We would

like the actuator output M to follow Md. Figure 2 shows the adaptation scheme.

3 Solution

3.1 Reference Model for Actuator

To apply the MRAC approach to the actuator (2), a reference model for the actuator
is needed as follows. The computed desired moments Md (3) will be represented as
the states of the reference model

x_ m ðtÞ ¼ Am xm ðtÞ þ Bm um ðtÞ

(4)
Md ðtÞ ¼ xm ðtÞ

where Am and Bm represent the desired dynamics for the actuator to follow. um(t)
represents the command input to the reference model of the actuator and is required
for the MRAC.

3.2 Inverse Reference Model

However, we do not know the input um. So the approach here is to estimate the
unknown um knowing xm. The unknown um can be represented as the output of an
waveform shaping model, i.e.,
x_ u ¼ Au xu þ wu
(5a)
um ¼ Cu xu
17 Adaptive and Neural Learning for Biped Robot Actuator Control 217

where Au and Cu represent approximate waveform characteristics and wu is a sparse

and small shaping input [13].
An estimate of the of um can be found using an observer of the form
^_ mu ¼ Amu ^
x xmu þ Kmu ½xm  Cmu ^xmu 
(5b)
^m ¼ ½0 Cu  ^
u xmu
     
xm Am Bm Cu 0
where, xmu ¼ ; Amu ¼ ; Bmu ¼ and Cmu ¼ [I 0]; 0 and
xu 0 Au I
I are null and identity matrix. Kmu is chosen such that Amu  Kmu Cmu has
exponentially stable eigenvalues. We refer (5b) as the inverse reference model,
^m .
xm is given and input um is estimated by the inverse model as the output u

3.3 MRAC Scheme

3.3.1 Configuration of MRAC Actuator

The adaptive actuator scheme is shown in Fig. 3, where the reference and control
models are specified by
x_ m ¼ Am xm þ Bm u
^m
(6)
ua ¼ Lxa þ N^ um
The adaptation algorithm in the MRAC will adjust the gains L and N based on
Lyapunov stability criteria as illustrated in Fig. 3.

yd Inverse Reference Model

Md Input ûm
Gc(s)
Predictor
y
q
da(q; q̇ ; ¿ext) F
ua M
Actuator Biped
N Σ
Dynamics Dynamics
q; q̇
L

Adaptation
Mechanism

Cqq

Fig. 3 The MRAC with the Neural Network

input predictor
218 P.K. Vempaty et al.

3.3.2 Error Dynamics

The error e ¼ xm  xa between the actuator torque xa ¼ M and the desired torque
xm ¼ Md behave according to
e_ ¼ Am e þ ½Am  Aa  Ba Lxa
(7)
þ ½Bm  Ba N^ _ text Þ
um  da ðq; q;

3.3.3 Lyapunov Stability Analysis

Define a candidate for a Lyapunov function as

v ¼ eT Pe
h i
þ trace ðAm  Aa  Ba LÞT QðAm  Aa  Ba LÞ (8)
h i
þ trace ðBm  Ba NÞT RðBm  Ba NÞ

where P ¼ PT > 0, Q ¼ QT > 0 and R ¼ RT > 0 are positive definite matrices. Then
v_ ¼ e_ T Pe þ eT P_e (9)

_ we choose
Based on the analysis of v,
Ba L_ ¼ Q1 PexTa
(10)
Ba N_ ¼ R1 Pe^
uTm
_ text). We next choose an S ¼ ST > 0
_ eT[PAm þ AmTP]e  2eTAmTda(q, q,
so that v¼
and solve P from
PAm þ ATm P ¼ S (11)

We arrive at v¼ _  eTSe þ 2eTAmTda(q, q, _ text), where v_ is negative under the

assumption e Se > 2eTAmTda(q, q,
T
_ text), which implies that the magnitude of error
should be larger than the disturbance. In practice this means that the system should be
persistently excited. Hence we conclude that the overall dynamic system comprising
of (4) and (7) has a candidate function that satisfies the Lyapunov stability criterion.

4 MRAC for Walking Biped Actuators

4.1 Dynamics of Walking Biped

The bipedal model has five links with four pin joints as shown in Fig. 4. One link
represents the upper body and two links are for each lower limb. The biped has two
17 Adaptive and Neural Learning for Biped Robot Actuator Control 219

hip joints, two knee joints and two ankles at the tips of the lower limbs. There is an
actuator located at each joint and all of the joints are considered rotating only in the
sagittal plane. The system contains five links and seven degrees of freedom,
selected according to Fig. 4a and given by

q ¼ ½x0 ; y0 ; a; bL ; bR ; gL ; gR T (12)

The coordinates (x0, y0) fix the position of the center of mass of the torso, and the
rest of the coordinates describe the joint angles. The link lengths are denoted by
(h0, h1, h2) and masses by (m0, m1, m2). The centers of mass of the links are located
at the distances (r0, r1, r2) from the corresponding joints.
The model is actuated with four moments; two of them acting between the torso
and both thighs, and two at the knee joints, as illustrated in Fig. 4b.

M ¼ ½ML1 ; MR1 ; ML2 ; MR2 T (13)

The walking surface is modelled using external forces that affect the leg tips, as
shown in Fig. 4b.

T
F ¼ FLx ; FLy ; FRx ; FRy (14)

a b
m0
y
(x0; y0)
x
r0
h0

®
r1
m1 h1
¯R MR1
¯L r2 ML1

m2
°L h2 °R MR2
FLx ML2 FRx

FLy FRy

Fig. 4 (a) Biped robot with the corresponding seven coordinates q. (b) Biped robot with the
moments M and reaction forces F
220 P.K. Vempaty et al.

When the leg should touch the ground, the corresponding forces are switched on to
support the leg. As the leg rises, the forces are zeroed.
Using Lagrangian mechanics, the dynamic equations for the biped system
can be derived as shown in (1) where A(q) 2 ℜ77 is the inertia matrix and
_ M, F 2 ℜ71 is a vector containing the right hand sides of the seven partial
bq, q,
differential equations. The closed form formulas for both A and b are presented
in [4].

4.2 Computation of Desired Moments

The desired human movement angles will be measured and represented by

yd ¼ ½ am ðbL  bR Þ gLm gRm T (15)

An output y is constructed from the biped feedback information q as

y ¼ ½a ðbL  bR Þ gL gR  T ¼ C y q (16)

The desired torque is derived as the output of a controller that operates on the error
between yd and y. That is
 
Md ðsÞ ¼ GC ðsÞ yd ðsÞ  yðsÞ
(17)
Md ¼ ½Md1 Md2 Md3 Md4 T

4.3 Dynamics of Actuators

We note that the actuator states xa ¼ M are the torques that will drive the biped
robot. The goal is to find ua such that M ! Md. We assume that DC motors are used
as actuators. It follows that (2) can be decoupled into individual motors represent-
ing the first-order stator-rotor dynamics (aai, bai) that generates output torque (xai)
while subject to disturbance (dai), that is
xa ¼ ½xa1 xa2 xa3 xa4 T ; ua ¼ ½ua1 ua2 ua3 ua4 T
Aa ¼ diagfaa1 aa2 aa3 aa4 g; Ba ¼ diagfba1 ba2 ba3 ba4 g
_ text Þ ¼ ½da1
da ðq; q; da2 da3 da4  T

where Aa, Ba and da are the uncertain parameter values.

17 Adaptive and Neural Learning for Biped Robot Actuator Control 221

4.4 Configuration of MRAC Actuator

The control specified by

ua ¼ Lxa þ N^ um
L ¼ diagfl1 ; l2 ; l3 ; l4 g; (18)
N ¼ diagfn1 ; n2 ; n3 ; n4 g

It follows from the Lyapunov design that the gains are adjusted according to
1 1
li ¼ p ðMdi  xai Þxai ; ni ¼  p ðMdi  xai Þ^umi (19)
bai qi i bai r i i

4.5 Convergence Analysis of MRAC

The convergence of the MRAC depends on the following dynamics

e_ i ¼ ami ei þ bai xai ~li  bai u _ text Þ
^mi n~i  dai ðq; q;
~l_i ¼  pi xai ei ; p^ umi (20)
~
ni ¼ i e i
bai qi bai r i
where ei ¼ xm  xai, l~i ¼ li  l^i , n~i ¼ ni  n^i , and li∗ and ni∗ are the true values of
parameters li and ni, whose dynamics2 are characterized by a third order polynomial
px p u^2
with coefficients given by  ami ; iqiai þ irimi and 0. We would assign the first two
coefficients as the convergence characteristics. The first coefficient in turn defines
ami. We next choose a large value for si to ensure that the Lyapunov rate is satisfied,

_ t ext Þ < 0
v_ i ¼ si e2i  2ami ei dai ðq; q; (21)

We can now compute pi, qi, and ri for the algebraic Lyapunov function as

si 2pi x2a 2pi u2m

pi ¼  ; qi ¼ ; ri ¼
2ami c2 c2

The analysis here was used to assist tuning of the parameters.

4.6 Neural Network Learning

The purpose of the NN is to eventually replace the necessity of human input (yd).
After the MRAC converges, the NN is trained to generate an estimate ^yd of yd using
222 P.K. Vempaty et al.

y and y_ as inputs. The basis for the mapping comes from the closed-loop system
formed by (1), (3) and The closed-loop characteristic can be expressed in a generic
_ q
form by f ðq; q; €; xa ; yd ; FÞ ¼ 0. The NN is an approximate mapping of f1(y, y,
_ yd)
¼ 0, where f1 is a reduced version of f.

5 Simulation Results

The biped movements from (16) are measured and used to generate the desired
torques from (17), (18) and (19) were implemented as the MRAC scheme. Three
sets of simulation runs are shown below. The first simulation does not include the
disturbance torque da(q, q,_ text). We introduced the disturbance in the second
simulation [13]. The third simulation shows the estimation performance of the
NN in predicting the reference command.

5.1 First Simulation (Without Disturbance)

5.1.1 Results of the Biped Moments M and Desired Biped Moments Md

Figure 5 shows MR1 converging to Md2 for one step cycle. Similarly, it can be noted
that the rest of the biped moments converge to the reference model’s behavior.
Therefore, this gives the biped walker a human-like gait as described by Md.

5.1.2 Results of the Biped Walking Simulation

Figure 6 plots the stable torso height y0 (see Fig. 4) of the biped walking as the result
of the adapted torques M.

5.2 Second Simulation (Disturbance)

5.2.1 Results of the Biped Walking with Torque Disturbance

Equation (2) includes disturbance torques da(q, q,_ text) which represents torque
feedback and other external moments. This causes the dynamics of the actuator to
vary. Figure 7 shows the biped torso height y0, recovering from the impact due to
17 Adaptive and Neural Learning for Biped Robot Actuator Control 223

Fig. 5 MR1 tracking Md2

Fig. 6 Height of the biped torso y0

the external disturbance to the biped walker, introduced by four impulses at 0.05,
0.3, 1.34, and 2.2 s with a magnitude of 5 and 6 Nm.

5.3 Third Simulation (Neural Network Estimation)

Figure 8 shows an 8-15-1 neural network model with y & y_ as the inputs and
estimate ^yd3 of the reference command signal yd3 as the output.
224 P.K. Vempaty et al.

_ text),
Fig. 7 Height of the biped torso without disturbance and with disturbance torques da(q, q,
applied at t ¼ 0.05, t ¼ 0.3, t ¼ 1.34, and t ¼ 2.2s

Fig. 8 Tracking performance of trained gL pattern to the reference pattern

6 Conclusions

In this paper, we presented an MRAC technique to ensure that the actuators reliably
produce desired torques necessary for a walking robot. An observer was used to
predict an anticipated state of the desired torque, thus causing the adaptive actuators
to anticipate motion torques. We provided a proof to show that the MRAC scheme
results in a stable system in the sense of Lyapunov when errors between the desired
and actuator torques are significant. Also, the convergence analysis for tuning p, q,
and r is provided. Simulation results verify that the system is robust when tested
with various parameters and unknown coefficients. This paper also investigates and
presents a neural network estimation of the human gait reference inputs. In the
future, we plan to implement the MRAC and neural network schemes to control a
real-time biped walker with transport delay.
17 Adaptive and Neural Learning for Biped Robot Actuator Control 225

References

1. E. Bobasu, D. Popescu, Adaptive nonlinear control algorithms of robotic manipulators.

Proceedings of the 7th WSEAS International Conference on Automation & Information,
2006, pp. 83–88
2. R. Chalodnan, D.B. Grimes, R.P.N. Rao, Learning to walk through imitation. Proceedings of
the 20th International Joint Conference on Artificial Intelligence, 2007, pp. 2084–2090
3. S. Ha, Y. Han, H. Hahn, Adaptive gait pattern generation of biped robot based on humans gait
pattern analysis. Proc. Int. J. Mech. Sys. Sci. Engr. 1(2), 80–85 (2008)
4. O. Haavisto, H. Hyotyniemi, Simulation tool of biped walking robot model. Master’s thesis,
Helsinki University of Technology, Espoo, Finland, 2004
5. J. Hu, J. Pratt, C.-M. Chew, H. Herr, G. Pratt, Adaptive virtual model control of a bipedal
walking robot. Int. J. Artif. Intel. Tools 8(3), 337–348(1999)
6. S. Kajita, F. Kanehiro, K. Kaneko, K. Fijiwara, K. Yokoi, H. Hirukawa, A realtime pattern
generator for biped walking. Proceedings of the IEEE International Conference on Robotics &
Automation, vol. 1, 2002, pp. 31–37
7. S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Harada, K. Yokoi, H. Hirukawa, Biped
walking pattern generation by using preview control of zero-moment point. Proceedings of
the IEEE International Conference on Robotics & Automation, vol. 2, 2003, pp. 1620–1626
8. P. Manoonpong, T. Geng, T. Kulvicius, B. Porr, F. Worgotter, Adaptive, fast walking in a biped
robot under neuronal control and learning. PLos Computat. Biol. 3(7), 1305–1320(2007)
9. J. Morimoto, G. Cheng, C.-G. Atkeson,G. Zeglen, A simple reinforcement learning algorithm
for biped walking. Proceedings of the IEEE Conference on Robotics and Automation, vol. 3,
2004, pp. 3030–3035
10. J. Nakanishi, J. Morimoto, G. Endo, G. Cheng, S. Schaal, M. Kawato, Learning from
demonstration and adaptation of biped locomotion with dynamical movement primitives.
Proceedings of the 4th IEEE/RAS International Conference on Humanoid Robotics, vol. 2,
2004, pp. 925–940
11. J. Pratt, C.-M. Chew, A. Torres, P. Dilworth, G. Pratt, Virtual model control: an intuitive
approach for bipedal locomotion. Int. J. Robot. Res. 20(2), 129–143 (2001)
12. A. Takanobu, S. Jun, K. Shigenobu, Reinforcement learning for biped walking using human
demonstration towards the human like walking of biped robot. Chino Shisutemu Shinpojiumu
Shiryo, 32, 393–398 (2005)
13. P.K. Vempaty, K.C. Cheok, R.N.K. Loh, Model reference adaptive control for actuators of a
biped robot locomotion. Proceedings of the World Congress on Engineering and Computer
Science, San Francisco, USA, vol. 2, 2009, pp. 983–988
14. K. Wolff, P. Nordin, Evolutionary learning from first principles of biped walking on a
simulated humanoid robot. Proceedings of the Business and Industry Symposium of the
Simulation Technologies Conference, 2003, pp. 31–36
15. C. Zhou, Q. Meng, Dynamic balance of a biped robot using fuzzy reinforcement learning
agents. Fuzzy Sets Syst. Arch. 134, 169–187 (2003)
Chapter 18
Modeling, Simulation, and Analysis
for Battery Electric Vehicles

Wei Zhan, Make McDermott, Behbood Zoghi, and Muhammad Hasan

Abstract Steady state and dynamic vehicle models are derived for analyses of
requirements on motors and batteries for battery electric vehicles. Vehicle level
performance requirements such as driving range, maximum cruise speed, maximum
gradeability, and maximum acceleration are used to analyze the requirements on
motors and batteries including motor power and torque, battery weight and specific
energy. MATLAB simulation tools are developed to allow validation of these
vehicle level performance requirements for a given set of motor/battery.

1 Introduction

Even though the internal combustion engine (ICE) is currently still the dominant
power source for automobiles, the cost of fuel and more stringent government
regulations on greenhouse gas emissions have led to more active interest in hybrid
and electric vehicles. Hybrid vehicles have better gas mileage than ICE-powered
vehicles. But they still have greenhouse gas emissions and the dual power sources
make them more complex and expensive. Battery electric vehicles (BEV) have
zero emission with a single power source that makes their design, control, and
maintenance relatively simple compared to hybrid vehicles. In addition, the
wide use of BEVs will reduce dependence on foreign oil, lower the cost per mile
of driving, and can potentially reduce the cost of electricity by using the vehicle-
to-grid power capability.

W. Zhan (*)
Department of Engineering Technology and Industrial Distribution, Texas A&M University,
College Station, TX 77843-3367, USA
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 227

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_18, # Springer ScienceþBusiness Media B.V. 2010
228 W. Zhan et al.

The main limitations on BEVs lie in the battery technology. The low energy and
power densities of batteries compared to hydrocarbon fuels significantly reduce the
driving range and other vehicle level performances. Initial cost is another factor that
slows the commercialization of electric vehicles. However, the latest developments
in battery technologies are making BEVs more and more attractive. From lead-acid
batteries to nickel metal-hydride (NiMH) batteries, lithium-ion cell technology [1],
and the latest nano-technology based batteries [2], the energy and power densities
have improved drastically.
The U.S. government is investing heavily to support battery technologies and
infrastructure for electric vehicles. It has set a target of one million electric vehicles
on U.S. roads by 2012. Tax credits up to $7,500 for buyers of plug-in electric
vehicles are offered by the U.S. government. The private sector is also investing
billions of dollars in electric vehicles. GM and Ford are both planning to roll out
plug-in electric vehicles in 2010. All these developments point to a trend toward
electric vehicles in the auto industry.
Not only are an increasing number of new BEVs being manufactured, there is
also significant interest in converting existing ICE-powered vehicles to electric
power. A Google search for “conversion to EV” results in millions of websites
and books, many of them providing Do it Yourself kits with focus on removal
and addition of components [3]. There is increasing interest in academia in the
development of BEVs; for example, an undergraduate senior design project devel-
oped an electric vehicle conversion [4]. However, many of these efforts lack
consideration of detailed system design requirements. Most of the components
for conversions are selected to provide similar output to that of the ICE or
by using the vehicle weight to determine the energy and power requirements. The
conversion to electric propulsion is a complex process and requires analysis that can
be very different from that of an ICE-powered vehicle [5]. Many performance
objectives impose conflicting demands. If not designed carefully, the resulting
electric vehicle can have many problems such as driving range shorter than
expected, battery and motor lacking enough power for desired acceleration, and
safety-related design problems.
In this paper, first principle models are derived for electric vehicles and used to
establish quantitative design requirements. Software tools are developed in
MATLAB [6, 7] to allow users to quickly determine expected vehicle level
performances for a given set of motors and batteries. They can also be used to
conduct trade-off studies for many design parameters.

2 Steady State Analysis

One of the system level requirements for BEV is the driving range at constant
speed. This requirement can be used to derive motor power and battery power and
energy requirements. Since the vehicle is assumed to be moving at a constant speed,
steady state analysis can be use to study this problem.
18 Modeling, Simulation, and Analysis for Battery Electric Vehicles 229

Fig. 1 Road load force

components Wsin(Θ)
DA RX

Θ W

The energy required to move a vehicle is determined by the distance it travels

and the force it has to overcome. The road load force the vehicle must overcome to
move the given distance has three components [8, 9], as illustrated in Fig. 1:
l The component of the gravity force in the direction of travel, if it is an inclined
path (Wsin(y))
l The aerodynamic drag (DA)
l The rolling resistance (Rx)

2.1 Projected Gravity Force

The gravity force is decomposed into two components, one in the direction of travel
and the other in the direction perpendicular to the surface. In order to move the
vehicle up the inclined surface, the vehicle must overcome the gravity force
component in the direction of travel. This is given by

Wx ¼ WsinðyÞ (1)

where W is the gravity force, y is the angle of the inclined surface, and Wx is the
component of the gravity force in the direction of travel.

2.2 Aerodynamic Drag

The drag is a function of speed for any given vehicle. At low speed the drag force is
negligible. At high speed, the drag becomes a significant factor. For simplicity, a
semi-empirical model is used here [9]

1
DA ¼ rV 2 CD A (2)
2

where V is the vehicle speed (ft/s), A is the frontal area of the vehicle (ft2), CD is the
aerodynamic drag coefficient, DA is the aerodynamic drag (lb), r is the air density
230 W. Zhan et al.

(lb-s2/ft4). A nominal value of r ¼ 0.00236 (lb-s2/ft4) and an estimated value of

0.45 for CD are used in this paper.

2.3 The Rolling Resistance

Rolling resistance of the tires is a major vehicle resistance force. It is the dominant
motion resistance force at low speed (<50 mph). The rolling resistance can be
modeled as the load on the tires multiplied by the coefficient of rolling resistance fr:

Rx ¼ f r W (3)

The coefficient of rolling resistance is affected by tire construction, tire temper-

ature, vehicle speed, road surface, and tire pressure. For instance, the rolling
resistance coefficient changes as the temperature changes. To simplify our analysis,
the coefficient of rolling resistance is assumed constant. We use 0.015 as the
nominal value.

2.4 Power Required

Based on the above analysis, the power required to drive the vehicle at a given
speed V (mph) is given by the total road load forces multiplied by the vehicle
speed, i.e.,

HP ¼ 0:00267ðDA þ Rx þ Wx ÞV (4)

where Wx can be calculated using (1), DA is given by (2), Rx can be calculated using
(3) with fr ¼ 0.015, and 0.00267 is the conversion factor to horsepower, HP. To
calculate these quantities, we need the following inputs:
l Vehicle speed (mph)
l Vehicle weight (including trailer if there is one) (lb)
l Frontal area of the vehicle, (including trailer if there is one) (ft2)
l Aerodynamic drag coefficient (including trailer if there is one)
l Coefficient of rolling resistance
l Surface incline angle (degree)

2.5 Energy Required

Energy is power integrated over time. If the total distance traveled is long enough,
the initial acceleration and final deceleration have negligible effect on the total
18 Modeling, Simulation, and Analysis for Battery Electric Vehicles 231

energy calculation. Also, since this is a steady state analysis the aerodynamic drag
is constant. Noting that Wx ¼ Wsin (y) and V ¼ dx/dt, it follows that
ð ð ð
dx
Wx dt ¼ WsinðyÞdx ¼ WsinðyÞdx ¼ WDh
dt
where Dh is the change in elevation between the starting and ending points. Thus,
the energy required to move a vehicle for a distance of d (miles) at a speed V (mph)
with a change in elevation of Dh (miles) is given by
EðkWhÞ ¼ 0:00267½ðDA þ Rx Þd þ WDh  0:746ðkWÞ=HP
(5)
¼ 0:002½ðDA þ Rx Þd þ WDh:
Define y* as the average slope; i.e.,

siny ¼ Dh=d

and

Wx  ¼ Wsiny

Then the trip energy becomes

EðkWhÞ ¼ 0:002ðDA þ Rx þ Wx  Þd: (6)

If speed is not constant, Eqs. (5) and (6) do not apply and the power consumed to
overcome drag must be evaluated as an integral.
The energy calculations in (6) can be converted to MJ (1 kWh ¼ 3.6 MJ):
EðMJÞ ¼ 0:0072ðDA þ Rx þ Wx  Þd: (7)

2.6 Battery Specific Energy

The total energy required for driving a vehicle at constant speed over a given range
can be used to derive the requirement on battery specific energy Dse (MJ/kg).
Let the battery weight be Wb (lb) and the battery/motor delivery efficiency be .
The total available energy ET (MJ) from the battery/electric motor is determined by

ET ðMJÞ ¼ 0:455ðkg=1bÞ  Dse ðMJ=KgÞ  Wbð1bÞ  : (8)

This amount must be greater than or equal to the total energy required as given
in (7), i.e.,

0:0072ðDA þ Rx þ Wx Þd  0:455 Dse Wb : (9)

232 W. Zhan et al.

From this, one can solve for Dse required to travel a given distance

DA þ Rx þ Wx
Dse  0:0158 d: (10)
Wb

Alternatively, we can calculate the maximum distance dmax the vehicle can
travel when the specific energy is given

Wb
dmax ¼ 63:29 Dse : (11)
DA þ Rx þ Wx

Note that the battery weight Wb is part of the vehicle weight W, which is used in
the calculation of Rx and Wx. Denoting the vehicle weight without the battery by W0
(lb), we have

W ¼ W0 þ Wb (12)

Combining (9) and (12), we get

Wb ½63:29Dse  dðsin y þ fr Þ  d½DA þ ðsin y þ fr ÞW0 : (13)

Since the right hand side is positive, and the battery weight must be positive, we
must have

63:29Dse
d< (14)
sin y þ fr

The right hand side of (14) provides a theoretical upper bound for the distance a
vehicle can travel with infinite battery weight (energy) regardless of the speed,
vehicle weight, and aerodynamic drag.
Under the assumption that (14) holds, one can determine the weight of a battery
in order to travel a distance d

DA þ ðsiny þ fr ÞW0
Wb  d (15)
63:29Dse  dðsiny þ fr Þ

One can conclude that, as long as (14) holds, a sufficiently heavy battery would
always enable the vehicle to travel a given distance d. In practice, there are other
constraints such as the volumetric limitation for the battery and vehicle load
capacity.
Using (12) in (10) yields
  
DA W0 d
Dse  0:0158 þ ðfr þ sinðy ÞÞ 1 þ (16)
Wb Wb 
18 Modeling, Simulation, and Analysis for Battery Electric Vehicles 233

Design constraint on battery and driving range (at 40 mph)

3000

2500

2000
Wbmin (lbs)

1500

d = 150 miles

1000

d = 100 miles

500

0
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
Specific Energy Density (MJ / kg)

Fig. 2 Design constraint among specific energy, driving range, and battery weight for a given speed

The relationship between minimum battery weight and specific energy is plotted
in Fig. 2 for different values of driving range. The efficiency for battery and motor
is assumed to be 75% [10] and the speed is 40 mph. For a given driving range, the
battery weight and specific energy must be chosen so that the point is above the
curve corresponding to the driving range.
Similarly, one can plot the design constraints for any fixed driving range and
various speed.

2.7 Maximum Cruise Speed

The maximum cruise speed that the vehicle is required to maintain imposes
requirements on the power delivered by the motor and batteries. The example
below is for a cruise speed of 80 mph.
Using the following vehicle parameters in the steady state model: vehicle speed ¼
80 mph; vehicle weight (including motors, but without engine, transmission, battery)
¼ 4,000 lb; frontal area of the vehicle ¼ 34 ft2; air temperature ¼ 59 F; atmospheric
234 W. Zhan et al.

Power requirement with battery weight = 1000 lbs

70

65
Motor Power Requirement (kWatt)

60

55

50

45

40

35

30
75 80 85 90 95
Max Cruise Speed (mph)

Fig. 3 Motor power as a function of max vehicle speed

pressure ¼ 29.92 in-Hg; aerodynamic drag coefficient ¼ 0.45; coefficient of rolling

resistance ¼ 0.015; efficiency ¼ 75%; surface incline angle ¼ 0 ; battery weight ¼
1,000 lbs, one can calculate the requirement on motor power for a maximum cruise
speed of 80 mph to be greater than 40 kW. The power requirement on the motor is
plotted as a function of maximum cruise speed in Fig. 3. If two motors are used, then
each motor should have rated power greater than 20 kW. The corresponding power
requirement on the batteries can be derived by dividing the motor power by the
efficiency. With a maximum cruise speed of 80 mph, the battery power can be
calculated as greater than 53 kW. This can be translated to the current/voltage
requirements on the batteries. For example, if the battery voltage output is 300 V,
then the current must be greater than 53,000/300 ¼ 177 A.
Figure 3 shows that the power is very sensitive to the maximum cruise speed.
One can also find the impact of the battery weight on the power requirement based
on maximum cruise speed. Simulation result shows that a 200% increase of battery
weight from 500 to 1,500 lbs only results in about 5% change in the battery power
requirement. A 53 kW total motor power can cover all realistic battery weights.
A similar conclusion holds for battery power required vs. battery weight: the impact
of the battery weight on the power required based on maximum speed is insignifi-
cant. It can be calculated that a 71 kW battery power is sufficient for any realistic
battery weight for an 80 mph cruise speed.
18 Modeling, Simulation, and Analysis for Battery Electric Vehicles 235

The maximum cruise speed scenario is used to determine the rated power for
batteries and motors. In other words, the battery and motor are required to provide
the power over a long period of time.

3 Dynamic Analysis

When the vehicle is driven over a short period of time or the time spent in
accelerating/decelerating is a significant portion of the total time, steady state
analysis is not adequate. Instead, dynamic analysis is needed. The dynamic model
developed in this section will be used to derive requirements on motor and battery
outputs.
The forces acting on the vehicle are illustrated in Fig. 4, where
l W is the gravity force.
l Rxf and Rxr are front and rear rolling resistant forces and Rxf þ Rxr¼Rx.
l Wf and Wr are front and rear normal forces.
l DA is the aerodynamic drag.
l LA is the aerodynamic lift.
l Fx is the tractive force (rear wheel drive is assumed).
Newton’s Second Law is applied in the direction of the vehicle movement and
the direction perpendicular to the road surface.

W
Fx  W sin y  DA  a  Rx ¼ 0 (17)
g

Wf þ Wr þ LA  W cos y ¼ 0 (18)

where the aerodynamic lift force is given by

1
LA ¼ rV 2 CL A (19)
2

LA
L/2 b
c DA

Wsin(Θ) ha
ma h
Wcos(Θ)

Wf
W
Wr Rxf
Fig. 4 Vehicle dynamics Rxr Fx
model Θ
236 W. Zhan et al.

The typical range [6] of values for the aerodynamic lift coefficient is CL ¼
0.0–0.5. The lift force is applied at the center of the wheel base.
A moment equation about the contact point at the front wheels can also be written.

Wr ðb þ cÞ  DA ha  ðWa=g þ W sin yÞh þ LðLA =2Þ  W cosðyÞb ¼ 0 (20)

There are two scenarios for the traction force Fx:

l Limited by the motor output (power limited)
l Limited by the road surface friction coefficient (traction limited)

3.1 Power Limited

During acceleration, the motor, transmission, and wheel dynamics have significant
impact on the vehicle acceleration.
Newton’s Second Law can be applied to the motor to get

Tm  T t ¼ I m a m (21)

where Tm is the total motor torque (Nm) (if two motors are used, the torque from
each motor needs to be multiplied by 2), Tt is the torque input to the drive train
(transmission and differential) (Nm), Im is the motor total rotational inertia (kg m2),
and am is the motor angular acceleration (rad/s2).
The torque delivered at the output of the drive train is amplified by the gear ratio
of the transmission times the gear ratio of the differential, but is decreased by the
torque required to accelerate the gears and shafts. If the drive train inertia is
characterized by its value on the input side, we have the following equation:

Tt  Tw =G ¼ It am (22)

where G is the gear ratio of the combined transmission and differential, i.e., the
ratio between the angular velocities of the motor shaft and the driven wheels, Tw is
the torque at the output of the drive train, and It is the rotational inertia for the drive
train.
Applying Newton’s Second Law to the driven wheels, one gets

T w  F x r ¼ I w aw (23)

where Fx is the total tractive force for the two driven wheels, r is the tire radius, Iw is
the rotational inertia of the wheels (Iw includes the rotational inertia of everything
down stream of the drive train), and is the wheel acceleration (rad/s2) of the wheels.
By the definition of gear ratio, we have

Gaw ¼ am (24)
18 Modeling, Simulation, and Analysis for Battery Electric Vehicles 237

Equations (21)–(24) could be combined to solve for the tractive force available
at the ground. Recognizing that for power limited operation the vehicle accelera-
tion, a, is the wheel rotational acceleration, aw, times the tire radius, yields:

G ðIm þ It ÞG2 þ Iw
Fx ¼ Tm  a (25)
r r2
where r is the radius of the tire.
The effect of mechanical losses can be approximated by including an efficiency
factor, t, in the first term in the right hand side of (26), as a result we have

G ðIm þ It ÞG2 þ Iw
Fx ¼  t Tm  a (26)
r r2

In order to see more clearly the effect of the rotational inertia on the vehicle
acceleration, (17) and (26) are combined to get the following

G
ðM þ Mr Þa ¼  Tm  Wsiny  DA  Rx (27)
r t

where M ¼ W/g and Mr is the equivalent mass of the rotating components, given by

ðIm þ It ÞG2 þ Iw
Mr ¼ (28)
r2

Clearly, the effect of the rotating components on the vehicle acceleration is

equivalent to increasing the vehicle mass by the amount in (28).

3.2 Traction Limited

In this case, the tractive force calculated by (26) exceeds the maximum tractive
force the road surface and the tire can generate, which is determined by

Fx ¼ mWr (29)

where m is the surface friction coefficient. For truck tires, on dry asphalt, m is
approximately 1.
From (17), (20), and (29), we can solve for the maximum tractive force Fxmax
 
m L
Fx max ¼ DA ha  ðDA þ Rx Þh  LA þ Wb cos y (30)
L  mh 2

In the model, these two cases can be combined by calculating the minimum of
the two forces in (26) and (30).
238 W. Zhan et al.

3.3 0–60 mph

The requirement on 0–60 mph time determines the maximum outputs from the
batteries and the motors. A 10 s 0–60 mph time is used as a vehicle acceleration
requirement. During the 10 s while the vehicle accelerates to 60 mph from a
standing start, the maximum power generated by the motor and battery can be
significantly higher than the rated values.
Figure 5 shows the characteristic of a typical AC induction motor. This can be
used together with the vehicle dynamics model to derive the maximum power
requirements on the motor and battery. The combined model can be used
to determine the 0–60 mph time. Based on the simulation result, one can find if a
specific motor/battery combination meets the requirement of 0–60 mph time.
In addition to the parameters used in the derivation of power requirements for
maximum speed, the following parameters are used:
Gear ratio ¼ 6; Motor inertia ¼ 0.36 kg m2 ¼ 3.18 lb in s2; Rotational inertia of
drive train ¼ 0.045 kg m2 ¼ 0.40 lb in s2; Rotational inertia of each driven wheel ¼
1.69 kg m2 ¼ 14.92 lb in s2; Drive train efficiency ¼ 96%; Wheel base ¼ 126 in;
CG height ¼ 3.3 ft; Height of aerodynamic drag ¼ 3.5 ft; Radius of loaded wheel ¼
13.82 in; Weight percentage on front wheels ¼ 55%; Motor torque and motor
efficiency curves given in Fig. 5; battery power limit ¼ 140 kW; surface mu ¼ 0.95;

Typical AC induction motor characteristics

200
power
180 efficiency
Torque output (Nm), efficiency (%), power (kW)

torque
160

140

120

100

80

60

40

20

0
0 2000 4000 6000 8000 10000 12000
Motor speed (rpm)

Fig. 5 Motor power, torque, and efficiency as a function of speed

18 Modeling, Simulation, and Analysis for Battery Electric Vehicles 239

where the gear ratio is the ratio between the angular velocities of the motor and the
driven wheels.
The simulation result in Fig. 6 shows that the vehicle can accelerate from 0 to 60
mph in 9.5 s.
By varying the battery weight and repeating the simulation, the 0–60 mph time
as a function of the battery weight is plotted in Fig. 7. It can be seen that the battery
weight has a significant impact on the 0–60 mph time.

3.4 Maximum Gradeability

Maximum gradeability is defined as the largest surface incline that a vehicle can
overcome. It is an important vehicle level performance requirement. Lower level
requirements on motor torque and battery power can be developed based on the
maximum gradeability requirement.
In the analysis of maximum gradeability, the following conditions hold:

a  0; DA ¼ 0; La ¼ 0

Battery weight = 1000 lbs

70

60

50
Vehicle speed (mph)

40

30

20

10

0
0 2 4 6 8 10 12
Time (seconds)

Fig. 6 Maximum vehicle acceleration

240 W. Zhan et al.

10.6

10.4

10.2

10
0-60 mph time (seconds)

9.8

9.6

9.4

9.2

8.8

8.6
500 600 700 800 900 1000 1100 1200 1300 1400 1500
Battery weight (lbs)

Fig. 7 Impact of battery weight on maximum acceleration

There are two cases one must consider: traction limited and power limited.
When the vehicle is traction limited, from (17), (20), and (29) we have
m
Fx ¼ W ðh sin y þ b cos yÞ  W sin y þ Rx (31)
L
From the above inequality, we can solve for the maximum surface incline when
it is traction limited
 
1 mb  Lfr
y  tan (32)
L  mh
When the vehicle is power limited, from (17) and (26), we have
G
Tm  W sin y þ Rx (33)
r
Note that t¼1 is assumed for maximum gradability analysis. Combining (32)
and (33), one can plot the maximum grade as a function of motor torque.

4 Conclusion

Several design requirements for electric vehicles are discussed in this paper. System
level design requirements are used to derive requirements on motor power and
torque and battery power, weight and specific energy based on simulation tools
18 Modeling, Simulation, and Analysis for Battery Electric Vehicles 241

developed using first principle models. A design constraint involving battery

weight, specific energy, and driving range is derived. The maximum cruise speed
requirement results in requirements on the rated motor torque, power, and battery
power. For a given motor and battery, the maximum acceleration can be simulated
to see if the 0–60 mph time requirement is met. The maximum gradeability
requirement can be used to derive requirements on maximum motor torque/battery
current. A software package EVSim [11] has been developed based on the analysis
in this paper. The simulation tool allows further investigation of design trade-offs
among different parameters. Future research includes the optimization of vehicle
performance and cost using the simulation tools developed in this paper, sensitivity
studies for the design parameters in the simulation model, and incorporating motor
dynamics [12] into our model.

References

1. G.A. Nazri, G. Pistoia, Lithium Batteries: Science and Technology (Kluwer/Plenum, Boston,
2004)
2. C.K. Chan, H. Peng, G. Liu, K. McIlwrath, X.F. Zhang, R.A. Huggins, Y. Cui,
High-performance lithium battery anodes using silicon nanowires. Nat. Nanotechnol. V(3)
31–35 (2008)
3. T. Lucas, F. Riess, How to Convert to an Electric Car (Crown Publishers, New York, 1980)
4. J. Dave, J. Dong, Conversion of an existing car to a rechargeable electric vehicle, ASEE
Annual Conference, 2009
5. I. Husain, Electric and Hybrid Vehicles (CRC Press, Boca Raton, FL, 2003)
6. R. Pratap, Getting Started with MATLAB 7: A Quick Introduction for Scientists and Engineers
(Oxford University Press, New York, 2006)
7. The MathWorks, Inc. MATLAB®7 Getting Started Guide (2008)
8. M. Ehsani, Y. Gao, S.E. Gay, A. Emadi, Modern Electric, Hybrid Electric, and Fuel Cell
Vehicles (CRC Press LLC, London, 2005)
9. T.D. Gillespie, Fundamentals of Vehicle Dynamics (Society of Automotive Engineers, Inc.,
Warrendale, PA, 1992)
10. Formula Hybrid Rules, SAE International, August 26, 2008. Available at http://www.formula-
hybrid.org/pdf/Formula-Hybrid-2009-Rules.pdf. Last accessed on 29 Dec 2009
11. W. Zhan, M. McDermott, M. Hasan, EVSim Software Design (Aug 2009)
12. W. Gao, E. Solodovnik, R. Dougal, Symbolically-aided model development for an induction
machine in virtual test bed. IEEE Trans. Energy Conver. 19(1) 125–135 (March 2004)
Chapter 19
Modeling Confined Jets with Particles and Swril*

Osama A. Marzouk, and E. David Huckaby

Abstract We present mathematical modeling and numerical simulations, using the

finite volume method, of a coaxial particle-laden airflow entering an expansion in a
vertical pipe. An Eulerian approach is used for the gas (air) phase, which is modeled
by the unsteady Favre-averaged Navier–Stokes equations. A Lagrangian model is
used for the dispersed (particles) phase. The results of the simulations using three
implementations of the k
e turbulence model (standard, renormalization group –
RNG, and realizable) are compared with measured axial profiles of the mean gas-
phase velocities. The standard model achieved the best overall performance. The
realizable model was unable to satisfactorily predict the radial velocity; it is also the
most computationally-expensive model. The simulations using the RNG model
predicted extra recirculation zones.

1 Overview

The use of computational fluid dynamics (CFD) to accurately model energy

production systems is a challenging task [1]. Of current interest, due to ever-
increasing energy demands, are coal-based energy systems such as pulverized
coal (PC) boilers and gasifiers with an emphasis on systems which provide for
carbon capture and storage (e.g. PC-oxyfuel). Turbulence and particle sub-models

*The support of this work by U.S. DOE Existing Plants Emissions and Capture program is
gratefully acknowledged. The first author was supported in part by an appointment to the U.S.
Department of Energy (DOE) Postgraduate Research Program at the National Energy Technology
Laboratory administered by the Oak Ridge Institute for Science and Education.
O.A. Marzouk (*)
United States Department of Energy, National Energy Technology Laboratory, 3610 Collins Ferry
Road, Morgantown, WV 26507-0880, USA
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 243

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_19, # Springer ScienceþBusiness Media B.V. 2010
244 O.A. Marzouk and E.D. Huckaby

are one of many sub-models which are required to calculate the behavior of these
gas–solid flow systems.
The particle-laden swirling flow experiment of Sommerfeld and Qiu [2] was
selected as a test-case to assess the performance of three implementations of
the k
e turbulence model for gas–solid flows. Previous numerical investigations
of this experiment include, Euler–Lagrange (EL)/Reynolds-averaged Navier–
Stokes (RANS-steady) [3], EL/large eddy simulations (LES) [4], Euler–Euler
(EE)/RANS-unsteady [5], and EE/LES [6]. The extensive experimental measure-
ments make this experiment a good test-case for gas–solid CFD.
A schematic of the experiment is shown in Fig. 1. The coaxial flow consists of
a primary central jet, laden with particles at a loading of 0.034 kg-particles/kg-air
and an annular secondary jet with a swirl number of 0.47 based on the inlet velocity.
Coaxial combustors have a similar configuration to this system. Generating a
swirling flow is an approach used to stabilize combustion and maintain a steady
flame [7]. Swirl entrains and recirculates a portion of the hot combustion products.
It also enhances the mixing of air and fuel. The inlet swirl number was calculated as
the ratio between the axial flux of angular momentum to the axial flux of linear
momentum
R Rsec
2 r Uy U x r 2 dr
S¼ 0
R Rsec (1)
Dcyl 0 r U 2x r dr

where Ux and Uy are the axial and tangential (swirl) velocities, Rsec ¼ 32mm is the
outer radius of the swirling secondary jet, and Dcyl ¼ 197mm is the inner diameter

Fig.1 Illustration of the coaxial flow (with swirl and particles)

19 Modeling Confined Jets with Particles and Swril 245

of the pipe into which the jets enter. The inlet Reynolds number (Re) is approxi-
mately 52,400 based on the outer diameter of the secondary jet, thus

^ x ð2 Rsec Þ
rU
Re ¼ (2)
m

where r is the density, m is the dynamic viscosity, and Ûx is an equivalent axial
velocity that accounts for the total volume flow rate from both primary and
secondary jets. The particles were small spherical glass beads, with a density of
2,500kg/m3. The beads injected according to a log-normal distribution with a mean
number diameter of 45 mm.

2 Gas Phase and Turbulence Models

The continuity and momentum equations for the resolved gas-phase fields are
expressed and solved in Cartesian coordinates as

@ r @ðr U j Þ
þ ¼0 (3)
@t @ xj

@ðr Ui Þ @ðr U i Uj Þ @ p @ðsij þ tij Þ

þ ¼
þ þ r gi þ Sp (4)
@t @ xj @ xi @ xj

where Ui is the velocity vector, p is the pressure, sij and tij are the viscous and
Reynolds (or turbulent) stress tensors, gi is gravitational vector (we only have
g1 ¼ 9.81m/s2), and Sp is a source term accounting for the momentum from the
dispersed phase. As for Newtonian fluids, sij is calculated as

sij ¼ 2m Sdev
ij

where Sijdev is the deviatoric (traceless) part of the strain-rate tensor Sij
 
1 @ Ui @ Uj
Sij ¼ þ
2 @ xj @ xi
 
1 @ Ui @ Uj 2 @ Uk
Sij ¼
dev
þ
dij
2 @ xj @ xi 3 @ xk

The tensor tij is not resolved directly. Instead, its effects are approximated using the
gradient transport hypothesis
  2
tij ¼ 2mt Sdev
ij
r k dij (5)
3
246 O.A. Marzouk and E.D. Huckaby

where mt is the turbulent (or eddy) viscosity. The spherical tensor on the right-hand
side is not considered here [3, 8]. Different eddy-viscosity turbulence models propose
different strategies to calculate mt. In the case of k
e models, mt is calculated as

k2
mt ¼ Cm r (6)
E
where k is the turbulent kinetic energy per unit mass and e is its dissipation rate.
They are calculated by solving two coupled transport equations. The forms of these
equations vary depending on the model implementation. We consider here three
implementations, which are described in the following subsections.

2.1 Standard k
e Model

The standard k
e model refers to the Jones–Launder form [9], without wall
damping functions, and with the empirical constants given by Launder and Sharma
[10]. The k and e equations are
  
@ðr kÞ @ðr Uj kÞ @ mt @ k
þ ¼ mþ þP
rE (7)
@t @ xj @ xj sk @ xj

  
@ðr EÞ @ðr U j EÞ @ mt @ E E
þ ¼ mþ þ ðCE1 G
CE2 rEÞ
@t @ xj @ xj sE @ xj k
 
2 @ Uk

CE1 þ CE3 r E (8)
3 @ xk

where P is the production rate of kinetic energy (per unit volume) due to the
gradients in the resolved velocity field

@ Ui
P ¼ tij
@ xj

which is evaluated as

2 @ Uk
P ¼ G
rk
3 @ xk

with
  !
@ Ui 1 @ Uk 2
G¼ 2mt Sdev ¼ 2mt Sij Sij

ij
@ xj 3 @ xk
19 Modeling Confined Jets with Particles and Swril 247

The addition of Ce3 in the last term on the right-hand side of (19.8) is not included in
the standard model. It was proposed [8, 11] for compressible turbulence. However,
we will refer to this implementation as the standard model. The model constants are

Cm ¼ 0:09; sk ¼ 1:0; sE ¼ 1:3; CE1 ¼ 1:44; CE2 ¼ 1:92; CE3 ¼
0:33

2.2 Renormalization Group (RNG) k
e Model

The RNG model was developed [12, 13] using techniques from the renormalization
group theory with scale expansions for the Reynolds stress. The k and e equations
have the same form in (7) and (8), but the constants have different values. In
addition, the constant Ce1 is replaced by C∗
e1, which is no longer a constant, but is
determined from an auxiliary function as

ð1
=0 Þ
CE1 ¼ CE1

1 þ b 3

where

k pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
¼ 2 Sij Sij
E

is the expansion parameter (ratio of the turbulent time scale to the mean-strain time
scale). The model constants are

Cm ¼ 0:0845; sk ¼ 0:7194; sE ¼ 0:7194; CE1 ¼ 1:42

CE2 ¼ 1:68; CE3 ¼
0:33; 0 ¼ 4:38; b ¼ 0:012

2.3 Realizable k
e Model

The realizable k
e model was formulated [14] such that the calculated normal
(diagonal) Reynolds stresses are positive definite and shear (off-diagonal) Reynolds
stresses satisfy the Schwarz inequality. Similar to the RNG model, the form of the k
equation is the same as the one in (7). In addition to altering the model constants,
the two main modifications lie in replacing the constant Cm used in calculating the
eddy viscosity in (6) by a function, and in changing the right-hand side (the
production and destruction terms) of the e equation. The last term in (8) is dropped.
With this, the e equation becomes
248 O.A. Marzouk and E.D. Huckaby

  
@ðr EÞ @ðr U j EÞ @ m @E E2
þ ¼ mþ t þ C1 r E S
CE2 r pffiffiffiffiffiffiffiffiffiffiffiffiffiffi (9)
@t @ xj @ xj sE @ xj k þ ðm=rÞE

where
 
 1
C1 ¼ max 0:43; ; Cm ¼
þ5 A0 þ AS ðU  k=EÞ
 
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 @ Ui @ Uj
U ¼ Sij Sij þ Oij Oij ; Oij ¼

2 @ xj @ xi

pffiffiffi 1 pffiffiffi 
AS ¼ 6 cosðfÞ; arccos 6 Wf¼
3
   
pffiffiffi Sij Sjk Sik 1 1
W ¼ min max 2 2 ;
pffiffiffi ; pffiffiffi
S3 6 6

and Oij is rate-of-rotation tensor. The model constants are

sk ¼ 1:0; sE ¼ 1:2; CE2 ¼ 1:9; A0 ¼ 4:0

3 Dispersed Phase

The Lagrangian equations of motion of a particle (in vector form) are

dx
¼u (10a)
dt

du
m ¼f (10b)
dt

where m is the mass of the particle, u is the particle velocity, and f is the force acting
on the particle. In this study; the drag, gravity, and buoyancy are considered, thus
the force f has the following form [15]:

p d2
f¼
r CD ju
U jðu
U Þ þ m g
r 8 g (11)
8

where d is the particle diameter, CD is the drag coefficient (which is a function of

the particle Reynolds number, Red, as will be described later), 8 is the particle
volume and U∗ is the instantaneous fluid velocity
19 Modeling Confined Jets with Particles and Swril 249

U  U þ U0 (12)

The vector U is the resolved velocity of the fluid (interpolated at the particle
location) which is calculated after solving the governing equations of the flow,
coupled with the turbulence model. The fluctuating velocity, U0 , is estimated using
the discrete random walk algorithm [16, 17]. In this algorithm, uncorrelated eddies
are generated randomly, but the particle trajectory is deterministic within an eddy.
The fluctuating velocity affects the particle over an interaction time, Tinterac, which
is the shortest of the eddy life time (Lagrangian integral time scale of turbulence),
Teddy, and the residence or transit time, Tresid. The latter is the time needed by the
particle to traverse the eddy. These characteristic times are calculated as

k
T eddy ¼ (13a)
E

k3=2
T resid ¼ Cresid (13b)
E ju
U
U 0 j

T interac ¼ min T eddy ; T resid (13c)

In (19.13b), U0 is lagged from the previous time step, and Cresid ¼ 0.093/4 ¼
0.16432. The turbulence information, thus the characteristic times in (13), are
updated every time step to account for the fact that the turbulence encountered by
a particle in its trajectory is not homogeneous.
The drag coefficient for the spherical particles is determined from the
following two-region formula [11], which is very similar to the Schiller–Nau-
mann[18] expression for Red  1,000, and uses a constant Newton drag coefficient
for Red > 1,000
8  
< 24 1 2=3
1 þ Red ; Red  1000
CD ¼ Red 6 (14)
:
0:424; Red > 1000

where the particle Reynolds number is defined as

rju
U jd
Red ¼ (15)
m

Combining (19.10a) and (19.11), and using m ¼ rppd3/6 (rp is the particle
density), the particle equations of motion become

dx
¼u (16a)
dt
250 O.A. Marzouk and E.D. Huckaby

!
du u
U r
¼
þ 1
g
dt t rp

where

4 rp d rp d2 24
t¼  ¼
3 r CD ju
U j 18 m Red CD

is the momentum relaxation time of the particle.

The particle position is tracked using the algorithm described by Macpherson
et al. [19]. The algorithm consists of a series of substeps in which the particle
position, x, is tracked within a cell using a forward (explicit) Euler scheme followed
by integrating the particle momentum equation using a backward (implicit) Euler
scheme to update the particle velocity, u. When calculating the resultant force due
to the particle on the fluid phase, the algorithm takes into account the particle
residence time in each cell. Interaction with the wall is represented through elastic
frictionless collisions.

4 Simulation Settings

The problem is treated as axisymmetric (although the results are mirrored in some
figures for better visualization). The domain starts at the expansion location with
x ¼ 0 in Fig. 1 and extends to x ¼ 1.0m. The domain is a 3D wedge (full-opening
angle 5∘), with a Frontal Area of 1.0m, 0.097m, and 240 and 182 mesh points in the
axial and radial directions, respectively. The mesh is nonuniform both axially and
radially, with finer resolution near walls and between the two jets. The mesh has
40,080 cells. The inlet condition for the velocity in the primary (inner) jet is
specified in terms of the mass flow rate (9.9g/s). For the secondary (outer) jet, the
inlet velocity is specified using the experimental velocity profile. A zero-gradient
condition is applied to the pressure at the inlet. The specific turbulent kinetic
energy, k, is set to 0.211m2/s2 and 0.567m2/s2 in the primary and secondary jets,
respectively; and the dissipation rate, e, is set to 0.796m2/s3 and 3.51m2/s3 in the
primary and secondary jets, respectively. The inlet k was estimated assuming 3%
turbulence intensity (the experimental value was not specified, but 3% is a reason-
able medium-turbulence level [20]) and the inlet e was then estimated from [21]

e ¼ C3=4
m k =l
1:5
(17)

where the standard value 0.09 is used for Cm, and l is the turbulence length scale,
which is approximated as 10% of the pipe diameter (l ¼ 0.02m). At the outlet,
zero-gradient conditions are applied for all variables except the pressure, where a
19 Modeling Confined Jets with Particles and Swril 251

constant value of 105 N/m2 is imposed. At the walls, the wall-function treatment is
used for the turbulence, and a zero-gradient condition is used for the pressure.
The PISO (pressure implicit splitting of operators) scheme is used to solve the
governing flow equations. A variable time step is adjusted dynamically to limit the
maximum convective CFL to 0.3. The backward Euler scheme is used for the time
integration of the flow equations. Upwind differencing is used for the convective
terms. Linear (second-order central difference) interpolation is used to find the mass
fluxes at the face centers of the cells from the cell-center values, and is also used for
the diffusion terms.
The particle mass flow rate is 0.34 g/s, which corresponds to 0.00472 g/s for our
case of 5 wedge. The particles are injected at a velocity of 12.5m/s, which is the
nominal axial inlet velocity in the primary jet.
We used version 1.5 of the finite volume open source code OpenFOAM [22, 23]
to perform all the simulations presented here. Similar particles are grouped into a
parcel (or a particle cloud), where they have a common velocity, to reduce the
amount of computation for the dispersed phase. The number of particles per parcel
is not uniform.

5 Results

The simulated flow time is 0.6 s for the results presented in this chapter. We have
found that this time interval is sufficient for all particles to traverse the domain and
for the gas phase to achieve stationary conditions. The last 0.1 s is used to obtain the
mean gas-phase velocities.
Figure 2 shows three snapshots of the parcels after 0.05, 0.1, and 0.15 s using the
standard k
e model (the diameters of the particles associated with a parcel are evenly
scaled by a factor of 100). This figure illustrates that the model captures well the
expected dynamics of the dispersed phase. The larger particles (with larger inertia)
maintain their axial motion and penetrate the central recirculation bubble. They are
not affected strongly by the swirl and radial velocity of the gas phase. Smaller
particles are entrained due to smaller relaxation times, and directed to the walls.
The mean axial, radial, and tangential velocities of the gas phase are shown in
Fig. 3. The negative mean axial velocity along the centerline and the walls identify
the regions of recirculation. The strong variations in all velocities are confined to a
distance of 150 mm after the inlet. The axial and tangential velocities exhibit initial
decay, whereas the radial velocity increases at the upstream boundary of the central
recirculation bubble.
A comparison between the mean streamlines obtained with the three turbu-
lence models is given in Fig. 4. Besides the central bubble, there are secondary
recirculation zones (due to the sudden expansion) at the top of the pipe. The
standard model gives the shortest recirculation bubble, with best agreement with
the experimental results. The realizable model gives a longer bubble, but with a
qualitatively similar structure. The RNG model resolves, in addition to the central
252 O.A. Marzouk and E.D. Huckaby

Fig. 2 Parcel motion with the standard k
e model at t ¼ 0.05s (left), t ¼ 0.10s (middle), and
t ¼ 0.15s (right)

Fig. 3 Mean gas-phase velocities near the inlet with the standard k
e model

and two secondary recirculation zones, two noticeable tertiary recirculation zones
at the beginning (top) of the central bubble. This feature was not reported in the
experimental results.
19 Modeling Confined Jets with Particles and Swril 253

Fig. 4 Comparison of the mean streamlines with 3 k
e models: standard (left), RNG (middle),
and realizable (right)

The mean gas-phase velocity fields are sampled at different axial stations. We
compare the mean velocities obtained using the three turbulence models with the
measured values at two stations in Fig. 5, located at x ¼ 3 and 112 mm. The first
station is located close to the inlet, thus upstream of the central bubble, whereas the
second one spans the bubble. At x ¼ 3 mm, all models predict axial and tangential
velocities that are in agreement with the measured ones. The radial velocity using
the realizable model shows considerable disparity, with excessively-negative
(inward) velocity in the region away from the jets, followed by an outward flow
near the wall, which is opposite to the inward flow observed experimentally. The
standard model has a slightly better agreement with the measurements than the
RNG model at this axial location. At x ¼ 112 mm, the RNG predictions deviate
considerably from the measurements. This is a direct consequence of the tertiary
recirculation shown in Fig. 4. As in the earlier station, the standard and realizable
models provide similar results except for the radial velocity, with the realizable
model failing to capture the measured peak at r  80 mm. The standard model
provides better prediction of the axial velocity in the vicinity of the wall than the
other two models.
On a computing machine with two processors: quad core Intel Xeon L5335
2.00 GHz, a simulation interval of 0.5 s required 17.13 h CPU time for the standard
model, 19.44 h for the RNG model, and 24.81 h for the realizable model. The
standard model has the lowest computational demand due to its relative simplicity.
The realizable model is the most computationally-expensive implementation, with
CPU time equal to 145% and 128% of the CPU times in the case of the standard and
RNG models, respectively.
254 O.A. Marzouk and E.D. Huckaby

At 3 mm after inlet At 3 mm after inlet

25
Experiment
2

Mean Radial Velocity (m / s)

20 Standard
Mean Axial Velocity (m / s)

RNG
Realizable
15
0

10

–2
5
Experiment
Standard
0 RNG
–4 Realizable
–5
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
Radial Distance (m) Radial Distance (m)

At 3 mm after inlet At 112 mm after inlet

20 9
Experiment Experiment
Mean Tangential Velocity (m / s)

Standard
Mean Axial Velocity (m / s)
15 Standard
RNG RNG
6 Realizable
Realizable
10
3
5

0
0

–5 –3
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
Radial Distance (m) Radial Distance (m)

At 112 mm after inlet At 112 mm after inlet

2 3
Experiment
Mean Tangential Velocity (m / s)

Standard
Mean Radial Velocity (m / s)

RNG
Realizable
2
1

Experiment
0 Standard
0 RNG
Realizable

0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
Radial Distance (m) Radial Distance (m)

Fig. 5 Comparison of the mean gas-phase velocities at x ¼ 3 and 112 mm

6 Conclusions

We simulated a coaxial airflow with a particle-laden primary jet and a swirling

annular jet, entering a sudden expansion. Three implementations of the k
e model:
standard, renormalization group (RNG), and realizable were applied. The standard
model had the best overall performance based on the mean gas-phase velocities.
The RNG model showed considerable deviations from the measurements in some
regions. The main drawback of the realizable model was its erroneous prediction of
19 Modeling Confined Jets with Particles and Swril 255

the radial velocity. The main differences in the predicted velocity profiles were
related to the different flow structures and mean streamlines. We should point out
that our finding that the more complex models did not outperform the simpler one
should not be considered a universal statement. In fact, more work is needed to
investigate the reasons of this unexpected outcome. This includes examining the
effect of Ce3, and the effect of dropping the spherical component in the modeled
Reynolds stresses (which causes violation of one of the realizability constraints).

References

1. T.F. Wall, Combustion processes for carbon capture. P. Combust. Inst. V31(N1), pp. 31–47
(2007)
2. M. Sommerfeld, H.-H. Qiu, Detailed measurements in a swirling particulate two-phase flow
by a phase-Doppler Anemometer. Int. J. Heat Fluid Fl. V12(N1), pp. 20–28 (1991)
3. M. Sommerfeld, A. Ando, D. Wennerberg, Swirling, particle-laden flows through a pipe
expansion. J. Fluids Eng. V114(N4), pp. 648–656 (1992)
4. S.V. Apte, K. Mahesh, P. Moin, J.C. Oefelein, Large-eddy simulation of swirling particle-
laden flows in a coaxial-jet combustor. Int. J. Multiphase Flow V29, pp. 1311–1331 (2003)
5. Y. Yu, L.X. Zhou, C.G. Zheng, Z.H. Liu, Simulation of swirling gas-particle flows using
different time scales for the closure of two-phase velocity correlation in the second-order
moment two-phase turbulence model. J. Fluids Eng. V125(N2), pp. 247–250 (2003)
6. M. Boileau, S. Pascaud, E. Riber, B. Cuenot, L.Y.M. Gicquel, T.J. Poinsot, M. Cazalens,
Investigation of two-fluid methods for large eddy simulation of spray combustion in gas
turbines. Flow Turbul. Combust. V80(N3), pp. 291–321 (2008)
7. G. Monnot, in Principles of Turbulent Fired Heat, Editions Technip, 1985
8. F.P. K€arrholm, Numerical modelling of diesel spray injection, turbulence interaction and
combustion. Ph.D. Dissertation, Chalmers University of Technology, G€ oteborg, Sweden,
2008
9. W.P. Jones, B.E. Launder, The prediction of laminarization with a two-equation model. Int. J.
Heat Mass Tran. V15(N2), pp. 301–314 (1972)
10. B.E. Launder, B.I. Sharma, Application of the energy-dissipation model of turbulence to the
calculation of flow near a spinning disk. Lett Heat Mass Tran., V1(N2), pp. 131–138 (1974)
11. P.A.N. Nordin, Complex chemistry modeling of diesel spray combustion. Ph.D. Dissertation,
Chalmers University of Technology, G€ oteborg, Sweden, 2001
12. V. Yakhot, S.A. Orszag, Renormalization group analysis of turbulence: 1. Basic theory. J. Sci.
Comput. V1(N1), pp. 3–51 (1986)
13. V. Yakhot, S.A. Orszag, S. Thangam, T.B. Gatski, C.G. Speziale, Development of turbulence
models for shear flows by a double expansion technique. Phys. Fluids A. V4(N7), pp.
1510–1520 (1992)
14. T.-H. Shih, W.W. Liou, A. Shabbir, J. Zhu, A new k
e eddy-viscosity model for high
reynolds number turbulent flows – model development and validation. Comput. Fluids V24
(N3), pp. 227–238 (1995)
15. S. Elghobashi, On predicting particle-laden turbulent flows. Appl. Sci. Res. V52(N4), pp.
309–329 (1994)
16. D. Milojević, Lagrangian stochastic-deterministic (LSD) predictions of particle dispersion in
turbulence. Part Part Syst. Charact., V7(N1-4), pp. 181–190 (1990)
17. D. Lakehal, On the modelling of multiphase turbulent flows for environmental and hydrody-
namic applications. Int. J. Multiphase Flow V28(N5), pp. 823–863 (2002)
18. L. Schiller, A. Naumann, Ûber die grundlegenden Berechungen bei der Schwerkraftaufber-
eitung. Zeit Ver Deut Ing V77, pp. 318–320 (1933)
256 O.A. Marzouk and E.D. Huckaby

19. G.B. Macpherson, N. Nordin, H.G. Weller, Particle tracking in unstructured, arbitrary poly-
hedral meshes for use in CFD and molecular dynamics. Commun. Numer. Meth. Eng. V25
(N3), pp. 263–273 (2009)
20. C. Sicot, P. Devinanta, S. Loyera, J. Hureaua, Rotational and turbulence effects on a wind
turbine blade. Investigation of the stall mechanisms. J. Wind Eng. Ind. Aerod. V96(N8-9), pp.
1320–1331 (2008)
21. J.H. Ferziger, M. Perić, in Computational Methods for Fluid Dynamics, 3rd edn. (Springer,
2002)
22. OpenFOAM: the open source CFD toolbox. Available from:http://www.openfoam.org
23. H.G. Weller, G. Tabor, H. Jasak, C. Fureby, A tensorial approach to computational continuum
mechanics using object-oriented techniques. Comput. Phys. V12(N6), pp. 620–631 (1998)
Chapter 20
Robust Tracking and Control of MIMO
Processes with Input Saturation and
Unknown Disturbance

Ajiboye Saheeb Osunleke and Mingcong Deng

Abstract In this chapter, the design of robust stabilization and output tracking
performance of multi-input multi-output processes with input saturations and
unknown disturbance are considered. The proposed control technique is the robust
anti-windup generalized predictive control (RAGPC) scheme for multivariable
processes. The proposed control scheme embodies both the optimal attributes of
generalized predictive control and the robust performance feature of operator-based
theoretic approach. As a result, a strongly robust stable feedback control system
with disturbance rejection feature and good tracking performance is achieved.

1 Introduction

The control of multivariable systems is paramount in nearly all industrial and

allied processes. This is so because of the need to optimize outputs at the expense
of limited available materials. In designing MIMO control systems, systems and
control engineers have to deal with some challenges which impose limitations on
the performance of MIMO systems.
In recent past, several techniques for control of MIMO systems have been
proposed and good reviews are available [1–3]. Among these techniques, the predic-
tive control approach is distinctive in its application to control of MIMO processes
which exhibit some of these performance limited characteristics such as input satu-
ration and disturbance. Predictive control strategy is now widely regarded as one of the
standard control paradigms for industrial processes and has continued to be a central

A.S. Osunleke (*)

Division of Industrial Innovation Sciences, Graduate School of Natural Science and Technology,
Okayama University, 3-1-1 Tsushima-Naka, 3-chome, Okayama 700-8530, Japan
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 257

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_20, # Springer ScienceþBusiness Media B.V. 2010
258 A.S. Osunleke and M. Deng

research focus in recent times [1, 4, 5. In practice, control systems have to deal with
disturbances of all kinds, such as stepwise load disturbances, high frequency sensor or
thermal noise, input saturations etc. These undesirable effects are systematically dealt
with using predictive control.
The new proposed control strategy under the generalized predictive control
scheme is known as multivariable robust anti-windup generalized predictive
control (MRAGPC). It is the multivariable extension of the robust anti-windup
generalized predictive control (RAGPC) scheme for SISO systems in earlier work
[8]. Like RAGPC scheme, MRAGPC also shares the optimal and robust perfor-
mance attributes characteristic of a good control strategy. This work proposes a
design procedure for Multivariable systems having both input saturation and dis-
turbance. The proposed MRAGPC design procedure can be implemented in four
different modes depending on the nature of problem at hand and the design control
objective. The design application entails firstly, the design of compensator to
compensate the effect of minimal phase and strictly properness as these are
characteristics of MIMO industrial processes. Secondly is to construct stable
coprime factors for the compensated system. Subsequently, anti-windup controllers
are designed to achieve a stable feedback closed loop control system. This proce-
dure is completed by the design of a tracking operator using operator-based
theoretic approach. In order to estimate the unknown disturbance in the system,
unknown disturbance estimation mechanism for MIMO system is also proposed.
Based on the outlined procedure, the robust performance of the closed-loop system
in the presence of input saturation and disturbances is ensured.
In this proposed design, the coupling effect of interactions between components
of the MIMO systems is curtailed by the appropriate choice of MRAGPC design
parameters. The improved performance of the proposed scheme is confirmed by
simulation of the model of a two-input two-output MIMO system.

2 MRAGPC Design Scheme

In order to control MIMO processes with all the design limitations, MRAGPC
scheme is implemented in three different modes depending on the type and nature
of problems at hand.
Mode 1 – known as Anti-windup mode, is implementable for the case of input
saturation but with no disturbances load on the input to the system.
Mode 2 – otherwise called compensation mode, is implementable for the case of
input saturation, non-minimal phase and disturbance load in the input to the system.
Mode 3 – referred to as the fault detection mode, for system with both input
saturation and unknown disturbance.
In this paper, all these modes of MRAGPC application are implemented.
20 Robust Tracking and Control of MIMO Processes with Input Saturation 259

2.1 MRAGPC Problem Formulation

The proposed linear multivariable system model is given by Deng et al. [4].

A½ pYð pÞ ¼ B½ pUð pÞ þ C½ pVð pÞ (1)

where Y( p), U( p), V( p) are m  1 output, q  1 input and p  1 disturbance

vectors respectively. B[ p] is a m  q polynomial matrix while A[ p] and C[ p] are
m  m diagonal polynomial matrices. As before, the polynomial matrices A[ p] and
B[ p], A[ p] and C[ p] are coprime. The elements of C[ p] are with a degree of one
less or equal to that of the corresponding elements of A[p] and are chosen by the
designer. The model in (1) corresponds to a Left Matrix Fraction Description
(LMFD) described as

H0 ðpÞ ¼ A1 ½ pB½ p (2)

Equation (2) can be equivalently expressed as

H0 ð pÞ ¼ D0 1 ð pÞN0 ð pÞ (3)

where P0( p) is the nominal multivariable plant, D0(p) and N0(p) are coprime
factors of P0( p) and are defined as
1 1
~ ½ pB½
N0 ð pÞ ¼ D ~ p; D0 ð pÞ ¼ D
~ ½ pA½
~ p (4)
~
where DðpÞ is a Hurwitz polynomial matrix with degree equal to that of AðpÞ. ~
~
The polynomial matrix BðpÞ is the same as B(p) for non-strictly proper and minimal
system. The control input uj(t) is subject to the following constraint (Fig. 1)

umin;j  uj ðtÞ  umax;j ðj ¼ 1; 2; : : : ; mÞ (5)

The constraint (7) is equivalently expressed as

0 1
sð
u1 ðtÞÞ
B .. C
uðtÞ ¼ sðu1 Þ ¼ @ . A; (6)
sð
um ðtÞÞ

u max
σ(v)
u(t) y(t)
v
umin

LIMITER
Fig. 1 Input saturation model
260 A.S. Osunleke and M. Deng

where u1 is the ideal controller output vector. The objective is to design stable
controllers using coprime factorization and Youla-Kucera parameterization [5] for
the above process.

2.2 Controllers Parameterization

The proposed design scheme for the MIMO systems follows four steps as high-
lighted below.

2.2.1 Design of Stable Feedback MCGPC Controllers V ( p) and U ( p)

with Input Constraint Part (Fig. 2)

By adapting the MCGPC design method to the anti-windup design approach, the
proposed controllers are given by Youla-Kucera parameterization as follows.

UðpÞ ¼ XðpÞ þ QðpÞDðpÞ (7)

VðpÞ ¼ YðpÞ  QðpÞNðpÞ (8)

where Q(p) ∈ RH1 is a design parameter matrix for ensuring a strongly stable
feedback controllers in the above and is given by

QðpÞ ¼ U1
d ð pÞUn ð pÞ (9)

Δ(t)∈U

MCAGPC Closed-Loop D0–1

e (t)∈U
r(t) ∈Y + + + u1(t) σ (v) u (t)∈U +
D0–1 N0
– – + +

Ip-V

Fig. 2 Proposed MIMO control structure

20 Robust Tracking and Control of MIMO Processes with Input Saturation 261

N( p), D( p) are the coprime factor representation of D0–1 ( p), X( p), Y( p) ∈ RH1
are operators chosen to satisfy the Bezout Identity

XðpÞNðpÞ þ YðpÞDðpÞ ¼ Im (10)

^ XðpÞ
NðpÞ ^ ^ YðpÞ
þ DðpÞ ^ ¼ Im (11)

^
XðpÞ; ^
YðpÞ; XðpÞ; YðpÞ 2 RH1 (12)

^ pÞ and Dð
where the coprime factorization N( p) and D( p), Nð ^ pÞof the CAGPC
Plant can be chosen as follows.

H0 ð pÞ ¼ D1 ~ 1 1 ~
0 ð pÞ ¼ D½ pA ½ p ¼ A ½ pD½ p (13)

From (13), defining the stable closed-loop characteristic polynomial matrices

^ c ½p, then we choose N(p) and
of the process without input constraint as Tc ½p and T
^
D(p), NðpÞ ^
and DðpÞ as

^
NðpÞ ¼ D½ ^
^ 1 ½ pD½
~ pT ~ p; NðpÞ
^ ¼ T1 ~
c c ½ pD½ p (14)

^ T
DðpÞ ¼ A½p
1^
^ ½pD½p;
~ DðpÞ ¼ T1
c c ½ pA½ p (15)

XðpÞ ¼ Ke Re þ Ke Ce ðpÞFðpÞ (16)

YðpÞ ¼ Im þ Ke Ce ð pÞGð pÞ (17)

Tc[p] is the closed-loop characteristic polynomial matrix for the control system
in Fig. 2 and is given by

~ p
Tc ½ p ¼ A½ p þ Ke Lð pÞ þ Ke Re D½ (18)

where Ke, Ce, Re, G[p] and F[p] are given in [4].

2.2.2 Robust Stability of D0–1 (p) with Input Constraint

The predictive anti-windup part is quasi-linear with respect to the effect of input
constraint. The closed loop transfer function of this part of the control structure is
always stable and invertible. Therefore, it can be seen that the plant retains a robust
rcf and it is stable. Then, we can design a Bezout operator S(p) to satisfy the
following Bezout Identity [7]

Rð pÞD0 ð pÞ þ Sð pÞN0 ð pÞ ¼ Im (19)

262 A.S. Osunleke and M. Deng

2.2.3 Tracking Operator Design

Here a tracking operator M(p) for reference signal r(t) such that plant output y(t)
tracks to the setpoint signal r(t) as depicted in Fig. 3. Based on the proposed lemma
and proof in [7], operator M(p) is designed to satisfy

N0 ð pÞðMð pÞrð pÞ þ Rð pÞDð pÞÞ ¼ Im rð pÞ (20)

2.2.4 Decoupling of the Control System

In the proposed control scheme, the possible interactions between different control
loops are identified as input-input, disturbance-input and input-output. In order to
minimize the effect of these interactions, the following design steps are proposed.
l Choosing larger value for control orders and/or smaller value for the maximum
prediction horizons and/or using reference models for the output in the design
step 1.
l Design a decoupling matrix GD( p) such that loop interactions are decoupled.
The elements of GD( p) can be selected in a number of ways to achieve optimally
decoupled control system. This is further investigated in the future works.

3 Additional Design Schemes for MRAGPC

In this section, we present two other design features to enhance the performance of
the proposed method for MIMO processes similarly to that obtained for SISO [8]
namely, the robust parallel compensator and unknown fault detection schemes.

3.1 Robust Parallel Compensator (RPC) Scheme

for MIMO Processes

In an extension to the design of robust parallel compensator designed for SISO

system [9], we present in this section, an equivalent form of this compensator for

Δ(t)∈U

R0

r(t)∈Y + u(t)∈U w(t)∈U y(t)∈Y

M Im N0
+

Fig. 3 Robust tracking structure for MIMO system

20 Robust Tracking and Control of MIMO Processes with Input Saturation 263

MIMO processes. Here, a RPC, F( p) (see Fig. 4) is designed for all classes of system
which exhibit non-minimal phase and non-strictly properness. The condition for this
design to satisfy static output feedback is the same as in SISO [10] that is:
l The process must be inversely stable.
l Its relative degree must be zero i.e. non-strictly properness.
l The leading coefficient must be positive.
For the process represented by (2), we propose the following GRPC design
scheme.

Fð pÞ ¼ F1 d
d ½ pFn ½ pPp (21)

where
Fn[ p] and Fd [ p] are (g  i)th and (g i + 1)th order monic stable polynomial
matrices respectively and are as given as

Fn ½ p ¼ bgk;i;j pgk þ    þ b0;i;j ; i ¼ 1; 2; . . . ; m; j ¼ 1; 2; . . . ; q (22)

Fd ½ p ¼ agkþ1;i;j pgkþ1 þ    þ a0;i;j ; i ¼ 1; 2; . . . ; m; j ¼ 1; 2; . . . ; q (23)

and Pdp is a diagonal matrix m  q of element pd where d is the degree of

polynomial in p.
~ pÞ be the augmented system matrix defined as
Let Hð

~ pÞ ¼ Hð pÞ þ Fð pÞ
Hð (24)

Equivalently,

Hð ~ 1 ½ pB½
~ pÞ ¼ A ~ p (25)

~
Augmented Plant H
Δ(t)∈U
F

y(t)∈Y
r(t)∈Y + e(t)∈U σ(v) +
H
– +

Anti-windup part
b∈U

Fig. 4 Robust parallel compensator design for MIMO system

264 A.S. Osunleke and M. Deng

In summary, the coefficients bk,i,j > 0, ak,i,j > 0 in (25) above are determined
such that
~
1. B½p, ~ ∈ RH1 for sufficiently large b0,i,j.
A½p
2. The augmented plant Hð ~ pÞ has the same steady state gain as the nominal plant
Hð pÞ.
~
3. There exists positive constants ak,i,j such that the augmented system HðpÞ in (25)
will satisfy the sufficient condition of SOF for any small positive integer value of
d and a0,i,j > 0.
For the RAGPC design scheme, it is just sufficient to choose d ¼ 1 and g as
minimum as possible while the unknown coefficients are designed satisfying the
above stated conditions using Kharitonov’s theorem [11]. Without loss of generality,
~ pÞ have a relative degree equal zero and a positive
it is obvious from (25) that Hð
~ pÞbeing SOF stabi-
leading coefficient thereby satisfying all the conditions of Hð
lized and almost strictly positive real (ASPR).

3.2 Unknown Disturbance Estimation Scheme

for MIMO Processes

As noted in earlier work [2], the unknown disturbance of the system is estimated as
fault signals acting on the process input. If stable low-order operators S0 and R0 can
be designed such that the pair [S0, R0]  [S, R] (see Fig. 5). Then the following
equations enable us to detect the fault signal [2]

u0 ¼ R0 ud ðtÞ þ S0 ðya ÞðtÞ
(26)
yd ¼ D0 ðu0 ÞðtÞ

Then as shown in Fig. 3, the fault signal is estimated as

~ ¼ absðR1 ðu0 ðtÞ  rm ðtÞÞÞ

D (27)

Detected Fault
signal
~
+ R–1 Δ (t)
Fault signal
Δ(t) u 0(t)
R0 + S0
R
Compensated system ya(t)∈Y
r(t)∈Y u 0(t) +e(t)∈U w(t)∈U y(t)∈Y
M1 ~ ~
+ R–1 D−1 N
r m (t) –
Anti-windup part
b∈U
S

Fig. 5 Unknown disturbance estimation structure for MIMO system

20 Robust Tracking and Control of MIMO Processes with Input Saturation 265

4 Simulation Examples

The Mode 1 and Mode 2 implementation of MRAGPC are considered in this

section. The process models used are the same. The objective is to control this
MIMO system with and without disturbance load into the system.

4.1 Control of MIMO System Without Disturbance

The parameters of the 2-input 2-output system are given in Table 1 below.

4.2 Control of MIMO System with Disturbance

The process model (4.1) is considered here with disturbance load in the input. The
process and control parameters are as shown in Table 2.

Table 1 Simulation parameters for MIMO process without disturbance

System parameters:
A11 ¼ p2 þ p þ 0.1 A22 ¼ p2 þ 0.5p þ 0.06 B11 ¼ 0.2p þ 10
B12 ¼ 0.5p þ 2 B21 ¼ 0.1p þ 1.2 B22 ¼ 0.25p þ 6
C11 ¼ C22 ¼ p þ 0.8 m¼q¼2
Controller parameters:
Ny1 ¼ Ny2 ¼ 10 Nu1 ¼ Nu2 ¼ 3 rm1 ¼ rm2 ¼ 0.5
Un11 ¼ 0.3 Un22 ¼ 0.2 u1,1 ¼ u1,2 ¼ 0
u2,1 ¼ u2,2 ¼ 0.1 r1 ¼ 10, r2 ¼ 2 l1 ¼ l2 ¼ 0.1
Ud ¼ Iq T1,1 ¼T1,2 ¼ 0 T2,1 ¼T2,2 ¼ 6

Table 2 Simulation parameters for MIMO process with disturbance

System parameters:
A11 ¼ p2 þ p þ 0.1 A22 ¼ p2 þ 0.5p þ 0.06 B11 ¼ 0.2p þ 10
B12 ¼ 0.5 p þ 2 B21 ¼ 0.1 p þ 1.2 B22 ¼ 0.25 pþ 6
C11 ¼ C22 ¼ p þ 0.8 m ¼ q ¼ 2 D1 ¼ D2¼ 20
sin(1.8t)
Compensator parameters:
d ¼ 1, g ¼ 1 b0,11 ¼ 1.21 b0,12 ¼ 0.64 b0,21 ¼ 0.64 b0,22 ¼ 0.06
a,0,11 ¼ 1.476 a,0,12 ¼ 4.51 a,0,21 ¼ 0.88 a,0,22 ¼ 0.004
a,1,11 ¼ 3.105 a,1,12 ¼ 1.25 a,1,21 ¼ 1.25 a,1,22 ¼ 0.112
Controller parameters:
Ny1 ¼ Ny2 ¼ 12 Nu1 ¼ Nu2 ¼ 3 rm1 ¼ rm2 ¼ 1 Un11 ¼ Un22
¼ 0.45
u1,1 ¼ u1,2 ¼ 0 u2,1 ¼ u2,2 ¼ 0.1 r1 ¼ 10, r2 ¼ 2 l1 ¼ l2 ¼ 0.1
Ud ¼ Iq T1,1 ¼ T1,2 ¼ 0 T2,1 ¼ T2,2 ¼ 6
266 A.S. Osunleke and M. Deng

0.1

u1

0.05

0
0 20 40 60 80 100 120
Time(s)
0.1
u2

0.05

0
0 20 40 60 80 100 120
Time(s)

Fig. 6 Controlled inputs step responses for MIMO system without disturbance load

10

controlled output
reference input
y1

0
0 20 40 60 80 100 120
Time(s)

6
controlled output
referenced input
4
y2

0
0 20 40 60 80 100 120
Time(s)

Fig. 7 Step responses of controlled outputs to step references of +10 and +2 respectively for
MIMO process without disturbance load
20 Robust Tracking and Control of MIMO Processes with Input Saturation 267

0.1

u1

0.05

0
0 20 40 60 80 100 120
Time(s)

0.1
u2

0.05

0
0 20 40 60 80 100 120
Time(s)

Fig. 8 Controlled inputs step responses for MIMO system with disturbance load

10

controlled output
reference input
y1

0
0 20 40 60 80 100 120
Time(s)

6 controlled output
reference input
y2

0
0 20 40 60 80 100 120
Time(s)

Fig. 9 Step responses of controlled outputs to step references of +10 and +2 respectively for
MIMO process with disturbance load
268 A.S. Osunleke and M. Deng

Simulation results for the two cases considered are shown in Figs. 6 – 9. The
control inputs saturate at 0.1 and the controlled outputs tracks to the given step
reference inputs þ10 and þ2 for the two outputs respectively.

5 Conclusion

It is evident that the present control performance is better than the earlier work [12]
and that the coupling effects have been considerably checked by chosen appropriate
design parameters. Also, the designing of unknown disturbance estimation scheme
confers another important feature on MRAGPC in serving as a fault detection
scheme for MIMO processes.

References

1. H. Demircioglu, P.J. Gawthrop, Multivariable continuous-time generalized predictive control

(MCGPC) Automatica 28(4), 697–713 (1992)
2. M. Deng, A. Inoue, K. Edahiro, Fault detection in a thermal process system with input
constraints using a robust right coprime factorization approach. Proc. IMechE Part I Syst.
Control Eng. 221(part I),819–831 (2007a)
3. A. Nassirhand, S.A. Hosseini, Tracking and decoupling of multivariable non-linear systems.
Int. J. Model. Identification Contr. 6(4): 341–348 (2009)
4. M. Deng, A. Inoue, N. Ishibashi, A. Yanou, Application of an anti-windup multivariable
continuous-time generalized predictive control to a temperature control of an aluminum plate.
Int. J. Model. Identification Contr. 2(2), 130–1733 (2007b)
5. K. Zhou, J.C. Doyle, K. Glover, Robust and Optimal Control (Prentice-Hall, London, 1996)
6. A.S. Osunleke, M. Deng, A. Inoue, A robust anti-windup generalized predictive controller
design for systems with input constraints and disturbances. Int. J. Innovative Comput. Inf.
Contr. 5(10B), 3517–3526 (2009a)
7. M. Deng, A. Inoue, K. Ishikawa, Operator-based nonlinear feedback control design using
robust right coprime factorization. IEEE Trans. Automatic Contr. 51(4), 645–648 (2006)
8. A.S. Osunleke, M. Deng, A. Yanou, A design procedure for control of strictly proper non-
minimum phase processes with input constraints and disturbance. Intl. J. Model. Identification
Contr. (2010) (in press)
9. A.S. Osunleke, M. Deng, A. Inoue, A. Yanou, Adaptive robust anti-windup generalized
predictive control (RAGPC) of non-minimum phase systems with input constraint and
disturbance, in Proceedings of IEEE AFRICON 2009 (Nairobi, Kenya, 2009c)
10. M. Deng, Z. Iwai, I. Mizumoto, Robust parallel compensator design for output feed-back
stabilization of plants with structured uncertainties. Syst. Contr. Lett. 36, 193–198 (1999)
11. V.L. Kharitonov, Asymptotic stability of an equilibrium position of a family of systems of
differential equations. Differentsialnye uravneniya 14, 2086–2088 (1978)
12. A.S. Osunleke, M. Deng, A. Yanou, MRAGPC control of MIMO processes with input
constraints and disturbance, Lecture notes in engineering and computer science, World
Congress on Engineering and Computer Science, 20–22 Oct 2009b, pp. 872–877
Chapter 21
Analysis of Priority Rule-Based Scheduling
in Dual-Resource-Constrained Shop-Floor
Scenarios

Bernd Scholz-Reiter, Jens Heger, and Torsten Hildebrandt

Abstract A lot of research on scheduling manufacturing systems with priority

rules has been done. Most studies, however, concentrate on simplified scenarios
considering only one type of resource, usually machines. In this study priority rules
are applied to a more realistic scenario, in which machines and operators are dual-
constrained and have a re-entrant process flow. Interdependencies of priority rules
are analyzed by long-term simulation. Strength and weaknesses of various priority
rule combinations are determined at different utilization levels. Further insights are
gained by additionally solving static instances optimally by using a mixed integer
linear program (MILP) of the production system and comparing the results with
those of the priority rules.

1 Introduction

Shop-scheduling has attracted researchers for many decades and is still of big
interest, because of its practical relevance and its np-complete character. Multi-
or dual-resource problems are significantly less analyzed, despite being more
realistic. Scheduling with priority rules is quite appealing for various reasons
(see Section 2.2) and often used, especially when considering more realistic and
complex manufacturing systems.
This work analyses the quality of priority rules in the dual-resource constrained
case. Different combinations of rules are tested in order to analyze their interde-
pendencies. The analysis combines simulation with the optimal solution of static
instances. To evaluate the results of the simulation, a mixed integer linear program
(MILP) has been developed to calculate optimal solutions. The paper is organized

B. Scholz-Reiter (*)
Bremer Institut f€ur Produktion und Logistik (BIBA) at the University of Bremen, Hochschulring
20, D-28359 Bremen, Germany
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 269

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_21, # Springer ScienceþBusiness Media B.V. 2010
270 B. Scholz-Reiter et al.

as follows: Section 2 gives a short literature review about shop scheduling, priority
rules and dual constrained scheduling. A problem description follows in Section 3.
Section 4 analyses small instances in comparison with optimal solutions and
Section 5 describes a long term simulation study. The paper ends with a conclusion
and description of future research.

2 Literature Review

In this chapter a literature review is given considering the job shop problem, priority
rules and dual-constrained scheduling.

2.1 Shop Scheduling

Haupt [1] gives a definition of the scheduling or sequencing problem as “the

determination of the order in which a set of jobs (tasks) (i|i ¼ 1,. . ., n) is to be
processed through a set of machines (processors, work stations) (k|k ¼ 1,. . ., m).”
Usual abstractions of real-world scheduling problems are the job-shop and flow-
shop problem. With both problem types the number of operations and their
sequence are known in advance and fixed. In a Section scenario each job can
have its own route whereas in the classical flow-shop all jobs share a common
routing. Due to their high complexity optimal solutions can only be calculated for
small problems. New, flexible approaches and different solutions need to be
applied. Decentralized planning and scheduling as well as smart autonomous
items are, in our opinion, a very promising approach to this (e.g. [2]). Decentralized
decisions can be based on local decision rules. This is where priority rules come
into play.

2.2 Priority Rules

Priority rules are applied to assign a job to a resource. This is done each time the
resource gets idle and there are jobs waiting or a new job arrives. The priority rule
assigns a priority to each job. This priority can be based on attributes of the job, the
resource or the system. The job with the highest priority is chosen to be processed
next.
Priority-scheduling rules have been developed and analyzed for many years
[1, 3]. They are widely used in industry, especially in complex manufacturing
systems, e.g. semiconductor manufacturing. Their popularity is derived from the
fact that they perform reasonably well in a wide range of environments, are
relatively easy to understand and need minimal computational time.
21 Analysis of Priority Rule-Based Scheduling in Dual-Resource-Constrained 271

Depending on the manufacturing system and the various objectives (e.g. flow
time, tardiness, etc.) no single rule has been found, which outperforms all others [4].
As a result, there are approaches to find new priority rules according to the system’s
current state (e.g. [3, 5]).

2.3 Multi/dual-Resource Constrained Scheduling

Most research has been done on the machine-only constrained problem, where
machines are the only limiting resource. Nevertheless, closer to the ‘real’ world are
multi- or dual-constrained (DRC) problems, where more than one resource restricts
the output of the system and impacts shop performance. Gargeya and Deane [6]
defined the multiple resource constrained job shop as “a job shop in which two or
more resources are constraining output. The resources may include machines, labor
and auxiliary resources. Dual-constrained job shops are constrained by two resources
(e.g. machine and laborers). Dual-constrained job shops are thus a specific type of
multiple resource constrained job-shop.”
To solve the multi-resource constrained problem different approaches were
proposed. Mati and Xie [7] developed a greedy heuristic. This heuristic is guided
by a genetic algorithm in order to identify effective job sequences. Dauzère-Pérès
et al. [8, 9] developed a disjunctive graph representation of the multi-resource
problem and proposed a connected neighborhood structure, which can be used to
apply a local search algorithm such as taboo search. Patel et al. [10] proposed a
genetic algorithm for dual resource constrained manufacturing and compared
different priority rules against different performance measures. In the study of
Chen et al. [11], an integer optimization formulation with a separable structure is
developed where both machines and operators are modeled as resources with finite
capacities. By relaxing resource capacity constraints and portions of precedence
constraints, the problem is decomposed into smaller sub-problems that are effec-
tively solved by using a dynamic programming procedure.

3 Problem Description

ElMaraghy et al. [12] defined the machine/worker/job scheduling problem as:

“Given process plans for each part, a shop capacity constrained by machines and
workers, where the number of workers is less than the number of machines in the
system, and workers are capable of operating more than one machine, the objective
is to find a feasible schedule for a set of job orders such that a given performance
criteria is optimized.” An interesting object of investigation in DRC is the interac-
tion effect of the resource constraints and how they impact the performance
measure like the mean flow time.
The experimental design used in this study corresponds to the MiniFab scenario
[13, 14], which is shown schematically in Fig. 1. For practical reasons the model
has been simplified, yet the main properties remain.
272 B. Scholz-Reiter et al.

Ma Mc
1 2 3 Me 7
start end

Mb Md

4
6

Fig. 1 Process flow of the 5 Machine 6 step production model

Table 1 Processing and (un)loading times used

Machine Loading Processing Unloading
Step1 Ma/Mb 15 55 25
Step2 Mc/Md 20 25 15
Step3 Me 15 45 15
Step4 Ma/Mb 20 35 25
Step5 Mc/Md 15 65 25
Step6 Me 10 10 10

The MiniFab scenario is a simplification of a semiconductor manufacturing

system, but still contains the characteristics of such a system. It consists of five
machines with 6-step re-entrant processes. Operators are used to load and unload
machines. Machines Ma and Mb as well as Mc and Md are parallel identical
machines sharing a common buffer. To be able to get optimal solutions, we do
not consider sequence-dependent setup times, parallel batching and machine break
downs.
Operations at the machines are divided into three parts: loading, processing and
unloading. Processing starts directly after the machines are loaded. Unloading
begins after processing, but if there is no operator available for unloading, the
machine stays idle. In Table 1 processing, loading and unloading times are listed.
The processing and (un)loading times as well as the machine and sequence settings
are fixed. The machines select a job out of the waiting queue and call an operator to
load them (point in time a in Fig. 2). If no operator is available the machines wait until
eventually they are loaded (b-c) and proceed with processing (c-d).
If an operator is available, unloading is performed immediately (e-f), otherwise
there is an additional waiting period (d-e). The operators are not needed during
processing and can work on other machines during that time period. We vary the
number of operators from 0 to 3 to create scenarios constraining the available
capacity in different ways. The four resulting scenarios are:
l NO_OPERATOR: operators are not considered at all in this case, operations
require machines only (five machines, zero operators).
l MACHINE_CONSTRAINED: three operators are available, making machines
the more critical resource. (five machines, three operators).
21 Analysis of Priority Rule-Based Scheduling in Dual-Resource-Constrained 273

W = waiting;
Operator L U L = loading;
P = processing;
U = unloading;
Machine W L P W U a, b, c,
a b c d e f d, e, f : points in time

Fig. 2 Gantt chart illustration of machine-operator assignment

l OPERATOR_CONSTRAINED: there is just one operator; therefore operators

are much more critical than machines. In fact in this scenario the operator
becomes the capacity-constraining factor, reducing the maximum long-term
system capacity to 6.857 instead of 13.714 orders per day (five machines, one
operator).
l DUAL_CONSTRAINED: there are two operators in the model, resulting in
machines being roughly as critical as operators. (five machines, two operators).
We consider FIFO (First In buffer First Out), FSFO (First in System first Out),
Rnd (Random) and SPT (Shortest Processing Time first) as sequencing rules for the
machines. Besides FIFO, FSFO, Rnd and SPT we consider two additional decision
rules for operators: MQL (longest Machine Queue Length first) and SSPT (Shortest
Step Processing Time first). With MQL, priority is given to machines with a long
queue of waiting jobs. SSPT is similar to SPT, but instead of considering the
processing time of the complete operation, only the length of the required load or
unload processing steps are used. Furthermore, we vary the tie-breaker rule which is
used as a secondary criterion if priorities of the primary rule are equal. Using FIFO,
FSFO and Rnd no equal priorities can occur, so there is no need for a tie-breaker
rule in these cases.
To decide which job to process next, we use a two-step procedure, first using the
machine rule to select a job out of the waiting jobs in front of a machine. This sends
a processing request to the operator pool. The operator rule is then used to decide
between requests of different machines. If no operator is available immediately, the
machine is kept in a waiting state.
In this study we only consider the common performance measure mean flow
time. Our experiments are divided into two parts. Firstly we have experiments with
only a few jobs, which enable us to compare optimal schedules with the schedules
the priority rules provide (Section 4). Secondly we perform a long term simulation
study, with a time span of ten years to analyze the priority rules in detail (Section 5).

4 Experiments with Static Instances

To determine not only the differences between priority rules and their combina-
tions, it is interesting to analyze their performance in comparison to optimal
solutions and evaluate their performance in general.
274 B. Scholz-Reiter et al.

4.1 Experimental Design

We simulated instances from the same scenario with only 2–50 jobs in the system.
All jobs were released at the same time and known in advance. This simplification
makes it possible to calculate optimal solutions for the instances with few jobs in
the system. For larger instances feasible schedules (solutions of the MILP) could be
found, but they were not proven optimal. Due to the complexity of the model, there
still was a gap between the found solution (upper bound) and the theoretically
possible solution (lower bound).
The optimum results are calculated with CPLEX 11 solving a MILP-formulation,
which is an extension of the advanced job-shop formulation used by Pan and Chen
[15]. Operators and re-entrant processes were added to their flexible job-shop
MILP, so that the MiniFab model as described above can be solved and optimal
schedules are calculated. The notation used in the model is as follows:

F Mean flow time Z Binary variables

m Number of machines ri,j,g 1 if operation Oi,j requires
machine group g,
0 otherwise
n Number of jobs Si 1 if Oi,j is processed on machine
k, 0 otherwise, binary
Ji Job number i si Starting time of Oi,j
Mk Machine number k Fg Group g of parallel machines
N Number of operations QQ, QQ2 Variables for optim. expressions
a Operator action: 1 ¼ load., OSi,j,k,a Starting time of operator h doing
2 ¼ unload. a with operation Oi,j
H Number of operators OOi,j,k,a 1 if operator h does a with Oi,j
Oi,j operation j of Ji bigM A very large positive number
MinimizeF (1)
(OSi,j,h,2 þ utj)  bigM(1  ooi,j,h,2)  si,j þ 1, (2a)
i ¼ 1,2...n;j ¼ 1,2...N  1;h ¼ 1..H
si;j þ lti;j þ pi;j  osi;j;h;2 þ bigMð1  ooi;j;h;2 Þ; (2b)
i ¼ 1; 2:::n; j ¼ 1; 2:::N; h ¼ 1::H
bigMð3  vi;j;k  vii;jj;k  ooii;jj;h;2 þ bigMð1  Zi;j;ii;jj Þ (3)
þ sii;jj osi;j;h;2  uti;j ¼ Qi;j;ii;jj;k;h ;
1  i  ii  n
Qi;j;ii;jj;k;h  bigMð7  2vi;j;k  2vii;jj;k  2ooi;j;h;2 Þ þ sii;jj (4)
osi;j;h;2 þ uti;j þ si;j  osii;jj;h;2  utii;jj
P 1  i  ii  n; j; jj ¼ 1; 2:::N; h ¼ 1::H; k ¼ 1::m
k2Fg vi;j;k ¼ ri;j;g (5)
i ¼ 1; 2::n; j ¼ 1; 2:::N
osi;j;h;1  si;j  ð1  ooi;j;h;1 ÞbigM; (6)
i ¼ 1; 2::n; j ¼ 1; 2:::N; h ¼ 1::H
osi;j;h;1  si;j þ ð1  ooi;j;h;1 ÞbigM; (7)
i ¼ 1; 2::n; j ¼ 1; 2:::N; h ¼ 1::H
PH
h¼1 ðooi;j;h;a Þ ¼ 1;
(8)
i ¼ 1; 2::n; j ¼ 1; 2:::N; h ¼ 1::H; a ¼ 1::2
(9)
(continued)
21 Analysis of Priority Rule-Based Scheduling in Dual-Resource-Constrained 275

osi;j;h;a  ooi;j;h;a bigM;

i ¼ 1; 2::n; j ¼ 1; 2:::N; h ¼ 1::H; a ¼ 1::2
osi;j;h;1 þ lti;j  ð3  Yi;j;ii;jj;h;1 (10)
ooii;jj;h;1  ooi;j;h;1 ÞbigM  osii;jj;h;1 ¼ QQi;j;ii;jj;h
1  i  ii  n; j; jj ¼ 1; 2:::N; h ¼ 1::H;
QQi;j;ii;jj;h  bigMð5  2ooii;jj;h;1  2ooi;j;h;1 Þ  ltii;jj  lti;j (11)
1  i  ii  n; j; j ¼ 1; 2:::N; h ¼ 1::H;
ooi;j;h;1 þ lti;j  ð3  Yi;j;ii;jj;h;2 (12)
ooii;jj;h;2  ooi;j;h;1 ÞbigM  osii;jj;h;2 ¼ QQ2i;j;ii;jj;h
1  i  ii  n; j; jj ¼ 1; 2:::N; h ¼ 1::H;
QQ2i;j;ii;jj;h  bigMð5  2ooii;jj;h;2  2ooi;j;h;1 Þ  lti;j  uti;j (13)
1  i  ii  n; j; jj ¼ 1; 2:::N; h ¼ 1::H;
P
F ¼ ni¼1 ðosi;N;h;2 þuti;N  bigMð1  ooi;N;h;2 ÞÞ=n (14)
h ¼ 1::H;

number of jobs
[mean flow time in minutes] 2 3 4 5 6 7 8 9 10 20 50
optimum 483 520 568 622 678 733 785 870 918
gap (lower/upper bound)
OPERATOR

0% 0% 0% 0% 0% 8% 0% 27% 37%
best rule 483 520 568 628 684 739 799 861 908 1424 2989
NO_

worst rule 483 520 568 637 716 794 881 956 1040 1940 4616
span best-worst 0% 0% 0% 1% 5% 7% 10% 11% 15% 36% 54%
FSFO SSPT[FSFO] 483 520 568 628 684 739 799 861 908 1424 2989
optimum 483 520 568 622
CONSTRAINED CONSTRAINED CONSTRAINED

gap (lower / upper bound) 0% 0% 0% 0%

MACH_

best rule 483 520 568 628 684 739 809 869 919 1437 3039
worst rule 483 520 568 637 716 804 905 998 1095 2085 5050
span best-worst 0% 0% 0% 1% 5% 9% 12% 15% 19% 45% 66%
FSFO SSPT[FSFO] 483 520 568 628 684 739 809 869 930 1437 3042
optimum 483 524 569 640 746 838 930
gap (lower / upper bound) 0% 1% 0% 3% 9% 19% 16%
DUAL_

best rule 483 525 589 655 708 781 854 912 967 1527 3173
worst rule 483 530 620 724 823 916 999 1135 1229 2325 5500
span best-worst 0% 1% 5% 11% 16% 17% 17% 24% 27% 52% 73%
FSFO SSPT[FSFO] 483 528 598 656 708 789 875 924 979 1547 3173
optimum 483 520 568 929 1088 1272
gap (lower / upper bound) 0% 0% 0% 33% 38% 47%
best rule 535 643 756 917 1013 1118 1225 1351 1456 2524 5695
OP_

worst rule 548 685 842 1041 1243 1423 1616 1807 2022 4024 10010
span best-worst 2% 6% 11% 13% 23% 27% 32% 34% 39% 59% 76%
FSFO SSPT[FSFO] 540 643 803 918 1013 1118 1229 1351 1486 2541 5695

Fig. 3 Results of static scenarios in minutes

In our study, we were able to calculate some optimal solutions, which in most
cases were only slightly better than the best rule combination. In Fig. 3 we list the
solver results with the remaining gap and the performance of the best and worst rule
combination taken from any of the 72 combinations (see Section 3). For a detailed
analysis of the static runs, we chose the FSFO ShortestOpStepLength [FSFO] rule,
which seems to be the best performing one. In Fig. 3 the corresponding graphs are
plotted. The FSFO ShortestOpStepLength [FSFO] rule is indicated by the black
276 B. Scholz-Reiter et al.

line. The grey area defines the corridor between best and worst rule solutions. Black
dots correspond with the solver results, gaps between found solutions and lower
bounds are printed as vertical lines.

4.2 Analyses of Static Instances

About four or five jobs need to be in the system and first differences between
good and bad rule combinations can be found. An example is the DUAL_
CONSTRAINED case with five jobs in the system: The rule combination FSFO
(machine) ShortestOpStepLength [FSFO] (operator) has a mean flow time of 656
min and the combination FIFO (machine) FSFO (operator) has 724 min. This is
already a difference of more than 10%. The difference to the solver solution is only
around 2%.
The results of the scenario with NO_OPERATORS and the MACHINE_RES-
TRICTED scenario are very similar as Figs. 3 and 4 show. When there are more

a b
[minutes] [minutes]
1100 1100

1000 1000

900 900

800 800

700 700

600 600

500 500

400 400
2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10
[# Jobs] [# Jobs]
NO_OPERATOR MACHINE_CONSTRAINED

c d
[minutes] [minutes]
1300 2200
1200 2000
1100 1800
1000 1600
900 1400
800 1200
700 1000
600 800
500 600
400 400
2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10
DUAL_CONSTRAINED [# Jobs] OPERATOR_CONSTRAINED [# Jobs]

Fig. 4 Results for all static cases. Black line: FSFO ShortestOpStepLength [FSFO] rule, grey
area: best/worst rule combination result, black points: Solver results with gap (line)
21 Analysis of Priority Rule-Based Scheduling in Dual-Resource-Constrained 277

jobs in the system the performance differences between the rule combinations arise.
In the DUAL_CONSTRAINED case, the performance differences between rule
combinations are much higher and arise in smaller scenarios. This result was
expected, because in the dual constrained case the interdependencies between the
machine priority rule and the operator priority rule are the strongest and amplify
each other.
In the OPERATOR_CONSTRAINED case the shop is restricted the most,
which leads very quickly to high mean flow times. The differences between
the rule combinations are also very high in this case. The solution the solver
provided for the cases with three and four jobs indicate that either the rule
performance or the order of selection (machines first, operators second) do not
perform very well in this scenario. It seems more likely that the order of selection is
responsible for this effect. The operator is clearly the bottleneck and the scheduling
should be arranged in a way that utilizes him as best as possible, which is clearly
not the case.
In instances with more jobs we were not able to proof the same effect, because
the solver provided no optimal solutions even after days of calculations. The
solutions found were comparable to those of the used rules.

5 Long-Term Simulation

To validate the results from our static analyses, an extensive simulation study was
performed with a time horizon of 10 years.

5.1 Long-Term Simulation

To assess the performance of the various rule combinations in the long term, we
simulate the system under three load conditions for each of the four scenarios: 70%,
80% and 90% bottleneck utilization. Given the scenario and load level, we
determine the appropriate arrival rate based on a static analysis of system capacity.
Inter-arrival times follow an exponential distribution. We simulate a time span
of 10 years, ignoring data from the first year. Altogether we determine system
performance for 864 different parameter settings (4 scenarios, 3 load levels, 72 rule
combinations). For the different parameter settings we use common random num-
bers as a variance reduction technique [16]. The results presented are the averages
of mean flow times achieved in 20 independent replications.

5.2 Analysis of Long-Term Simulations

Figure 5 shows graphically the scenario, load level setting and the best rule result
achieved in this case. The four lines correspond to our scenarios:
278 B. Scholz-Reiter et al.

3000

2500
mean flowtime

2000

1500

1000

500
1
0 0.8
1 0.8 0.6
0.6 0.4 a t ion
0.4 0.2 utiliz
machine ut 0.2 0 0 ator
ilization oper

NO_OPERATOR MACHINE_CONSTRAINED DUAL_CONSTRAINED OPERATOR_CONSTRAINED

Fig. 5 Best rule performance for the four scenarios and three load level settings

l NO_OPERATOR (dotted line)

l MACHINE_CONSTRAINED (dashed line with short dashes)
l DUAL_CONSTRAINED (dashed line with long dashes)
l OPERATOR_CONSTRAINED (dot-and-dashed line)

Selecting one of the scenarios fixes the achievable combination of machine and
operator utilization along the respective line. To give an example: if MACHINE_-
CONSTRAINED is chosen and a load level of 90% bottleneck utilization (in this
case: machine utilization), then this results in an operator utilization of 60%. The
grey bar at this point shows the best flow time achieved, i.e. out of the 72 rule
combinations investigated.
The best results can obviously be achieved if no operator constraints are present
at all. This corresponds to the dotted line (NO_OPERATOR) in our simulation; we
also have the lowest flow times in this scenario, compared to the other scenarios
with the same load level. The DUAL_CONSTRAINED case (long dashes) on the
other hand is the most difficult, operator and machines are equally critical.
Figure 6 lists our simulation results in more detail. Mean flow time as measured
in our simulation experiment is expressed there as a flow factor, i.e. flow time
divided by total processing time. The best and worst rule performances for a given
scenario and load level are highlighted. To summarize our results, FSFO is the best
machine rule irrespectively of the scenario and load level. Results concerning the
best operator rule are a bit more complex: the MACHINE_CONSTRAINED and
OPERATOR_CONSTRAINED scenarios where the operator rule SSPT[FSFO],
i.e. Shortest Step Processing Time first with FSFO is used as a tie breaker, the best
results are gained if used together with FSFO as a machine rule.
21 Analysis of Priority Rule-Based Scheduling in Dual-Resource-Constrained 279

NO_OPERATOR MACH_CONSTRAINED OP_CONSTRAINED DUAL_CONSTRAINED

machine rule operator rule (0 Operator) (1 Operator) (2 Operator) (3 Operator) Mean
[tie breaker] [tie breaker] utilisation of bootleneck (flow
utilisation of bootleneck utilisation of bootleneck utilisation of bootleneck
factor)
70% 80% 90% 70% 80% 90% 70% 80% 90% 70% 80% 90%
FIFO FIFO 1,42 1,74 2,71 1,49 1,94 3,71 2,26 3,40 8,26 2,22 6,28 so 3,22
FIFO Rnd 1,42 1,74 2,71 1,50 1,95 3,81 2,29 3,48 8,93 2,27 6,90 so 3,36
FIFO SSPT[FSFO] 1,42 1,74 2,71 1,46 1,83 3,08 1,88 2,50 4,44 1,76 2,70 so 2,32
FSFO SSPT[FIFO] 1,32 1,52 2,13 1,34 1,56 2,25 1,83 2,31 3,68 1,54 2,00 so 1,95
FSFO SSPT[FSFO] 1,32 1,52 2,13 1,34 1,56 2,25 1,76 2,21 3,51 1,58 2,27 so 1,95
FSFO SPT[FIFO] 1,32 1,52 2,13 1,35 1,59 2,41 1,92 2,68 7,26 1,56 2,61 so 2,39
Rnd Rnd 1,42 1,73 2,69 1,49 1,94 3,87 2,27 3,45 8,91 2,29 7,99 so 3,46
Rnd SSPT[FSFO] 1,42 1,73 2,69 1,45 1,81 3,05 1,87 2,49 4,44 1,77 2,79 so 2,32
SPT[FIFO] MQL[FIFO] 1,37 1,65 2,44 1,40 1,70 2,59 2,02 2,77 5,23 1,73 2,53 13,14 2,31
SPT[FIFO] SPT[FSFO] 1,37 1,65 2,44 1,41 1,73 2,69 2,40 3,94 12,06 2,15 5,23 so 3,37
SPT[FSFO] MQL[FSFO] 1,37 1,65 2,44 1,40 1,70 2,59 1,90 2,63 5,10 1,73 2,55 13,20 2,28
SPT[FSFO] SPT[FSFO] 1,37 1,65 2,44 1,41 1,73 2,69 2,40 3,94 12,06 2,15 5,23 so 3,37
SPT[Rnd] MQL[FSFO] 1,37 1,65 2,44 1,40 1,70 2,59 1,90 2,64 5,11 1,73 2,54 13,71 2,28
SPT[Rnd] SPT[FIFO] 1,37 1,65 2,44 1,41 1,72 2,69 2,35 3,77 10,48 2,12 5,00 so 3,18
flowfactor best rule 1,32 1,52 2,13 1,34 1,56 2,25 1,76 2,21 3,51 1,54 2,00 6,56 1,95
flowfactor worst rule 1,42 1,74 2,71 1,50 1,95 3,87 2,40 3,94 12,06 2,29 7,99 14,65 3,46
flowtime span ((worst-best) / best) 8,2% 14,7% 27,3% 12,2% 24,8% 72,5% 36,0% 78,0% 243,0% 48,6% 299,6% 123,5% 77,4%

Fig. 6 Results for selected (worst/best) rule combinations investigated. Except for the last row
(span) all values are flow factors, i.e. mean flow time divided by the processing time (445 min).
The last column contains a mean performance averaged over all scenario/load level combinations
(but excluding DUAL_CONSTRAINED/90%, see text); so (system overflow) more incoming than
outgoing jobs

The FIFO rule performs quite badly in our experiments. In most cases it does not
yield better results than random selection, i.e. choosing an arbitrary job to process
next.
The DUAL_CONSTRAINED scenario SSPT[FIFO] is the best, i.e. SSPT with
tie breaker FIFO, for the 70% and 80% load levels. This changes if load is further
increased to 90%. Most rule combinations are not able to handle this high load for
both machines and operators - the system is running full, meaning more jobs are
entering the system than leaving it. Cases are marked with “so” in the table. Only
combinations with MQL as an operator rule are able to handle this high load. The
best results in this case are produced by the combination of the FSFO machine rule
and MQL[FIFO] as an operator rule. The MQL - rule prefers machines with long
queues, which means that operators go to highly loaded machines first. Although
this rule combination gives by far the best results for this case, the increase in
average flow time is very high.
Comparing the results for MACHINE_CONSTRAINED and NO_OPERATOR,
the increase in mean flow time is only small, especially for lower load levels. If only
flow time estimates are of interest, it seems viable to simplify the model by ignoring
the operator requirements.
The experiments with small instances show that in the OPERATOR_
CONSTRAINED cases, the optimal solutions are much better than the results
produced by our heuristics. In the long term simulation, the best rule results are
higher for scenarios with moderate load levels of 70% and 80% compared to the
DUAL_CONSTRAINED. Only the 90% load level results are smaller than its
DUAL_CONSTRAINED equivalent, but still about 56% higher than in the
MACHINE_CONSTRAINED scenario. This seems to be an effect of the heuristic
procedure used (machines choose next job first, then operators choose between the
machines that are ready to process). In future research more simulation runs with a
different heuristic setup will be performed to analyze this effect in more detail.
280 B. Scholz-Reiter et al.

In summary, all four scenarios show that the performance of priority rules differs
tremendously in some cases. The interdependencies of the rules, especially in the
DUAL_CONSTRAINED scenarios, lead to high performance differences. Other
system factors, e.g. the utilization rate, also affect the results.

6 Conclusion and Further Research

The paper analyses a priority-rule based approach to shop-floor scheduling consid-

ering both machines and operators. Our dual-constrained scenario was derived from
semiconductor manufacturing. To gain insights into priority rule performance we
use two approaches: simulation of small problem instances and its comparison with
optimal solutions and simulation of long term system behavior.
Especially in shop environments, where more than one resource is constraining
the system’s performance, proper selection and combination of priority rules is
crucial. The long term simulation has also shown that in high utilization scenarios,
many rule combinations are not able to handle the workload at all.
Comparison of known optimal solutions and best rule results indicates only a
small difference between the optimum and the best rule result. The differences
between well performing rule combinations and unsuitable combinations is much
higher than the differences to optimal solutions. They show that well adjusted
priority rules provide good solutions.
It is interesting to see, that the rule combinations’ performance depends on the
utilization level of the system. This brings up the idea to develop a control system,
which selects or develops priority rules automatically, considering system para-
meters and adopting to changing shop floor conditions.

Acknowledgements This research is funded by the German Research Foundation (Deutsche

Forschungsgemeinschaft DFG), under grant SCHO 540/17-1.

References

1. R. Haupt, A survey of priority rule-based scheduling. OR Spektrum, 11(1), 3–16 (1989)

2. B. Scholz-Reiter, M. G€ orges, T. Philipp, Autonomously controlled production systems-influ-
ence of autonomous control level on logistic performance. CIRP Ann. Manuf. Technol. 58(1),
395–398 (2009)
3. C.D. Geiger, R. Uzsoy, Learning effective dispatching rules for batch processor scheduling.
Int. Journal Prod. Res. 46(6), 1431–1454 (2008)
4. C. Rajendran, O. Holthaus, A comparative study of dispatching rules in dynamic flowshops
and jobshops. Eur. J. Oper. Res. 116(1), 156–170 (1999)
5. K.-C. Jeong, Y.-D. Kim, A real-time scheduling mechanism for a exible manufacturing
system: using simulation and dispatching rules. Int. J. Prod. Res. 36(9), 2609–2626 (1998)
6. V.B. Gargeya, R.H. Deane, Scheduling research in multiple resource constrained job shops: a
review and critique. Int. J. Prod. Res. 34, 2077–2097 (1996)
21 Analysis of Priority Rule-Based Scheduling in Dual-Resource-Constrained 281

7. Y. Mati, X. Xie, A genetic-search-guided greedy algorithm for multi-resource shop

scheduling with resource flexibility, IIE Trans. 40(12), 1228–1240 (2007)
8. S. Dauzère-Pérès, W. Roux, J.B. Lasserre, Multi-resource shop scheduling with resource
flexibility. Eur. J. Oper. Res. 107(2), 289–305 (1998)
9. S. Dauzère-Pérès, C. Pavageau, Extensions of an integrated approach for multi-resource shop
scheduling, IEEE Transactions on Systems, Man. and Cybernetics, Part C: Applications and
Reviews (2003)
10. V. Patel, H.A. ElMaraghy, I. Ben-Abdallah, Scheduling in dual-resources constrained
manufacturing systems using genetic algorithms, 7th IEEE International Conference on
Emerging Technologies and Factory Automation. Proceedings. ETFA 99 (1999)
11. D. Chen, P.B.c Luh, Optimization-based manufacturing scheduling with multiple resources,
setup requirements, and transfer lots. IIE Trans. 35, 973–985 (2003)
12. H. ElMaraghy, V. Patel, I.B. Abdallah, Scheduling of manufacturing systems under dual-re-
source constraints using genetic algorithms. J. Manuf. Syst. 19(3), 186–201 (2000)
13. G. Feigin, J. Fowler, R. Leachman Masm test data sets (2009). http://www.eas.asu.edu/
masmlab. Last accessed on 28 May 2009
14. M.K. El Adl, A.A. Rodriguez, K.S. Tsakalis Hierarchical modeling and control of re-entrant
semiconductor manufacturing facilities, Proceedings of the 35th Conference on Decision and
Control Kobe, Japan (1996)
15. J.C.-H. Pan, J.-S. Chen, Mixed binary integer programming formulations for the reentrant job
shop scheduling problem. Comput. Oper. Res. 32(5), 1197–1212 (2005)
16. A.M. Law, Simulation Modeling and Analysis, 4th edn. (McGraw-Hill, Boston, MA, 2007)
Chapter 22
A Hybrid Framework for Servo-Actuated
Systems Fault Diagnosis

Seraphin C. Abou, Manali Kulkarni, and Marian Stachowicz

Abstract A hybrid fault diagnosis method is proposed in this paper which is based
on analytical and fuzzy logic theory. Analytical redundancy is employed by using
statistical analysis. Fuzzy logic is then used to maximize the signal- to-threshold
ratio of the residual and to detect different faults. The method was successfully
demonstrated experimentally on hydraulic actuated system test rig. Real data and
simulation results have shown that the sensitivity of the residual to the faults is
maximized, while that to the unknown input is minimized. The decision of whether
‘a fault has occurred or not?’ is upgraded to ‘what is the severity of that fault?’ at the
output. Simulation results show that fuzzy logic is more sensitive and informative
regarding the fault condition, and less sensitive to uncertainties and disturbances.

1 Introduction

Hydraulic systems are very commonly used in industry. Like any other system these
systems too are prone to different types of faults. Proportional valves are much less
expensive in hydraulic control applications; they are more suitable for industrial
environments. Since proportional valves do not contain sensitive, precision com-
ponents, they offer various advantages over servo valves because they are less
prone to malfunction due to fluid contamination. However, these advantages are
offset by their nonlinear response characteristics. Since proportional valves have
less precise manufacturing tolerances, they suffer from performance degradation.
Larger tolerances on spool geometry result in response nonlinearities, especially in
the vicinity of neutral spool position. Proportional valves lack the smooth flow
properties of “critical center” valves, a condition closely approximated by servo

S.C. Abou (*)

Mechanical and Industrial Engineering Department, University of Minnesota Duluth, 1305 Ordean
Ct, Duluth, MN 55812, USA
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 283

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_22, # Springer ScienceþBusiness Media B.V. 2010
284 S.C. Abou et al.

valves at the expense of high machining cost. As a result, small changes in spool
geometry (in terms of lapping) may have large effects on the hydraulic system
dynamics [1]. Especially, a closed-center spool (overlapped) of proportional valve,
which usually provides the motion of the actuator in a proportional hydraulic
system, may result in the steady state error because of its dead-zones characteristics
in flow gain [1]. Figure 1 illustrates the characteristics of proportional valve.
Continuous online monitoring of fault in hydraulic system becomes increasingly
important day-by-day.
The characteristics of the proportional valve with dead-zones, g(u) is described
as follow (Fig. 1):
8
< aðu  bÞ
> if u  b
gðuÞ ¼ 0 if  b  u  b (1)
>
:
aðu þ bÞ if u <  b

where b; a  0; a represents the slope of the response outside the dead-zone, while
the width of the dead-zone equals 2b.
Remarkable efforts have been devoted to develop controllers. However, PID
controllers are not robust to the parameter variation to the plants being controlled.
Moreover, it takes time for the automatically self tuned PID controllers to adapt
themselves up to their final stable state (steady state). The fault detection problem
can be solved using different approaches like Wald’s Sequential Test, as in [1]
which is a conventional approach or using innovative approaches like genetic
algorithms as in [2], neural networks as in [3, 4], fuzzy logic as in [5] etc. each
having its own advantages and disadvantages.
Human experts play central roles in troubleshooting or fault analysis. In power
systems, it is required to diagnose equipment malfunctions as well as disturbances.
The information available to perform equipment malfunction diagnosis is most of
the time incomplete. In addition, the conditions that induce faults may change with
time. Subjective conjectures based on experience are necessary. Accordingly, the

qi qo Proportional valve
1 Q (l / min)
A B
g (u)
QA
u
p T xsp a
ps
2 –b b u
QB
M

Pump 1. Proportional valve

pr 2. Pump with Pressure
Regulator

Fig. 1 Characteristics of 4/3 closed-center proportional valve

22 A Hybrid Framework for Servo-Actuated Systems Fault Diagnosis 285

expert systems approach has proved to be useful. As stated previously, fuzzy theory
can lend itself for representation of knowledge and the building of an expert system.
In this paper we used fuzzy logic to detect the severity of fault at the output. The
concept of fuzzy logic was first introduced by Professor Lofti Zadeh [6], who
represented the vagueness of human concepts in terms of linguistic variables.
After the introduction of fuzzy sets, their applications to solve real world problems
were concentrated [7, 8]. Reference [9] concentrates on robust fault detection on an
aircraft flight control system. A model based fault diagnosis method for an indus-
trial robot is proposed in [10]. Residuals are calculated by the observer using a
dynamic robot model and later evaluated using fuzzy logic. In this paper we
demonstrate a similar model based approach for evaluating of severity of fault in
the hydraulic actuator using fixed threshold approach [11]. The objective know-
ledge on the system is represented by mathematical modeling (calculating the
residuals using nonlinear observer) [1], while the subjective knowledge is repre-
sented using fuzzy logic (fuzzy rules and membership functions).

2 System Under Consideration

The schematic of the hydraulic system under consideration, the mathematical

model and the design of nonlinear observer can be found in [1]. General nonlinear
dynamical systems can be described as follows:
(
f ðx; y; u; tÞ ¼ A0 x þ BhðxÞ þ ðxÞ
(2)
y ¼ CT x

where hðxÞ ¼ lðxÞ þ ’ðxÞu þ bðt  TÞfðxÞ; ’ and l are unknown smooth functions.
x ¼ ðx1 ; x2 ; x3 ; :::; xn Þ is the state vector, x 2 Rn , u and y 2 R; z ¼ y^ is the observer
output vector; ðxÞ represents the uncertainty in the system dynamics that may
include parameter perturbations, external disturbances, noise, etc. All long this
study, we consider abrupt fault at time T. As a result, bðt  TÞ is the time profile
of failures as shown in Fig. 2.
fðxÞ is another function which represents a failure in the system. Mathematical
description of the nonlinear observer as follows:
(
^ þ Kðy  C^
^x_ ¼ Ax þ BhðxÞ xÞ
(3)
z ¼ y^ ¼ CT x^

^
where hðxjy ^ ^
f Þ ¼ lðxjyf Þ þ ’ðxjy f Þu þ fðxjyf Þ; K is the observer gain matrix
selected as follows A ¼ A0  KCT , which is strictly a Hurwitz matrix.
In the discrete time system, consider from (2) and (3) eðkÞ ¼ yðkÞ  zðkÞ. The
actual state of the system y(k) is known through the sensors. The residual e(k) is
calculated as follows:
286 S.C. Abou et al.

b (t – T)
1

⎧1 for t ≥ T
b(t – T) = ⎨
⎩0 otherwise
0.8

0.6

0.4

0.2

0
0 0.1 T 0.4 0.6 0.8 1
time [s]

Fig. 2 Function bðt  TÞ in (2)

eðkÞ ¼ Mp yðkÞ  Mp zðkÞ (4)

Mp is an identity matrix of size m  n, Mp ¼ I41

It is perceived that the performance of the actuated system is selected based on
four parameters having a range of value from zero (0) to one (1). The elements of
the state vector z  [v Pi Po xsp]T are: velocity x_ ¼ v, input pressure Pi, output
pressure Po and xsp spool displacement. The residual of these four can be measured.
In this paper we have concentrated on the velocity residual and the identity matrix
Mp ¼ [1 0 0 0].
Theoretically, these residuals should be zero under no fault condition. However,
in practical context, due to noise, inexact mathematical modeling and system
nonlinearity, this residual is never zero even under no fault condition. Reference
[1] uses a conventional method called Wald’s Sequential Test to detect fault. In this
method, the cumulative residual error is calculated over a period of time and fault is
detected using the fixed threshold concept. This conventional method has some
disadvantages. A value just below the threshold is not considered as a fault while
some value just above the threshold will be considered as a fault. This can also lead
to missing alarms and false triggers. This information could be potentially mis-
guiding to the operators working on the hydraulic system. This is the drawback of
binary logic.
The conventional method is rigid and does not consider a smooth transition
between the faulty and the no fault condition. The probability assignment procedure
is heuristic and depends on the number of Zeros/Ones in the failure signature. This
22 A Hybrid Framework for Servo-Actuated Systems Fault Diagnosis 287

does not give any information about the fault in between the thresholds. In order to
take care of this condition we try to replace this binary logic by multi-valued one
using fuzzy logic. Evaluating these residuals using fuzzy logic replaces the yes/no
decision of fault by the severity of fault at the output.

3 Role of Fuzzy Logic

From the point of view of human-machine cooperation, it is desirable that faults

classification process would be interpretable by humans in such a way that experts
could be able to evaluate easily the classifier solution. Another interest of an
intelligent interface lies in the implementation of such a system in a control
room. Operators have to be informed very quickly if a fault is occurring. They
have to understand what exactly the process situation is, in order to make the right
counteraction if possible or to stop the system if necessary. For instance, as shown
in Fig. 3, the fuzzification vector can be assigned to an index of a color map,
representing a color code, by defuzzification. Figure 3 depicts the overall architec-
ture of the hydraulic system fault detection where u(t) is the control input.
The mapping of the inputs to the outputs for the fuzzy system is in part
characterized by a set of condition ! action rules (if-Then) form:

If premise then consequent (5)

The inputs of the fuzzy system are associated with the premise, and the outputs
are associated with the consequent. As already seen, the difference between the
expected state z(k) and the actual state of the system y(k) gives the residual e(k).

u(k) y(k)
Hydraulic plant
+ e(k)
∑ e(k)
–
Observer – +
z(k)
Threshold
Fuzzification
Rule-base

Inference
Mechanism

Defuzzification

Fig. 3 The structure of Fault

Fuzzy fault diagnosis Severity
288 S.C. Abou et al.

The value Pof residual is added over a period of time which gives the cumulative
residual e(k). This value is subtracted from the predicted threshold and is called
cumulative residual difference.
The lower the value of this cumulative residual difference, higher is the fault
severity, indicating that the cumulative residual is approaching the threshold and
vice versa. The threshold is determined through observations. It will vary depen-
ding upon the fault tolerance of the application in which the hydraulic system is
used. Even if there is no fault, the modeling errors or noise drive several residuals
beyond their threshold. This is usually indicated by all suspect residuals being
weak. The residual is bounded between the upper and the lower threshold. As soon
as it approaches these thresholds, the fault severity increases. Thus, the residual and
the cumulative residual difference are given as two inputs to the fuzzy logic
controller. Based on these two inputs, the controller decides the fault severity at
the output. One of the equations of fuzzy equality can be written as:

P
n
minfmA ðiÞ; mB ðiÞg
SA;B ¼ n
i¼1
 (6)
P Pn
max mA ðiÞ; mB ðiÞ
i¼1 i¼1

where n is the dimension number of the discrete space, mA is the membership

function of fuzzy set A, mB is the membership function of fuzzy set B.
Suppose that a fuzzy rule set to be detected F represents the current working
class of the actuated hydraulic system, and the other fuzzy reference rule set Fi
stands for one working class of the system. Since both the fuzzy reference rule sets
and fuzzy rule set to be detected have the same hypersphere subspaces, the Eq. (6)
can be used for their contrasts. As a result, in this study, the approximate measure-
ment of the fuzzy reference rule set Fi can be expressed by:

jei ðkÞj
si ðkÞ ¼ (7)
ei

where ei(k) is the ith residual; si(k) is the residual-to-threshold ratio.

Obviously si(k) is greater than or equal to 1 if the test is fired on the residual and
si(k) is less than 1 if it did not.

4 Design of Fuzzy Logic Controller

On one hand, note that the fuzzy reference rule sets almost cover all the faults;
while on the other, the fuzzy rule set to be detected may bring forward the undefined
symptoms which cannot be distinguished from fuzzy reference rule sets, as shown
in Fig. 3. How can we solve this problem?
22 A Hybrid Framework for Servo-Actuated Systems Fault Diagnosis 289

To solve this problem, we include in the evaluation procedure an additional

credit degree of unknown classes which is expressed as follows:
 n
eei ¼ 1  max SA;B i¼1

Obviously; 0  SA;B  1 (8)

4.1 Inputs

Figure 4 illustrates the actual and the estimate velocities. The difference is due to
the error introduced in the actual system by adding random noise to the velocity
during simulation.
The plot of residual, cumulative residual, cumulative residual difference along
with the thresholds can be seen in Figs. 5 and 6. As seen earlier, the residual and
the cumulative residual difference are the two inputs to the fuzzy logic controller.
Fault isolation thresholds are very important parameters; their values are also
decided by the statistical analysis of the fault credit degrees [13]. As for unknown
fault type, consulting fault isolation thresholds are selected upon our knowledge
on the system. The detection results of the normal data of the space propulsion

0.35
Observed Shaft Velocity
Actual Shaft Velocity
0.3

0.25
Velocity (m/s)

0.2

0.15

0.1

0.05

0
0 0.2 0.4 0.6 0.8 1
Time (s)

Fig. 4 Graph showing Actual velocity and observed velocity vs time

290 S.C. Abou et al.

0.2
Residual (m/s)
0.15 Upper Threshold
Lower Threshold
0.1

0.05
Residual (m/s)

–0.05

–0.1

–0.15

–0.2
0 200 400 600 800 1000
Number of Observations

Fig. 5 Graph showing ‘Residual’ along with the upper and lower thresholds vs ‘number of
observations’

30
Cumulative Residual (m/s)
25 Upper Threshold
Lower Threshold
Cumultive Residual (m / s)

20 Cumulative Residual Difference (m/s)

15

10

–5
0 200 400 600 800 1000
Number of Observations

Fig. 6 Cumulative residual and the cumulative residual difference along with the upper and lower
thresholds vs the number of observations
22 A Hybrid Framework for Servo-Actuated Systems Fault Diagnosis 291

system are shown in Fig. 5. Because normal credit degree does not exceed the
threshold, the detection results are that no fault exists, and working conditions are
normal.

4.2 Membership Functions

The first input which is residual is divided into seven membership functions
namely, Big Negative (BN), Negative (N), Small Negative (SN), Zero (Z),
Small Positive (SP), Positive (P) and Big Positive (BP) shown below Fig. 7.
Similarly, we developed five membership functions for the second input which
is cumulative residual difference. They are Large Negative (LNeg), Medium
Negative (MNeg), Small Negative (SNeg), Zero (Zero) and Positive (POS) as
seen in the following Fig. 8. As already seen there are four parameters which can
be used to calculate the residuals. Among them the velocity is the most concerned
parameter in this case of study.
Hence, the velocity residual is selected to determine the fault severity at the
output. The membership functions for the output i.e. fault severity are F0, F1, F2,
F3, F4, F5 and F6 where F0 represents the lowest fault severity and F6 represents
the highest fault severity Fig. 9.
The shapes of the membership functions which are triangular and trapezoidal
were selected based on the simple guidelines suggested in [12]. This can be seen
in Fig. 9.

BN N SN Z SP P BP
1
Degree of membership

0.5

–0.1 –0.08 –0.06 –0.04 –0.02 0 0.02 0.04 0.06 0.08 0.1
First input ‘Residual’

Fig. 7 Membership functions for the first input ‘Residual’

292 S.C. Abou et al.

LNeg MNeg SNeg Zero Pos

Degree of membership

0.5

–10 –8 –6 –4 –2 0 2 4 6 8 1
Second input ‘Cumulative Residual Difference’

Fig. 8 Membership functions for the second input ‘Cumulative Residual Difference’

F0 F1 F2 F3 F4 F5 F6
1
Degree of membership

0.5

0 10 20 30 40 50 60 70 80 90 100
Output Variable ‘Fault Severity’

Fig. 9 Membership functions for the output ‘Fault Severity’

4.3 Rule-Based Inference

Inference rules were developed which relate the two inputs to the output. They are
summarized in the Table 1. As seen from the table, there are in all 35 rules. For
example, if the residual is Big Positive (BP) and the cumulative residual difference
is Large Negative (LNeg) then the output fault severity is the highest (F6). Simi-
larly, if the residual is Zero (Z) and the cumulative residual difference is Positive
(Pos) then the output fault severity is the lowest (F0).
22 A Hybrid Framework for Servo-Actuated Systems Fault Diagnosis 293

Table 1 Rule based inference

Residual
Cumulative residual

BN NEG SN Z SP POS BP
difference

POS F3 F2 F1 F0 F1 F2 F3
Zero F4 F3 F2 F1 F2 F3 F4
SNeg F5 F4 F3 F2 F3 F4 F5
MNeg F6 F5 F4 F3 F4 F5 F6
Lneg F6 F6 F5 F4 F5 F6 F6

4.4 Defuzzification

After converting the crisp information into fuzzy the last step is to reverse that.
Converting the fuzzy information to crisp is known as defuzzification. The center of
area/centroid method was used to defuzzify these sets which can be represented
mathematically as follows:

Sfi  mðfi Þ
Defuzzified value ¼ (9)
Smðfi Þ

Where fi is the fault severity at the output and m(fi) is the output membership
function.

4.5 Rule Viewer

The rules can also be seen from the rule viewer using the fuzzy logic toolbox in
MATLAB software. When the residual is 0.01, it is far away from both the upper and
lower thresholds (almost at the center) and hence, has lower fault severity. Also, the
cumulative residual difference is nine which means the difference between the
actual value of cumulative residual and threshold is high i.e. cumulative residual is
far away from the threshold. Hence, the fault severity should be low. A combination
of these values of residual and cumulative residual gives fault severity percentage of
9.96% which is low. Similarly, when the residual is 0.089 it indicates that it is very
close to the threshold. A cumulative residual difference of 9 indicates that the
threshold has been already crossed by the cumulative residual (hence it is negative).
Both of these conditions lead to a very high fault severity of 98.4%. This can be seen
with the help of the rule viewer facility in the fuzzy logic toolbox. These examples
are shown in Figs. 10 and 11 respectively with the help of rule viewer.

5 Simulation

This simulation was carried out in MATLAB SIMULINK using fuzzy logic
controller from the fuzzy logic toolbox as shown in Figs. 10 and 11. The upper
subsystem represents the actual system (actual state of the hydraulic system) and
294 S.C. Abou et al.

Fig. 10 Test Results for low fault severity

the lower subsystem is the nonlinear observer (which predicts the state of the
system). The SIMULINK diagram is the implementation of the block diagram
shown in Fig. 1. The simulation is carried out for a unit step input. Fault is
introduced in the actual system by adding noise to the velocity in the actual system
and different fault severities are tested at the output.

6 Conclusion

With the looming reliability challenges in future new technologies, the ability to
provide practicing engineers online information about the systems health is vital for
decision-making. The approach proposed in this study showed that fuzzy logic,
when used in combination with analytical methods like non linear observer, can
enhance the output. Simulation results clearly demonstrated what we all know:
whatever fault type the plant generates, its symptoms always depart from the
characteristics of the fuzzy reference rule set standing for the normal working
condition. Moreover while fuzzy rule sets are set up, the fuzzy reference rule set
generated is used for representing the normal working condition, which is supposed
to be the first fuzzy reference rule set. Thus, with the credit degree representing the
normal working condition, we judged whether the plant working condition is
22 A Hybrid Framework for Servo-Actuated Systems Fault Diagnosis 295

Fig. 11 Test results for high fault severity

normal, further obtain the fault degree. This study helped to assure that the plant
fault existed and to report the system conditions. Future work will be developed to
identify the fault type and predict the equipment remaining life.

Acknowledgement The present work has been performed in the scope of the activities of Grant In
Aid project. The financial support of the University of Minnesota, Grant In Aid FY 2008 is
gratefully acknowledged.

References

1. H. Khan, S.C. Abou, N. Sepehri, Nonlinear observer-based fault detection technique for
electro hydraulic servo positioning systems. Mechatronics 15(9), 1037–1059 (Nov 2005)
2. D. Xuewu, L. Guangyuan, L. Zhengji, Discrete-time robust fault detection observer design: a
genetic algorithm approach, in IEEE Proceedings of the 7th World Congress on Intelligent
Control and Automation, Chongqing, June 2008
3. G. Qian-jin Guo, Y. Hai-bin Yu, X. Ai-dong, Modified Morlet wavelet neural networks for
fault detection. Int. Conf. Control Automat. (ICCA 2005) 2, 1209–1214 (June 2005)
296 S.C. Abou et al.

4. Q. Sheng, X.Z. Gao, Z. Xianyi, State-of-the-art in soft computing-based-motor fault diagno-

sis. IEEE 2, 1381–1386 (June 2003)
5. V. Pedro, R. Jover, A. Antero, Detection of stator winding fault in induction motor using fuzzy
logic. Appl. Soft Comput. 8(2), 1112–1120 (Mar 2007)
6. L.A. Zadeh, Fuzzy sets, Inf. Control 8, 338–35, 1965
7. M.M. Jerry, Fuzzy logic systems for engineering: a tutorial. IEEE 83(3), 345–377 (Mar 1995)
8. C. Stephen, J. C. John, C. Sitter. E. Fooks, Perspectives on the industrial application
of intelligent control, in IEEE Proceedings of the 34th Conference on Decision & Control,
vol. 1. New Orleans, LA, 1995
9. T. Curry, E.G. Collins Jr., M. Selekwa, Robust fault detection using robust l1 estimation and
fuzzy logic. IEEE, 2, 1753–1758 (June 2001)
10. H. Shneider, P.M. Frank, Observer-based supervision and fault detection in robots. IEEE 4(3),
274–282 (May 1996)
11. K. Manali, S.C. Abou, S. Marian, Fault detection in hydraulic system using fuzzy logic, In
Proceedings of the World Congress on Engineering and Computer Science 2009 Vol. II;
WCECS 2009, San Francisco, USA, 20–22 Oct 2009
12. Y. John, L. Reza, Fuzzy Logic: Intelligence, Control and Information. (Prentice Hall, Englewood
Cliffs, NJ, 1999), pp. 22–27
13. C.T. Shing, P.L. Chee, Application of an adaptive neural network with symbolic rule
extraction to fault detection and diagnosis in a power generation plant. IEEE Trans. Energy
Convers. 19(2), 369–377 (June 2004)
Chapter 23
Multigrid Finite Volume Method for FGF-2
Transport and Binding

Wensheng Shen, Kimberly Forsten-Williams, Michael Fannon, Changjiang

Zhang, and Jun Zhang

Abstract A multigrid finite volume method has been developed to accelerate the
solution process for simulating complex biological transport phenomena, involving
convection, diffusion, and reaction. The method has been applied to a computational
model which includes media flow in a cell-lined cylindrical vessel and fibroblast
growth factor-2 (FGF-2) within the fluid capable of binding to receptors and heparan
sulfate proteoglycans (HSPGs) on the cell surface. The differential equations were
discretized by the finite volume method and solved with the multigrid V-cycle
algorithm. Our work indicates that the multigrid finite volume method may allow
users to investigate complex biological systems with less CPU time.

1 Introduction

Basic fibroblast growth factor (FGF-2) is the prototypical member of the family of
fibroblast growth factors [14]. It has been demonstrated to be important in normal
physiology and pathologies in cancer development process [4]. In addition to
its target cell surface tyrosine kinase receptor, FGF-2 also binds with high affinity
to heparan sulfate proteoglycan (HSPG), which consists of a protein core with
O-linked carbohydrate chains that are polymers of repeating disaccharides [6].
HSPGs are present on almost all cell surfaces and generally in much higher
numbers than FGF surface receptors. FGF-2 is present in circulation and its
presence in elevated levels in blood is used in clinical settings as criteria for
treatment strategies such as interferon alpha therapy for infantile hemangioma
[5]. Although FGF-2 binding interactions have been the subject of numerous
studies in static systems, far less is known about their behavior in flow.

W. Shen (*)
SUNY College at Brockport, Brockport, NY 14420, USA
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 297

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_23, # Springer ScienceþBusiness Media B.V. 2010
298 W. Shen et al.

The interaction between biochemical reaction and mass transfer is of particular

interest for the real time analysis of biomolecules in microflow systems. Glaser [12]
proposed a two-dimensional (2D) computer model, which was greatly simplified
by assuming that the flow is parallel and diffusion is perpendicular to the surface,
to study antigen-antibody binding and mass transport in a microflow chamber.
Myszka et al. [15] used a computer model similar to that of [12], to simulate the
binding between soluble ligands and surface-bound receptors in a diagnostic
device, BIACORE, for both transport-influenced and transport-negligible binding
kinetics. A 2D convection and diffusion equation was used by Chang and Hammer
[2] to analytically investigate the effect of relative motion between two surfaces on
the bindings of surface-tethered reactants. We take a different approach by solving
the 2D convection diffusion equation directly.

2 Methods

2.1 Mathematical Model

A coupled nonlinear convection-diffusion-reaction model for simulating growth

factor binding under flow conditions has recently been developed by [10, 18] and is
used here. This model can be applied to predict the time-dependent distribution of
FGF-2 in a capillary with complicated binding kinetics on the tube surface.
For simplicity, we consider a basic model, where FGF-2 is the only ligand in the
solution, and cells which line the capillary express both FGF-2 receptors (FGFRs)
and HSPGs. As illustrated in Fig. 1, FGF-2 binds to FGFR and HSPG to form
complexes of FGF-2-FGFR and FGF-2-HSPG. The resulting complexes may
continue binding to produce either their dimers or FGF-2-FGFR-HSPG triads.
The FGF-2-FGFR-HSPG triads may further bind to each other to generate FGF-
2-FGFR-HSPG dimers. Consequently, the concentration of FGF-2 in the solution is
affected due to the molecular binding process.The 2D time-dependent equations for

P G
R C C2
G2 T T2
L

G2 T T2
G
R P C C2

Fig. 1 Sketch of growth factor binding to receptors and HSPGs and the formation of various
compounds on the surface of a capillary. The symbols in the sketch are as follows: L¼FGF-2,
R¼FGFR, P¼HSPG, C¼FGF-2-FGFR complex, G¼FGF-2-HSPG complex, C2¼FGF-2-FGFR
dimer, G2¼FGF-2-HSPG dimer, T¼FGF-2-FGFR-HSPG complex, and T2¼FGF-2-FGFR-HSPG
dimer. The arrows represent velocity vectors, which are uniform in the entrance and evolve to
parabolic later on
23 Multigrid Finite Volume Method for FGF-2 Transport and Binding 299

fluid mass and momentum transfer in conservative form for incompressible flow
can be written as:

@r 1 @ðrrvr Þ @ðrvz Þ
þ þ ¼ 0; (1)
@t r @r @z
   
@ðrvr Þ 1 @ðrrvr vr Þ 1 @ @vr @ðrvr vz Þ @ @vr @p vr
þ  rm þ  m ¼ m 2 ; (2)
@t r @r r @r @r @z @z @z @r r
   
@ðrvz Þ 1 @ðrrvr vz Þ 1 @ @vz @ðrvz vz Þ @ @vz @p
þ  rm þ  m ¼ ; (3)
@t r @r r @r @r @z @z @z @z

where r is the density, m the dynamic viscosity, p the dynamic pressure, and vr and
vz are velocity components in the radial and axial directions, respectively. In the
above equation set, the independent variables are time t, radial coordinate r, and
axial coordinate z.
The mass of each species must be conserved. If binding or reactions occur within
the fluid, the coupling of mass transport and chemical kinetics in a circular pipe can
be described by the following equations:
 
@ ðrfi Þ 1 @ðrrufi Þ @ðrvfi Þ 1 @ @ ðrfi Þ
þ þ ¼ Kr r
@t r @r @x r @r @r
  (4)
@ @ ðrfi Þ
þ Kx þ Fi ðf1    fn Þ; 1  i  n;
@x @x

where fi is the concentration of species i, u and v are the radial and longitudinal
components of velocity, Kr and Kz the molecular diffusion coefficients, and Fi the
rate of change due to kinetic transformations for each species i. The basic model
however has only FGF-2 within the fluid (fi is simply FGF-2) and, thus, reactions
(i.e., binding and dissociation) occur only on the capillary surface. That is to say
that Fi is valid merely on the tube surface. The reactants and products involved in
the chemical kinetics include FGF-2, FGFR, HSPG, FGF-FGFR complex and its
dimer, FGF-HSPG complex and its dimer, FGF-HSPG-FGFR complex and its
dimer, with a total of nine species (n ¼ 9) [10].

2.2 Collocated Finite Volume Discretization

A cell-centered finite volume approach is applied to discretize the partial differen-

tial equations. Eqs. 1–4 can be re-organized in the form of convection diffusion
equations and written as
   
@ 1 @ @f @ @f
ðrfÞ þ yrvf  yG þ ruf  G ¼ S; (5)
@t y @y @y @x @x
300 W. Shen et al.

where x and y are axial and radial coordinates, u and v are the velocity components
in the axial and radial directions, respectively. It is worth noticing that the mass
conservation equation is a special case of Eq. 5 in which f, G, and S are taken as
f ¼ 1, G ¼ 0, and S ¼ 0.
To achieve higher order temporal accuracy, we use a quadratic backward appro-
ximation for the time derivative term. Such arrangement gives us second order
temporal accuracy. Integrating Eq. 5, the corresponding finite volume equations can
be derived,

3ðrfÞnþ1  4ðrfÞnP þ ðrfÞn1

P P
þ ðJ e  J w Þ þ ðJ n  J s Þ ¼ SC þ SP fP þ Ssym ;
Dt
(6)
where Je, w, n, s is the convection-diffusion flux at each of the four interfaces of the
control volume P, with J e;w ¼ Fe;w  De;w , J n;s ¼ Fn;s  Dn;s , SC and SP are the
results of source term linearization, and Ssym is the contribution from axisymmetric
coordinates, with Ssym ¼ G yv2 for the momentum equation of v. The approxima-
tion of convective flux is critical for an accurate solution. We use a so-called
deferred correction technique [8]. This technique calculate higher-order flux explic-
itly using values from the previous iteration. Using deferred correction, which is a
combination of the first order upwind differencing and the second order central
differencing, the convective flux is written as a mixture of upwind and central
differences, Fe;w ¼ Fue;w þ lðFce;w  Fue;w Þo , where Fue ¼ maxððruÞe Dr j ; 0:ÞfP þ
minððruÞe Dr j ; 0:ÞfE , Fuw ¼ maxððruÞw Dr j ; 0:ÞfW þ minððruÞw Dr j ; 0:ÞfP , Fce ¼
ðruÞe Dr j ð1  ae ÞfP þ ðruÞe Dr j ae fE , Fcw ¼ ðruÞw Dr j ð1  aw ÞfW þ ðruÞw Dr j aw fP ,
l is a parameter set as l ¼ 0  1, and the superscript (o) indicates taking the
value from the previous iteration, which will be taken to the right hand side and
treated as a part of the source term. The same can be applied to convective
flux in radial direction. The interpolation factors are defined as ae ¼ xxEe x xP ,
P

r P r s K x yj ðfE fP Þ
xw
aw ¼ xxPPx W
, an ¼ rrNn r
rP , and as ¼ r P r S . The diffusion fluxes are De ¼
P
xE xP ,
K y ðf f Þ
Dw ¼ x xj P x
P
W
W
, Dn ¼ Kr xirðf N fP Þ
N r P
, and Ds ¼ Kr xirðf P fS Þ
P r S
. The notations of spatial
discretization in Eq. 6 is illustrated in Fig. 2, where the uppercase letters indicate the
center of the control volumes, and the lowercase letters indicate the interfaces
between neighboring control volumes.
Substituting numerical fluxes into Eq. 6 and collecting terms, a set of algebraic
equations are obtained of the following form
AS fS þ AW fW þ AP fP þ AE fE þ AN fN ¼ b: (7)

The coefficients of Eq. 7 consist of a pentadiagonal matrix, which is solved using an

efficient linear system solver that will be discussed later. The coefficients and the
right-hand side in Eq. 7 are given by
K r Dxi r s
AS ¼  maxððrvÞs Dxi ; 0:Þ  ; (8a)
rP  rS
23 Multigrid Finite Volume Method for FGF-2 Transport and Binding 301

Δ x i−1 Δ xi Δ x i+1
j+1

N
Δ r j+1

n j

W w e
P E Δ rj

j−1
s

Δ r j−1
S

i−1 i i+1
x

Fig. 2 Finite volume notation of control volumes in axisymmetrical coordinates

K x Dr j r P
AW ¼  maxððruÞw Dr j ; 0:Þ  ; (8b)
xP  xW

K r Dxi r n
AN ¼ minððrvÞn Dxi ; 0:Þ  ; (8c)
rN  rP

K x Dr j r P
AE ¼ minððruÞe Dr j ; 0:Þ  ; (8d)
xE  xP

3rr P Dxi Dr j
AP ¼  ðAW þ AS þ AE þ AN Þ; (8e)
2Dt
 
4:0rfnP rfn1
b ¼ ðSC þ SP Þr P Dxi Dr j þ  P
r P Dxi Dr j
2:0Dt 2:0Dt
 lðFce  Fue  Fcw þ Fuw þ Fcn  Fun  Fcs þ Fus Þ: (8f)

2.3 Multigrid Methods

With the finite volume method on a structured grid, the multigrid version in 2D can
be constructed such that each coarse grid control volume is composed of four
302 W. Shen et al.

control volumes of the next finer grid [8]. To do this, a grid generator that is able to
generate multiple grids is used. The grid generator takes input of the number of grid
levels, the number of lines along each boundary, the coordinates of starting and
ending points for each line, the line type, etc., for the coarsest grid. The data of the
rest levels of grids are computed automatically by subdividing each control volume
in finer ones and saved as binary files to be loaded by the flow solver.
The algebraic Eq. 7 can be written more abstractly as

Af ¼ b; (9)

where A is a square matrix, f the unknown vector, and b the source vector. After the
kth outer iterations on a grid with spacing h, the intermediate solution satisfies the
following equation,

Akh fkh  bk ¼ r kh ; (10)

where rhk is the residual vector after the kth iteration. Once the approximate solution
at the kth iteration is obtained, we restrict the approximate solution as well as the
residual to the the next coarse grid by the restriction operators Ih2h and Iˆh2h. An
approximate solution to the coarse grid problem can be found by solving the
following system of equations,

A2h f12h ¼ A2h ðI 2h ^2h k

h Þ  Ih rh : (11)

After the solution on the coarse grid is obtained, the correction Df ¼ f12h  f02h is
transfered to the fine grid by interpolation (prolongation), where f2h 0
¼ Ih2hfhk. The
difference of Ih and Iˆh is as follows, Ih takes the mean value of states in a set of
2h 2h 2h

cells, but Iˆh2h performs a summation of residuals over a set of cells. The value of fhk
is updated by

fkþ1
h ¼ fkh þ I h2h Df; (12)
h
where I2h is a prolongation operator. This procedure is repeated until the approxi-
mate solution on the finest grid converges using the multigrid V-cycle algorithm
[16]. The relaxation technique used in the multigrid method is Stone’s strong
implicit procedure (SIP), which is a modification from the standard ILU decompo-
sition [19]. This paper has adapted the multigrid method for incompressible
Navier–Stokes equations provided by [16], and extended it to include the mass
transport. Since the concentration of ligand is very small, in the order of 1011 to
1010 M, we may assume that the momentum and mass transfer equations are
independent to each other. Consequently the momentum transfer equation can be
solved first to obtain the velocity distribution, which is then put into the mass
transfer equation.
23 Multigrid Finite Volume Method for FGF-2 Transport and Binding 303

3 Results

The flow model in the current paper is the well-known incompressible Navier–
Stokes equation, which has been widely used to solve low speed fluid mechanics
problems. A convection-diffusion equation is applied to address the mass transport
of growth factor in solution. A similar mass transport model has been used by [12]
to describe the biological interaction of antigen and antibody, and by [15] to
investigate the interactions of a variety of biomolecules in BIACORE, where the
flow cell was of rectangular geometry. However, application of the 2D Navier–
Stokes equation to the study of growth factor binding to cell surface receptors and
HSPGs lining a cylindrical capillary has not, to the best of our knowledge, been
reported. The numerical procedure and simulation results for a single grid is
presented in a previous paper [18], while the main purpose of the current paper is
to demonstrate a multigrid solution to the convection-diffusion-reaction model of
mass transport and chemical kinetics of protein ligands.
The dimensions of the capillary under consideration were length L ¼ 0. 1 m and
radius R ¼ 0. 00035 m, corresponding to that of a Cellmax Cartridge System
bioreactor (FiberCell Systems, Inc., Frederick, MD, USA). The ratio of length
over diameter is 286. To simulate the capillary flow, four types of boundary
conditions are used in the numerical simulation: inlet boundary to the left-hand
side, outlet boundary to the right-hand side, symmetry boundary at the bottom, and
impermeable wall boundary at the top. For inlet boundary, all quantities are
prescribed, and the incoming convective flux is calculated as well. For outlet
boundary, zero gradient along the grid line is applied. A three-level multigrid
method is considered, as shown in Fig. 3, and the number of control volumes are
40  5, 80  10, and 160  20, respectively. Each time the grid is refined, one
control volume is divided into four smaller control volumes. The ligand interactions
within the capillary were modeled as a problem of mass transport with a reactive
boundary condition. For the mass transport equation, the boundary conditions are:
given concentration on the west, zero flux on the east, symmetrical condition on the
south, and reaction boundary on the north. An existing model of biochemical
reactions on the capillary surface from [10] is used in our simulations (illustrated in
Fig. 1). The reaction rates for all species are represented by a set of ordinary
differential equations (ODEs), as shown in Table 1, and the parameters are primarily
determined by experiments [10]. The system of ordinary differential equations are
solved by a variable-coefficient ODE solver VODE for stiff and nonstiff systems of
initial value problems [1]. The code is developed to solve the time-dependent
Navier–Stokes and convection-diffusion-reaction equations with a particular appli-
cation in growth factor transport and binding. The computation is performed on a
Sun-Blade-100 machine with a single 500 MHz SPARC processor and 2 GB mem-
ory. A typical numerical solution is plotted in Fig. 4, where the concentration
distribution of fibroblast growth factor inside the capillary is shown at an instant
time of t ¼ 10 min. The solution in Fig. 4 corresponds to the finest grid arrangement
in the multigrid system. Note that the ligand concentration in the figure is non-
304 W. Shen et al.

Fig. 3 Computational grids for capillary flow and protein transport: (a) first level grid, (b) second
level grid, (c) third level grid

dimensionalized with respect to the inlet concentration on the west boundary. To

demonstrate the impact of surface binding on ligand transport, two numerical experi-
ments have been conducted, one without surface binding, Fig. 4(a), and the other with
surface binding to receptors and heparan sulfate proteoglycans, Fig. 4(b). As
expected, without surface binding, displayed in Fig. 4(a), the ligand moves with the
flow by convection and diffusion, and its concentration has a uniform distribution
along the radial direction in a portion of the capillary from x ¼ 0 to roughly x ¼ 0.
05 m. One may further predict that a uniform concentration distribution in the whole
capillary will be obtained after t ¼ 20 min. It clearly shows in Fig. 4(b) that the
concentration of ligand in the capillary is spatially reduced by a great margin down
along the capillary as well as in the radial direction due to biochemical reactions on
the surface. In Fig. 4, the front of ligand transport is dissipative, which may be
explained by the relative importance between convection and diffusion. The Peclet
number is defined as the ratio of convective transport to diffusive transport,
Pe ¼ uL=D, where u is the velocity, L is the characteristic length, and D is the
diffusion coefficient. The characteristic length can have a significant impact on the
value of Pe. Under current simulation, the velocity is u ¼ 0:0867 mm=s, the diffusion
coefficient is D ¼ 9:2  107 cm2 =s. Using the capillary radius as the characteristic
length, the resulting Peclet number is Pe ¼ 330. This value indicates a convective
dominated process however the value is somewhat intermediate indicating diffusion
does make a not insignificant contribution to the overall process.
Figure 5(a) displays the convergence history for capillary flow in a single grid
with the finest grid spacing. In the current calculation, the energy equation is not
considered. The results clearly show that the residuals are reduced effectively in the
first 200 outer iterations, and after that, they cannot be further eliminated no matter
how many more outer iterations are performed. The results also indicate that the
pressure equation has a relatively large residual (Fig. 5(a)).
23 Multigrid Finite Volume Method for FGF-2 Transport and Binding 305

Table 1 The rate of concentration change due to protein interactions (the initial concentration of
surface variables are R0 ¼ 1:6  104 receptors=cell and P0 ¼ 3:36  105 sites=cell)
Reaction rate Parameters
d½R kRf ¼ 2:5  108 M1 min1 , kRr ¼ 0:048 min1
¼ kRf ½L½R þ kRr ½C þ kTr ½T
dt
 kc ½R½G  kint ½R þ V R kTr ¼ 0:001 min1 , kc ¼ 0:001 min1 ð#=cellÞ1
kint ¼ 0:005 min1 , VR ¼ 80 sites min 1
d½P kPf ¼ 0:9  108 M1 min1 , kPr ¼ 0:068 min1
¼ kPf ½L½P þ kPr ½G þ kTr ½T
dt
 kc ½C½P  kint ½P þ V P kTr ¼ 0:001 min1 , kc ¼ 0:001 min1 ð#=cellÞ1
kint ¼ 0:005 min1 , VP ¼ 1, 680 sites min 1
d½L kRf ¼ 2:5  108 M1 min1
V ¼ kRf ½L½R þ kRr ½C þ kTr ½T
dt
 kPf ½L½P  kRr ½G kPf ¼ 0:9  108 M1 min1
kRr ¼ 0:048 min1 , kTr ¼ 0:001 min1
d½C kRf ¼ 2:5  108 M1 min1
¼ kRf ½L½R  kRr ½C  kc ½C½P
dt
 kc ½C2 þ 2kuc ½C2   kint ½C kRr ¼ 0:048 min1 , kc ¼ 0:001 min1 ð#=cellÞ1
kuc ¼ 1 min1 , kint ¼ 0:005 min1
d½C2  kc 2 1 1 1
¼ ½C  kuc ½C2   kD k c ¼ 0:001 min ð#=cellÞ , k uc ¼ 1 min
int ½C2 
dt 2
1
kDint ¼ 0:078 min
d½G kf ¼ 2:5  10 M1 min1
R 8
¼ kRf ½L½R þ kRr ½C þ kTr ½T
dt
 kc ½R½G  kint ½R kRr ¼ 0:048 min1 , kTr ¼ 0:001 min1
kc ¼ 0:001 min1 ð#=cellÞ1 , kint ¼ 0:005 min1
d½G2  kc 2 kc ¼ 0:001 min1 ð#=cellÞ1 , kuc ¼ 1 min1
¼ ½G  kuc ½G2   kint ½G2 
dt 2
1
kDint ¼ 0:078 min
d½T
¼ kc ½R½G þ kc ½C½P  kr ½T
T k c ¼ 0:001 min 1
ð#=cellÞ1 , kTr ¼ 0:001 min1
dt
 kc ½T2 þ 2kuc ½T 2   kint ½T kuc ¼ 1 min1 , kint ¼ 0:005 min1
d½T 2  kc 2 kc ¼ 0:001 min1 ð#=cellÞ1 , kuc ¼ 1 min1
¼ ½T  kuc ½T 2   kD int ½T 2 
dt 2
1
kDint ¼ 0:078 min

A comparison was then made between the single-grid and multigrid computa-
tions, shown in Fig. 5(b), where the number of iterations means the number of outer
iterations on the finest grid (160  20). Only the residual from the pressure equation
is shown as it had the largest value for the single-grid (Fig. 5(a)), but similar
differences were found for the other residuals (data not shown). In single-grid
computation, the pressure residual can only be reduced to the order of 103 while
in the case of the three-level multigrid computation, it can be reduced to the order
of 105 within less than 100 outer iterations. Table 2 lists the number of outer
iterations, CPU-time, and speed-up ratio for the capillary flow with various mesh
size and different linear system solvers. In Table 2, the recorded CPU-time for
multigrid is the total computing time for all levels of grids that are involved. For all
three solvers studied (SIP, BiCGSTAB, and GMRES), a great amount of savings in
306 W. Shen et al.

a
0.0625 0.1875 0.3125 0.4375 0.5625 0.6875 0.8125 0.9375
0.0003
0.0002
0.0001
R

0
–0.0001
–0.0002
–0.0003
0 0.025 0.05 0.075 0.1
X
b
0.0625 0.1875 0.3125 0.4375 0.5625 0.6875 0.8125 0.9375
0.0003
0.0002
0.0001
R

0
–0.0001
–0.0002
–0.0003
0 0.025 0.05 0.075 0.1
X

Fig. 4 A snapshot of ligand concentration distribution at t ¼ 10 min in the capillary at the

temperature of 37 C and the uniform inlet velocity of u ¼ 5:2 mm=min. (a) Surface binding is not
considered. (b) Ligands bind to receptors and HSPGs on the capillary surface

a b
100 100

10–1
10–1
10–2
Presure Residual
Residual

10–2 10–3

10–3 10–4

10–5
10–4
10–6

10–5 0 10–7 0
10 101 102 103 104 10 101 102 103 104
No. of Iterations No. of Iterations

Fig. 5 Comparison of convergence history between single-grid and multigrid computations. (a)
Residuals of pressure and velocities in single-grid computation, where dashed line is the U
residual, dashed-dotted-dotted line the V residual, solide line the P residual. (b) Comparison of
pressure residuals between single-grid and multigrid computation with different linear system
solvers, where dashed-dotted line is for single grid, solid line for multigrid with SIP, dashed line
for multigrid with BiCGSTAB, and dashed-dotted-dotted line for multigrid with GMRES

CPU time has been observed by using multigrid. For example, the CPU time needed
for a 160  20 grid using the SIP method was 176.45 seconds for the single-grid and
5.39 for the multigrid. This difference in time was due to the significant difference
23 Multigrid Finite Volume Method for FGF-2 Transport and Binding 307

Table 2 Performance comparison of multigrid method with different linear system solvers
Solver Mesh No. outer iterations CPU-time (s) Speed-up
Single-grid Multigrid Single-grid Multigrid ratio
SIP 40  5 234 234 0.41 0.41 1.0
80  10 898 28 6.87 0.64 10.7
160  20 6,000 54 176.45 5.39 32.7
BiCGSTAB 40  5 114 114 0.35 0.35 1.0
80  10 432 26 5.24 0.56 9.36
160  20 6,000 50 174.66 5.43 32.16
GMRES 40  5 101 101 0.22 0.22 1.0
80  10 389 23 5.43 0.55 9.87
160  20 6,000 52 186.75 5.82 32.08

in the number of outer iterations required which was 6,000 for the single-grid and
only 54 for the multigrid. This can result in a speed-up ratio for the capillary flow of
over 30 for the finest grid. Three iterative solvers, SIP [19], BiCGSTAB [20] and
GMRES [17] are used to solve the inner linear system and their performance
are compared. In this particular case with a five diagonal coefficient matrix,
BiCGSTAB and GMRES solvers do not have obvious advantages over SIP in
reducing the number of outer iterations of Navier–Stokes equations in the multigrid
computation.

4 Discussions

The binding kinetics and signaling pathways of FGF-2 are very complicated, and
many mathematical models have been proposed to predict the behavior of FGF-
2 [3, 7, 9–11]. These models include pure reaction models [7, 10, 11], where a
system of ordinary differential equations is provided and it is assumed that the fluid
movement of FGF-2 does not impact the solution, and reaction-diffusion models
[3, 9], which are relatively more complex and the movement of FGF-2 molecules
from fluid to cell surface is modeled by diffusion. The diffusion model is valid only
if the fluid is quiescent, which is not consistent with the actual biological environ-
ment of moving bio-fluids. Our model is unique, in which a coupled convection-
diffusion-reaction model is applied, and the motion of bio-fluids is fully considered.
We model the process of growth-factor binding and dissociation under flow as
fluid flow in capillary by incompressible Navier–Stokes equation, protein transport
by convection-diffusion, and local bioreaction on capillary surface. The flow is
assumed to be laminar, and both the fluid density and viscosity are taken as con-
stants. This is due to the fact that the fluid in the bioreactor fibers is mainly water
with some additives (FiberCell Systems, Inc., Frederick, MD, USA). The impact of
fluid flow on FGF-2 binding is through the transport equations in the coupled non-
linear convection-diffusion-reaction model, where the flow velocity affects the
308 W. Shen et al.

distribution of FGF-2 in the solution, and that of FGFR, HSPG, and their com-
pounds on capillary surface.
The solution of Navier–Stokes equations consists of two loops. The inner
iteration handles each of the individual equations of momentum, energy, and
turbulent kinetics if necessary, and the outer iteration deals with the coupling and
nonlinearity. For unsteady flow using implicit discretization, the linear equations
need not be solved very accurately at each outer iteration. Usually a few iterations
of a linear system solver is enough. More accurate solution will not reduce the
number of outer iterations but may increase the computing time [21].
A colocated grid system is used to discretize the Navier–Stokes and transport
equations. The naive way of colocated arrangement makes the solution of Navier–
Stokes equations not very pleasant due to the checkerboard pressure distribution.
Therefore staggered grid was once considered as a standard way for the calculation
of incompressible flow [13]. A remedy was proposed to deal with the pressure-
velocity coupling on colocated grids by eliminating the oscillations in the pressure
field [8]. Our work uses a colocated grid and adopts the remedy to filter out the
oscillations. A colocated arrangement is attractive when using a non-orthogonal
grid, complex geometry, and multigrid methods.
Developing efficient solvers for nonlinear system of equations arising from fluid
flow and mass transfer problems is of practical interest. When nonlinear partial
differential equations are discretized in a particular spatial mesh, the result is a
set of ordinary differential equations, with time as the independent variable. The
temporal terms may be further discretized using either an explicit or an implicit
scheme. In an explicit method, a large number of independent equations, one for
each control volume, have to be solved. In the case of an implicit discretization, we
have to solve a large set of simultaneous equations.

f ð~ uÞo
uÞ  f ð~
uÞ ¼
Fð~ þ Fð~
uÞ ¼ 0 (13)
Dt

uÞo is the value at the previous time step. If the number of components is
where f ð~
m and the number of control volumes is n, the number of independent variables
becomes mn. Such large sets of nonlinear equations are generally solved by a
variant of Newton’s method, where a sequence of linear systems are solved [8].
The great advantage of the Newton iteration is its quadratic convergence. However,
Newton iteration converges only if the initial guess is close to the solution. Instead,
in this paper, the method of Picard iteration is applied, in which the nonlinear
convective term and source term are linearized by using values from the previous
outer iterations. This kind of linearization requires many more iterations than a
coupled technique such as Newton-like linearization, but an initial guess close to
the solution is not critical for the convergence. The number of outer iterations,
however, can be substantially reduced by using multigrid techniques, as shown
previously.
23 Multigrid Finite Volume Method for FGF-2 Transport and Binding 309

5 Concluding Remarks

In conclusion, this paper presents a multigrid computation strategy for the numeri-
cal solution of FGF-2 transport and cellular binding within a capillary (i.e.,
cylindrical geometry). The capillary flow is predicted using the incompressible
Navier–Stokes equations by the finite volume method with collocated mesh
arrangement with the assumption that the capillary is a circular pipe. To reduce
the computation cost, a multigrid V-cycle technique, which includes restriction and
prolongation to restrict fine grid solution to coarse grid, and to interpolate coarse
grid solution to fine grid, is applied for the nonlinear Navier–Stokes equations. The
multigrid method is extended to solve the mass transport equation. Computational
results indicate that the multigrid method can reduce CPU time substantially. The
computed profile of FGF-2 distribution in the capillary is presented for a set of
conditions and no advantage with regard to CPU time or the magnitude of residue
has been observed among the three linear system solvers being used.

References

1. P. Brown, G. Byrne, A. Hindmarsh, VODE: a variable coefficient ODE solver. SIAM J. Sci.
Stat. Comput. 10, 1038–1051 (1989)
2. K. Chang, D. Hammer, The forward rate of binding of surface-tethered reactants: effect of
relative motion between two surfaces. Biophys. J. 76, 1280–1292 (1999)
3. C.J. Dowd, C.L. Cooney, M.A. Nugent, Heparan sulfate mediates bFGF transport through
basement membrane by diffusion with rapid reversible binding. J. Biol. Chem. 274,
5236–5244 (1999)
4. S. Elhadj, R.M. Akers, K. Forsten-Williams, Chronic pulsatile shear stress alters insulin-like
growth factor-I (IGF-I) bonding protein release in vitro. Annal. Biomed. Eng. 31, 163–170
(2003)
5. A. Ezekowitz, J. Mulliken, J. Folkman, Interferon alpha therapy of haemangiomas in
newborns and infants. Br. J. Haematol. 79(Suppl 1), 67–68 (1991)
6. M. Fannon, M.A. Nugent, Basic fibroblast growth factor binds its receptors, is internalized,
and stimulates DNA synthesis in Balb/c3T3 cells in the absence of heparan sulfate. J. Biol.
Chem. 271, 17949–17956 (1996)
7. M. Fannon, K.E. Forsten, M.A. Nugent, Potentiation and inhibition of bFGF binding by
heparin: a model for regulation of cellular response. Biochemistry, 39, 1434–1444 (2000)
8. J.H. Ferziger, M. Peri\acute{c}, in Computational Methods for Fluid Dynamics (Springer,
Berlin, 1999)
9. R.J. Filion, A.S. Popel, A reaction-diffusion model of basic fibroblast growth factor interac-
tions with cell surface receptors. Annal. Biomed. Eng. 32, 645–663 (2004)
10. K. Forsten-Williams, C. Chua, M. Nugent, The kinetics of FGF-2 binding to heparan sulfate
proteoglycans and MAP kinase signaling. J. Theor. Biol. 233, 483–499 (2005)
11. F.M. Gabhann, M.T. Yang, A.S. Popel, Monte Carlo simulations of VEGF binding to cell
surface receptors in vitro. Biochim. Biophys. Acta 1746, 95–107 (2005)
12. W. Glaser, Antigen-antibody binding and mass transport by convection and diffusion to a
surface: a two-dimensional computer model of binding and dissociation kinetics. Analyt.
Biochem. 213, 152–161 (1993)
310 W. Shen et al.

13. F. Harlow, J. Welsh, Numerical calculation of time dependent viscous incompressible flow
with free surface. Phys. Fluid 8, 2182–2189 (1965)
14. N. Itoh, The Fgf families in humans, mice, and zebrafish: their evolutional processes and roles
in development, metabolism, and disease. Biol. Pharm. Bull. 30, 1819–1825 (2007)
15. D. Myszka, X. He, et al., Extending the range of rate constants available from BIACORE:
interpreting mass transport-influenced binding data. Biophys. J. 75, 583–594 (1998)
16. E. Schreck, M. Peri\acute{c}, Computation of fluid flow with a parallel multigrid solver. Int.
J. Numer. Meth. Fluid 16, 303–327 (1993)
17. Y. Saad, M. Schultz, GMRES: a generalized minimal residual algorithm for solving nonsym-
metric linear systems. SIAM J. Sci. Stat. Comput. 7, 856–869 (1986)
18. W. Shen, C. Zhang, M. Fannon et al., A computational model of FGF-2 binding and HSPG
regulation under flow condition. IEEE Trans. Biom. Eng. 56, 2147–2155 (2009)
19. H. Stone, Iterative solution of implicit approximations of multidimensional partial differential
equations. SIAM J. Numer. Anal. 5, 530–558 (1968)
20. A. Van der Vorst, Bi-CGSTAB: a fast and smoothly converging variant of Bi-CG for the
solution of nonsymmetric linear systems. SIAM J. Sci. Stat. Comput. 13, 631–644 (1992)
21. J. Zhang, H. Sun, J.J. Zhao, High order compact scheme with multigrid local mesh refinement
procedure for convection diffusion problems. Comput. Meth. Appl. Mech. Engi. 191,
4661–4674 (2002)
Chapter 24
Integrated Mining Fuzzy Association Rules
For Mineral Processing State Identification

Seraphin C. Abou and Thien-My Dao

Abstract Mineral processes are multi-variable, power-intensive and strongly cou-

pled with large delay and nonlinearities. The properties of controllability, observ-
ability and theory of minimal realization for linear systems are well understood and
have been very useful in analyzing such systems. This paper deals with analogous
questions for nonlinear systems with application to mineral processing. A method
that can control and provide accurate prediction of optimum milling condition and
power consumption, water and chemical additive requirement is developed for
mineral plants operation. A fuzzy mining algorithm is proposed for extracting
implicit generalized knowledge on grading process performance as qualitative
values. It integrates fuzzy-set concepts and generalized data mining technologies
to achieve this purpose. Using a generalized similarity transformation for the error
dynamics, simulation results show that under boundedness condition the proposed
approach guarantees the global exponential convergence of the error estimation.
Although the nominal performance of the process is improved, the robust stability
still is not guaranteed to fully avoid the mill plugging.

1 Introduction

Data mining is the process for automatic extraction of high level knowledge from
the information provided by sensors. A control-engineering method for fuzzy
control was provided in [1]. In mineral processing, understanding how to modify
the rheological characteristics of fine particle systems is a key for the process
performance. These characteristics include particle settling, pH, bulk/carrier fluid
viscosity, particulate flocculation or dispersion, attrition, pipe/fitting/impeller

S.C. Abou (*)

Mechanical and Industrial Engineering Department, University of Minnesota Duluth, 1305
Ordean Ct, Duluth, MN 55812, USA
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 311

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_24, # Springer ScienceþBusiness Media B.V. 2010
312 S.C. Abou and T.-M. Dao

wear, degradation of flocculated or friable solids and the pumpability of the slurry.
Moreover, fine particle systems exhibit a range of rheological properties that
influence processing and handling. The properties, settling behavior and the rheo-
logy of fine particulate slurries are determined by both the physical properties and
surface chemistry of the particles. Figure 1 depicts two stages structure considered
for developing both useful fuzzy association rules and suitable membership func-
tions from quantitative values: (1) adaptive learning procedure to learn the mem-
bership functions, and (2) a method to mine fuzzy association rules.
A wet grinding plant shown in Fig. 2 has been analyzed with the objective of
evaluating the effects of many variables on particle size reduction in continuous
grinding processes. Detailed phenomenological model that describes the charge
behaviour has been developed and validated against real data [2]. Indeed, mineral

Learning Membership Functions Fuzzy Association Rules Mining

Defined
Database Learning Membership
Process Functions

Predefined
Fitness Fuzzy
Membership Database
(Evaluation module) Mining
Functions

Membership Fuzzy Association

Functions Rules

Fig. 1 Scheme for fuzzy association rules and membership functions

Motor

Ball
Mill

Disturbances
+ i(t)
Controller Motor
w r (t)–
w (t)
p update
law
Sensor

Fig. 2 Motor servo-system of the ball mill

24 Integrated Mining Fuzzy Association Rules For Mineral Processing 313

processes present non-linear/chaotic dynamic behaviour. Considerable efforts have

been developed in controlling such systems [3, 4]. In [2], a comprehensive model
integrating physical mechanisms and fundamental understanding of the charge
behaviour was developed. The poor quality of fine grinding is due to lacks of an
appropriate control of the power draw of the mill. This causes increase of energy
consumption, and production cost [5]. A classical PID controller is worthless to
efficiently monitor the mineral plant. Through the past few decades, advanced
control methods are being developed to address the system as a whole, dealing
with all the problems that arise out of the complexity of the breakage process,
its distributed nature, the networked interactions of the components [6–8]. Never-
theless, several approaches mainly depend on the model of the controlled plant.
Fuzzy control does not need the model of the controlled plant. In general, the fuzzy
logic rules are obtained from the knowledge of experts and operators [3, 9, 10].
Since its introduction in 1993 [11], the task association rule mining finds interesting
relationships among a large set of data items [4]. In this paper the association rules
mining algorithm is adopted to design a fuzzy logic controller as a control law.

2 Grinding Process Modelling

Besides in batch mode operation, grinding circuit can operate in continuous or

fed-batch mode. As shown in Fig. 1, the motor load is strongly influenced by
the filling percentage, the speed, the mill geometry and other relevant material
properties such as stiffness and the coefficient of friction, etc.
The theoretical position of the charge at different rotation speeds was first
derived by Davis [5] based on the total force balance. We are interested in the
constitutive characteristics of the charge motion defined by a function f (x, u) that
better describes continuous grinding phenomena:


@ @ @mi
½xð:Þ ¼ Cn ð:Þ  Cn ð:Þmi (1)
@t @z @z

where, mi (kg), is particle mass of size i

The left side term of Eq. (1) expresses the rate of mineral production. Without
lacking the physical sense for the process, we can write:

Cn ð:Þ ¼ Cn ðx; uÞ (2)

Based on volumetric analysis, the fraction of the total mass broken within a tiny
volume dV of the charge is assumed to be s(t) that is determined as follows:
ðð ð
sðtÞ ¼ arc dV (3)
V
314 S.C. Abou and T.-M. Dao

Where rc is the charge bulk density, a is defined as a mass volume of material of

classes i; Therefore, the flow rate of particle through the mill is:
ðð ð ðð ð
ds @ðarc Þ dðdVÞ
¼ dV þ arc (4)
dt @t dt
V V

However, as the mass could not be transferred by conduction phenomena, the

mass flux vanishes, so that we could write:
ðð ððð
ds
¼ J~i  dF
~þ #sp dV (5)
dt
F V

where, J~i ¼ longitudinal diffusion flux of the mass in class I; # ¼ piecewise

parameter; sp ¼ local fine particle.

3 The Controller Design

Wet communition (grinding) and particle size classification can be effected

by viscosity, particle size distribution, fines concentration etc. For example, hydro-
cyclone classification is effected by the presence of a yield stress (the minimum
force required for slurry to flow). An interlaced approach, called platform-based
control design, can be developed using fuzzy logic controller based on association
rules mining [11]. The main problem of the grinding process lies in developing a
sufficiently simple and reliable method for measuring the load in the Ball mill
during operation. Although the fineness distribution capability of the particle can
not be measured directly, it is obviously related to ball mill load, the stability of the
grinding and the floatation processes.
Figure 3 depicts the characteristics of ball mill in function of the ball mill load, l.
Functions m(l) and p(l) represent the grinding capability and the driving motor

p(l ) m(l )
[kW ] [ton/h]
Pmax p(l )
Unstable
region

Mmax
m(l )
l [%]
Fig. 3 Static characteristics 0
of ball mill L1 L2
24 Integrated Mining Fuzzy Association Rules For Mineral Processing 315

power, respectively. Our motivation is to design a fuzzy system to produce alarms

and reconfigure the control variables if critical conditions occur in the process
(pH, rheology, degradation of flocculated solids, slurry pumpability, etc.). Assume
the fuzzy system uses x1(t), and x2 (t) as inputs, and its output is an indication of
what type of warning condition occurred along with the certainty that this warning
condition has occurred.
The alarm signals represent certain warnings characterized by the decision
regions shown in Fig. 4. For instance the density of the slurry is greater than the
upper threshold limit; this occurs if: x1 ðtÞ > a1 and x1 ðtÞ  x2 ðtÞ þ a3 . Small
changes in chemical properties (in terms of lapping) may have large effects on
the mineral quality and the grinding system dynamics. In the next section, we
consider the system as multi-state system. The grinding process is complex ill-
defined process consisting of n elements, any element j, 1  j  n can have kj
different states with corresponding performance rates (levels), which can be repre-
sented by the ordering set as follows [3]:
 
gj ¼ gj1 ; . . . gjij ; . . . ; gjkj (6)

where
 gjij is the performance rate (level) of the element j in the state ij;
ij 2 1; 2; . . . ; kj . The performance rate Gj(t) of element j at any instant t  0 is
a random variable that takes its values from gj : Gj ðtÞ 2 gj .Thus, the probabilities
associated with different states for the element j can be represented by a set:
 
pj ¼ pj1 ; . . . ; pjkj ; pj ! 1  j  n

The mapping gjij ! pjij is usually called the probability mass function as defined
for multi-state system [3]. However, for some multi-state systems, evaluating
precisely the state probability and performance rate of an element is difficult. As
above pointed out for the cause of the deterioration of the grinding quality, let
define the error er and change of error ec at sampled times k as follows:

pðkÞ  pðk  1Þ
er ðkÞ ¼ and ec ðkÞ ¼ er ðkÞ  er ðk  1Þ (7)
Vm ðkÞ  Vm ðk  1Þ

x2(t)

a2
−a2
Fig. 4 Decision regions for
a3 a1 x1(t)
the rheology state
316 S.C. Abou and T.-M. Dao

where p is the power and can be measured by means of measuring the electric
current of the ball mill.
The noise and modeling errors will cause different distortions to the probability
assignments in the different models. Assume the probability distribution sd of
performance rates for all of the system elements at any instant t  0 and system
structure function as follows:

fðG1 ðtÞ . . . Gn Þ (8)

Accordingly, the total number of possible states or performance rates of the
system are:
Y
n
pp ¼ kj (9)
j¼1

Let Ln ¼ fg11 ; . . . ; g1k1 g      fgn1 ; . . . ; gnkn g be the space of

 possible com-

binations of performance rates for all system elements and M ¼ g1 ; . . . ; gpp be
the space of possible values of entire system performance levels. The transform
fðG1 ðtÞ . . . Gn ðtÞÞ : Ln ! M which maps the space of performance rates of system
elements into the space of system’s performance rates, is the system structure
Qn
function [10]. The probability of the system state is given as: si ¼ sjij ; the
performance rate for state i is: j¼1

gi ¼ fðgni1 ; . . . ; gnin Þ (10)

For the system under consideration, the estimation of a single number for the
probabilities and performance levels is very difficult. To avoid the “dimension
damnation”, the model is reduced to decrease the computational burden. Fuzzy
logic controllers (FLC) have the advantage to be robust and relatively simple to
design since they do not require the knowledge of the exact model. The fuzzy
system that serves the implementation of general decision-making in tuning the
driving motor power of the mill is presented in Fig. 5.

Interface
Signal
conditioner
Motor T

w
+ Ra La
C + J,
Servo- Va ia Vb
if D
Controller Ampli –
– q
D/A

Fig. 5 Hardware setup

24 Integrated Mining Fuzzy Association Rules For Mineral Processing 317

In this study, the proposed fuzzy logic controller has three inputs and outputs.
The tuned control surface is nonlinear corresponding to the property of controlled
plant l, or and np are the measured value of the ball mill load, the rotation speed
and the inlet negative pressure, respectively. In addition, p is the measured value of
the ball mill driving motor power. el, eo and enp which are input variables of the
fuzzy logic controller, represent the error of l, or and np respectively. ul, uw and unp
are the output variables of the fuzzy logic controller, which are usually used to
control the raw ore feeder, the driving motor speed and the recycle air damper,
respectively. Therefore, the probability distribution sd of the system is:

Y
3
fðg1i1 ; :::: ; g3i3 Þ; sd ¼ sjij (11)
j¼1

Furthermore, the max-min algorithm is used in fuzzy logic inference, and the
defuzzification is accomplished by the largest of maximum method.

3.1 Fuzzy Logic Controller

In the fuzzy control design methodology, we made use of heuristic information as

from an operator who has acted as a “human-in-the-loop” controller for the process.
To develop a set of rules applicable to mineral processing control, the practical
expertise is drawn on our knowledge performed through extensive mathematical
modeling, analysis, and development of control algorithms for diverse processes.
Then we incorporate these into a fuzzy controller that emulates the decision-making
process of the human, Fig. 6.
The function fðÞ, (11) is strictly defined by the type of connection between
elements in the reliability logic-diagram sense, i.e. on the structure of the logic-
diagram representing the system/subsystem. Despite the fact that the universal
generating function resembles a polynomial, it is not a polynomial because: (i) its
exponents are not necessary scalar variables, but can be arbitrary mathematical
objects (e.g. vectors); (ii) operators defined over the universal generating function

Fuzzy Controller
Defuzzification

Inference
Fuzzification

rc (t) Mechanism u (t) Grinding y(t)

Process

Rule-base

Fig. 6 Fuzzy controller architecture

318 S.C. Abou and T.-M. Dao

can differ from the operator of the polynomial product (unlike the ordinary gen-
erating function technique in which only the product of polynomials is defined)
[12]. Figure 7 idealizes a general flow transmission through out the system (e.g.,
ore, particle size, fluid flow, energy). For instance, consider a flow transmission
system shown in Fig. 7 which consists of three elements. As a result, the system
performance rate which is defined by its transmission capacity can have several
discrete values depending on the state of control equipments.
Assume the element 1, (the rheology of the slurry) has three states with the
performance rates: g11 ¼ 1.5, g12 ¼ 1, and g13 ¼ 0; the corresponding probabilities
are respectively: sp11 ¼ 0:8, sp12 ¼ 0:1 and sp13 ¼ 0:1. The element 2, (the pH) has
three states with the performance rates g21 ¼ 2, g22 ¼ 1.5, g23 ¼ 0 and the corresponding
probabilities sp21 ¼ 0:7, sp22 ¼ 0:22 and sp23 ¼ 0:08. The element 3, (the density)
has two states with the performance rates g31 ¼ 4, g32 ¼ 0 and the corresponding
probabilities sp31 ¼ 0:98 and sp32 ¼ 0:02. According to (9) the total number of
possible combinations of the states of elements are pp ¼ 3  3  2 ¼ 18.
In order to obtain the output sp for entire system with the arbitrary structure
function fð:Þ [9] used a general composition operator @f over individual universal
z-transform representations of n system elements:
8
>
> Xkj
>
> ¼ @ ð ðzÞ; . . . ; ðzÞ Þ ¼ sd jij :Z gjij
>
<
UðzÞ f u 1 u n and uðzÞ
i¼1
! (12)
>
> X k1 X k2 X
kn Yn
>
> UðzÞ ¼ ... fðg
sd j :z j ;...;g Þ
>
ji nin
:
i1 i2 in j¼1
where U(z) is z-transform representation of output performance distribution for
the entire system.
Figure 6 illustrates the basic control structure. The scheme includes a classical
PID control structure together with fuzzy corrector. The fuzzy corrector uses the
command input rc(t) and the plant output y to generate a command signal uc(t),
described by the following equations:
8
< eðtÞ ¼ rc ðtÞ  yðtÞ
>
DeðkÞ ¼ eðkÞ  eðk  1Þ (13)
>
:
mðkÞ ¼ F½eðkÞ; DeðkÞ

U1(z)

U3(z)

U2(z)

Fig. 7 A flow transmission structure

24 Integrated Mining Fuzzy Association Rules For Mineral Processing 319

Fig. 8 The fuzzy sets BN N SN Z SP P BP

1

e
0
–10 –5 0 5 10
BN N SN Z SP P BP
1

De
0
–10 –5 0 5 10
BN N SN Z SP P BP
1

output
0
–10 –5 0 5 10

In the above, e(k) is the position error between the command input rc(t) and the
process output y(k); De(k) is the change in position error. The term F[e(k), De(k)] is
a nonlinear mapping of e(k) and De(k) based on fuzzy logic. The term m(k)
represents a correction term. The control u(k) is applied to the input of the grinding
circuit. The purpose of the fuzzy corrector is to modify the command signal to
compensate for the overshoots and undershoots present in the output response when
the load dynamics has unknown nonlinearities.
Let x1 ; x2 ; x3 ; y1 ; y2 and y3 represent l, or, np, ul, uo and unp respectively. The
expertise and knowledge method used to build a rule base and membership func-
tions provide the description of e(k) and De(k) as inputs, and m(k) as the output. The
unified fuzzy inverse is given as[0.1, 0.1], Fig. 8. The fuzzy states of the inputs
and the output, all are chosen to be equal in number and use the same linguistic
descriptors: Big Negative (BN), Negative (N), Small Negative (SN), Zero (Z), Small
Positive (SP), Positive (P) and Big Positive (BP). Figure 8 illustrates the member-
ship functions. Despite the operator expertise and knowledge at the level of the
inference rules and the membership functions, some defects may appear. The
degree of membership of each value of attribute ik in any of its fuzzy sets is directly
based on the evaluation of the membership function of the particular fuzzy set with
the value of ik as input.
To improve the conventional FLC, association rules mining algorithms are
used to find the optimal membership functions. This is achieved based on the
membership functions depicted in Fig. 9.

3.2 Association Rules Miming Algorithm

Specified fuzzy linguistic terms in fuzzy association rules can be given only
when the properties of the attributes are estimated. In real life, contents of columns
(i.e., values of attributes) may be unknown and meaningful intervals are usually
320 S.C. Abou and T.-M. Dao

BN N SN Z SP P BP
1

Degree of membership

0.5

–0.1 –0.08 –0.06 –0.04 –0.02 0 0.02 0.04 0.06 0.08 0.1
First input ‘Residual’

Fig. 9 Membership functions

not concise and crisp enough. In this paper, the target is to find out some interesting
and potentially useful regularities, i.e., fuzzy association rules with enough support
and high confidence.We use  the following
 form for fuzzy association rules.
Let be x1i ; . . . ; xki and y1i ; . . . ; yki the antecedent set and the consequence set
respectively,
 1in a database.  A fuzzy
 1 association  rule mining is expressed as:
If X ¼
 i x ; . . . ; x
i
k
is A ¼
 i a ; . . . ; a
i
k
! Then
Y ¼ y1i ; . . . ; yki is B ¼ d1i ; . . . ; dki . Here, X and Y are disjoint sets of attri-
butes called item-sets, i.e., X  I; Y  I and X \ Y ¼ f. A and B contain the fuzzy
sets associated with corresponding attributes in X and Y, respectively.
Let m(.) represent the membership value of each element of the antecedent and
the consequence set. Under fuzzy taxonomies, using the measurements could result
in some mistakes. Consider for instance the following conditions:
1. mðxki Þ  mðyki Þ and mðxki Þ  mðym i Þ
2. mðxki Þ < mðyki Þ and mðxki Þ < mðym
i Þ

In the first condition, the confidence under fuzzy taxonomies of the two rules
is equal, while in the second, the coverage of the two rules is equal. This
situation gives rise to the following question: which rule can be judged as best
evidence rule?
In the proposed method the algorithm iterations alternate between the generation
of the candidate and frequent item-sets until large item-sets are identified. The
fuzzy support value of item-set Z is calculated as:
P Q
ti 2T zj 2Z mðaj 2 F; ti ðzj ÞÞ
SðZ; FÞ ¼ (14)
nT
24 Integrated Mining Fuzzy Association Rules For Mineral Processing 321

where nT is the number of transactions in the database. Also, to ensure the

accuracy of rules base, the consequent strength measure, c is used to estimate the
mined rules as:
!
X
n Y
kX Y
kY n Y
X kY
c¼ mðyki Þ ^ mðym
i Þ  mðyji Þ (15)
i¼1 k¼1 m¼1 i¼1 j¼1

Based on the literature review, the fuzzy-PID control is implemented in such

way that the fuzzy system has c_ (derivative of the PID output) as input and returns
an output g_ that has a nonlinear dependence upon the input. The dependence upon c_
ensures that changes in the steady value of c do not change the closed-loop
response characteristics of the fuzzy-PID controller. Hence the simplest form of
fuzzy-PID controller is:
 
dg 1 dc
¼ f a (16)
dt b dt
where a is an input scaling factor and b is an output scaling factor. The input
magnitude is scaled to unity by:
  1
dc
a¼ max   (17)
dt
 
 
The evaluation of max dc
dt  is normally estimated from one or more PID runs.

The final system implemented is presented as:

8 1
< ’BðjBjð’  1Þ þ 1Þ
> for jBj  1
f ðBÞ ¼ 1 for B  1 (18)
>
:
1 for B  1

The graph of the function f ðBÞ is presented in Fig. 10. It depicts sigmoidal
dependence on z. As a result, it may corrupt all slowly varying signals of the
process. In order to suppress the noise and get the nonlinear control surface, the
parameters of input membership functions and output membership functions of
the fuzzy rule base are tuned by a nonlinear optimization. In this study a sequential
quadratic programming (SQP) algorithm as presented in Matlab is used.
This approach represents state-of-the-art in non-linear programming methods,
because a non-linearly constrained problem can often be solved in fewer iterations
using SQP than an unconstrained problem. One of the reasons is that due to the
limits on the feasible area, the optimizer can make well informed decisions regard-
ing directions of search and step length. Note that widely overlapping membership
functions give good numerical results. However, fuzzy membership functions
322 S.C. Abou and T.-M. Dao

f (ζ) vs ζ with different values of j

2

1.5
j=1

1
j=8

0.5
j=2
j=4
f (ζ) 0

–0.5

–1

–1.5

–2
–2 –1.5 –1 –0.5 0 0.5 1 1.5 2
ζ

Fig. 10 f(z) with varying f

lacking any physical interpretation and loosing locality are possible. To avoid this,
different kinds of constraints may be put on the optimization: kinds of constraints,
inequality constraints and parameter bounds. SQP efficiently solve this constrained
nonlinear optimization problem in which the objective function and constraints may
be nonlinear functions of variables.

4 Simulation Results

Through available actuators it is possible to adjust the speed of the grinding circuit
feeders, the fluid added to the mill and sump, the pump and the driving motor speed.
The level of the sump is controlled by adjusting the set-point of pump velocity. The
water flow to the flotation circuit is kept proportional to the load reference. As
mentioned, the above process is subject to many disturbances, being the hardness of
the raw feeding ore the most significant one. This exerts a strong influence on the
quality of the grinding product. Modeling of the process is further complicated by
the fact that different operation points are defined by changes in mineral hardness.
Based on a reasonably simple transfer function, the adjustable input ports of the
driving motor and entrance negative pressure of the breakage circuit are initialized
with step signal. The strongly coupling characteristics of variables implies that if
regulating one controlled variable is attempted with one manipulated variable, the
24 Integrated Mining Fuzzy Association Rules For Mineral Processing 323

1.2

0.8
Outputs

0.6

PID Controller
0.4
Fuzzy Logic Controller
Fuzzy Logic Controller
0.2
And change amplitude after 25s

0
0 5 10 15 20 25 30 35 40 45
time [s]

Fig. 11 Fuzzy logic controller step response

other controlled variables will, more likely than not, be influenced in an undesired
fashion. Using the aforementioned fuzzy logic rules the simulated step response of
the grinding circuit is shown in Fig. 11. The amplitude changes are followed
smoothly to reach the steady state. To ensure both controller stability and perfor-
mance, although guidelines exist, the control procedure will require a number of
iterations until a solution is found, satisfying all requirements.
The existence of large time delays, time-varying parameters and nonlinearities are
some of the other difficulties generally encountered in mineral processing control.
As illustrated in Fig. 12, one of the major attractions of the proposed methodol-
ogy is that it can be readily used to design robust multivariable controllers that take
account of all control loop interactions while still guaranteeing all users specified
stability and performance requirements. As shown in Fig. 12, in the presence of
disturbances the fuzzy logic controller model presents a response better than the
conventional PID model.

5 Conclusion

Understanding how to modify rheological characteristics of fine particle is a key for

mineral process performance. Constrained optimization methods with integrated
association rules mining based on a nonlinear, fuzzy dynamic control scheme were
designed to improve the grinding process. The proposed structure of fuzzy logic
controller combines the advantages of the fuzzy networks approach and the associa-
tion mining rules. The proposed algorithms can control the process in a better way
324 S.C. Abou and T.-M. Dao

Best Fuzzy-logic control for physical model

1.4

1.2

0.8
Outputs

0.6

0.4

0.2

0
0 2 4 6 8 10 12 14 16 18 20
time [s]

Fig. 12 Tuned fuzzy logic controller using physical model

than the conventional PID approach. Based on simulation results it can be deduced
that the fuzzy controller is faster than the conventional controller in the transitional
state, and also presents a much smoother signal with less fluctuation at steady state.
The proposed method has a strong adaptability and can overcome nonlinear and
strong coupling features of mineral processing in a wide range.

References

1. S.C. Abou, M.D. Thien, Fuzzy logic controller based on association rules mining: application
to mineral processing. International Conference on Modeling, Simulation and Control
(ICMSC’09), San Francisco, USA, 20–22 Oct 2009
2. S.C. Abou, Contribution to ball mill modeling. Doctorate Thesis, Laval University, Quebec,
Canada, 1998
3. G. Levitin, Universal Generating Function and Its Applications (Springer, Berlin, 2005)
4. S. Morell, The prediction of powder draws in wet tumbling mills. Doctorate Thesis, University
of Queensland, 1993
5. E.W. Davis, Fine crushing in ball mills. AIME Trans. 61, 250–296 (1919)
6. A. Desbiens, K. Najim, A. Pomerleau, D. Hodouin, Adaptive control-practical aspects and
application to a grinding circuit. Optim. Control Appl. Methods 18, 29–47 (1997)
7. R.K. Rajamani, J.A. Herbst, Optimal control of a ball mill grinding circuit: II. Feedback and
optimal control. Chem. Eng. Sci. 46(3), 871–879, 1991
24 Integrated Mining Fuzzy Association Rules For Mineral Processing 325

8. L. Zhai, T. Chai, Nonlinear decoupling PID control using neural networks and multiple
models. J. Control Theory Appl. 4(1), 62–69, 2006
9. W. Kuo, R. Wan, Recent advances in optimal reliability allocation. IEEE Trans. Syst. Man
Cybern. Part A Syst. Hum. 37, 143–156, 2007
10. H.J. Zimmermann, Fuzzy Set Theory and its Application, 2nd edn. Kluwer Academic,
Dordrecht, 1991
11. R. Agrawal, T. Imielinski, A. Swami, Mining association rules between sets of items in
large databases. International conference on management of data (ACM SIGMOD’93),
Washington, DC, 1993
12. J. Guan, Y. Wu, Repairable consecutive-k-out-of-n: F system with fuzzy states. Fuzzy Sets
Syst. 157, 121–142 (2006)
13. J. Huang, M.J. Zuo, Y. Wu, Generalized multi-state k-out-of-n: G systems. IEEE Trans.
Reliab. 49(1), 105–111 (2000)
Chapter 25
A Combined Cycle Power Plant Simulator:
A Powerful, Competitive, and Useful Tool
for Operator’s Training

Eric Zabre, Edgardo J. Roldán-Villasana, and Guillermo Romero-Jiménez

Abstract In this chapter we present the development of a combined cycle

power plant simulator for operator’s training. This simulator has been designed
and developed by the Simulation Department of the Instituto de Investigaciones
Eléctricas. This is one of the several technological developments carried out by
the Simulation Department, and it is part of a full scale simulators group that
belongs to the Comisión Federal de Electricidad that offers the electrical service
in the whole country. The simulator is currently under testing and evaluation
by the final user, before entering on service at the National Center for Opera-
tor’s Training and Qualification. Tendencies of these development and impact
within the operators’ scope as well as some results and future works are also
presented.

1 Introduction

The Instituto de Investigaciones Eléctricas (IIE, Mexican Electric Research Insti-

tute, founded in 1975) has been the right hand as research and development
institution of the Comisión Federal de Electricidad (CFE, Mexican electrical util-
ity company) offering technical innovations and developments. The Simulation
Department (SD), one of the technical areas of the IIE, has developed a series of
software programs and applications to generate full scale simulators for operators’
training. In this chapter, a full scope simulator of a combined cycle power plant
(CCPP) is presented and its different parts are described within training purposes.

E. Zabre (*)
Instituto de Investigaciones Eléctricas, Simulation Department, Reforma 113, Col. Palmira,
Cuernavaca, Morelos, Mexico 62490
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 327

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_25, # Springer ScienceþBusiness Media B.V. 2010
328 E. Zabre et al.

The development of simulators’ technology involves many areas and, to reach

the goals, different specialists were required.
Engineers of software, processes modeling, control, communications network,
maintenance and operators have shared the same commitment.

2 Antecedent

According to Webopedia’s Online Dictionary, a combined cycle power plant is

defined as “the coexistence of two thermodynamic cycles in the same system”, and
simulation as “the process of imitating a real phenomenon with a set of mathemati-
cal formulas; in addition to imitating processes to see how they behave under
different conditions, simulations are also used to test new theories”.
Generally speaking, fluids of work: the gas and the steam water, product of
combustion and a heat recovery steam generator, respectively, produce energy in
the so called combined cycle. A combined cycle power plant may be very large,
typically rated in the hundreds of mega watts. If any combined cycle power plant
operates only the gas turbine to generate electricity, and diverts the exhaust gases,
there is a substantial loss of efficiency. A typical large gas turbine is in the low 30%
efficiency range, but combined cycle plants can exceed 60% of global efficiency,
because the waste heat of the gas turbine is used to make steam to generate
additional electricity through a steam turbine, getting a high efficiency of electricity
generation over a wide range of loads [1, 2]. The most modern combined cycle
power plants implement innovative strategies in order to reduce emissions.
In 2006, the IIE’s SD developed a gas turbine full scope simulator (GTS) [3]
based on the combined cycle power plant “El Sauz”, for the Centro Nacional de
Adiestramiento y Capacitación (CENAC, National Center for Operator’s Training
and Qualification), using proprietary technology of the IIE. The GTS is formed by a
computer platform, an instructor console (IC), a set of interactive process diagrams
(IPD), and mathematical models of the plant systems.
In this chapter, the development of a combined cycle power plant simulator
(CCPPS) is presented. It is based on 1, 2, and 3 unit’s package from the Chihuahua
II CCPP [1, 2, 4] (also known as “El Encino”, located at the north of the Mexican
territory, generates 450 MW and went into operation on May, 2001). The CCPPS
has a configuration of two gas turbines (GT1 and GT2), two heat recovery steam
generators (HRSG1 and HRSG2), and one steam turbine. One of the gas turbines
derives from the actual CENAC0’s GTS; the second one is a simplified gas turbine
with a reduced scope, this means that it has the values of the main processes
variables, in such a way to simulate the necessary variables to interact with the
steam turbine models.
Nowadays, combined cycle power plants (CCPP) represent the most frequent
form of energy production in México [5, 6]. Figure 1 shows a typical diagram of the
main components of a CCPP (with only one HRSG and without auxiliary systems,
for simplicity). An advantage of this type of plant is the possibility of constructing
25 A Combined Cycle Power Plant Simulator 329

Cycle
Heat Recovery chimney
Steam Generator
Natural gas supply

Substation Gas
Main turbine Steam high & low Electric B
transformer pressure turbine generator
Transmission Steam Transmission
tower Air tower
Vacuum system cooling
deareator condenser

Main
Demineralized
transformer
B plant
Fan Condenser
tank Water
Circulating water pump well
Condensate
pump Raw water
tank

Fig. 1 Combined cycle power plant diagram

them in two stages. The first one corresponds to the gas unit, which can be finished
in a brief term and immediately initiate its operation; later it is possible to finish the
construction of the steam unit, and then the combined cycle may be completed.

3 Architecture Configuration

3.1 Software Architecture

In the CCPPS several configuration modules are identified, as shown in Fig. 2. All
modules are strongly related to a real time system (RTS) that control and sequence
the different tasks running in the simulator. The modules are the following:
l Instructor Console® (IC). This is an important man machine interface (MMI) of
the instructor, which controls and coordinates the scenarios seen by the operator
under training. These scenarios could be presented to the operator as a normal
plant operation or under any malfunction of a particular system or equipment. A
more detailed functional description of this interface is given below.
l CONINS. A module to retrieve all the static information during simulation
session, for example: malfunctions, remote functions, local instrumentation,
external parameters, etc.
l Operator Console (OC). This system is a replica of the real operation station
and allows the operator to monitor and control the plant parameters from
three different sets of graphical interfaces (GI): gas turbine, steam turbine and
330 E. Zabre et al.

Real Time
SABADAMAS BDSIM System Instructor
Console Operator
Console

Mathematical CONINS Special

IPD
models tasks
Auxiliary
Gas Steam electrical
turbine turbine board

Alarms and graphical

variable tendencies

Fig. 2 Conceptual software architecture of the simulator

auxiliary emulated electrical external boards. These three GI have a special

display for the alarms and graphical variable tendencies.
l IPD. This module executes all interactive processes diagrams or graphical
interfaces, accessed from the IC or from the OC.
l Mathematical models. This system has a set of mathematical models that rep-
resent all the simulated systems of the plant. The RTS coordinates the execution
sequence of each mathematical model, which runs sequentially in two proces-
sors within a 1/10 s execution frame.
l The special task module which coordinates other applications like environmen-
tal plant noises such as spinning turbine, start-up or shutdown pumps, etc.
l BDSIM. A module to store the simulator information in a data base using
Microsoft Access™ interfaces and programs written on SQL language. The
data base can be updated by means of external (off-line) tool as data base
administrator system, named SABADAMAS®. It is dedicated to receive/update
any information required by the executive system. The CCPPS uses and man-
ages about 27,000 global variables.

3.1.1 Instructor Console

The IC is the MMI, which is constituted by functions as shown in Fig. 3. The

simulation session is controlled since the IC. The instructor has a main menu of
functions. The main functions of the IC are the following:
1. Run/Freeze. The instructor may start or freeze a dynamic simulation session.
The mathematical models of the control and process respond to any action of the
trainee in a very similar way as it occurs in the real plant.
2. Initial conditions. The instructor may select an initial condition to initiate the
simulation session, either from the general pre-selected list of initial conditions
or from their own catalogue (each instructor has access up to 100 initial condi-
tions). Also in this function, it is possible to create an initial condition or to erase
25 A Combined Cycle Power Plant Simulator 331

INSTRUCTOR: Instituto de Investigaciones Eléctricas SIMULADOR: SCCC

Control Condiciones iniciales Funciones de instrucción Seguimiento Miscelanea Salir

ESTACION-1 PRUEBA 70 TG#1 y TGSIMPLIFICADA 150 MW, TV 152.6 MW

Fig. 3 Part of the main display of the Instructor Console

an old one. The simulators gets an automatic snapshoot every 15 s and may be
configured up to a frequency of 10 min.
3. Simulation speed. Since the beginning of the simulation session the simulator is
executed in real time, but the instructor may execute the simulator up to two or
three times faster than real time. For example, slow thermal processes such as
the turbine iron-heating can be simulated to occur faster. To slow the simulation
speed (up to ten times) may be useful to study some fast events with detail.
4. Malfunctions. It is used to introduce, modify, or remove a simulated failure of
plant equipment, for example: pumps trips, heater exchanger tube breaking,
electrical switch opening, valve obstruction, etc. The instructor has the option to
define the malfunction initial time and its permanence time as well as the
evolution time.
5. External parameter. The external conditions such as atmospheric pressure and
temperature, voltage and frequency of the external system, among others can be
modified by the instructor.
6. Repetition. Simulation sessions may be repeated as many times as the instructor
considers, including the trainees actions.
7. Actions register. This is an automatic registration of the actions carried out by
the trainee in order to be re-played exactly in the same sequence and time with
the purpose of analyzing what the trainee did during the simulation session.
8. Development tools. The simulator has implemented some others helpful tools to
use during simulation session development, for example: to monitor and change
on line any selected list of global variable, tabulate any selected list of variable
and plot them.

3.1.2 Operator Console

The OC is a replica of the real operation station, and it is formed by the interfaces
communication system and it consists of three main displaying sets of plant inter-
faces: gas turbine, combined cycle and auxiliary electrical external board interface.
From there it is possible to display any logical or analogical plant parameter
and to operate and handle some manual or automatic controls related to the plant
332 E. Zabre et al.

operation. Its objective is to maintain the optimal functional conditions, to monitor

or graphic tendencies values of diverse parameters, and to re-establish the normal
conditions, among others functions.

Gas Turbine Interface

The gas turbine interface (GTI), is a reproduction of the dynamic screens on which
it is possible to monitor and control the gas turbine parameters, including: ready to
start/trips; vibration analysis; combustion flashback; lube oil system; trend over-
view; turbine cooling system; emissions; synchronization; simplified gas turbine
(SGT) model; etc. [7].

Combined Cycle Interface

The combined cycle interface (CCI) is a translation of the real operator console,
known as engineering console. An example is shown in Fig. 4.

Auxiliary Electrical External Board Interface

The auxiliary electrical external board interface represents a couple of screens that
lead all the necessary external instrumentation to synchronize the electrical unit.

3.2 Software Platform

The CCPPS has Windows XP™ as operating system, it was programmed using
MS Visual Studio 2005; Fortran Intel™ for the mathematical models; Flash and

POP-UP LINEA AP/PI DEL HRSG 1

01 HAH41
AA093 026.4
01 HAH31 barg
AA901 DOMO PI No. 1
– 0002.9 mm 000%
01 HAH21
01 HAH21 AA900 01LCQ6
AA901 AA403
01 LCQ61
AA401
AP
000%

Fig. 4 CCI High and intermediate pressure of the HRSG1 (detail)

25 A Combined Cycle Power Plant Simulator 333

VisSim™ for the gas turbine screens; the steam turbine screens were translated from
the CCPP control, and C# was used to program the modules of the simulation
environment. The simulation environment (called MAS®, proprietary software of
the IIE [8]) has three main parts: the real time executive, the operator module, and
the console module. Each module runs in a different personal computer (PC), and all
of them are communicated by means of a TCP/IP protocol. The modules of the MAS
are programmed on C# under MS Visual Studio™ software development platform.

3.3 Hardware Architecture

As Fig. 5 shows, the CCPPS is constituted by four PC interconnected through a fast

Ethernet local area network. Each PC has a mini-tower PentiumD™ processor with
3.6 GHz, 1 GB of RAM memory, 40 GB HD, and Windows XP™ as operating
system. Figure 6 shows a schematic of this architecture.

Control
Room
Operation Operation
Station1 (OS1) Station2 (OS2)

Lan switch
Instructor
Maintenance Console (IC)
node (MN)

Fig. 5 Hardware architecture

Fig. 6 View of the Combined Cycle Power Plant Simulator on site.

334 E. Zabre et al.

The Instructor Console PC, also named simulation node (SN) has two 2000 flat
panel monitors, two operator stations (0S1 and 0S2) PC, each one has also 2000 flat
panel monitors, but the 0S1 has two additional 4200 flat panel monitors as auxiliary
monitors.
The maintenance node (MN) is used to make any modifications to the software
or process or control models and are tested and validated by an instructor before to
install it in the simulator. Figure 6 shows the real and final simulator architecture as
described above.

4 Modeled Systems

The control models were translated into C# by means of a graphical software tool,
while the system processes were programmed on Fortran Intel™.

4.1 Control System

Generally speaking, the distributed control system (DCS) is a set of PLC in where
are allocated the algorithms to control, in an automatic or semi-automatic way, all
the systems of a power plant. These control algorithms are organized in components
with specific function or task, for example: PID controllers, high/low detectors,
timers, memories set/reset, etc. This organization is represented by means of a
network of these components, which communicates information through connec-
tions, see Fig. 7. These networks are organized in a hierarchical way, in the bottom
levels there are the basic elements like AND, OR, NOT gates, in the middle level

Modulos
LD pcs1 AND
02611 00003
pcs2

LD pcs3 NOT
04001 hrsg1 00002

hrsg2
AND
deh 00004
OR
das 00001
gwc1
LD gwc2
02808

Fig. 7 Control diagram (detail) for speed control on the digital electro-hydraulic control (EHC)
system
25 A Combined Cycle Power Plant Simulator 335

are diagrams, and finally in the top level are modules. Then, with a set of modules, a
DCS was built [9].

4.2 DCS Model for Real-Time Simulation

For the gas turbine, the control models were developed by drawing the diagrams
(considering the original ones provided by the designer) into VisSim which gen-
erates the code of the drawing into standard C language routines (one per each
diagram).
For the control of the steam processes the DCS model was developed with C#
programming language, and taking into account the modularization made in the
power plant of reference. The main control modules modeled were:
l pcs1 and pcs2, auxiliary control system part 1 & 2
l pcs3, electric network control
l hrsg1 and hrsg2, control of heat recovery system generator, part 1 & 2
l deh, digital electro-hydraulic control system
l das, digital acquisition system

4.3 The Graphic Visualization Tool

The graphic visualization tool (GVT) is a software application developed to

visualize components in diagrams of the Simulator Control Model. This tool was
very useful during the simulator development and adjustments because it allows to
verify and to visualize signals, states, inputs, outputs and parameters of components
on line.
The GVT allows to disable diagrams, modules or components in such a way that
the components may be isolated to verify its behavior without the influence of the
entire control model.

4.4 Processes System

There are 14 different system processes modeled of the previous gas part plus 26
systems processes modeled of the steam part which together are the mathematical
representation of the behavior of the CCPP. All models were developed using
Fortran Intel™ in a methodology proposed by the IIE [3], which is summarized
in a few steps: (a) to obtain and classify process information directly from the power
plant; (b) to analyze information and state a conceptual model to be simulated;
(c) to simplify the analyzed information to show only the simulated equipments
336 E. Zabre et al.

Table 1 Classified common groups of the plant systems

Group Systems included
Water Condensed water, including air-condenser; HRSG water supply
Steam HRSG1 and HRSG2 water; HP, IP, and LP water and steam
Turbine HP, IP, and LP steam turbine; turbine metals
Electric and generator Electric (generator gas unit1, unit2, and steam unit); primary and
secondary regulation for the three generators
Auxiliaries Auxiliary steam; cooling water; control oil; unit efficiencies; generator
cooling air; turbine seal steam; lubricating oil
Auxiliaries minimized Chemical’s dosage to the cycle water-steam; instrument and service air;
HRSG1 and HRSG2 drain; steam drain; chemical analysis; potable
water; services water; water of sanitary garbage; demineralized
water; common drain; and chemical drain. This group contains a set
of simplified models with a reduced scope, because it is not
necessary to use the complete physical representation
Gas Fuel gas; gas turbine; compressed air; combustor; air for instruments;
lubrication oil; control oil; water-ethylene glycol cooling; electrical
network; generator; generator cooling with hydrogen; turbine
(metals temperatures and vibrations); performance calculations
(heat rate and efficiencies) and a simplified gas turbine

and its nomenclature; (d) to justify all possible assumptions; and (e) to obtain the
flow and pressure network with an operational point, namely at 100% of load. The
plant systems were classified on groups as listed on Table 1.

5 Project Control

The Project was divided in 10 main phases and 50 tasks. Time, work, and resources
were loaded and controlled with Microsoft Project Manager™ Some overwork was
needed in order to accomplish the programmed dates. The main sources of delays
were some discrepancies that needed to be solved when the local tests were
performed and the delay in the delivering of the plant information.
The project lasted seventeen months. The CFE participated actively in getting
the real plant data and defining with the SD the data not available. In particular CFE
created the control diagrams that were no part of the DCS and helped in performing
the plant operation on the simulator to determine and adjust the models behavior.
There was a head of project, one responsible of the software activities, one for
the hardware, one for the process modelling and one for the control models. A total
of seventeen researchers participated during peak stages of the development of the
project. All activities were closely followed in order to assure a punctual and
precise ending. Weekly revisions were performed with the heads of area. One
summarized report was delivered to the costumer monthly and a full revision was
performed every two months. The quality system implanted in the SD allowed the
opportune correction actions when needed.
25 A Combined Cycle Power Plant Simulator 337

6 Results

The simulator validation was carried out proving its response against the very
detailed operation acceptance simulator test procedures elaborated by specialized
personnel from CFE.
As example of the models behavior, some results of the GTS are presented. The
customer provided plant data for both, an automatic start up procedure and the
down load operation. The simulator results were compared with these data. No data
for other transients were available. Figure 8 presents a comparison between the
expected and simulator speed.
In Fig. 9, the x-axis is the sum of the turbine speed (rad/s) and then power
produced (MW). Pressure in the combustor, its temperature and the temperature

3600

3000
Speed (RPM)

2400

1800
Plant
1200
Simulator
600

0 150 300 450 600 750 900

Time (sec)

Fig. 8 Turbine speed plant and simulator comparison

1400 700

1200 Comb Press Original Comb Press Sim

600
Comb Shell Temp Original Comb Shell Temp Sim
Temperature (°C)

1000
Pressure (KPa)

Exhaust Temp Original Exhaust Temp Sim 500

800 400

600 300

400 200

200 100

0 0
0 66 184 314 417 912
Velocity + Power (rad / s + MW)

Fig. 9 Plant and simulator variables comparison

338 E. Zabre et al.

in the exhaust were chosen to exemplify the results. While the turbine is increasing
its speed the power is zero and when the plant produces power (up to 150 MW) the
speed is constant at 376.8 rad/s. In the graphic (with a non-linear x-axis) may be
observed that the behavior of the simulator variables is similar than the real
variables values.
Besides, the simulator was probed in all the normal operation range, from
cold start conditions to full charge, including the response under malfunctions
and abnormal operation procedures. In all cases the response satisfied the ISA-
S77.20-1993 Fossil-Fuel Power Plant Simulators Functional Requirements norm.

7 Conclusions

The CCPPS validation is been carried out by CENAC specialized personnel under
rigorous acceptance simulator testing procedures, and with a lot of experience in the
use of simulators, to ensure that this simulator will fulfill the performance specified
by the end user in order to have one more useful tool to train the future generation of
operators of the combined cycle power plants. The results demonstrate than the
simulator is a replica, high-fidelity, plant specific simulator for operator training.
Realism is provided by the use of DCS screens emulation.
This simulator had several adjustments during its development, to assure that its
behavior and dynamics will operate similarly to those obtained in the real plant.
An additional aspect that gives a lot of certitude and robustness to the simulator
behavior is that the simulator has the same DCS that the Chihuahua CCPP. Only
minimum adjustments and changes in the controller parameters of regulation loops
were required, which means that the process models that were obtained by the SD
engineers reproduce correctly the real behavior of the reference plant. The com-
bined cycle power plants have powerful features that include high thermal effi-
ciency, low installed cost, wide range of burning fuels, short installation cycle,
compared to conventional thermal, nuclear, and steam plants, as well as low oper-
ation costs, it is expected that this simulators will be so strongly claimed for training
purposes in a near future.

8 Future Works

As future works, we have considered some projects that might improve and
robustness this development, with the purpose of giving a better approach to the
control room of the power plant. Among these are the following: (1) to standardize
the MMI’s simulator to original CCPP interface view, because it is the real
equipment the Chihuahua combined cycle power plant has, so it will be very val-
uable as well as useful to this power plant trainee operators; (2) to expand the
simplified gas turbine model; (3) to include thermal regimen of a combined cycle
25 A Combined Cycle Power Plant Simulator 339

power plant; and (4) to incorporate a touch screen keyboard or replace the virtual
keyboard by a physical one with the same functionality that the used in the control
room.

Acknowledgment The simulators developed by the SD of the IIE have been projects carried out
thanks to the effort, enthusiastic and very professional hard work of researchers and technicians of
the IIE as well as the co-operation of operators, trainers and support personnel of CFE training
center. This project was carried out thanks to the financial support of CFE; R. Cruz was the
technical manager as client counterpart; A. Matı́as took part as specialized operator during the
simulator tests.
About the SD personnel’s labor, it is necessary to mention that all researchers participants
contributed with their valuable experience during design steps, implementation and final tests
concerning to this simulator.

References

1. L. Bo, M. Shahidehpour, Short-term scheduling of combined cycle units. IEEE Trans. Power
Syst. 19(3), 1616–1625 (Aug. 2004)
2. J.Y. Shin et al., Analysis of the dynamic characteristics of a combined-cycle power plant.
Energy Int. J. 27(12), 1085–1098 (Dec 2002). ISNN 0360-5442
3. J. Roldán-Villasana Edgardo, M.J. Cardoso, Y. Mendoza-Alegrı́a, Modeling methodology for
operator’s training full scope simulators applied in models of a gas-turbine power plant, in
Memorias del 9o. Congreso Interamericano de Computación Aplicada a la Industria de
Procesos, 25 al 28 de agosto, Montevideo, Uruguay, 2009
4. Mitsubishi Heavy Industries, Chihuahua Combined Cycle Power Plant Begins Commercial
Operation in Mexico, http://www.mhi.co.jp/en/ power/news/sec1/2001_sep_08.html, Sept
2001
5. CFE Comisión Federal de Electricidad, Una empresa de clase mundial, Centrales, http://www.
cfe.gob.mx/QuienesSomos/queEsCFE/ estadisticas/Paginas/Indicadoresdegeneración.aspx,
2009
6. D.L. Chase, GE power systems, Combined-Cycle Development Evolution and Future, Sche-
nectady, NY, 1998 http://www.gepower.com/prod_serv/products/tech_docs/en/downloads/
ger4206.pdf
7. E.J. Roldán-Villasana, Y. Mendoza-Alegrı́a, J. Zorrilla-Arena, Ma.J. Cardoso, EMS 2008,
Development of a Gas Turbine Full Scope Simulator for Operator’s Training, in Second
UKSIM European Symposium on Computer Modeling and Simulation, Liverpool, UK, 8–10
Sept 2008
8. L.A. Jiménez-Fraustro, Desarrollo del Medio Ambiente de Simulación para Simuladores en
Tiempo Real basado en sistema operativo Windows XP, Reporte Interno, Gerencia de Simula-
ción, IIE, México, 2005
9. G. Romero-Jiménez, V. Jiménez-Sánchez, E.J. Roldán-Villasana, WASET 2008, in Proceed-
ings of World Academy of Science, Engineering and Technology, vol. 30, July 2008, ISSN
1307-6884, Graphical Environment for Modeling Control Systems in Full Scope Training
Simulators, Paris, France, 5–7 July 2008, pp. 792–797
Chapter 26
Texture Features Extraction in Mammograms
Using Non-Shannon Entropies

Amar Partap Singh and Baljit Singh

Abstract This paper deals with the problem of texture-features-extraction in

digital mammograms using non-Shannon measures of entropy. Texture-features-
extraction is normally achieved using statistical texture-analysis method based on
gray-level histogram moments. Entropy is important texture feature to measure the
randomness of intensity distribution in a digital image. Generally, Shannon’s
measure of entropy is employed in various feature-descriptors implemented so
far. These feature-descriptors are used for the purpose of making a distinction
between normal and abnormal regions in mammograms. As non-Shannon entropies
have a higher dynamic range than Shannon’s entropy covering much wider range
of scattering conditions, they are more useful in estimating scatter density and
regularity. Based on these considerations, an attempt is made to develop a new type
of feature-descriptor using non-Shannon’s measures of entropy for classifying
normal and abnormal mammograms. Experiments are conducted on images
of mini-MIAS (Mammogram Image Analysis Society) database to examine its
effectiveness. The results of this study are quite promising for extending the work
towards the development of a complete Computer Aided Diagnosis (CAD) system
for early detection of breast cancer.

1 Introduction

Cancer remains one of the frequently occurring fatal diseases affecting men and
women throughout the world. Among the cancer diseases, breast cancer is espe-
cially a concern in women. According to the statistics, breast cancer is one of the

A.P. Singh (*)

Department of Electronics & Communication Engineering, SLIET, Longowal, 148106, Sangrur,
Punjab, India
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 341

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_26, # Springer ScienceþBusiness Media B.V. 2010
342 A.P. Singh and B. Singh

major causes for the increase in mortality among middle-aged women in both the
developed and developing countries. However, the etiologies of breast cancer are
unknown and no single dominant cause has emerged. Still, there is no known way
of preventing breast cancer but early detection allows treatment before it is spread
to other parts of the body. However, it is evident that the earlier breast cancer is
found, the better chance a women gets for a full recovery. Moreover, the early
detection of breast cancer can play a very important role in reducing the morbidity
and mortality rates.
Mammography [1, 2] is the single most effective, reliable, low cost and highly
sensitive method for early detection of breast cancer. Mammography offers high-
quality images at low radiation doses and is the only widely accepted imaging
method for routine breast cancer screening. It is recommended that women at the
ages of 40 or above should have a mammogram every 1–2 years. Although
mammography is widely used around the world for breast cancer detection, there
are some difficulties when mammograms are searched for signs of abnormality
by expert radiologists. Such difficulties are that mammograms generally have low
contrast compared with normal breast structure and sign of early disease are often
small or subtle. This is the main cause of many missed diagnoses that can be
mainly attributed to human factors such as subjective or varying decision criteria,
distraction by other image features, or simple oversight. As the consequences of
errors in detection or classification are costly and since mammography alone can
not prove that a suspicious area is tumorous, malignant or benign the tissue has to
be removed for closer examination using breast biopsy techniques. Nevertheless, a
false-positive detection causes unnecessary biopsy. On the other hand, in false-
negative detection an actual tumor remains undetected. Thus, there is a significant
necessity for developing methods for automatic classification of suspicious areas
in mammograms, as a means of aiding radiologists to improve the efficacy of
screening programs and avoid unnecessary biopsies.
A typical mammogram contains a vast amount of heterogeneous information
that depicts different tissues, vessels, ducts, chest skin, breast edge, the film and the
x-ray machine characteristics [3–5]. In order to build a robust diagnostic system
towards correctly classifying abnormal and normal regions of mammograms,
present all the available information that exits in mammograms to the diagnostic
system so that it can easily discriminate between the abnormal and normal tissues.
However, the use of all the heterogeneous information, results to high dimensioned
features vectors that degrade the diagnostic accuracy of utilized systems signifi-
cantly as well as increase their computational complexity, Therefore, reliable
feature vectors should be considered the reduce the amount of irrelevant informa-
tion thus producing robust mammographic descriptors of compact size [6–8].
Texture is one of the important characteristics used in classifying abnormal and
normal regions in mammograms. The texture of images refers to the appearance,
structural and arrangement of the parts of an object within the image [9, 10].
Images used for diagnostic purposes in clinical practice are digital. A two dimen-
sional digital image is made up of little rectangular blocks or pixels (picture
elements). Each is represented by a set of coordinates in space and each has a value
26 Texture Features Extraction in Mammograms Using Non-Shannon Entropies 343

representing the gray-level intensity of that picture element in space. A feature value
is a real number, which encodes some discriminatory information about a property
of an object. Generally speaking, texture feature extraction methods can be classified
into three major categories, namely, statistical, structural and spectral. In statistical
approaches, texture statistics such as the moments of the gray level histogram, or
statistics based on gray level co-occurrence matrix are computed to discriminate
different textures [11]. For structural approaches, texture primitive, the basic element
of texture, is used to form more complex texture pattern by grammar rules which
specify the generation of texture pattern. Finally, in spectral approaches, the textured
image is transformed into frequency domain. Then, the extraction of texture features
can be done by analyzing the power spectrum. Various texture descriptors have
been proposed in the past. In addition to the aforementioned methods, Law’s texture
energy measures, Markov random field models, texture spectrum etc. are some other
texture descriptors [12–14].
This paper deals with the problem of statistical approaches to extract texture
features in digital mammogram. Gray level histogram moments method is normally
used for this purpose. Entropy [15–17] is an important texture feature, which is
computed based on this method, to build a robust descriptor towards correctly
classifying abnormal and normal regions of mammograms. Entropy measures
the randomness of intensity distribution. In most feature descriptors, Shannon’s
measure is used to measure entropy. However, in this paper, non-Shannon measures
are used to measure entropy. Non-Shannon entropies have a higher dynamic range
than Shannon entropy over a range of scattering conditions, and are therefore more
useful in estimating scatter density and regularity [18].

2 Gray Level Histogram Moments

Regarding the statistical approach for describing texture, one of the simplest
computational approaches is to use statistical moments of the gray level histogram
of the image. The image histogram carries important information about the content
of an image and can be used for discriminating the abnormal tissue from the local
healthy background. Considering the gray level histogram {hi, i ¼ 0, 1, 2, . . .,
Ng1}, where Ng is the number of distinct gray levels in the ROI (region of interest).
If n is the total number of pixels in the region, then the normalized histogram of the
ROI is the set {Hi, i ¼ 0, 1, 2, . . ., Ng1}, where Hi¼ hi/n. Generally, Shannon’s
measure is used to estimate entropy in most of the descriptors. Shannon’s entropy
(S) is a degree of randomness and is defined as

X
N g 1

S¼ Hi log2 ðHi Þ

i¼0
344 A.P. Singh and B. Singh

However, non-Shannon’s entropies have a higher dynamic range than Shannon

entropy covering a wider range of scattering conditions. Important non-Shannon
measures of entropies include Renyi’s entropy (R), Havrda and Charvat’s Entropy
(HC) and Kapur’s entropy (K). These entropies are defined as

!
X
N g 1
1
R¼ log2 Hi a
1a i¼0

a 6¼ 1; a > 0

!
X
N g 1
1 a
HC ¼ Hi  1
1a i¼0

a 6¼ 1; a > 0

g 1
NP
Hi a
1
Ka;b ¼ log i¼0
b  a 2 NPg 1
Hi b
i¼0
a 6¼ b; a > 0; b > 0

3 Experimental Results

Experiments were conducted on images of mini-MIAS database (Mammogram

Image Analysis Society) database, UK. For implementation, MATLAB 7.0 was
used. The mammograms of about 200 images were considered for simulating the
proposed work. For the completeness of description several experimental results are
depicted in Tables 1 and 2 at different values of a and b. The exact values of
parameters are of less significant and hence the range values have been considered
for true or best classification
The basic classification of normal, benign and malignant mammogram images
based on the values of the texture parameters are shown in Tables 3 and 4. The
original images and their corresponding histograms for basic classification are
shown in Figs. 1–3. Graphical representation of these classifications is shown in
Figs. 4 and 5. In order to validate the results, uncompressed fatty, fatty and non-
uniform breast mammogram images are considered to examine the versatility of the
proposed feature descriptor. These images and their corresponding histograms are
shown in Figs. 6–8. These classification results are shown in Tables 5 and 6.
Graphical representation of these classifications is shown in Figs. 9 and 10.
26 Texture Features Extraction in Mammograms Using Non-Shannon Entropies 345

Table 1 Entropy measure of mammograms based on gray level histogram moments at a ¼ 0.5 for
Renyi, a ¼ 0.7 for Havrda & Charvat and for Kapur a ¼ 0.5, b ¼ 0.7
Image samples S R HC K a, b
a ¼ 0.5 a ¼ 0.7 a ¼ 0.5
b¼0.7
Mam1 4.7658 6.9242 8.7940 7.9945
Mam2 5.1761 6.9060 9.0540 7.7959
Mam3 5.2400 7.0073 9.3439 7.8821
Mam4 5.1644 6.7166 8.6331 7.5718
Mam5 4.8436 6.7857 8.5198 7.8131
Mam6 5.1044 6.7977 8.7193 7.7227
Mam7 5.5805 7.0348 9.7094 7.7459
Mam8 4.2316 6.6505 7.8027 7.9252
Mam9 2.9738 6.1737 6.0264 7.9867
Mam10 4.1991 6.4753 7.3892 7.7603
Mam11 4.5951 6.7681 8.2881 7.9115
Mam12 4.9022 6.8990 8.8035 7.9257
Mam13 4.9559 6.8047 8.6226 7.7983
Mam14 4.4958 6.7285 8.1625 7.8909
Mam15 4.9878 6.9086 8.9161 7.8832

Table 2 Entropy measure of mammograms based on gray level histogram moments at a ¼ 0.1 for
Renyi, a ¼ 0.5 for Havrda & Charvat and for Kapur a ¼ 0.1, b ¼ 0.9
Image samples S R HC K a, b
a ¼ 0.1 a ¼ 0.5 a ¼ 0.1
b ¼ 0.9
Mam1 4.7658 7.7796 20.0410 8.0925
Mam2 5.1761 7.7505 19.9023 8.0219
Mam3 5.2400 7.7346 20.6845 7.9934
Mam4 5.1644 7.6608 18.5109 7.9306
Mam5 4.8436 7.7167 19.0077 8.0217
Mam6 5.1044 7.7035 19.0954 7.9825
Mam7 5.5805 7.6587 20.9020 7.8736
Mam8 4.2316 7.7460 18.0460 8.1189
Mam9 2.9738 7.6251 14.9927 8.1298
Mam10 4.1991 7.7053 16.8654 8.0857
Mam11 4.5951 7.7552 18.800 8.0900
Mam12 4.9022 7.7795 19.8491 8.0820
Mam13 4.9559 7.7529 19.1466 8.0516
Mam14 4.4958 7.7473 18.5954 8.0910
Mam15 4.9878 7.7747 19.9222 8.0668

Table 3 Classification mammograms based on Shannon entropy, Renyi’s entropy at a ¼ 0.5,

Havrda & Charvat’s entropy at a ¼ 0.7 and Kapur’s entropy at a ¼ 0.5, b ¼ 0.7
Mammogram Average S Average R Average HC Average K a, b
classification a ¼ 0.5 a ¼ 0.7 a ¼ 0.5
b ¼0.7
Normal 4.2315 6.4321 7.6275 7.6789
Benign 4.4236 6.7543 8.235 7.7678
Malignant 4.9875 7.0024 8.9678 7.8123
346 A.P. Singh and B. Singh

Table 4 Classification mammograms based on Shannon entropy, Renyi’s entropy at a ¼ 0.1,

Havrda & Charvat’s entropy at a ¼ 0.5 and Kapur’s entropy at a ¼ 0.1, b ¼ 0.9
Mammogram Average S Average R Average HC Average K a, b
classification a ¼ 0.1 a ¼ 0.5 a ¼ 0.1
b ¼0.9
Normal 4.2315 7.6975 18.2013 7.8020
Benign 4.4236 7.7204 19.5750 7.9652
Malignant 4.9875 7.7824 20.2542 8.0125

× 104

JM
6

0 50 100 150 200 250

Fig. 1 Normal Mammogram:-mdb006 and its Histogram

× 104
7

6
51
JM
5

0 50 100 150 200 250

Fig. 2 Benign Mammogram:-mdb312 and its Histogram

26 Texture Features Extraction in Mammograms Using Non-Shannon Entropies 347

× 104
6

0 50 100 150 200 250

Fig. 3 Malignant Mammogram:-mdb028 and its Histogram

10

8
Normal
6 Benign

4 Malignant

0
Shannon Renyi Havrda & Kapur
Charvat

Fig. 4 Graphical representation of classification of mammogram images based on Shannon

entropy, Renyi’s entropy at a ¼0.5, Havrda & Char vat’s entropy at a ¼0.7 and Kapur’s entropy
at a ¼0.5, b ¼0.7

25

20

15 Normal

Benign
10
Malignant
5

0
Shannon Renyi Havrda & Kapur
Charvat

Fig. 5 Graphical representation of classification of mammograms based on Shannon entropy,

Renyi’s entropy at a ¼ 0.1, Havrda & Charvat’s entropy at a ¼ 0.5 and Kapur’s entropy at a ¼0.1,
b ¼0.9
348 A.P. Singh and B. Singh

× 104
7

6
JM

0
5

0 50 100 150 200 250

Fig. 6 Uncompressed Fatty image:-mdb008 and its Histogram

× 104

6
JM

0 50 100 150 200 250

Fig. 7 Fatty image:-mdb005 and its Histogram

From the above results, it can be inferred that non-Shannon’s entropies (Renyi,
Havrda and Charvat, and Kapur) are useful parameters for the classification of
normal and abnormal mammogram images. Specially, Havrda and Charvat is most
suitable for this purpose. Experimental results indicate that, in non-Shannon entro-
pies, the values of constants a, b play important role in classification. So, selection
of suitable values of these constants is necessary. Havrda and Charvat entropy at
a ¼ 0.5 gives the best suitable results for classification. This in turn also helps in the
diagnosis of palpable or microcalcifications or tumors in the breast. This work may
be employed to develop a Computer Aided Decision (CAD) System for early
detection of breast cancer.
26 Texture Features Extraction in Mammograms Using Non-Shannon Entropies 349

× 104

7
JM
6
0.1

0 50 100 150 200 250

Fig. 8 Non-uniform Fatty image:-mdb003 and its Histogram

Table 5 Classification of uncompressed fatty, fatty and non-uniform fatty breasts based on
Shannon entropy, Renyi’s entropy at a ¼ 0.5, Havrda & Charvat’s entropy at a ¼ 0.7 and Kapur’s
entropy at a ¼ 0.5, b ¼ 0.7
Mammogram Average S Average R Average HC Average K a, b
classification a ¼ 0.5 a ¼ 0.7 a ¼ 0.5
b ¼0.7
Uncompressed fatty 4.9878 6.9086 8.9161 7.8832
Fatty breast 5.1761 6.9060 9.2540 7.7959
Non-uniform fatty breast 4.5951 6.7681 8.2881 7.9115

Table 6 Classification of uncompressed fatty, fatty and non-uniform fatty breasts based on
Shannon entropy, Renyi’s entropy at a ¼ 0.1, Havrda & Charvat’s entropy at a ¼ 0.5 and Kapur’s
entropy at a ¼ 0.1, b ¼ 0.9
Mammogram classification Average S Average R Average HC Average K a, b
a ¼0.1 a ¼0.5 a ¼0.1
b¼0.9
Uncompressed fatty 4.9878 7.7747 19.9222 8.0668
Fatty breast 5.1761 7.7505 20.2023 8.0219
Non-uniform fatty breast 4.5951 7.7552 18.8800 8.0900

4 Conclusions and Future

In this paper, an attempt is made to develop a new features descriptor based on non-
Shannon’s entropies (Renyi, Havrda and Charvat, and Kapur) for classifying
normal and abnormal mammogram images. Experiment results have demonstrated
that Havrda and Charvat entropy based feature works satisfactorily for classifying
normal, benign and malignant digital mammograms including uncompressed-fatty,
fatty and non-uniform fatty breast mammograms.
350 A.P. Singh and B. Singh

10
8 Uncompressed
fatty
6
4
Fatty breast
2
0
Non-uniform fatty
n

yi

at

r
pu
no

en

rv
breast

Ka
an

ha
R
Sh

C
&
a
rd
av
H

Fig. 9 Graphical representation of classification of uncompressed fatty, fatty breast and non-
uniform fatty breast based on Shannon entropy, Renyi’s entropy at a ¼0.5, Havrda & Charvat’s
entropy at a ¼0.7 and Kapur’s entropy at a ¼0.5, b ¼0.7

25
20
15 Uncompressed
fatty
10
5 Fatty breast
0
Non-uniform fatty
n

yi

at

r
pu
no

en

rv

breast
Ka
an

ha
R
Sh

C
&
a
rd
av
H

Fig. 10 Graphical representation of classification of uncompressed fatty, fatty and non-uniform

fatty breasts based on Shannon entropy, Renyi’s entropy at a ¼ 0.1, Havrda & Charvat’s entropy at
a ¼ 0.5 and Kapur’s entropy at a ¼0.1, b ¼0.9

In future work, Havrda and Charvat entropy based feature will be included with
other existing statistical features to build a robust features descriptor for developing
a complete Computer Aided Diagnosis (CAD) for early detection of breast cancer.
The work is under further progress in this direction.

References

1. H.S. Sheshadri, A. Kandaswamy, Experimental investigation on breast tissue classification

based on statistical feature extraction of mammogram, computerized. Med. Imaging Graph.
31, 46–48 (2007)
2. I. Christoyianni, A. Koutras, E. Dermatas, G. Kokkinakis, Computer aided diagnosis of breast
cancer in digitized mammograms. Comput. Med. Imaging Graph. 26, 309–319 (2002)
26 Texture Features Extraction in Mammograms Using Non-Shannon Entropies 351

3. B. Verma, P. Zhang, A novel neural-genetic algorithm to find the most significant combina-
tion of features in digital mammograms. Appl. Soft Comput. l(7), 612–625 (2007)
4. S.-K. Lee, C.-S. Lo, C.-M. Wang, P.-C. Chung, C.-I. Chang, C.-W. Yang, P.-C. Hsu, A
computer-aided design mammography screening system for detection and classification of
microcalcifications. Int. J. Med. Inf. 60, 29–57 (2000)
5. H.D. Cheng, X. Cai, X. Chen, L. Hu, X. Lou, Computer-aided detection and classification of
microcalcifications in mammograms: a survey. Pattern Recogn. 36, 2967–2991 (2003)
6. H.D. Cheng, M. Cui, Mass lesion detection with a fuzzy neural network. Pattern Recogn. 37,
1189–1200 (2004)
7. H.D. Cheng, X.J. Shi, R. Min, L.M. Hu, X. P. Cai, H.N. Du, Approaches for automated
detection and classification of masses in mammograms. Pattern Recogn. 39, 646–668 (2006)
8. M.E. Mavroforakis, H.V. Georgiou, N. Dimitropoulos, D. Cavouras, S. Theodoridis, Mam-
mographic masses characterization based on localized texture and dataset fractal analysis
using linear, neural and support vector machine classifiers. Artif. Int. Med. 37, 145–162
(2006)
9. B. Verma, J. Zakos, A computer-aided diagnosis system for digital mammograms based on
fuzzy-neural and feature extraction techniques. IEEE Trans. Inf. Technol. Biomed. 5(1),
(Mar 2001)
10. J.K. Kim, H.W. Park, Statistical textural features for detection of microcalcifications in
digitized mammograms. IEEE Trans. Med. Imaging 18(3), (Mar 1999)
11. J.K. Kim, J. Mi Park, K.S. Song, H.W. Park, Texture analysis and artificial neural network for
detection of clustered microcalcifications on mammograms, in Neural Networks for Signal
Processing [1997] VII, Proceedings of the 1997 IEEE Workshop, Amelia Island, Florida,
USA, pp. 199–206
12. A Karahaliou, S Skiadopoulos, I Boniatis, P Sakellaropoulos, E Likaki, G Panayiotakis, L
Costaridou, Texture analysis microcalcifications on mammograms for breast cancer diagnosis.
Brit. J. Radiol. 80, 648–656 (2007)
13. Y.-Y. Wan, J.-X. Du, D.-S. Huang, Z. Chi, Y.-M. Cheung, X.-F. Wang, G.-J. Zhang,
Bark texture feature extraction based on statistical texture analysis, in Proceedings of 2004
International Symposium on Intelligent Multimedia, Video and Speech Processing, Hong
Kong, Oct 20–22, 2004
14. M.R. Chandraratne, Comparison of three statistical texture measures for lamb grading, in 1st
International Conference on Industrial and Information Systems,ICIIS2006 Sri Lanka, 8–11
Aug 2006
15. J.N. Kapur, Measure of Information and Their Applications (Wiley Eastern Limited,
New Delhi, 1994)
16. L.I. Yan, F. Xiaoping, L. Gang, An application of Tsallis entropy minimum difference on
image segmentation, in Proceeding of the 6th World Congress on Intelligent Control and
Automation, Dalian, China, 21–23 June 2006
17. M. Portes de Albuquerque, I.A. Esquef, A.R. Gesualdi Mello, Image thresholding using
Tsallis entropy. Pattern Recogn. Lett. 25(2004), 1059–1065 (2004)
18. R. Smolikova, M.P. Wachowiak, G.D. Tourassi, A. Elmaghraby, J.M. Zurada, Characteriza-
tion of ultrasonic back scatter based on generalized entropy, in Proceeding of the 2nd Joint
EMBS/BMES Conference, Houston, TX, USA, 23–26 Oct 2002
Chapter 27
A Wideband DOA Estimation Method Based
on Arbitrary Group Delay

Xinggan Zhang, Yechao Bai, and Wei Zhang

Abstract Direction-of-arrival (DOA) estimation is an important algorithm in array

processing. Most traditional DOA estimation methods focus on narrow-band
sources. A new wideband DOA estimation approach based on arbitrary group
delay is proposed in this paper. In the proposed algorithm, echo gain at each
direction is calculated based on digital group delay compensation. Digital group
delay can realize time delay precisely, and has no restriction on delay step. The
proposed method is able to operate arbitrary waveform and is suitable for linear
frequency modulation signals, nonlinear frequency modulation signals, and even
carrier free signals. Simulations with Gaussian function are carried out to show the
effectiveness of the proposed algorithm.

1 Introduction

DOA estimation is an important algorithm in array signal processing systems [1, 2].
Various estimation methods such as Maximum Likelihood (ML), Multiple Signal
Classification (MUSIC) [3], and Estimation of Signal Parameters via Rotational
Invariance Technique (ESPRIT) [4] were proposed in recent two decades. How-
ever, most of these methods are based on narrow band models. Several extensions
to the wideband sources are also proposed [5], for example, the Coherent Signal-
Subspace Processing (CSSP) [6], Wideband Cyclic MUSIC [7] and Wideband
ESPRIT [8]. A frequency domain algorithm based on ML criterion called AML
method has been proposed for source localization and DOA estimation [9, 10]. A
method proposed by Ward [11] performs beam-space processing using time-
domain frequency-invariant beamformers (TDFIB). In this paper, we propose a

X. Zhang (*)
School of Electronic Science and Engineering, Nanjing University, Nanjing, P. R. China
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 353

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_27, # Springer ScienceþBusiness Media B.V. 2010
354 X. Zhang et al.

new wideband DOA estimation approach based on arbitrary group delay. This
method is able to operate arbitrary waveform and is suitable for linear frequency
modulation signals, nonlinear frequency modulation signals, and even carrier free
signals.

2 Method of Digital Group Delay

When the object direction is far from the normal direction, the delay between array
elements is so large that signals need to be aligned to get a better distance resolution
and a higher power gain. The time delay compensation is usually operated in two
steps: the integral sampling period part and fractional sampling period part. The
integral part can be simply carried out by register shifting. The subsample time
delay is compensated by interpolation commonly, and the compensation effect lies
on the interpolation function. Dense sampling, in which sampling frequency is
much higher than the Nyquist frequency is also used in some situation to achieve a
higher accuracy. Unfortunately, this method brings too much data redundancy and
raises a higher demand to the sampling device. Especially, when the bandwidth of
ultra-wideband signals is very high, for example 1 GHz, it is hard to sample with
frequency several times higher, and the expense is also too huge.
Suppose that the wideband signal is s(t), and the delayed signal is s(tt).
According to the time shifting property of continuous time Fourier transform, we
can get:


sðt  tÞ ¼ F1 FfsðtÞgejot (1)

If the sampling frequency is so high that the spectral aliasing can be ignored,
the digital signal delay can be implemented by
n o
sðnTs  tÞ ¼ IDFT DFT fsðnTs Þgej NTs
2pkt
(2)

where Ts is the sampling period, and s(nTs) is the digital signal. The time delay t in
Eq. (2) need not to be integral multiple of the sampling period. N is the length of the
operated signal, consisting of original signal and INT(t/fs) padded zeroes, where
INT denotes rounding operation. Zeros are padded at the tail of the original signal
to avoid cyclic shifting of the original signal.
Gaussian modulated sine signal is used to demonstrate the method of digital
group delay. The analytical expression of the signal is

tT0 2
pðTg Þ
sðtÞ ¼ e cos½2pf ðt  T0 Þ (3)

where Tg, T0 and f are the time width parameter of Gaussian window, the time
center and the frequency of sine signal, respectively. Tg and f were chosen to be 50
27 A Wideband DOA Estimation Method Based on Arbitrary Group Delay 355

and 0.1. T0 was chosen to be 0 before signal is delayed. The results of simulation of
digital group delay with MATLAB are shown in Figs. 1 and 2. In simulations, the
signal delayed for 3.7 samples by the method of digital group delay is compared
with the theoretical waveform obtained by setting T0 to 3.7 in Eq. (3).

1
Waveform of
0.8 Digital Group
Delay
0.6
Theoretical
0.4 Waveform

0.2
Amplitude

– 0.2

– 0.4

– 0.6

– 0.8

–1
–50 0 50 100
Time / Samples

Fig. 1 Comparison between waveform of digital group delay and theoretical waveform

× 10–6
1

0.8

0.6

0.4

0.2
error

– 0.2

– 0.4

– 0.6

– 0.8

–1
– 50 0 50 100
Time / Samples

Fig. 2 Error between waveform of digital group delay and theoretical waveform
356 X. Zhang et al.

It can be seen that digital group delay can realize time delay precisely for
approximate band limit signals from Figs. 1 and 2. In addition, digital group delay
operates the integral sampling period part and fractional sampling period part
together, and has no restriction on delay step. Furthermore, this method has no
requirement for the signal waveform, and is suitable for linear frequency modulation
signals, nonlinear frequency modulation signals, and even carrier free signals.

3 DOA Estimation Based on Digital Group Delay

A simple time delay manifests itself as the slope of an additive linear phase in the
phase spectrum, which is an additional group delay

1 d’
tg ð f Þ ¼  ¼ t (4)
2p df

where tg is a constant.
Consider an equally spaced linear array composed of Ne identical elements.
The distance between consecutive sensors is d. The relative delay of the ith
element is

id sin a
ti ¼ (5)
c

where a is the beam direction, and c is the velocity of electromagnetic wave.

The target echoes of array elements are aligned according to each direction. At
the direction a, signals are aligned completely after group delay compensation, and
the gain after match filter reaches the maximum. In contrast, at other directions,
echo signals can not be totally aligned, and the gain decreases.
DOA estimation based on arbitrary group delay compensates the echo signals
according to different beam directions, and finds the target direction by locating
the maximum of the match filter gain. Considering that the searching directions
are discrete and the accurate is limited, we fit a convex parabola around the
peak and obtain the required direction by calculating the symmetry axis of the
parabola.

4 Simulation

Gaussian function [12] (GGP, generalized Gaussian pulse) was employed in the
simulation (using MATLAB) to demonstrate the effectiveness of the proposed
algorithm. One of forms based on experimentation with UWB impulse radiators
and sensors is:
27 A Wideband DOA Estimation Method Based on Arbitrary Group Delay 357

Table 1 The RMSE in the SNR (dB) 10 0 10 20 30

DOA estimation versus SNR RMSE ( ) 1.4457 0.4584 0.1627 0.0579 0.0187

–5

– 10
Gain / dB

– 15

– 20

– 25

– 30
– 100 – 80 – 60 – 40 – 20 0 20 40 60 80 100
Direction / °

Fig. 3 DOA estimation for two targets at 15 and 30 , SNR ¼ 20 dB

E0 n 4pðttDT0 Þ2 o
 ae4pa ð DT Þ
2 tt0 2
f ðtÞ ¼ e (6)
1a

where E0 is peak amplitude at time t ¼ t0, a is scale parameter. In the simulation,

E0 ¼ 1, DT ¼ 2 ms and a ¼ 10. In addition, element amount Ne, distance between
consecutive sensors d and the sampling frequency were chosen to be 100, 0.4 m and
100 MHz, respectively.
The root mean square error (RMSE) in the DOA estimation versus signal to
noise ratio (SNR) is shown in Table 1. The target is fixed at direction 45.5 . The
simulation for each SNR was performed 1,000 times.
From Table 1, it is easily seen that the proposed algorithm has a very small
RMSE, and the proposed algorithm also performs well at low SNR.
Figure 3 shows the resolution performance of the proposed algorithm. In the simula-
tion, E0, DT, a and the sampling frequency were chosen to be 1, 2 ms, 200 and 1 GHz,
respectively. The echo contains two targets located at 15 and 30 , and SNR is 20 dB.

5 Conclusion

The digital group delay can realize time delay precisely for approximate band limit
signals. In this paper, we apply the digital group delay method to DOA estimation.
The simulations show the effectiveness of the proposed algorithm. The proposed
358 X. Zhang et al.

algorithm has no requirement for the signal waveform, and is suitable for wideband
and ultra-wideband radars, sonar systems, and microphone arrays.

References

1. M.R. Azimi-Sadjadi, A. Pezeshki, N. Roseveare, Wideband DOA estimation algorithms for

multiple moving sources using unattended acoustic sensors. IEEE Trans. Aerosp. Electron.
Syst. 44, 1585–1599 (Oct 2008)
2. Y.-S. Yoon, L.M. Kaplan, J.H. McClellan, TOPS: new DOA estimator for wideband signals.
IEEE Trans. Signal Process. 54, 1977–1989 (June 2006)
3. R. Schmidt, Multiple emitter location and signal parameter estimation. IEEE Trans. Antennas
Propag. 34, 276–280 (Mar 1986)
4. R. Roy, A. Paulraj, T. Kailath, Direction-of-arrival estimation by subspace rotation methods -
ESPRIT, IEEE International Conference on ICASSP, Apr 1986, pp. 2495–2498
5. L. Yip, C.E. Chen, R.E. Hudson, K. Yao, DOA estimation method for wideband color signals
based on least-squares Joint Approximate Diagonalization, IEEE Sensor Array and Multi-
channel Signal Processing Workshop, July 2008, pp. 104–107
6. H. Wang, M. Kaveh, Coherent signal-subspace processing for the detection and estimation of
angles of arrival of multiple wide-band sources. IEEE Trans. Acoust. Speech Signal Process.
33, 823–831 (Aug 1985)
7. K.D. Mauck, Wideband cyclic MUSIC, ICASSP 4, 288–291 (Apr 1993)
8. B. Ottersten, T. Kailath, Direction-of-Arrival estimation for wideband signals using the
ESPRIT algorithm, IEEE Trans. Acoust. Speech Signal Process. 38, 317–327 (Feb 1990)
9. J.C. Chen, R.E. Hudson, K. Yao, Maximum-likelihood source localization and unknown
sensor location estimation for wideband signals in the near Field, IEEE. Trans. Signal Process.
50, 1843–1854 (Aug. 2002)
10. J.C. Chen, L. Yip, J. Elson, H. Wang, D. Maniezzo, R.E. Hudson, K. Yao, D. Estrin, Coherent
acoustic array processing and localization on wireless sensor networks, Proc. IEEE 91,
1154–1162 (Aug. 2003)
11. D.B. Ward, Z. Ding, R.A. Kennedy, Broadband DOA estimation using frequency invariant
beamforming, IEEE Trans. Signal Process. 46, 1463–1469 (May 1998)
12. W. Min, W. Shunjun A time domain beamforming method of UWB pulse array, IEEE Int.
Radar Conf. 697–702 (May 2005)
Chapter 28
Spatial Speaker Spatial Positioning of Synthesized
Speech in Java

Jaka Sodnik and Sašo Tomažič

Abstract In this paper, we propose a “Spatial speaker”, an enhancement of a Java

FreeTTS speech synthesizer with the additional function of spatial positioning of
both the speaker and the listener. Our module enables the reading of an arbitrary
text from a file or webpage to the user from a fixed or changing position in space
through normal stereo headphones. Our solution combines the following modules:
FreeTTS speech synthesizer, a custom made speech processing unit, MIT Media
Lab HRTF library, JOAL positioning library and Creative X-Fi sound card. The
paper gives an overview of the design of the “Spatial Speaker” and proposes three
different practical applications of such a system for visually impaired and blind
computer users. Some preliminary results of user studies confirmed the system’s
usability and showed its great potential also in other types of applications and
auditory interfaces. The entire system is developed as a single Java class which can
be imported and used in any Java application.

1 Introduction

The most important qualities of modern speech synthesis systems are their natural-
ness and intelligibility [1]. By naturalness we mean how closely the synthesized
speech resembles real human speech. Intelligibility, on the other hand, describes the
ease with which the speech is understood. The maximization of these two criteria is
the main research and development goal in the TTS field.
At the moment, numerous examples of TTS software can be found on the
market, for example stand-alone speech programs that convert all types of
text inserted through an input interface into high quality realistic speech output

J. Sodnik (*)
University of Ljubljana, Faculty of Electrical Engineering, Tržaška 25, 1000 Ljubljana, Slovenia,
Europe
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 359

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_28, # Springer ScienceþBusiness Media B.V. 2010
360 J. Sodnik and S. Tomažič

(TextAloud [2], VoiceMX STUDIO [3], etc.). Many of them can save the readings
as a media file that can later be played on demand. On the other hand, there are also
several speech synthesis libraries that can be included and used in various program
languages, for example FreeTTS, a TTS library written in Java programming
language [4].
In this article, we propose an extension to the FreeTTS software that enables
text-to-speech conversion in a 3D audio environment. Our solution enables the
positioning of the voice (i.e. the text reader) at any arbitrary spatial position in
relation to the listener. The relative spatial positions of the speaker and the listener
can be updated dynamically in real time while performing text-to-speech conver-
sion. For example, the text can be read from a specific spatial position in relation to
the listener or the position can be dynamically changed and updated.
We believe there are numerous potential applications of such technology, from
screen readers for visually impaired computer users to advanced auditory interfaces
for desktop and mobile devices. In the next chapter, we summarize some related
work which uses both spatial audio and TTS.

2 Related Work

In the past, spatial auditory interfaces were mostly used to enrich audio-only
applications for visually impaired computer users.
Mynatt proposed a general methodology for transforming the graphical inter-
faces into non-visual interfaces [5] by using 3D audio output techniques. Various
salient components of graphical interfaces were transformed into auditory com-
ponents with the so-called auditory icons. The final goal was to present object
hierarchies with sound.
The transformation of MS-Windows interface into spatial audio was proposed
by Crispien [6] and Sodnik [7], both transforming the hierarchical navigation
scheme into a “ring” metaphor. Sodnik used speech output for explaining various
commands to the user.
Nowadays, one of the most commonly used computer applications is a web
browser. Web browsers are used to represent extensive and complex web content
comprising of text, pictures, movies, animations, etc. Visually impaired computer
users use web browsers together with screen readers which process pages sequen-
tially and read the content. Several researchers proposed different and more
advanced auditory versions of web browsers.
The HearSay solution, for example, enabled flexible navigation and form-filling
and also context-directed browsing through extensible VoiceXML dialog interface
[8]. The Csurf application was intended to help with the identification of relevant
information on web pages [9]. When the user followed a web link, CSurf captured
the content and preprocessed it in order to identify the relevant information based
on statistical machine-learning models. The content of the web page was then read
to the user from the most to the least relevant section.
28 Spatial Speaker 361

Roth et al. proposed AB-Web, an audio-haptic internet browser for visually

impaired users [10]. AB-Web created an immersive 3D sound environment, into
which HTML objects were mapped. Web elements such as text, link or image were
represented by corresponding auditory elements whose simulated 3D locations
depend on their original locations displayed on the web browser. According to
our understanding, only non-speech items were positioned in space, while TTS
mechanisms were used without spatial positioning. The audio rendering was based
on Intel library RSX using HRTF technology.
Goose and Moller proposed a 3D audio-only interactive web browser with a new
conceptual model of the HTML document and its mapping into 3D audio space
[11]. The audio rendering techniques included the use of various non-speech
auditory elements as well as a TTS synthesizer with multiple voices, intonation,
etc. The rendering was performed by regular PC audio hardware and commercially
available 3D audio software toolkit (MS DirectSound, Intel RSX and Aureal). The
authors’ initial idea was to simulate the act of reading a physical page from top to
bottom. Due to very poor localization of up-down direction (elevation) with the
audio equipment selected, the authors decided to project the moving source to the
left–right direction (azimuth). The HTML document was therefore read sequen-
tially from left to right. Different synthesized voices were used to indicate different
portions of the document: header, content and links.
Synthesized speech seems extremely suitable also for various interfaces of
mobile devices, as these are often used in situations when the visual interaction is
limited or even impossible: when walking down the street, riding a bicycle, driving
a car, etc.
An example of an auditory interface for a communication device in a vehicle
was proposed and evaluated by Sodnik et al. [12]. The interface was based
on speech commands positioned in space and manipulated through a custom
interaction device. The audio rendering was preformed locally on the mobile device
itself by commercially available hardware and software (Creative X-Fi and
OpenAL).
A server-based 3D audio rendering for mobile devices was proposed by Goose
et al. [13]. In this case, a server received a request and dynamically created multiple
simultaneous audio objects. It spatialized the audio objects in 3D space, multi-
plexed them into a single stereo audio and sent over an audio stream to a mobile
device. A high quality 3D audio could therefore be played and experienced on
various mobile devices. The same authors also proposed a system called Confer-
encing3, which offers virtual conferencing based on 3D sound on commercially
available wireless devices [14].
One of the earliest solutions in this field, which is closely related to our work,
was proposed by Crispien et al. [15]. It also combined a screen reader and spatial
sound. It was based on a professional audio processing system called Beachtron,
which was used for spatial positioning of synthesized speech. The system enabled
the user to identify the position of spoken text parts in relation to their visual
representation on the screen.
362 J. Sodnik and S. Tomažič

2.1 Our Research Contribution

The aim of our research was the development of a spatial speaker (i.e. synthesized
voice) that can be positioned in arbitrary positions in space and manipulated in real
time. The solution can be used in any of the abovementioned applications. Since
the majority of auditory interfaces is designed to be used by a large group of users
(e.g. visually impaired computer users or users of various mobile devices), gener-
ally available hardware and software should preferably be used for audio rendering.
Our solution is based on Java platform, standard Creative X-Fi soundcard [16] and
some additional components which are also available off the shelf or even for free.
The major drawback of all general spatial audio rendering equipment is its poor
localization accuracy due to the use of simplified HRTF libraries. In our experi-
ment, we tried to improve the localization accuracy with some additional prepro-
cessing of the sound signals.
Our work is different from previous audio rendering techniques in a number of
ways:
l It combines software and hardware spatial audio rendering techniques.
l It improves the accuracy of elevation localization by adding additional HRTF
processing.
l Due to its simplicity, the system can be used in various desktop or mobile
platforms (hardware processing is optional and can be ignored if 3D sound
hardware is not available).
l We propose some innovative ideas for the practical use of such a system and we
already give some results of preliminary evaluation studies and user feedback.
In the following chapter, we describe the system design. All major components
of the system are listed here along with the most important design problems and
solutions. We also propose some future work and additional improvements in this
field.

3 System Design and Architecture

The system architecture consists of five major modules: four pieces of software
(libraries and Java classes) and a standard sound card with hardware support for 3D
sound. The software part is based on Java programming language with some
additional external plug-ins. The five modules are:
l FreeTTS: speech synthesis system written entirely in Java
l JOAL: implementation of the Java bindings for OpenAL API [17, 18]
l HRTF library from MIT Media Lab (measurements of KEMAR dummy head)
[19]
l A custom made signal processing module
l Creative Sound Blaster X-Fi ExtremeGamer sound card
28 Spatial Speaker 363

Fig. 1 System design

Figure 1 shows the basic concept of the system.

The FreeTTS module reads an arbitrary text from the file and synthesizes
the samples of the spoken sound signal. The samples are then reformatted and
convoluted with the samples of external HRTF function in order to add the eleva-
tion information to the signal. The correct HRTF function has to be determined
according to the relative positions of the sound source and the listener. The result
of convolution is then reformatted again to fit the JOAL input buffer. JOAL is a
positioning library which determines the positions of the source and the listener and
adds the information on the azimuth of the source. All positions are defined in the
Cartesian coordinate system. The output of JOAL library is then sent directly to
the sound card with hardware support for 3D sound. The final positioning is done by
the hardware itself and the sound signal is then played through normal stereo
headphones.
In the following subchapters, we describe the individual modules and present the
development necessary for the entire system to function.

3.1 FreeTTS

FreeTTS is a speech synthesis program written in Java. It is based on Flite, a speech

synthesis engine developed at Carnegie Mellon University. Flite is further based
on two additional systems: Festival (University of Edinburgh) and FestVox (also
Carnegie Mellon University). FreeTTS comes with extensive API documentation.
The system includes three basic voices. Additional voices can be imported from
FestVox or MBROLA systems. In our system, we use kevin16, a 16 kHz diphone
male US voice, and three additional MBROLA US voices (two male and one
female). All voices enable the setting of different pitch or speaking rate and
therefore changing the voice characteristics.
With FreeTTS included in any Java application, one has to define the instance
of the class voice. The selected voice is determined through the class called
voiceManager. The speech synthesis can then be done simply by calling a method
speak (a method of voice class) with the text to be spoken as parameter. We wrote a
wrapper function which accepts additional parameters on pitch and speaking rate.
By default, Java uses its default audio player for playback, and the player then sends
the synthesized sound directly to the default sound card. In our case, we wanted the
364 J. Sodnik and S. Tomažič

output to be a buffer of sound samples for further processing. We had to develop a

custom audio player by implementing Java class AudioPlayer. Our audio player
named BufferPlayer outputs synthesized speech as an array of bytes. The output is
an array of 16-bit samples with little endian byte order. Each sound sample is
presented as signed integer with two’s complement binary form (amplitude value
between 215 and 215).

3.2 MIT Media Lab HRTF Library

HRTFs (Head Related Transfer Functions) are transfer functions of head-related

impulse responses (HRIRs) that describe how a sound wave is filtered by diffraction
and reflection of torso, head and pinna, when it travels from the sound source to the
human eardrum [20]. These impulse responses are usually measured by inserting
microphones into human ears or by using dummy heads. The measurements for
different spatial positions are gathered in various databases or libraries and can be
used as Finite Impulse Response (FIR) filters for creating and playing spatial
sounds through the headphones. A separate function has to be used for each
individual ear and each spatial position.
MIT Media Lab library, which is used in our system, contains HRIR measure-
ments for 710 spatial positions: azimuths from 180 to 180 and elevations from
40 to 90 . The MIT library is available online in various formats and sizes and it
contains individualized as well as generalized functions (measured with a dummy
head) [19]. We use the library in order to improve the elevation positioning of the
sound sources (Fig. 2).

Fig. 2 The elevation area of

the filters used from HRIR
library
28 Spatial Speaker 365

The elevation localization depends strongly on individualized human factors:

torso, shoulder and pinna shape [21]. Ideally, individualized HRIR should be used
for each user, but this is virtually impossible when the system is intended to be used
by a large number of users. We therefore use the generalized compact measure-
ments in wav format. Only 14 functions are used in our application: elevations from
40 to 90 at azimuth 0 . At azimuth 0 , the same function can be used for both
ears. The azimuth positioning is done by JOAL library.

3.3 Signal Processing Module

The MIT HRTF library contains functions in PCM format consisting of 16-bit
samples at 44.1 kHz sampling frequency. In order to be used with the output
of FreeTTS, the samples had to be down-sampled to 16 kHz. In Java, the HRIRs
are defined as arrays of bytes.
The filtering of synthesized samples inp with HRIRel is done by calculating the
convolution between the two arrays.

X
N 1
resel ½n ¼ inp½n  mHRIRel ½m (1)
m¼1

In order to calculate the correct convolution, both arrays have to be converted to

arrays of float numbers. After convolution, the resulting array resel is converted to a
16-bit array of unsigned samples with big endian byte order, appropriate for input to
JOAL library functions.

3.4 JOAL Library

JOAL is a reference implementation of the Java bindings for OpenAL API.

OpenAL is a cross-platform 3D audio API appropriate for use in various audio
applications. The library models a collection of audio sources moving in 3D space
that are heard by a single listener located somewhere in that space. The positioning
of the listener or the sources is done by simply defining their coordinates in the
Cartesian coordinate system, for example: Source1(x1,y1,z1), Source2(x2,y2,z2) and
Listener(x0,y0,z0).
The actual processing of input sounds is done by software (the library itself) or
hardware (if provided within the soundcard). At the beginning of the experiment,
we expected we would be able to use JOAL for all spatial positioning (i.e. azimuth
and elevation), but the elevation perception proved to be too poor to be used in our
application (similar findings were reported also by Goose et al. [9]). The addition of
366 J. Sodnik and S. Tomažič

the afore-mentioned MIT Media Lab HRTF library was an attempt to improve the
elevation coding in the signals.
JOAL provides the functions for reading external wav files which are then
positioned and played. The samples from wav files are then written to special
buffers and the buffers are attached to the sources with specific spatial character-
istics. In our case, the input to JOAL is an array of samples (the output of FreeTTS
convoluted with HRIR), which is then written directly to buffers.
By using only JOAL library for spatial positioning, the spatial arrangement of
the sources and the listener can be changed at any time. However, in our case only
the changes of the horizontal position can be preformed dynamically through
JOAL. The vertical changes, on the other hand, require preprocessing with new
HRIR.

3.5 Soundcard

Creative Sound Blaster X-Fi Extreme Gamer was used within the system. The
soundcard has a special DSP unit called CMSS-3D, which offers hardware support
for spatial sound generation. CMSS is another type of a generalized HRTF library
used for filtering input sounds. In general, the soundcard can be configured for
output to various speaker configurations (i.e. headphones, desktop speakers, 5.1
surround, etc.), but in our case the use of an additional HRTF library required the
playback through stereo headphones.
Creative Sound Card works well with JOAL (OpenAL) positioning library.
However, if no hardware support can be found for spatial sound, the latter is
performed by the library itself (with a certain degradation of quality).

4 Prototype Applications and Preliminary User Studies

The “Spatial speaker” is meant to be a part of various auditory interfaces and

systems used mostly by visually impaired users or the so called users-on-the-move:
drivers, runners, bikers, etc. In this paper, we describe the prototypes of three
different applications which are meant to be used primarily by visually impaired
computer users:
l Spatial audio representation of a text file
l Spatial story reader
l Multiple simultaneous files reader
We preformed some preliminary user studies with the three applications. Ten
users participated in the experiments. All of them reported normal sight and
hearing. The tests were preformed in a semi-noise environment with Sennheiser
HD 270 headphones. The headphones enable approximately 10–15 dB attenuation
28 Spatial Speaker 367

of ambient noise. Since the primary focus of this paper is the design of the “Spatial
speaker” system, only a short summary of the results and user comments is
provided.

4.1 Spatial Audio Representation of a Text File

A typical screen reader, the most important tool for visually impaired users to
access and read web content or word processing files, reads the content of the file
sequentially from top left to bottom right corner. Some screen readers describe the
positions of the elements on the page with the aid of the coordinate system. Our
prototype application reads the content of the page as if coming from different
spatial positions relative to the listener. The spatial positions are defined according
to the original position of the element on the page. The reader starts reading the
page at the left top corner and then it moves from left to right and follows the
positions of the text on the page. It finishes at the bottom right corner of the page.
For example, a text appearing on the left side of the page can actually be heard on
the left, the text on the right can be heard on the right, etc. In this way, the
application enables the perception of the actual physical structure of the web
page (Fig. 3).
In a short experiment, the users were asked to describe the approximate structure
of the page by just listening to its content. The majority reported that the positions
of the individual texts could be perceived clearly, but the perception of the entire
structure remained blurry, since no information on images, banners and other web
page elements was given.

4.2 Spatial Story Reader

The speech synthesizer can also be used to read various messages and notes to
mobile users or to read fiction to visually impaired people. The normal output of
various speech synthesizers is usually very monotonous and dull. We developed a
story reader which separates different characters in the story by assigning them a
location in space. Each specific character in the story as well as the narrator speaks
from a different spatial position according to the listener. The characters with their
corresponding texts can therefore be easily separated from one another. In addition,
the pitch of synthesized speech is changed for each character, which makes the
separation even better.
The input for the speech synthesizer is a normal txt file. All characters or voices
which are to be manipulated and positioned in space need to be tagged with
appropriate coordinates and information about the pitch in order to be recognized
by the system (Fig. 4).
368 J. Sodnik and S. Tomažič

Fig. 3 The basic concept of auditory representation of a text file

The users listened to five different stories. They were asked to subjectively
evaluate the system without focusing on the content of the stories. The majority
especially liked the additional separation of the characters by pitch. Several users
commented that the accompanying text explaining which of the characters said
something was in some cases almost redundant, since the designated spatial posi-
tions and the corresponding pitch of the speech clearly identified the individual
characters.

4.3 Multiple Simultaneous Files Reader

A human listener is unable to perceive and understand more than one speech source
at a time, unless the sources are coming from different spatial positions [22].
General text-to-speech converters can therefore be used with just one text input
28 Spatial Speaker 369

Fig. 4 Multiple voices at different spatial positions

Fig. 5 Two simultaneous

speech sources

and one spoken output. Our “Spatial speaker” can, however, produce multiple
outputs that are separated by their locations in space. We developed a test applica-
tion that reads an arbitrary number of text files and synthesizes speech flows at
different spatial positions relative to the speaker. We evaluated the intelligibility
with two and three sources. The users were positioned at the origin of the coordinate
system (i.e. 0 ,0 ,0 ) and the speech sources were positioned as follows (Fig. 5):
l Two sources: (30 ,0 ,0 ) and (30 ,0 ,0 )
l Three sources: (30 ,0 ,0 ), (0 ,0 ,0 ) and (30 ,0 ,0 )
The preliminary results show that a maximum of two simultaneous speech
sources can be perceived and understood simultaneously. The users reported that
370 J. Sodnik and S. Tomažič

with three sources it was still possible to concentrate on one individual source and
understand the content, but it was impossible to understand all of the speech sources
simultaneously.

5 Conclusion and Future Work

This paper describes the development and system design of a 3D speech synthe-
sizer. The system is based on Java platform and comprises of standard program-
ming libraries, custom developed modules and a sound card, all of which are
reprogrammed and modified to fit together. An external MIT Media Lab HRTF
library is used in order to improve the localization accuracy of the system. A
personalized HRTF library can also be measured and used for a specific user,
thus enabling an extremely accurate localization performance and spatial awareness
if required. The current system is realized as a single Java class with some
additional libraries and files and can therefore be imported and used in any Java
application. The system works with any soundcard, since hardware support for 3D
sound is optional and is automatically replaced by the software library if not
present.
In addition to its use as a typical speech synthesizer or TTS converter, the
“Spatial speaker” could also be used as a key component of various auditory
interfaces. It can, for example, be used effectively by visually impaired people or
mobile users-on-the-go. Three different prototype applications are provided in this
paper along with some preliminary results of evaluation studies. Only users with
normal sight participated in the preliminary experiments in order to get some
feedback and ideas for improvements. All three prototypes proved to be interesting
and innovative and potentially useful for visually impaired people. Our next step
will be an extensive user study with visually impaired computer users. We intend to
establish some major advantages of a 3D speech synthesizer over the non-spatial
one.

Acknowledgements This work has been supported by the Slovenian Research Agency within the
research project: “Assisting visually impaired people in using the computer with the aid of spatial
sound”.

References

1. R.A. Cole, Survey of the State of the Art in Human Language Technology (1996)
2. TextAloud (5 Feb 2010). http://www.nextuptech.com
3. VoiceMX STUDIO (5 Feb 2010). http://www.tanseon.com/products/voicemx.htm
4. FreeTTS (5 Feb 2010). http://freetts.sourceforge.net/docs/index.php
5. E.D. Mynatt, Transforming graphical interfaces into auditory interfaces. Doctoral Disserta-
tion, Georgia Institute of Technology, 1995
28 Spatial Speaker 371

6. K. Crispien, K. Fellbaum, A 3D-auditory environment for hierarchical navigation in

non-visual interaction, in Proceedings of ICAD96, 1996
7. J. Sodnik, S. Tomažič, Spatial auditory interface for word processing application, in Proceed-
ings of IEEE ACHI09 2009, pp. 271–276
8. Y. Borodin, J. Mahmud, I.V. Ramakrishnan, A. Stent, The HearSay non-visual web browser.
Proc. Int. Cross Discip. Conf. Web Accessibility, 225, 128–129 (2007)
9. J.U. Mahmud, Y. Borodin, I.V. Ramakrishnan, Csurf: a context-driven non-visual
web-browser, in Proceedings of 16th International Conference on World Wide Web, pp.
31–40, 2007
10. P. Roth, L.S. Petrucci, A. Assimacopoulos, T. Pun, Audio-haptic internet browser and
associated tools for blind and visually impaired computer users, in Workshop on Friendly
Exchanging Through the Net, pp. 57–62, 2000
11. S. Goose, C. Moller, A 3D audio only interactive Web browser: using spatialization to convey
hypermedia document structure, in Proc. of Seventh ACM International Conference on
Multimedia, pp. 363–371, 1999
12. J. Sodnik, C. Dicke, S. Tomažič, M. Billinghurst, A user study of auditory versus visual
interfaces for use while driving. Int. J. Human Comput. Stud. 66(5), 318–332 (2008)
13. S. Goose, S. Kodlahalli, W. Pechter, R. Hjelsvold, Streaming speech3: a framework for
generating and streaming 3D text-to-speech and audio presentations to wireless PDAs as
specified using extensions to SMIL, in Proceedings of the International Conference on World
Wide Web, pp. 37–44, 2002
14. S. Goose, J. Riedlinger, S. Kodlahalli, Conferencing3: 3D audio conferencing and archiving
services for handheld wireless devices. Int. J. Wireless Mobile Comput. 1(1), 5–13 (2005)
15. K. Crispien, W. Wurz, G. Weber, Using spatial audio for the enhanced presentation of
synthesised speech within screen-readers for blind computer users, Computers for
Handicapped Persons, vol. 860. (Springer, Berlin, Heidelberg, 1994), pp. 144–153
16. Creative (5 Feb 2010). http://www.soundblaster.com
17. JOAL (5 Feb 2010). https://joal.dev.java.net
18. OpenAL (5 Feb 2010). http://connect.creativelabs.com/openal/default.aspx
19. HRTF Measurements of a KEMAR Dummy-Head Microphone (5 Feb 2010). http://sound.
media.mit.edu/resources/KEMAR.html
20. C.I. Cheng, G.H. Wakefield, Introduction to head-related transfer functions (HRTF’s):
representations of HRTF’s in time, frequency, and space (invited tutorial). J. Audio Eng.
Soc. 49(4), 231–249 (2001)
21. J. Blauert, Spatial Hearing: The Psychophysics of Human Sound Localization, Revised edn.
(MIT Press, Cambridge, MA, 1997)
22. B. Arons, A review of the cocktail party effect, J. Am. Voice I/O Soc. 12, 35–50 (1992)
Chapter 29
Commercial Break Detection and Content Based
Video Retrieval

N. Venkatesh and M. Girish Chandra

Abstract This chapter presents a novel approach for automatic annotation and
content based video retrieval by making use of the features extracted during
the process of detecting commercial boundaries in a recorded Television (TV)
program. In this approach, commercial boundaries are primarily detected using
audio and the detected boundaries are validated and enhanced using splash screen
of a program in the video domain. Detected splash screen of a program at the
commercial boundaries is used for automatic annotation of recorded video which
helps in fast content based video retrieval. The performance and validity of our
approach is demonstrated using the videos recorded from different Indian Television
broadcasts.

1 Introduction

The Television (TV) remains and always has been one of the most powerful
medium of communication since its inception. Today, we are in the next generation
of television where, the additional hardware and software have completely changed
the way the television is watched. The recent technologies such as set-top box
coupled with the reduction of costs for digital media storage have led to the
introduction of personal video recorders (PVR) which are now fairly established
and fast growing part of the TV landscape.
In this context, audiovisual analysis tools that help the user to manage the huge
amount of data are very important to introduce the novel recording devices in the
highly competitive market. Among other analysis tools, detection of TV advertise-
ments is a topic with many practical applications. For instance, from the point of

N. Venkatesh (*)
Innovation Lab, Tata Consultancy Services, Plot no 96, Abhilash building, EPIP Industrial area,
white field main road, Bangalore 560066, India
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 373

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_29, # Springer ScienceþBusiness Media B.V. 2010
374 N. Venkatesh and M.G. Chandra

view of a TV end user, it could be useful to avoid commercials in personal

recordings. Also of great utility is to appropriately index and retrieve the large
number of recorded videos automatically.
Existing commercial detection approaches can be generally divided into two
categories: feature based and matching-based approaches. While the feature-based
approaches use some inherent characteristics of TV commercials to distinguish
commercials and other types of videos, the matching based methods attempt to
identify commercials by searching a database that contains known commercials.
The challenges faced by both approaches are the same: how to accurately detect
commercial breaks, each of which consits of a series of commercials and how to
automatically perform fast commercial recognition in real time.
Similarly, the video retrieval approaches can also be categorized into two
different approaches [1]. First category focuses mainly on visual features, such as
colour histograms, shapes, textures, or motion, which characterize the low-level
visual content. Although these approaches use automatically extracted features
representing the video content, they do not provide semantics that describe high-
level video information (concepts), which is much more appropriate for users when
retrieving video segments. The second one focuses on annotation based approaches
[2], which makes use of keyword annotation or text attribute to represent high level
concepts of the video content. Performance of this approach is affected by several
factors, including the fact that the search process depends solely on the predefined
attributed information. Furthermore, manual annotation process is very tedious and
time consuming. The challenges associated with the management of large database
of recorded programs involve: (1) knowledge-based methods for interpreting raw
data into semantic contents, (2) better query representations, and (3) incorporation
of relevant spatiotemporal relations.
Commercial break detection [3, 4] have based their strategies in studying the
relation between audio silences and black frames as an indicator of commercials
boundaries. The analysis is performed in either compressed [3] or uncompressed [4]
domain. In [5], specific country regulations about commercials broadcast is used as
a further clue. Another interesting approach is presented in [6], where the overlaid
text trajectories are used to detect commercial breaks. The idea here is that overlaid
text (if any) usually remains more stable during the program time than in the case of
commercials. In [7], Hidden Markov Model (HMM) based commercial detection is
presented with two different observations taken for each video shot: logo presence
and shot duration.
In many of the existing works, black frame [8] is used as the crucial information
for detecting the commercial in video processing. But, in Indian TV stations black
frames are absent and our methodology works without using black frames as video
clues. Our simple and accurate approach for commercial detection uses audiovisual
features. Initially, we find the commercial boundaries using a combination of
different audio features and the confidence level of the boundary detected will be
further enhanced and validated using splash screen (program name) video frames.
This detection is carried out around the commercial boundaries detected by audio.
29 Commercial Break Detection and Content Based Video Retrieval 375

Preprocessing and
Recorded TV Audio Data
Feature Extraction
Program

Video Data

Video Audio
Clue based based
Detection Detection

Automatic
Annotation Video Database

Retrieval Retrieved
Query Image
Process Videos

Fig. 1 System overview

Image matching of splash screen frame appearing at the start and end of the
commercial is performed using simple pixel intensity based matching.
Further, we make use of the information extracted during the commercial
detection process for automatic annotation and retrieval system which tries to fill
the gap between the low-level media features and high-level concepts for yielding
better performance using computationally simple procedures.
The overall scheme is depicted in the block diagram shown in Fig. 1. In the
remaining sections of this chapter the requisite details of the proposed method and
the relevant performance analysis results are provided.

2 Preprocessing and Feature Extraction

As discussed in the previous section the methodology adopted in this chapter is

based on two step approach combining both audio and video based approach
to accurately detect the commercial break boundaries in recorded TV program.
The rationale for this two-step approach is to reduce the search space for splash
screen detection during the video processing step, hence significantly reducing the
computational time.
376 N. Venkatesh and M.G. Chandra

In the following two subsections audio and video feature extraction issues are
provided.

2.1 Audio Feature Extraction

As a pre-processing step, the audio signal is extracted from the recorded TV

program and is segmented into audio frames of 23 ms duration with 50% overlap.
By banking on the general observation that the energy of the audio signal would
be comparatively high during the commercial break, we have opted for an energy
based feature. A simple energy feature is the usual short time energy given by
!
1 X
N 1
Ei ¼ x 2 ð nÞ (1)
N n¼0 i

where, i ¼ 1; 2; . . . ; F and N is the frame length (i.e., the number of samples in the
frame).
As evident from Eq. (1), the short term energy is computed per frame as the sum
of squares of the signal samples normalized by the frame length. But, using only
short time energy would result in large number of falsely detected commercial
boundaries and therefore we have come out with a new feature termed as Energy
Peak Rate (EPR), which is defined as the ratio of total number of energy peaks
exceeding the threshold T to the total number of frames (F) in the evaluation time
interval.
The requisite threshold is heuristically fixed as

T ¼ 0:5  maxðEi Þ (2)

The EPR is computed using the following expression

1X F
EPR ¼ Ep ð Ei Þ (3)
F i¼1

where, Ep ð xÞ is the binary-valued function, defined as

(
1 if x > T
E p ð xÞ ¼ (4)
0 else

Based on the elaborate experimentation, the evaluation time interval is chosen

as 5 s.
Using the suggested audio feature we will only be able to find the approximate
commercial break boundaries. As mentioned earlier, it is possible to enhance the
29 Commercial Break Detection and Content Based Video Retrieval 377

boundary detected by the audio features to arrive at the exact commercials bound-
aries using video processing. The methodology we have adopted is based on the
extraction of Harris corner points which is presented in the next section.

2.2 Video Feature Extraction

As a preprocessing step, we convert colour image to gray scale image on which the
subsequent processing is carried out. For detecting the exact boundaries, splash
screen (program name) frames are matched at the start and end of a commercial
break. A typical splash screen, including the program name (from one of the TV
recorded programs) is shown in Fig. 2. The matching is carried out using Harris
corner points, which are nothing but the points on the image where there is
significant change in the brightness in orthogonal directions [2].

2.2.1 Harris Corner Points

The Harris corner detector is a popular interest point detector because of its strong
invariance to rotation, scale, illumination variation and image noise [9]. The
Harris corner detector essentially finds a small number of points useful for matching
or finding correspondence and tracking [10]. This detector is based on the local
auto-correlation function of the image, which measures the local changes of the
image. cðx; y; Dx; DyÞ quantifies how similar is the image function Iðx; yÞ at point
(x,y) to itself, when shifted by ðDx; DyÞ[10]:

Fig. 2 Splash screen

(Program name)
378 N. Venkatesh and M.G. Chandra

X
cðx; y; Dx; DyÞ ¼ wðu; vÞ½Iðu; vÞ  Iðu þ Dx; v þ DyÞ2 (5)
ðu;vÞ2W ðx;yÞ

where, Wðx; yÞ is a window centered at point (x,y), and wðu; vÞ is either Gaussian.
Approximating the shifted function Iðu þ Dx; v þ DyÞ by the first-order Taylor
expansion, the autocorrelation can be in turn approximated as
" #
Dx
cðx; y; Dx; DyÞ ffi ½Dx; Dy Qðx; yÞ (6)
Dy

where Qðx; yÞ is a 2  2 matrix computed from the image partial derivatives

Ix and Iy :
" #
X Ix2 Ix Iy
Qðx; yÞ ¼ (7)
W Ix Iy Iy2

P P
where, wðu; vÞ is represented with its short-hand notation . The
ðu;vÞ2Wðx;yÞ W
requisite corner points are obtained as the local maxima of the corner response
function Hðx; yÞ given by

Hðx; yÞ ¼ Det Q  k ðTrace QÞ2 (8)

where, Det and Trace denote the determinant and trace of a matrix respectively and
k is a constant, usually set to the value 0.04 [10]. The local maximum in an image is
defined as a point greater than its neighbors in 33 or even 55 neighborhood.
Figure 3 shows the Harris corner points of the image depicted in Fig. 2.

Fig. 3 Harris corner points of

splash screen (program name)
29 Commercial Break Detection and Content Based Video Retrieval 379

3 Commercial Detection Scheme

The approximate commercial boundaries based on the audio feature are detected by
applying heuristics found by elaborate experimentation.

3.1 Audio Feature Based Detection

From Eq. (3) if Energy Peak Rate (EPR) exceeds the experimentally found thresh-
old y then we declare that frames of particular time interval lies in the commercial
break segment and hence we will be able to detect the approximate commercial
boundaries. The exact commercial boundaries will be detected by video processing
algorithm as outlined below.

3.2 Video Feature Based Detection

After detecting the approximate commercial boundaries using audio features, we

validate and further look for exact commercial boundaries using the splash screen
(program name) video frames matching at the start and end of commercial break.
The video processing is carried out to detect splash screen (program name) around
the boundaries detected by audio. Figure 2 shows an example of splash screen
(program name) appearing in one of the TV program video.
The image matching using the extracted features is carried out by simple pixel
intensity matching algorithm as explained below:
MP ¼ 0 where, MP is the number of pixels matched.
Step-1: P
Step-2: x;y if ðRðx; yÞ ¼ 1 & T ðx; yÞ ¼ 1Þ then MP ¼ MP þ 1
where, R(x, y) denotes the reference image and T ðx; yÞis the test image.
Step-3: PM ¼ M WR  100
P

Where, WR is the number of white pixels in reference template image R.

Step-4: If PM is greater than 80% then one can say that two images are similar.
It is to be noted that the step 2 is carried out using the Harris corner points of the
respective images for computationally efficient matching.

4 Mechanism for Automatic Annotation and Retrieval

4.1 Automatic Annotation

Once the commercial boundaries have been detected, the next step is to remove the
commercial part by retaining only the recorded TV program and during this
process, the splash screen (program name) frame detected in video based boundary
380 N. Venkatesh and M.G. Chandra

detection process is inserted at the beginning of the recorded TV program video

(without commercial) which is not going to affect the visual display because we are
inserting only one key frame at the beginning and normally video quality will be 25
frames per second. Additionally the key frame can be inserted with the useful
information like recording date, time and length of the video, which can be useful
during the retrieval process, for short listing the closest matching videos in response
to the query.

4.2 Content Based Video Retrieval

There will be a set of query images where each image is a splash screen (program
name) representing the recorded TV programs. It is to be noted that these set of
query images can be incorporated into the modern day set-top box and the asso-
ciated remote can be provided with a dedicated key to facilitate the display of query
images. Later, the query image will be compared with all the first frame of all the
videos and the best match is obtained through correlation matching as explained
below.
The correlation matching technique is simple and computationally fast and
hence it has been used to find the best match between the query image and the set
of images in the database. Query image (A) and the extracted key frame (B) from
the video are converted to binary images and then correlation coefficient is com-
puted between these two images using Eq. (9).
PP  
Ai ðm; nÞ  Ai Bðm; nÞ  B
ri ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
m n
   (9)
PP 2 PP 2
Ai ðm; nÞ  Ai Bðm; nÞ  B
m n m n

where, I ¼ 1,. . ., N, N is number of stored key frames in the database.

Ai is the key frame belonging to ith class of a database
Ai is the mean of Ai image
B is the query/test image
B is the mean of an image
Largest ri value indicates us the closest match of key frame to the query image.

5 Results and Discussion

In order to examine the proposed methodology and to carry out the relevant
experimentation, we have recorded eleven videos from seven different Indian TV
stations.
29 Commercial Break Detection and Content Based Video Retrieval 381

Table 1 Details of data

Video Video length (min:s) Commercial break duration (min:s) Class label
1 17:30 i.12:25–17:12 1
2 21:00 i. 08:10–11:45 1
ii. 20:10–21:00
3 07:09 i. 00:40–4:12 1
4 23:19 i. 04:52–52 2
ii. 17:20–22:28
5 21:17 i. 03:0–06:45 2
ii. 16:15–20:58
6 18:54 i. 06:42–13:38 3
7 18:26 i. 02:08–05:25 4
ii. 11:30–15:20
8 20:41 i. 03:00–06:35 5
ii. 14:20–17:30
9 19:55 i. 07:40–12:45 6
10 34:14 i. 03:46–09:10 7
ii. 14:45–19:25
iii. 27:53–31:10
11 33:24 i. 07:30–11:35 7
ii. 17:28–21:25
iii. 30:10–33:24

Videos are captured at 25 frames per second with 288  352 resolution and the
audio was extracted from these videos and stored in uncompressed wav format at
44,100 kHz with 16 bits per sample for further processing. Commercial details for
the recorded data are provided in Table 1. We have performed our experiments
using MATLAB® programming in Windows operating system. Videos with same
class label indicate that they are different episodes of same TV program which is
helpful during retrieval process.
From Table 3 we observe that using our simple video clues of matching splash
screen video frames at commercial boundaries for refining the boundaries detected
by audio we are able to increase the confidence level of the commercial boundary
detection and able to find the exact commercial boundaries.
One important observation to be made from Tables 2 and 3 is that false events
detected by audio processing is eliminated by the information obtained through
video processing step as the splash screen will not be present at the start and the end
of falsely detected commercial boundaries using audio. We can have a desired trade
off between number of falsely detected events and the number of missed events by
varying a threshold in audio based feature given by Eq. (2).
Once the commercial detection process is completed, the video will be auto-
matically annotated using the key frame with the additional information (see
Section 4.1) appended at the beginning of the video. Currently, the experimentation
has been carried out for seven classes (seven types of recorded TV programs). The
content based video retrieval is checked for all seven classes (query images) using
correlation classifier to match between query and reference images to retrieve the
closest matching automatically annotated videos. Retrieval performance results are
382 N. Venkatesh and M.G. Chandra

Table 2 Results using only audio feature

Video Total no of events No of false events No of missed event No of true events
detected detected detected detected
1 1 0 0 1
2 2 0 0 2
3 1 0 0 1
4 2 0 0 2
5 3 1 0 2
6 1 0 0 1
7 2 0 0 2
8 2 0 0 2
9 1 0 0 1
10 3 0 0 3
11 4 1 0 3

Table 3 Results using both audio and video feature

Video Total no of events No of false events No of missed event No of true events
detected detected detected detected
1 1 0 0 1
2 2 0 0 2
3 1 0 0 1
4 2 0 0 2
5 2 0 0 2
6 1 0 0 1
7 2 0 0 2
8 2 0 0 2
9 1 0 0 1
10 3 0 0 3
11 3 0 0 3

comparatively fast and accurate, with no error detected in the limited data set
considered.

6 Conclusion and Future Work

This chapter describes the commercial detection for recorded TV program using
audiovisual features. Further, automatic indexing and content based retrieval
algorithm which makes use of the information extracted during commercial detec-
tion is proposed. The system provides the user with complete solution of recording
TV program without advertisement and more importantly, database management of
the recorded TV programs is achieved through automatic annotation and fast content
based video retrieval. The proposed scheme is worked out keeping in mind the
simplicity and time efficiency of the used algorithms.
Our future work will be focused on testing the robustness of the methodology by
applying to wide variety of TV programs like news, serials, reality shows etc,
29 Commercial Break Detection and Content Based Video Retrieval 383

encompassing both Indian and international TV broadcasts. Further, in order to

demonstrate the efficacy of the scheme for content based retrieval, it is essential to
experiment with large video archives. Also of interest is the study of amenability of
the scheme and any requisite modifications for easy implementation on real-time
embedded systems (like TV set-top box).

References

1. C.A. Yoshitaka, T. Ichikawa, A survey on content-based retrieval for multimedia databases.

IEEE Trans. Knowl. Data Eng. 11(1), 81–93 (1999)
2. M.S. Kankanhalli, et al, Video modeling using strata-based annotation. IEEE Multimedia 7
(1), 68–73 (2000)
3. D.A. Sadlier, et al., Automatic TV advertisement detection from mpeg bitstream. J. Patt. Rec.
Soc. 35(12), 2–15 (Dec 2002)
4. Y. Li, C.-C. Jay Kuo, Detecting commercial breaks in real TV program based on audiovisual
information, in SPIE Proceedings on IMMS, vol. 421, Nov. 2000
5. R. Lienhart, C. Kuhmnch, W. Effelsberg, On the detection and recognition of television
commercials, in Proceedings of the IEEE Conference on MCS, Otawa, Canada, 1996
6. N. Dimitrova, Multimedia content analysis: the next wave, in Proceedings of the 2nd CIVR,
Illinois, USA, Aug 2003
7. A. Albiol, M.J. Fullà, A. Albiol, L. Torres, Commercials Detection using HMMs, 2004
8. A. Albiol, M.J. Fullà, A. Albiol, Detection of TV Commercials, in Proceedings of the
International Conference on Acoustics, Speech and Signal Processing, 2004, Rocky Mountain
Research Lab., Boulder, CO, private communication, May 1995, ed. C.J. Kaufman
9. C. Harris, M. Stephens, A combined corner and edge detector in Proceedings of the 4th Alvey
Vision Conference, 1988, pp. 147—151
10. Tomas Werner, (2 Oct 2008). http://cmp.felk.cvut.cz/cmp/courses/X33PVR/2009-2010ZS/
Docs/harris.pdf
11. N. Venkatesh, B. Rajeev, M. Girish Chandra, Novel TV commercial detection in cookery
program videos, in Proceedings of the World Congress on Engineering and Computer Science
2009 Vol II, WCECS 2009, San Francisco, USA, 20–22 Oct 2009
Chapter 30
ClusterDAM: Clustering Mechanism
for Delivery of Adaptive Multimedia Content
in Two-Hop Wireless Networks

Hrishikesh Venkataraman and Gabriel-Miro Muntean

Abstract In the recent years, there has been an increasing demand for streaming
high quality multimedia content over wireless networks. Demand for triple play
services (voice, video, data) by a number of simultaneously communicating
users often results in lack of acceptable multimedia quality. In addition, the large
transmission distance and the limited battery power of the hand-held wireless
device serves as a major bottleneck over transmitting video directly from the net-
work operator to the end-users. There has been no successful mechanism developed
till date that would provide live video streaming over wireless networks. In this
paper, clusterDAM – a novel cluster-based architecture is proposed for delivering
adaptive multimedia content over two-hop wireless networks. An intermediate
relay node serves as a proxy-client-server between the base station and the mobile
users which ensures that the multimedia delivery is adapted in real-time according
to the channel conditions and the quality of the multimedia content. An extensive
simulation-based analysis with different kinds of network traffic and protocols
indicate that the cluster-DAM architecture is significantly superior to the traditional
single-hop design, not only in terms of the perceived video quality, but also in terms
of the loss rate and the average bit rate.

1 Introduction

With the advent of third generation (3G) mobile systems, video conferencing and
downloading of movies/documentaries that requires huge file transfers of the order
of several tens of megabytes per second (Mbps) have been in great demand [1].
Multimedia transmission and VoD streaming requires high data rates which can be

H. Venkataraman (*)
S220, Performance Engineering Laboratory, School of Electronic Engineering, Dublin City
University, Glasnevin, Dublin 9, Ireland
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 385

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_30, # Springer ScienceþBusiness Media B.V. 2010
386 H. Venkataraman and G.M. Muntean

achieved either by increasing the bandwidth or by increasing the signal power at the
receiver. However, the total bandwidth of the network is controlled by the Govern-
ment and cannot be increased arbitrarily. Similarly, the wireless devices are energy-
constrained units; and hence, the transmission power of these devices cannot be
increased indiscriminately. Significantly however, the power required at the hand-
held device for multimedia decoding and playing is quite high; and this itself drains
the battery power considerably. In this scenario, an efficient mechanism to achieve
high-data rate communication is to use multihop transmission between the source
and destination node [2].
A hierarchical multihop design with a single central entity and several smaller
devices that would serve as intermediate relays is shown in Fig. 1. The relays reduce
the distance between a transmission-reception (Tx-Rx) pair which in turn reduces
the power requirement. At the same time, this increases the achievable maximum
data rate of a communicating link [3]. In addition, with the existence of different
kinds of wireless networks in today’s scenario (GPRS, CDMA2000, UMTS,
WiMAX, WLAN, etc.), the relays could serve as a switching point between the
different networks. In this context, the multihop design can efficiently model
the multiplicity in wireless networks. It has been shown [4] that time division
multihop cellular networks (MCN) offers the potential to integrate various multihop
infrastructure-based wireless networks. In fact, it has been shown in [4] that in
a multihop design with up to five hops, the spectral efficiency is significantly
increased as compared to the single-hop transmission.
A novel cluster-based design for two-hop hierarchical networks has been
recently proposed in [5], wherein, it has been shown that the cluster-based two-
hop design shows a significantly superior performance as compared with the state-
of-the-art resource algorithms. In this paper, Cluster-DAM - a novel cluster-based
architecture is proposed for delivery of adaptive multimedia content in two-hop
cellular networks. A double dumbbell topology that fits the two-hop cellular design

User
BS

PDA

Relays

User
In-
Computer home

Fig. 1 Next generation two-

hop hierarchical wireless Laptop
architecture Mobile
30 ClusterDAM: Clustering Mechanism for Delivery 387

is considered; and analyzed for efficient multimedia transmission. Different adap-

tive multimedia techniques, viz., LDAþ (enhanced loss delay adaptation protocol),
TFRCP (TCP friendly rate control protocol), RBAR (receiver based auto-rate
protocol) and QOAS (quality oriented adaptive design) are considered in the net-
work design, and their performance is analyzed for both single-hop (dumbbell
topology) and cluster-based two-hop design (double dumbbell topology). It is
shown that the QOAS-based wireless network with double dumbbell topology per-
forms significantly better than any other adaptive solutions and topology, not only
with regard to the video perceived quality, but also in terms of the bit rate and the
average loss rate.
The paper is organized as follows: 2 describes the cluster-DAM architecture over
two-hop wireless network model in detail, along with QOAS and other adaptive
solutions. 3 describes the simulation model, the set up, and the simulation scenarios.
The simulation results are explained in 4, and the conclusions are written in 5.

2 Cluster-Dam Architecture

2.1 Cluster-based Two-Hop Design for WiMAX Networks

The basic architecture of cluster-DAM is a cluster-based hierarchical two-hop

cellular network based on IEEE 802.16j standards. A multihop relay networks is
established between the server/base station (BS) and the end-user mobile stations
(MSs), as shown in Fig. 2. As per this design, there are six clusters in a coverage
area. The circular coverage area has a radius, r, and is divided into two layers. The
wireless nodes in the inner-layer communicate directly with the server; whereas the
wireless terminals in the outer-layer are grouped into several clusters (six clusters in
Fig. 2) [6]. In each cluster, a wireless terminal located at the boundary of the inner

Relay

MS MS
r/2
BS

MS

r cluster 1
Fig. 2 Cluster-based
hierarchical two-hop cellular
networks
388 H. Venkataraman and G.M. Muntean

Server to PCS PCS to Client -

Video Proxy
(relatively high bit rate) (relatively low bit rate)
Rate Adapted bit stream
Mobile Access Networks Adaptation
Wi-Fi / WLAN / WiMAX

Fig. 3 Proxy-client-server as video proxy be‘tween base station and end-users

and outer layer of the network region is selected as a cluster-head node, alternately
known as relay. The server always communicates with the users in the outer-layer
through the relay. Hence, the maximum transmission distance of a communicating
pair in the network is r/2 [7].
The relay node serves as a client during its communication with the BS (server);
and as a server while communicating with the end-users. Hence, the relay node can
be described as a proxy-client-server (PCS). In the adaptive multimedia delivery,
there would be several instances where there are multiple clients requesting for the
same video. In case of a single-hop design, this will require multiple encoding of a
single video stream, thereby creating unnecessary load on the BS. However, in the
two-hop design, the video stream could be stored in the PCS. After that, the adapted
stream could be transmitted individually to different clients. Hence, in this regard,
the PCS can also be considered as the database server. A two-hop design between
the BS and the end-users is shown in Fig. 3, wherein the PCS serves as the video
proxy. An adaptation framework incorporated at the PCS divides the incoming
multimedia stream from the server to several chunks of multimedia packets. The
size and length of each chunk depends upon the total size, length and the type of
video to be streamed.
In order to reduce the additional interference arising from the simultaneous
communication of multiple communicating pairs, a Protocol Model is considered
[8] for interference avoidance in the two-hop design. The reusability of the avail-
able spectrum resource (time slot in a time division multiple access system and a
frequency band in a frequency division multiple access system) is increased in the
cluster-based design by allowing two multihop clusters in the network to utilize
the same spectrum resource. It should be noted that the number of clusters in the
cluster-based design need not be always six [9]. But it should be an ‘even’ number
due to the basic principle of simultaneous transmission of communication pairs
located in the diametrically opposite clusters.

2.2 QOAS - Quality Oriented Adaptive Scheme

The primary aim of integrating QOAS with the cluster-based design in the IEEE
802.16j wireless model is to maintain a high end-user perceived quality even with
30 ClusterDAM: Clustering Mechanism for Delivery 389

an increase in the number of wireless devices in the network. QOAS relies on the
fact that the impact on the end-user perceived quality is greater in case of random
losses than that of controlled reduction in quality [10]. The system architecture of
the feedback-based QOAS includes multiple instances of the end-to-end adaptive
client and server applications [11]. Following ITU-T R. P.910 standard [12], a five
state model is defined for the multimedia streaming process. The QOAS client
continuously monitors the transmission parameters and estimates the end-user
perceived quality. The Quality of Delivery Grading Scheme (QoDGS) regularly
computes Quality of Delivery (QoD) scores that reflect the multimedia streaming
quality in current delivery conditions. These grades are then sent as feedback to the
server arbitration scheme (SAS). The SAS assesses the values of a number of
consecutive QoD scores received as feedback in order to reduce the effect of
noise in the adaptive decision process [13]. Based on these scores SAS suggests
adjustments in the data rate and other parameters.

2.3 Other Adaptive Solutions

With an increase in the demand for multimedia streaming in wireless networks, there
have been significant approaches researched in the recent past. TFRCP is a unicast
transport layer protocol, designed for multimedia streaming, and provides nearly the
same amount of throughput as that of TCP on wired networks. The TFRCP controls
rate based on network conditions expressed in terms of RTT and packet loss
probability [14]. Similar to TFRCP, LDAþ (enhanced loss delay adaptation) also
aims to regulate the transmission behavior of multimedia transmitters in accordance
with the network congestion state [15]. LDAþ uses RTP protocol for calculating loss
and delay and uses them for regulating transmission rates of the senders. LDAþ
adapts the streams in a manner similar to that of TCP connections. In comparison,
RBAR is a receiver based auto-rate mechanism. It is a MAC layer protocol and is
based on RTS/CTS mechanism [16]. The main feature of RBAR is that both channel
quality estimation and rate selection mechanism are on the receiver side. This allows
the channel quality estimation mechanism to directly access all of the information
made available to it by the receiver (number of multipath components, symbol error
rate, received signal strength, etc.) for more accurate rate selection.

3 Simulation Model and Testing

3.1 Dumbbell and Double Dumbbell Topology

In cluster-DAM, the delivery of adaptive multimedia content is achieved by using

a multimedia/web server. The web server acts as the multimedia source, which
transmits the multimedia content to all the wireless devices in its coverage area.
390 H. Venkataraman and G.M. Muntean

The end-users are the web-clients which receive the multimedia information.
Figure 4 shows a dumbbell topology for achieving the single-hop communication,
wherein, B1-B2 forms the bottleneck link. B1 is the multimedia source and
transmits information to the n clients C1, C2  L  Cn. In case of a two-hop
communication, there is an intermediate relay between the web source and the end-
user client. This can be represented by a double dumbbell topology, as shown in
Fig. 5. The multimedia server is represented by B0, whereas the diametrically
opposite relays are represented by B1 and B2. The end-users S1, S2, . . ., Sn on
one end and C1, C2, . . ., Cn on the diametrically opposite cluster are the multimedia
clients. The major advantage of the double dumbbell topology (cluster-based two-
hop design) over the dumbbell topology (single-hop wireless network) is the
hierarchical formation of the network, and the provision for peer-to-peer commu-
nication among the wireless nodes.

S1 C1

S2 C2

B2 B1

Sn Transmitter / Transmitter
Cn
Receiver / Receiver
Background Background
Traffic Traffic
Sender Receivers

Fig. 4 Dumbbell network topology for single-hop client-server wireless networks

S1 C1

S2 C2

B2 B0 B1

Sn Transmitter Transmitter Transmitter

Cn
/ Receiver / Receiver
Background Background
Traffic Traffic
Senders Receivers

Fig. 5 Double dumbbell network topology for two-hop client-server networks

30 ClusterDAM: Clustering Mechanism for Delivery 391

In the simulation environment, B1-B2 in the dumbbell topology and B0-B1, B0-
B2 links in the double dumbbell topology are the bottleneck links, with 2 Mbps
bandwidth and 100 ms latency. The other communicating links connected to the
end-users in the network are over-provisioned. Hence, the congestion in the traffic,
packet loss and delays occur mainly because of the bottleneck link. The buffering at
the ends of the bottle-neck link uses a drop-tail queue of size proportional to the
product of round trip time (RTT) and bottleneck link bandwidth.

3.2 Simulation Setup

The simulation setup consists of a number of mobile nodes distributed in the given
coverage area. There is a centrally located server which covers the whole area. In
order to increase the coverage area and the data rate provided to each node, two
fixed access points (APs) located equidistantly from the server and diametrically
opposite to each other are considered in the system design. The nodes located
between the server and the AP communicates with the server directly. However, the
mobile nodes that are located farther from the server and relatively closer to the AP
communicate with the server through the AP. The APs therefore act as proxy-
client-server between the actual client and server, as shown in Fig. 4. Hence, a
hierarchical structure exists between the server, the relays and the mobile nodes.
The system is simulated using the server and client instances inbuilt in network
simulator, NS-2. The length of all NS-2 simulations is 250s. The video streams are
modeled using Transform Expand Sample (TES) [17] and then encoded using
MPEG4. The primary reason for using MPEG4 is that it supports media streaming
and is suitable for home networking applications with its low bit rate as well as its
interoperability of audio and video signals [18]. The topology proposed assumes a
bandwidth of 5 MHz and 100 ms latency. In the above topology, B0-B1 and B0-B2
links are the bottleneck links. The other links connected to the mobile nodes are
over provisioned. Hence, the congestion in the traffic, packet loss and delays occur
mainly because of the bottleneck link. The buffering at the ends of the bottle-neck
link uses a drop-tail queue of size proportional to the product of round trip time
(RTT) and bottleneck link bandwidth. In each of the envisaged scenarios 95–99%
of the total available bandwidth is used for video and multimedia communication
and the remaining 1–5 is reserved for feedback purpose. A constant PLR of 107 is
assumed throughout the analysis. Similarly, a constant transmission delay of 10 ns
is assumed between the PCS and the mobile nodes.
A binary phase shift keying (BPSK) modulation technique is used at the physical
layer. A slow varying flat fading channel is assumed throughout the simulations.
In addition, a two-ray model with a lognormal shadowing of 4 dB standard devia-
tion is considered [19]. At the MAC layer, an IEEE 802.11 g based distributed
coordination function (DCF) is used. The simulation is done at the packet level and
the performance is evaluated in terms of average bit rate, loss rate and estimated
user perceived quality. The end-user perceived quality is measured by developing
392 H. Venkataraman and G.M. Muntean

a relationship between coding bitrate, packet loss ratio and user-level quality.
Traffic with different sizes and shapes are considered as in [20], so as to emulate
the real life scenario of varieties of traffic sources with different average bit rates.
The network performance is analyzed in terms of the perceived quality, the loss rate
and the average bit rate, not only for QOAS technique, but also compared with other
protocols, i.e., LDAþ, TFRCP and RBAR.

4 Results

Tables 1 and 2 present the performance results for UDP-CBR periodic traffic case,
using ‘dumbbell’ and ‘double dumbbell’ topology respectively. The traffic patterns
are kept identical for both single-hop and clustered two-hop scenario. It can be
Table 1 Performance of dumbbell topology – UDP CBR periodic traffic
Chars Traffic 1  0.6 1  0.6 1  0.8 1  0.8 1  1.0 1  1.0
(Mbps) 20 s on– 30 s on– 20 s on– 30 s on– 20 s on– 30 s on–
40 s off 60 s off 40 s off 60 s off 40 s off 60 s off
Perc. QOAS 3.908 3.809 3.942 3.608 3.809 3.565
video quality
(1–5) LDAþ 3.722 3.624 3.754 3.586 3.609 3.405
TFRCP 3.608 3.493 3.571 3.216 3.282 3.165
RBAR 3.621 3.467 3.279 3.011 2.945 2.829
Loss QOAS 0.18 0.18 0.14 0.14 0.10 0.10
rate (%) LDAþ 0.42 0.42 0.40 0.38 0.24 0.24
TFRCP 0.20 0.18 0.16 0.16 0.14 0.14
RBAR 0.22 0.20 0.18 0.16 0.14 0.14
Avg. bit-rate QOAS 0.726 0.734 0.776 0.763 0.802 0,802
(Mbps) LDAþ 0.713 0.726 0.736 0.734 0.772 0.772
TFRCP 0.708 0.718 0.741 0.740 0.762 0.772
RBAR 0.702 0.721 0.731 0.745 0.758 0.762

Table 2 Performance of double dumbbell topology – UDP CBR periodic traffic

Chars Traffic 1  0.6 1  0.6 1  0.8 1  0.8 1  1.0 1  1.0
(Mbps) 20 s on– 30 s on– 20 s on– 30 s on– 20 s on– 30 s on–
40 s off 60 s off 40 s off 60 s off 40 s off 60 s off
Perc. QOAS 4.209 4.032 4.112 3.879 3.989 3.691
video quality
(1–5) LDAþ 3.814 3.724 3.855 3.586 3.809 3.505
TFRCP 3.809 3.692 3.674 3.315 3.308 3.261
RBAR 3.693 3.423 3.404 3.016 3.124 3.083
Loss QOAS 0.08 0.08 0.04 0.04 0.01 0.01
rate (%) LDAþ 0.38 0.38 0.34 0.34 0.21 0.21
TFRCP 0.12 0.12 0.16 0.16 0.10 0.10
RBAR 0.18 0.18 0.16 0.16 0.12 0.12
Avg. bit-rate QOAS 0.732 0.732 0.756 0.756 0.812 0.813
(Mbps) LDAþ 0.756 0.756 0.788 0.787 0.793 0.793
TFRCP 0.726 0.738 0.757 0.766 0.782 0.792
RBAR 0.704 0.718 0.735 0.758 0.779 0.786
30 ClusterDAM: Clustering Mechanism for Delivery 393

observed that for all three kinds of periodic traffic considered, the perceived
quality obtained for the cluster-based two-hop design (double dumbbell topology)
is significantly superior to that obtained using the single-hop design (dumbbell
topology).
For example, in case of 1  0.6 Mbps traffic with 20 s on 40 s off, the
perceived quality obtained with QOAS in the single-hop model is 3.908, whereas
that obtained using the two-hop model is 4.209, an increase of 7.71%. In addition,
for the same traffic model, the perceived quality of QOAS in the single-hop design
is better than any other scheme (LDAþ) by 5.25% (QOAS has a Q of 3.908 whereas
LDAþ has a Q of 3.7025). However, in case of a cluster-based two-hop design, the
improvement in the perceived quality between QOAS and LDAþ is at least 9.97%,
almost twice the benefit obtained from ‘dumbbell’ topology. As seen from Tables 1
and 2, the improvement of double dumbbell topology over dumbbell topology is
observed consistently across different UDP-CBR traffics.
In a similar result, in case of UDP-CBR staircase traffic, the perceived quality
obtained from the ‘double dumbbell’ topology scores significantly over the ‘dumb-
bell’ scheme, as can be seen from Tables 3 and 4. For example, in case of 4  0.4
Mbps (Up 40 s steps), the perceived quality of QOAS using ‘dumbbell’ scheme is
3.808, whereas the same using ‘double dumbbell’ scheme is 4.298, an increase of
12.87%. It can be observed from Tables 3 and 4 that the improvement in the average
bit rate and the loss rate is also notably high in case of the double dumbbell
topology, as compared to the dumbbell scheme. Similarly, the improvement in
loss rate for QOAS is over six times (0.04 using ‘double dumbbell’ and 0.24 using
dumbbell). Importantly, this performance improvement remains consistent over
different scenarios of staircase traffic, as shown in Tables 3 and 4.
In the Internet environment, traffic is generally bursty, as in case of FTP traffic.
Tables 5 and 6 shows the performance for different FTP and WWW traffic

Table 3 Performance of UDP-CBR staircase traffic in dumbbell topology

Char. Traffic 1  0.6 1  0.6 1  0.8 1  0.8 1  1.0 1  1.0
(Mbps) 20 s on– 30 s on– 20 s on– 30 s on– 20 s on– 30 s on–
40 s off 60s off 40 s off 60 s off 40 s off 60 s off
Perc. QOAS 3.808 3.862 3.726 3.887 3.727 3.581
video quality
(1–5) LDAþ 3.683 3.677 3.496 3.579 3.561 3.321
TFRCP 3.585 3.481 3.176 3.202 2.871 2.687
RBAR 3.481 3.179 3.006 2.954 2.784 2.677
Loss QOAS 0.24 0.18 0.18 0.16 0.14 0.14
rate (%) LDAþ 0.46 0.42 0.40 0.28 0.24 0.24
TFRCP 0.26 0.18 0.18 0.18 0.14 0.13
RBAR 0.24 0.20 0.18 0.18 0.16 0.16
Avg. bit-rate QOAS 0.716 0.763 0.753 0.802 0.846 0.858
(Mbps) LDAþ 0.706 0.746 0.748 0.752 0.814 0.805
TFRCP 0.686 0.721 0.734 0.742 0.801 0.808
RBAR 0.671 0.711 0.725 0.738 0.762 0.801
394 H. Venkataraman and G.M. Muntean

Table 4 Performance of UDP-CBR staircase traffic-double dumbbell topology

Char. Traffic 1  0.6 1  0.6 1  0.8 1  0.8 1  1.0 1  1.0
(Mbps) 20 s on– 30 s on– 20 s on- 30 s on– 20 s on- 30 s on–
40 s off 60 s off 40 s off 60 s off 40 s off 60 s off
Perc. QOAS 4.298 4.124 4.061 4.098 3.961 3.835
video quality
(1–5) LDAþ 4.096 3.929 3.861 3.803 3.662 3.635
TFRCP 3.834 3.675 3.636 3.597 3.441 3.335
RBAR 3.721 3.485 3.241 3.144 3.004 2.955
Loss QOAS 0.04 0.00 0.00 0.06 0.01 0.00
rate (%) LDAþ 0.24 0.18 0.18 0.16 0.11 0.09
TFRCP 0.08 0.03 0.04 0.10 0.04 0.03
RBAR 0.08 0.14 0.10 0.12 0.12 0.12
Avg. bit-rate QOAS 0.736 0.809 0.818 0.712 0.797 0.817
(Mbps) LDAþ 0.725 0.789 0.768 0.682 0.767 0.807
TFRCP 0.705 0.787 0.798 0.682 0.767 0.784
RBAR 0.702 0.768 0.756 0.678 0.702 0.725

Table 5 Performance of TCP-FTP traffic using dumbbell topology

Char. Traffic 50 x 54 x 58 x 40 x 50 x 60 x
(Mbps) FTP FTP FTP WWW WWW WWW
Perc. QOAS 4.093 3.722 3.701 4.794 4.451 4.002
video quality
(1–5) LDAþ 3.901 3.725 3.612 4.393 4.160 3.922
TFRCP 3.609 3.473 3.305 4.203 4.065 3.682
RBAR 3.512 3.307 3.288 2.754 2.659 2.804
Loss rate QOAS 0.10 0.08 0.08 0.10 0.08 0.15
rate (%) LDAþ 0.20 0.20 0.18 0.14 0.14 0.18
TFRCP 0.14 0.14 0.14 0.14 0.14 0.19
RBAR 0.18 0.22 0.22 0.24 0.24 0.18
Avg. bit-rate QOAS 0.743 0.774 0.775 0.701 0.701 0.758
(Mbps) LDAþ 0.714 0.723 0.739 0.684 0.684 0.748
TFRCP 0.714 0.737 0.752 0.691 0.691 0.716
RBAR 0.693 0.716 0.739 0.752 0.752 0.781

Table 6 Performance of TCP-FTP traffic using double dumbbell topology

Char. Traffic 50 x 54 x 58 x 40 x 50 x 60 x
(Mbps) FTP FTP FTP WWW WWW WWW
Perc. QOAS 4.109 3.972 3.801 4.809 4.565 4.162
video quality
(1–5) LDAþ 4.003 3.902 3.712 4.599 4.361 4.062
TFRCP 3.809 3.672 3.501 4.409 4.265 3.862
RBAR 3.712 3.506 3.387 2.924 2.829 2.904
Loss QOAS 0.00 0.01 0.04 0.01 0.01 0.05
rate (%) LDAþ 0.10 0.10 0.14 0.11 0.12 0.15
TFRCP 0.10 0.10 0.12 0.04 0.04 0.09
RBAR 0.10 0.12 0.12 0.14 0.14 0.12
Avg. bit-rate QOAS 0.756 0.787 0.797 0.712 0.721 0.776
(Mbps) LDAþ 0.724 0.741 0.759 0.701 0.719 0.753
TFRCP 0.736 0.757 0.767 0.702 0.711 0.737
RBAR 0.702 0.722 0.754 0.769 0.769 0.791
30 ClusterDAM: Clustering Mechanism for Delivery 395

models for the duration of 250 s. Similar to the results obtained in case of UDP
transmission, the performance of the ‘double dumbbell’ topology is superior to that
obtained from the ‘dumbbell’ topology for both FTP and WWW traffic.

5 Conclusions

This paper proposes cluster-DAM, a resource-efficient cluster-based design for

delivery of adaptive multimedia content in IEEE 801.16j enabled two-hop cellular
network. A ‘double dumbbell’ topology is considered in the cluster-based two-hop
design so that the clients that are located diametrically opposite to the web server
could communicate simultaneously. The ‘state-of-the-art’ cluster-DAM solution -
QOAS combined with the cluster-based design not only results in superior multi-
media transmission as compared to single-hop network, but also outperforms other
protocols like LDAþ, TFRCP and RBAR. In addition, the loss rate of the video
frames is reduced, even up to a factor of 10, when the two-hop cluster-based design
is used. This is a significant result for next generation data-centric wireless trans-
mission systems. It demonstrates the feasibility of multimedia streaming over
distantly located wireless users. This would boost the network operator to incorpo-
rate both the aspects of the cluster-DAM model in designing the next generation
wireless networks.

Acknowledgements The authors would like to acknowledge Dr. Sinan Sinanovic and Prof.
Harald Haas from the University of Edinburgh for their input during the initiatory stages. Also,
the authors would like to thank Irish Research Council for Science Engineering and Technology
(IRCSET) for supporting this research work.

References

1. B. Li, H. Yin, Peer-to-peer Live Video Streaming on the Internet: Issues, Existing Approaches
and Challenges. IEEE Commun. Mag. 45(6), 94–99 (2007)
2. S. Toumpis, A.J. Goldsmith, Capacity regions for wireless ad hoc networks. IEEE Trans.
Wireless Commun. 2(4), 736–748 (2003)
3. H. Venkataraman, G.M. Muntean, Analysis of random data hopping in distributed multihop
wireless networks, in Proceedings of IEEE Region Ten Conference (TENCON), Hyderabad,
India, 18–21 Nov 2008
4. Y. Liu, R. Hoshyar, X. Yang, R. Tafazolli, Integrated radio resource allocation for multihop
cellular networks with fixed relay stations. IEEE J. Selected Areas Commun. 24(11),
2137–2146 (Nov 2006)
5. H. Venkataraman, S. Sinanovic, H. Haas, Cluster-based design for two-hop cellular networks.
Int. J. Commun. Netw. Syst. (IJCNS) 1(4), Nov 2008
6. H. Venkataraman, P. Kalyampudi, A. Krishnamurthy, J. McManis G.M. Muntean, Clustered
architecture for adaptive multimedia streaming in WiMAX based cellular networks, in
Proceedings of World Congress on Engineering and Computer Science (WCECS), Berkeley,
San Francisco, CA, Oct 2009
396 H. Venkataraman and G.M. Muntean

7. H. Venkataraman, P.K. Jaini, S. Revanth, Optimum location of gateways for cluster-based

two-hop cellular networks. Int. J. Commun. Netw. Syst. (IJCNS) 2(5) (Aug 2009)
8. P. Gupta, P.R. Kumar, The capacity of wireless networks. IEEE Trans. Inf. Theory 46(2)
388–404 (Feb 2000)
9. H. Venkataraman, S. Nainwal, P. Srivastava, Optimum number of gateways in cluster-based
two-hop cellular networks. AEUe J. Electron. Commun. Elsevier 64(4), 310–321 (April 2010)
10. G.M. Muntean, N. Cranley, Resource efficient quality-oriented wireless broadcasting of
adaptive multimedia content. IEEE Trans. Broadcasting 53(1), 362–368 (Mar 2007)
11. G.M. Muntean, P. Perry, L. Murphy, A New Adaptive Multimedia Streaming System for all-
IP Multi-Service Networks. IEEE Trans. Broadcasting 50(1), 1–10 (2004)
12. ITU-T Recommendation P.910, Subjective Video Quality Assessment Methods for Multime-
dia Applications, Sept 1999
13. G.M. Muntean, P. Perry, L. Murphy, A comparison-based study of quality-oriented video on
demand. IEEE Trans. Broadcasting 53(1), 92–102 (Mar 2007)
14. M. Miyabayashi, N. Wakamiya, M. Murata, H. Miyahara, MPEG-TFRCP: video transfer with
TCP-friendly rate control protocol, in Proceedings of IEEE International Conference on
Communications (ICC’01), vol. 1, Helsinki, Finland, June 2001, pp. 137–141
15. D. Sisalem, A. Wolisz, LDAþ TCP Friendly Adaptation: A Measurement and Comparison
Study, in ACM International Workshop on Network and Operating Systems Support for
Digital Audio and Video (NOSSDAV), Chapel Hill, NC, USA, June 2000
16. G. Holland, N. Vaidy, V. Bahl, A rate-adaptive MAC protocol for multihop wireless networks,
Rome, Italy, in Proceedings of ACM Mobile Computing and Networking (MOBICOM’ 01),
2001, pp. 236–251
17. A. Matrawy, I. Lambadaris, C. Huang, MPEG4 traffic,modeling using the transform expand
sample methodology, in Proceedings of 4th IEEE International Workshop on Networked
Appliances, 2002, pp. 249–256
18. H. Venkataraman, S. Agrawal, A. Maheshwari, G.M. Muntean, Quality-controlled prioritized
adaptive multimedia streaming in two-hop wireless networks, in Proceedings of World
Congress on Engineering and Computer Science (WCECS), Berkeley, San Francisco, CA,
Oct 2009
19. 3rd Generation Partnership Project (3GPP), Technical Specification Group Radio Access
Network, Selection procedures for the choice of radio transmission technologies of the
UMTS, 3GPP TR 30.03U, May 1998
20. Z. Yuan, H. Venkataraman, G.M. Muntean, iPAS: an user perceived quality based intelligent
prioritized adaptive scheme for IPTV in wireless home networks, in Proceedings of IEEE
Broadband Multimedia Systems and Broadcasting (BMSB) Conference, Shanghai, China,
24–26 Mar 2010
Chapter 31
Ranking Intervals in Complex Stochastic
Boolean Systems Using Intrinsic Ordering

Luis González

Abstract Many different phenomena, arising from scientific, technical or social

areas, can be modeled by a system depending on a certain number n of random
Boolean variables. The so-called complex stochastic Boolean systems (CSBSs) are
characterized by the ordering between the occurrence probabilities Pr{u} of the 2n
associated binary strings of length n, i.e., u¼(u1,. . .,un) 2 {0,1}n. The intrinsic order
defined on {0,1}n provides us with a simple positional criterion for ordering the
binary n-tuple probabilities without computing them, simply looking at the relative
positions of their 0s and 1s. For every given binary n-tuple u, this paper presents two
simple formulas – based on the positions of the 1-bits (0-bits, respectively) in u – for
counting (and also for rapidly generating, if desired) all the binary n-tuples v whose
occurrence probabilities Pr{v} are always less than or equal to (greater than or equal
to, respectively) Pr{u}. Then, from these formulas, we determine the closed interval
covering all possible values of the rank (position) of u in the list of all binary
n-tuples arranged by decreasing order of their occurrence probabilities. Further, the
length of this so-called ranking interval for u, also provides the number of binary
n-tuples v incomparable by intrinsic order with u. Results are illustrated with the
intrinsic order graph, i.e., the Hasse diagram of the partial intrinsic order.

1 Introduction

In many different scientific, technical or social areas, one can find phenomena
depending on an arbitrarily large number n of basic random Boolean variables. In
other words, the n basic variables of the system are assumed to be stochastic and
they only take two possible values: either 0 or 1. We call such a system: a complex

L. González
Research Institute IUSIANI, Department of Mathematics, University of Las Palmas de Gran
Canaria, Campus Universitario de Tafira, 35017 Las Palmas de Gran Canaria, Spain
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 397

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_31, # Springer ScienceþBusiness Media B.V. 2010
398 L. González

stochastic Boolean system (CSBS). Each one of the 2n possible elementary states
associated to a CSBS is given by a binary n-tuple u¼(u1,. . . , un) 2 {0,1}n of 0s and
1s, and it has its own occurrence probability Pr{(u1,. . . , un)}.
Using the statistical terminology, a CSBS on n variables x1,. . . , xn can be
modeled by the n-dimensional Bernoulli distribution with parameters p1,. . . , pn
defined by

Prfxi ¼ 1g ¼ pi ; Prfxi ¼ 0g ¼ 1  pi ;

Throughout this paper we assume that the n Bernoulli variables xi are mutually
statistically independent, so that the occurrence probability of a given binary string
u¼(u1,. . . , un) 2 {0,1}n of length n can be easily computed as

Y
n
Prfug ¼ pui i ð1  pi Þ1ui ; (1)
i¼1

that is, Pr{u} is the product of factors pi if ui¼1, 1pi if ui¼0 [6].
Example 1. Let n¼4 and u¼(0,1,0,1) 2 {0,1}4. Let p1¼0.1, p2¼0.2, p3¼0.3,
p4¼0.4. Then, using Eq. 1, we have

Prfð0; 1; 0; 1Þg ¼ ð1  p1 Þp2 ð1  p3 Þp4 ¼ 0:0504:

The behavior of a CSBS is determined by the ordering between the current

values of the 2n associated binary n-tuple probabilities Pr{u}. For this purpose, in
[1] we have established a simple positional criterion that allows one to compare two
given binary n-tuple probabilities, Pr{u},Pr{v}, without computing them, simply
looking at the positions of the 0s and 1s in the n-tuples u,v. We have called it the
intrinsic order criterion, because it is independent of the basic probabilities pi and it
intrinsically depends on the positions of the 0s and 1s in the binary strings.
In this context, the main goal of this paper is to determine, in a simple way, the
ranking interval of u for every fixed n1 and for every fixed binary n-tuple u. This
is defined as the closed interval that covers all possible ranks (i.e., positions) of u in
the list of all binary n-tuples when they are arranged by decreasing order of their
occurrence probabilities. Of course the position of each n-tuple u in such lists
depends on the way in which the current values {pi}ni¼1 of the Bernoulli parameters
affect the ordering between the occurrence probability of u and the occurrence
probability of each binary n-tuple v incomparable with u.
For this purpose, this paper has been organized as follows. In Section 2, we
present some previous results on the intrinsic order relation, the intrinsic order
graph and other related aspects, enabling non-specialists to follow the paper without
difficulty and making the presentation self-contained. Section 3 is devoted to
provide new formulas for counting (and generating) all the binary n-tuples intrinsi-
cally greater than, less than, or incomparable with u. In Section 4, we present our
31 Ranking Intervals in Complex Stochastic Boolean Systems 399

main result: We exactly obtain the ranking interval for each given binary string u,
and we illustrate our propositions with a simple, academic example. Finally, in
Section 5, we present our conclusions.

2 The Intrinsic Ordering

2.1 Intrinsic Order Relation on {0,1}n

Throughout this paper, the decimal numbering of a binary string u is denoted by

the symbol u(10, while the Hamming weight of a binary n-tuple u (i.e., the number
of 1-bits in u) will be denoted, as usual by wH{u}, i.e.,

X
n X
n
ðu1 ; . . . ; un Þ  uð10 ¼ 2ni ui ; wH ðuÞ ¼ ui ;
i¼1 i¼1

e.g., for n¼5 we have

ð1; 0; 1; 1; 1Þ  20 þ 21 þ 22 þ 24 ¼ 23; wH ð1; 0; 1; 1; 1Þ ¼ 4:

According to Eq. 1, the ordering between two given binary string probabilities
Pr{u} and Pr{v} depends, in general, on the parameters pi, as the following simple
example shows.
Example 2. Let n ¼ 3; u ¼ (0,1,1) and v¼(1,0,0). Using Eq. 1 we have

p1 ¼ 0:1; p2 ¼ 0:2; p3 ¼ 0:3 : Prfug ¼ 0:054< Prfvg ¼ 0:056;

p1 ¼ 0:2; p2 ¼ 0:3; p3 ¼ 0:4 : Prfug ¼ 0:096> Prfvg ¼ 0:084:

As mentioned in Section 1, to overcome the exponential complexity inherent to

the task of computing and sorting the 2n binary string probabilities (associated to a
CSBS with n Boolean variables), we have introduced the following intrinsic order
criterion [1], denoted from now on by the acronym IOC (see[5] for the proof).

Theorem 1 (The intrinsic order theorem). Let n  1. Suppose that x1 , . . . , xn

are n mutually independent Bernoulli variables whose parameters pi ¼
Prfxi ¼ 1gsatisfy

0<p1  p2      pn  0:5: (2)

Then the probability of the binary n-tuple v ¼ (v1,. . .,vn) is intrinsically less than or
equal to the probability of the binary n-tuple u ¼ (u1,. . .,un) (that is, for all set
{pi}ni¼1 satisfying (2)) if and only if the matrix
400 L. González

!
u1 . . . un
Muv :¼
v1 . . . vn

either has no ð10Þ columns, or for each ð10Þ column in Mvu there exists (at least) one
corresponding preceding ð01Þ column (IOC).
Remark 1. In the following, we assume that the parameters pi always satisfy
condition (2). Note that this hypothesis is not restrictive for practical applications
because, if for some i : pi > 12 , then we only need to consider the variable
xi ¼ 1  xi , instead of xi. Next, we order the n Bernoulli variables by increasing
order of their probabilities.
Remark 2. The ð01Þ column preceding to each ð10Þ column is not required to be
necessarily placed at the immediately previous position, but just at previous position.
Remark 3. The term corresponding, used in Theorem 1, has the following meaning:
For each two ð10Þ columns in matrix Mvu, there must exist (at least) two different ð01Þ
columns preceding to each other. In other words: For each ð10Þ column in matrix
Mvu, the number of preceding ð01Þ columns must be strictly greater than the number
of preceding ð10Þ columns.
The matrix condition IOC, stated by Theorem 1, is called the intrinsic order
criterion, because it is independent of the basic probabilities pi and it intrinsically
depends on the relative positions of the 0s and 1s in the binary n-tuples u,v.
Theorem 1 naturally leads to the following partial order relation on the set {0,1}n
[1]. The so-called intrinsic order will be denoted by “  ”, and we shall write v  u
(u  v) to indicate that v (u) is intrinsically less (greater) than or equal to u (v).
Definition 1. For all u,v 2 {0,1}n

v  u iff Prfvg  Prfug for all set fpi gni¼1 s.t. ð2Þ

iff Muv satisfies IOC.

From now on, the partially ordered set (poset, for short) ({0,1}n,  ) will be
denoted by In.

Example 3. For n¼3, we have 3  ð0; 1; 1Þ  4  ð1; 0; 0Þ and also

4  ð1; 0; 0Þ  3  ð0; 1; 1Þ because the matrices
   
1 0 0 0 1 1
and
0 1 1 1 0 0

do not satisfy IOC (Remark 3). Thus, (0,1,1) and (1,0,0) are incomparable by
intrinsic order, i.e., the ordering between Pr{(0,1,1)} and Pr{(1,0,0)} depends on the
parameters {pi}3i¼1, as Example 2 has shown.
31 Ranking Intervals in Complex Stochastic Boolean Systems 401

Example 4. For n¼5, we have 24(1,1,0,0,0)  5(0,0,1,0,1) because matrix

 
0 0 1 0 1
1 1 0 0 0

satisfies IOC (Remark 2).Thus, for all {pi}5i¼1 s.t. (2)

Prfð1; 1; 0; 0; 0Þg  Prfð0; 0; 1; 0; 1Þg:

Example 5. For all n1, the binary n-tuples

 n
  n

0; .^
..;0  0 and 1; .^
. . ; 1  2n  1

are the maximum and minimum elements, respectively, in the poset In. Indeed, both
matrices
   
0 ... 0 . . . unu1
and
u1 ... un ... 1 1
 
1
satisfy the intrinsic order criterion, since they have no columns!.Thus, for all
0
u 2 {0,1}n and for all {pi}ni¼1 s.t. (2)
n n o n n o
Pr 1; .^
..;1  Prfðu1 ; . . . ; un Þg  Pr 0; .^
..;0 :

Many different properties of the intrinsic order relation can be derived from its
simple matrix description IOC (see, e.g., [1–3]). For the purpose of this paper, we
need recall here the following necessary (but not sufficient) condition for intrinsic
order; see [2] for the proof.

Corollary 1. For all u,v 2 {0,1}n

uv ) wH ðuÞ  wH ðvÞ:

2.2 The Intrinsic Order Graph

Now, we present, in Fig. 1, the graphical representation of the poset In. The usual
representation of a poset is its Hasse diagram (see, e.g., [7] for more details about
posets and Hasse diagrams). This is a directed graph (digraph, for short) whose
vertices are the binary n-tuples of 0s and 1s, and whose edges go downward from u
to v whenever u covers v (denoted by u B v), that is, whenever u is intrinsically
greater than v with no other elements between them, i.e.,
402 L. González

0
|
1
|
0 2
| |
1 3 4
0 | |
| 2 5 8
0 | |
1 |
| 3 4 9
| 6
1 2 | |
|
| 5 7 10
3 | |
6 11 12
| |
7 13
|
14
|
15

Fig. 1 The intrinsic order graph for n¼1,2,3,4

uBv iff u v and there is no w 2 f0; 1gn s.t. u w v:

The Hasse diagram of the poset In will be also called the intrinsic order
graph for n variables. For all n2, in [3] we have developed an algorithm for
iteratively building up the digraph of In from the digraph of I1. Basically, In is
obtained by adding to In1 its isomorphic copy 2n1 þ I n1 : A nice fractal
property of In!
In Fig. 1, the intrinsic order graphs of I1, I2, I3 and I4 are shown, using the
decimal numbering instead of the binary representation of their 2n nodes, for a more
comfortable and simpler notation. Each pair (u,v) of vertices connected in the
digraph of In either by one edge or by a longer descending path (consisting of
more than one edge) from u to v, means that u is intrinsically greater than v, i.e.,
u v. On the contrary, each pair (u,v) of non-connected vertices in the digraph of In
either by one edge or by a longer descending path, means that u and v are
incomparable by intrinsic order, i.e., u  v and v  u.
Looking at any of the four graphs in Fig. 1, we can confirm the properties of the
intrinsic order stated by both Example 5 and Corollary 1.

2.3 Three Sets of Bitstrings Related to a Binary n-tuple

In the following two definitions, we present three subsets of {0,1}n related to each
binary n-tuple u 2 {0,1}n.
31 Ranking Intervals in Complex Stochastic Boolean Systems 403

Definition 2. For every binary n-tuple u 2 {0,1}n, the set Cu (the set Cu, respec-
tively) is the set of all binary n-tuples v whose occurrence probabilities Pr{v}
are always less (greater, respectively) than or equal to Pr{u}, i.e., according to
Definition 1, those n-tuples v intrinsically less (greater, respectively) than or
equal to u, i.e.,

Cu ¼ fv 2 f0; 1gn j u  vg; (3)

Cu ¼ fv 2 f0; 1gn j u  vg: (4)

Definition 3. For every binary n-tuple u 2 {0,1}n, Inc{u} is the set of all binary n-
tuples v intrinsically incomparable with u, i.e.,

IncðuÞ ¼ fv 2 f0; 1gn j u  v; u  vg ¼ f0; 1gn  ðCu [ Cu Þ: (5)

For instance, due to Example 5, we have, for all n1,

 n   n   n 
^
C 0;...; 0 ¼ f0; 1gn ¼ Cð1;...;
^ 1 ; Inc
n
Þ 0; .^
. . ; 0 ¼ fg ¼ Inc 1; .^
..;1

as Fig. 1 shows, for n¼1,2,3,4.

Remark 4. Note that, because of the reflexive and antisymmetric properties of the
intrinsic order relation  , we obviously have, for all n1 and for all u 2 {0,1} n

Cu \ Cu ¼ fug: (6)

The sets Cu and Cu are closely related via complementary n-tuples, as described
with precision by the following definition and theorem (see [4] for the proof).

Definition 4. The complementary n-tuple of a given binary n-tuple u2{0,1}n is

obtained by changing its 0s by 1s and its 1s by 0s

uc ¼ ðu1 ; . . . ; un Þc ¼ ð1  u1 ; . . . ; 1  un Þ:

The complementary set of a given subset S {0,1}n of binary n-tuples is the set of
the complementary n-tuples of all the n-tuples of S

Sc ¼ fuc j u 2 Sg:

Theorem 2. For all n  1 and for all u 2 {0,1}n

 c c
Cu ¼ Cu ; Cu ¼ ½Cuc c : (7)
404 L. González

3 Generating and counting the elements of Cu and Cu

First, we need to set the following nomenclature and notation.

Definition 5. Let n1 and let u 2 {0,1}n with Hamming weight wH(u)¼m. Then
(i) The vector of positions of 1s of u is the vector of positions of its m 1-bits,
displayed in increasing order from the left-most position to the right-most position,
and it will be denoted by
 1
u ¼ i11 ; . . . ; i1m n ; 1  i11 <    <i1m  n; 0<m  n:

(ii) The vector of positions of 0s of u is the vector of positions of its (nm) 0-bits,
displayed in increasing order from the left-most position to the right-most position,
and it will be denoted by
 0
u ¼ i01 ; . . . ; i0nm n ; 1  i01 <    <i0nm  n; 0  m<n:

Example 6. Let n¼6 and u ¼ 27  ð0; 1; 1; 0; 1; 1Þ. Then we have m ¼ wH ðuÞ ¼ 4,

n  m ¼ 2, and
 1  0
u ¼ i11 ; i12 ; i13 ; i14 6
¼ ½2; 3; 5; 6 16 ; u ¼ i01 ; i02 6
¼ ½1; 4 06 :

In [4], the author presents an algorithm for obtaining the set Cu, for each given
binary n-tuple u. Basically, this algorithm determines Cu by expressing the set
difference {0,1}nCu as a set union of certain half-closed intervals of natural
numbers (decimal representation of bitstrings). The next two theorems present
new, more efficient algorithms for rapidly determining Cu and Cu, using the binary
representation of their elements. Moreover, our algorithms allow us not only to
generate, but also to count the number of elements of Cu and Cu.

Theorem 3. Let u 2 {0,1}n, u6¼0, with (nonzero) Hamming weight wH{u} ¼ m (0 <
m  n). Let u ¼ {i11,. . .,im1}n1be the vector of positions of 1s of u. Then Cu is the set
of binary n-tuples v generated by the following algorithm:
The Algorithm for Cu
(i) Step 1 (Generation of the sequences {j11,. . .,jm1}) :
For j11 ¼ 1 to i11 Do:
For j12 ¼ j11 þ 1 to i21 Do:
                                
For j1m ¼ j1m1 þ 1 to im1 Do:
Write {j11,j21,. . .,jm1}
EndDo
                                
EndDo
EndDo
31 Ranking Intervals in Complex Stochastic Boolean Systems 405

(ii) Step 2 (Generation of the n-tuples v 2 Cu) :

For every sequence {j11,. . .,jm1} generated by Step 1, define the binary
n-tuple v ¼ {v1,. . .,vn} as follows
8
>
< 1 if i 2 j11 ; . . . ; j1m ;
vi ¼ 2 j11 ; . . . ; j1m ; i<j1m ;
0 if i= (8)
>
:
2
0; 1 if i= j11 ; . . . ; j1m ; i>j1m :

Moreover, the cardinality of the set Cu is given by the multiple sum

X X X
1 1 1
i1 i2 im
1
jC j ¼
u
 2njm : (9)
j11 ¼1 j12 ¼j11 þ1 j1m ¼j1m1 þ1

Proof. According to (3) and Corollary 1, we have

v 2 Cu , u  v ) m ¼ wH ðuÞ  wH ðvÞ;

that is, the Hamming weight of every n-tuple v2Cu is necessarily greater than or
equal to the Hamming weight m of u. So, all n-tuples v of Cu must contain at least
m 1-bits.
First, note that, according to Definition 1, u  v if and only if matrix Mvu satisfies
IOC. That is, either Mvu has no ð10Þ columns, or for each ð10Þ column in Mvu there exists
(at least) one corresponding preceding ð01Þ column. But, obviously, IOC can be
equivalently reformulated as follows: For each 1-bit in u there exists at least one
corresponding 1-bit in v, placed at the same or at a previous position. This reduces
the characterization of the elements v2Cu to the vectors of positions of 1-bits in u
and v and thus, according to the notation used in Definition 5-(i), we derive that v 2
Cu with wH ðvÞ ¼ m ¼ wH ðuÞ if and only if

1  j11  i11 ; j11 þ 1  j12  i12 ; . . . ; j1m1 þ 1  j1m  i1m ; (10)

but these are the sequences {j11,j21,. . .,jm1} generated by Step 1.

Second, note that the substitution of 0s by 1s in any n-tuple v such that u  v does
not avoid the IOC condition, because such a substitution changes the ð00Þ and ð10Þ
columns of matrix Mvu into ð01Þ and ð11Þ columns, respectively. Hence, to obtain all
the binary strings of the set Cu, it is enough to assign, in all possible ways, both
values 0, 1, to any of the njm1 (if jm1<n) null right-most components vj1m þ1 ; . . . ; vn
of all the binary strings v¼[j11,. . .,jm1]n1 with weight m, generated by (10). In other
words, we define v as described by (8) in Step 2.
1
Finally, since there are exactly 2njm different ways of assigning the values
vi¼0,1 for all jm1
<in, and since by this procedure we generate all elements of
C without repetitions, then the cardinality of Cu is given by (9).
u
□
406 L. González

Theorem 4. Let u 2 {0,1}n, u6¼2n  1, with Hamming weight wH(u) ¼ m (0  m <
0
n). Let u ¼ i10 ; . . . ; inm 0
n
be the vector of positions of 0s of u. Then Cu is the set of
binary n-tuples v generated by the following algorithm:
The Algorithm for Cu
(i) Step 1 (Generation of the sequences {j10,. . .,jnm0} :
For j10 ¼ 1 to i10 Do:
For j20 ¼ j10 þ 1 to i20 Do:
                                
For jnm 0
¼ jnm1
0
þ 1 to inm 0
Do:
Write {j1 ,j2 ,. . .,jnm }
0 0 0

EndDo
                                
EndDo
EndDo
(ii) Step 2 (Generation of the n-tuples v 2 Cu):
For every sequence {j10,. . .,jnm0} generated by Step 1, define the binary n-tuple
v ¼ (v1,. . .,vn ) as follows
8
< 0 if i 2 j1 ; . . . ; jnm ;
0 0
>
vi ¼ 1 if i= 2 j01 ; . . . ; j0nm ; i<j0nm ; (11)
>
:
2 j01 ; . . . ; j0nm ; i>j0nm :
0; 1 if i=

Moreover, the cardinality of the set Cu is given by the multiple sum

X
0
X
0
X
i1 i2 0
im
0
jCu j ¼  2njnm : (12)
j01 ¼1 j02 ¼j01 þ1 j0nm ¼j0nm1 þ1

Proof. For proving this theorem, it is enough to use Theorems 2 and 3. More
precisely,
 uc c due to the duality property stated by Theorem 2, we get that v 2 Cu iff v 2
c
C iff vc 2 Cu . Then the proof is concluded using the obvious fact that the 0-bits
and 1-bits in u and v become the 1-bits and 0-bits, respectively, in uc and vc. □
Remark 5. Note the strong duality relation between Theorems 3 and 4. The
statement of each theorem is exactly the statement of the other one after inter-
changing 1s and 0s, Cu and Cu. Indeed, due to Theorem 2,
 one can determine the set
c c
Cu (Cu, respectively) by determining the set ½Cuc c ( Cu , respectively) using
Theorem 4 (Theorem 3, respectively).

4 Ranking Intervals

For each given binary n-tuple u 2 {0,1}n, the next theorem provides us with the
closed interval covering all possible values of the rank (position) of u in the list of
all binary n-tuples arranged by decreasing order of their occurrence probabilities.
31 Ranking Intervals in Complex Stochastic Boolean Systems 407

We call this interval “the ranking interval of u”. The following is the main result of
this paper.
Theorem 5 (The ranking interval). Let n  1 and u 2 {0,1}n. Then the ranking
interval of u is

½ jCu j ; 2n þ 1  jCu j ; (13)

where |Cu| and |Cu| are respectively given by Eqs. 12 and 9. Moreover, the length of
the ranking interval of u coincides with the number of binary n-tuples incomparable
by intrinsic order with u, i.e.,

j IncðuÞj ¼ lð½ jCu j ; 2n þ 1  jCu j Þ ¼ 2n þ 1  jCu j  jCu j: (14)

Proof. On one hand, using Eq. 6 we have that u 2 Cu. Then, in the list of all
binary n-tuples arranged by decreasing order of their occurrence probabilities, at
least |Cu|1 binary strings are always below u. Thus, in that list, u is always among
the positions 1 (the first possible position) and 2n  ðjCu j  1Þ (the last possible
position), i.e.,

u 2 ½ 1 ; 2n þ 1  jCu j : (15)

On the other hand, using again Eq. 6 we have that u2Cu. Then, in the list of all
binary n-tuples arranged by decreasing order of their occurrence probabilities, at
least |Cu|1 binary strings are always above u. Thus, in that list, u is always among
the positions |Cu| (the first possible position) and 2n (the last possible position), i.e.,

u 2 ½ jCu j ; 2n : (16)

Hence, from Eqs. 15 and 16, we get

u 2 ½ 1 ; 2n þ 1  jCu j \ ½ jCu j ; 2n : (17)

Now, we prove that

1  jCu j  2n þ 1  jCu j  2n : (18)

First, the left-most inequality in Eq. 18, i.e., 1  |Cu| is obvious taking into
account that u 2 Cu (see Eq. 6).
Second, the right-most inequality in Eq. 18, i.e., 2n þ 1  jCu j  2n , that is, 1  |
C | is obvious taking into account that u 2 Cu (see Eq. 6).
u

Third, the central inequality in Eq. 18, i.e., jCu j  2n þ 1  jCu j, that is,
jCu j þ jCu j  2n þ 1 immediately follows from Eq. (6). That is, using the inclu-
sion-exclusion formula, we have

jCu j þ jCu j ¼ jCu [ Cu j þ jCu \ Cu j  jf0; 1gn j þ jfugj ¼ 2n þ 1:

408 L. González

Hence, from Eqs. 17 and 18, we get

u 2 ½ 1 ; 2n þ 1  jCu j \ ½ jCu j ; 2n ¼ ½ jCu j ; 2n þ 1  jCu j :

Finally, the last assertion of the theorem is an immediate consequence of Eq. 5

and the inclusion-exclusion formula

j IncðuÞj ¼ jf0; 1gn  ðCu [ Cu Þj ¼ jf0; 1gn j  jCu [ Cu Þj

¼ 2n  ðjCu j þ jCu j  jCu \ Cu jÞ ¼ 2n þ 1  jCu j  jCu j;

as was to be shown. □
The following example illustrates Theorems 3, 4 and 5.
Example 7. Let n¼5 and u ¼ 21  ð1; 0; 1; 0; 1Þ. Then, we have m ¼ wH ðuÞ ¼ 3,
n  m ¼ 2, and
 1  0
u ¼ i11 ; i12 ; i13 5
¼ ½1; 3; 5 15 ; u ¼ i01 ; i02 5
¼ ½2; 4 05 :

Using Theorems 3 and 4 we can generate all the 5-tuples v2Cu and all the 5-tuples
v2Cu, respectively. Results are depicted in Tables 1 and 2, respectively. The first
column of Tables 1 and 2, show the auxiliary sequences generated by Step 1 in
Theorems 3 and 4, respectively. The second column of Tables 1 and 2, show all the
binary 5-tuples v2Cu and v2Cu generated by Step 2 in Theorems 3 and 4, respec-
tively. The symbol “∗” means that we must choose both binary digits: 0 and 1, as
described by the last line of (8) and (11). According to (9) and (12), the sums of all
quantities in the third column of Tables 1 and 2, give the number of elements of Cu

Table 1 The set C(1,0,1,0,1) {j11,j21,j31} v∈Cu 1 1

2njm ¼ 25j3
and its cardinality
{1,2,3} (1,1,1,∗,∗)28,29,30,31 253 ¼ 4
{1,2,4} (1,1,0,1,∗)26,27 254 ¼ 2
{1,2,5} (1,1,0,0,1)25 255 ¼ 1
{1,3,4} (1,0,1,1,∗)22,23 254 ¼ 2
{1,3,5} (1,0,1,0,1)21 255 ¼ 1
P 1
jCu j ¼ 25j3 ¼ 10

Table 2 The set C(1,0,1,0,1) {j10,j20} 0 0

v∈Cu 2njnm ¼ 25j2
and its cardinality
{1,2} (0,0,∗,∗,∗)0,1,2,3,4,5,6,7 252 ¼ 8
{1,3} (0,1,0,∗∗)8,9,10,11 253 ¼ 4
{1,4} (0,1,1,0,∗)12,13 254 ¼ 2
{2,3} (1,0,0,∗∗)16,17,18,19 253 ¼ 4
{2,4} (1,0,1,0,∗)20,21 254 ¼ 2
P 0
jCu j ¼ 25j2 ¼ 20
31 Ranking Intervals in Complex Stochastic Boolean Systems 409

and Cu, respectively, using just the first columns, so that we do not need to obtain
these elements by the second columns.
So, on one hand, regarding the bitstrings intrinsically comparable with u, from
Tables 1 and 2 we have

Cu ¼ f21; 22; 23; 25; 26; 27; 28; 29; 30; 31g; jCu j ¼ 10;

Cu ¼ f0; 1; 2; 3; 4; 5; 6; 7; 8; 9; 10; 11; 12; 13; 16; 17; 18; 19; 20; 21g; jCu j ¼ 20

and then, using Eq. 13, the ranking interval of u¼21 is

½ jCu j ; 2n þ 1  jCu j ¼ ½ 20 ; 23 :

On the other hand, regarding the bitstrings intrinsically incomparable with u,

using Eq. 5, we have

IncðuÞ ¼ f0; 1g5  ðCu [ Cu Þ ¼ f14; 15; 24g;

so that Inc(u) is a 3-set, in accordance with Eq. 14, i.e.,

j IncðuÞj ¼ lð½ jCu j ; 2n þ 1  jCu j Þ ¼ lð½ 20 ; 23 Þ ¼ 3:

5 Conclusions

For any fixed binary n-tuple u, two dual algorithms for rapidly generating the set Cu
(Cu, respectively) of binary n-tuples that are always (i.e., intrinsically) less (more,
respectively) probable than u have been presented. These algorithms, exclusively
based on the positions of the 1-bits (0-bits, respectively) of u, also allow one to
count the elements of Cu (Cu, respectively) with no need to previously obtain them.
From the cardinalities of the sets Cu and Cu, we have exactly obtained the ranking
interval for the binary n-tuple u, i.e., the interval of all possible positions (ranks) of
u in the list of all the 2n binary n-tuples arranged by decreasing order of their
occurrence probabilities. By the way, the length of the ranking interval for u, gives
us the exact number of binary n-tuples v intrinsically incomparable with u. Appli-
cations can be found in many different scientific or engineering areas modeled by a
CSBS.

Acknowledgements This work was partially supported by Ministerio de Ciencia e Innovación

(Spain) and FEDER through Grant contract: CGL2008-06003-C03-01/CLI.
410 L. González

References

1. L. González, A new method for ordering binary states probabilities in reliability and risk
analysis. Lect. Notes Comp. Sci. 2329, 137–146, (2002)
2. L. González, N-tuples of 0s and 1s: necessary and sufficient conditions for intrinsic order. Lect.
Notes Comp. Sci. 2667, 937–946 (2003)
3. L. González, A picture for complex stochastic boolean systems: the intrinsic order graph. Lect.
Notes Comput. Sci. 3993, 305–312(2006)
4. L. González, Algorithm comparing binary string probabilities in complex stochastic Boolean
systems using intrinsic order graph. Adv. Complex Syst. 10(Suppl. 1), 111–143 (2007)
5. L. González, D. Garcı́a, B. Galván. An intrinsic order criterion to evaluate large, complex fault
trees. IEEE Trans. Reliab. 53, 297–305 (2004)
6. W.G. Schneeweiss, Boolean Functions with Engineering Applications and Computer Pro-
grams (Springer, New York, 1989)
7. R.P. Stanley, Enumerative Combinatorics, vol. 1 (Cambridge University Press, Cambridge,
1997)
Chapter 32
Predicting Memory Phases

Michael Zwick

Abstract The recent years, phase classification has frequently been discussed as a
method to guide scheduling, compiler optimizations and program simulations. In this
chapter, I introduce a new classification method called Setvectors. I show that the new
method outperforms classification accuracy of state-of-the-art methods by approxi-
mately 6–25%, while it has about the same computational complexity as the fastest
known methods. Additionally, I introduce a new method called PoEC (Percentage of
Equal Clustering) to objectively compare phase classification techniques.

1 Introduction

It is well known that the execution of computer programs shows cyclic, reoccuring
behavior over time. In one time interval, a program may get stuck waiting for I/O,
in another time interval, it may likely stall on branch misprediction or wait for the
ALU (Arithmetic Logic Unit) to complete its calculations. These intervals of specific
behavior are called phases. Phase classification is the method that analyzes programs
and groups program intervals of similar behavior to equal phases [5, 7, 8, 12].
To identify phases, programs are split into intervals of a fixed amount of
instructions. Then, these intervals are analyzed by some method to predict their
behavior. Intervals showing similar behavior are grouped together to form a specific
class of a phase. Therefore, a phase is a set of intervals that show similar behavior
while discarding temporal adjacency. The phase information of a program can be
used to guide scheduling, compiler optimizations, program simulations, etc.
Several phase classification techniques have been proposed the recent years,
many of which rely on code metrics such as Basic Block Vectors [12] and Dynamic

M. Zwick
Lehrstuhl f€ur Datenverarbeitung, Technische Universit€at M€unchen, Arcisstr. 21, 80333 M€
unchen,
Germany
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 411

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_32, # Springer ScienceþBusiness Media B.V. 2010
412 M. Zwick

Branch Counts [3]. Since code related methods only have low correlation with
memory hierarchy behavior, several memory related phase classification methods
have been proposed to predict L1 and L2 processor cache misses, such as wavelet
based phase classification [5], Activity Vectors [11] and Stack Reuse Distances [4].
In this chapter, I introduce a new phase classification technique called Setvector
method to predict L2 cache performance. On the basis of ten SPEC2006 bench-
marks, I show that the mean accuracy of the Setvector method outperforms the
Activity Vector method by about 6%, the Stack Reuse Distance method by about
18% and the wavelet based method by about 25%. Further, I introduce a new metric
called PoEC (Percentage of Equal Clustering) to objectively evaluate different
phase classification methods and make them comparable to one another.
The remainder of this chapter is organized as follows: Section 2 describes state-
of-the-art phase classification techniques, Section 3 presents the Setvector method,
Section 4 introduces PoEC as a methodology to evaluate phase classification
accuracy, Section 5 presents the results and Section 6 concludes this chapter.

2 Phase Classification Techniques

In this section, I describe state-of-the-art techniques I compare the Setvector

method to. As I focus on phase classification for L2 cache performance prediction,
I exclusively consider memory based techniques.
All the methods I apply were evaluated using the same tracefiles comprised of
memory references that were gathered using the Pin tool described in [9], as it has
been done in [5]. For cross validation, I used the MCCCSim (Multi Core Cache
Contention Simulator) simulator described in [14] to obtain the hitrates for each
interval. As I initially started the evaluation of the Setvector method in comparison
to the wavelet technique presented in [5], I chose the same interval size of 1 million
instructions as they did to make the results better comparable. All methods were
evaluated on ten SPEC2006 test benchmarks, each comprised of 512 million
instructions.

2.1 Wavelet Based Phase Classification

Huffmire and Sherwood [5] use Haar wavelets [13] to perform phase classification.
First, they create 16  16 matrices for each interval of 1 million instructions.
Therefore, they split each set of 106 instructions into 20 subsets of 106 =20 ¼ 50; 000
instructions forming 20 column vectors. They determine the elements of every such
column vector by calculating m ¼ ððaddress%MÞ  ð400=MÞÞ for each address in
the corresponding subset, where ‘%’ is the modulo operator and M ¼ 16k the modulo
size that has been matched to the L1 cache size. By iterating over each address of the
50k instructions, Huffmire and Sherwood fill up the rows of the column vectors
by summing up the occurences of each m (0m<400) in a histogram manner.
After having calculated each of the 20 400  1 column vectors, they scale the
resulting 400  20 matrix to a 16  16 matrix using methods from image scaling.
32 Predicting Memory Phases 413

In a second step, they calculate the Haar wavelet transform [13] for each 16  16
matrix and weight the coefficients according to [6].
In a third step, Huffmire and Sherwood apply the k-means clustering algorithm
[10] on the scaled wavelet coefficients and compare the clusters with the hitrates of
the corresponding intervals, gained by a cache simulator.
In my implementation of the wavelet technique, I followed Huffmire’s and
Sherwood’s description except for the following: I split each 106 instructions in
16 intervals of 64k instructions each to omit the scaling from 20 to 16 columns.
Additionally, I utilized the MCCCSim Simulator [14] that is based on Huffmire’s
and Sherwood’s cache simulator anyway. I implemented everything else as it as
been presented by [5].
Since the top-left element of the coefficient matrix corresponds to the mean
value of the original matrix, I also clustered all top-left elements of the coefficient
matrices and compared the results of the clustering process to the L2 hitrates of the
corresponding intervals.
As the wavelet transform seems not an obvious choice of algorithm for this
problem, yet it achieved good results shown by Huffmire and Sherwood, I decided
to replace the wavelet transformation by a SVD (Singular Value Decomposition)
M ¼ USVT in another experiment to probe if a more general method could find
more information in the data matrix. I clustered both columns and rows of U and V
respectively but could not find any impressive results, as I will show in Section 5.

2.2 Activity Vectors

Settle et al. [11] propose to use Activity Vectors for enhanced SMT (simultaneous
multi-threaded) job scheduling. The Activity Vectors are used like a phase classifi-
cation technique to classify program behavior with respect to memory hierarchy
performance. The Activity Vector method has been proposed as an online classifi-
cation method that relies on a set of on-chip event counters that count memory
accesses to so-called Super Sets. Note that Settle et al. include both access count
and cache miss information in their Activity Vectors. In my evaluation however,
I exclusively incorporate access count information.
I implemented the Activity Vector method in software and applied the same
tracefile data that I applied to all the other simulations mentioned in this chapter. To
use the Activity Vector method as a phase classification method, I clustered both the
vectors and the length of each vector and compared the clustering results with the
L2 cache performance of the corresponding intervals.
In Section 5, I show that the Activity Vector method on average achieves better
results than the wavelet based method.

2.3 Stack Reuse Distances

Beyls and D’Hollander [1] propose to use the so-called Stack Reuse Distance as a
metric for cache behavior. They define the Stack Reuse Distance of a memory
414 M. Zwick

access as the number of unique memory addresses that have been referenced
since the last access to the requested data. In Section 5, I show that on average,
the classification performance of the Stack Reuse Distance method lies between the
wavelet and the Activity Vector method, whereas its calculation takes a huge
amount of time.

2.4 Other Techniques

Although many other techniques for phase classification have been proposed such
as Basic Block Vectors [12], Local/Global Stride [8], Working Set Signatures [2]
etc., I omitted to compare the Setvector technique to those methods as it has been
shown that they are outperformed by other methods, for example the wavelet based
method [5].

3 Setvector Based Phase Classification

In this section, I describe the Setvector based phase classification method. Setvec-
tors are as easily derived as they are effective: For all addresses of an interval and
an n-way set-associative cache, determine the number of addresses with different
key that are mapped to the same cache set. That means: Given a L2 cache with 32 bit
address length that uses b bits to code the byte selection, s bits to code the selection
of the cache set and k ¼ 32  s  b bits to code the key that has to be compared to
the tags stored in the tag RAM, do the following:
l Extract the set number from the address, e.g. by shifting the address k bits to the
left and then unsigned-shifting the result kþb bits to the right.
l Extract the key from the address, e.g. by unsigned-shifting the address sþb bits
to the right.
l In the list for the given set, determine whether the given key is already present.
l If the key is already present, do nothing and process the next address.
l If the key is not in the list yet, add the key and increase the counter that
corresponds to that set.
More formally written: Starting with a tracefile
T ¼ fai j1  i  tg (1)

that contains t memory addresses ai, 1it that are grouped into intervals of a fixed
amount of instructions, split the tracefile into a set of access vectors ai, each
representing an interval of several ai:
T ¼ ½a1    ac  (2)
32 Predicting Memory Phases 415

Now, for each ai and caches with a row size of r byte and a way size of w byte,
derive the set vector
si ¼ ½s1    ss T (3)

by


a ða%wÞ  ðða%wÞ%rÞ a
Si ja 2 ai ; ¼ i; 2= Si (4)
r r r

X
si ¼ minð2n  1; 1Þ (5)
a=r2Si
with ‘%’ representing the modulo operator and n the maximum number of bits
allowed to code the set vector elements, if there should be such a size constraint.
This way, each element of a set vector contains for each corresponding interval the
number of addresses that belong to the same cache set, but have a different cache
key – saturated by n, the number of bits at most to be spent for each vector element.
Due to this composition, the Setvectors directly represent cache set saturation, a
measure that is highly correlated with cache misses.
In Section 5, I show that on average, the Setvector method outperforms all
methods mentioned in Section 2.

4 Metrics to Compare Phase Classification Techniques

Lau et al. [8] define the Coefficient of Variation (CoV) as a metric to measure the
effectiveness of phase classification techniques. CoV measures the standard devia-
tion as percentage of the average and can be calculated by
si
X
phases
averagei  intervalsi
CoV ¼ : (6)
i¼1
total intervals
Huffmire and Sherwood adapt this metric by omitting the division by the
average, resulting in the weighted standard deviation

X
phases
si  intervalsi
sweighted ¼ : (7)
i¼1
total intervals

Being derived from standard deviation, both CoV and sweighted denote better
clustering performance by smaller values. However, the CoV metric (as well as
sweighted) may describe the standard deviation of the L2 cache performance in each
phase, but not the correlation between L2 cache performance and the different
phases, what should be the key evaluation metric for phase classification. There-
fore, I developed the PoEC (Percentage of Equal Clustering) metric that can be
calculated as follows:
416 M. Zwick

Consider the cluster vector g as a vector that holds, for each index, the phase of
the corresponding 16 16 scaled matrix. In a first step, sort the elements of the
cluster vector g according to its phases, such that 8i21::indeces : gi  giþ1 . Then, in a
second step, calculate the percentage of equal clustering by
  Pindeces  
ðgh;i ¼¼ gx;i Þ
PoEC ¼ 2  min i¼1
; 0:5  0:5 (8)
indeces

This way, high correlation between L2 cache performance and the cluster
vectors result in PoEC values near 1 and low correlation corresponds to values
near 0, with 0PoEC1.
Figures 1 (CoV) and 2 (PoEC) show the difference between those metrics by
clustering some SPEC2006 benchmarks in two phases (“good ones” and “bad ones”)
using the wavelet method and plotting the phases (each ring corresponds to one
interval) against the L2 hitrate of the corresponding intervals. As L2 cache hitrates
of the same magnitude should be classified into the same class, a good clustering
is achieved if one class contains higher hitrates and the other contains lower
hitrates, as it is the case for the “milc”, “soplex”, “lbm” and “bzip2” benchmarks.
Figure 1 shows the clustering of the SPEC2006 benchmark programs sorted by
their CoV value (calculated according to formula 6). While analyzing Figure 1, one
can observe the following: There are benchmarks that achieve good clustering, such
as “soplex”, “milc”, “bizp2” and “lbm”. And there are benchmarks that do not
cluster well at all, such as “hmmer”, “libquantum”, “gobmk”. But the point is: The
clustering quality does not fit the CoV value the plots are arranged by. Although not
plotted in this place, the sweighted metric shows similar behavior.
In Figure 2, the programs are sorted according to their PoEC value (calculated
according to formula 8). Although the clustering is the same, this time the clustering
quality does fit the PoEC value the plots are arranged by.
As the PoEC metric is obviously more applicable to evaluate clustering perfor-
mance, I decided to omit both sweighted and CoV and to perform the evaluation of
the phase classification techniques using the PoEC metric.

5 Results

In this section, I present the results gathered from my simulations.

5.1 Classification Accuracy

Figure 3 shows the PoEC values for each of the mentioned methods for ten
SPEC2006 benchmarks. For the Activity Vectors, I clustered the vector as well
as the magnitude of the vector |Activity Vector|. For the Haar wavelet method, I
32 Predicting Memory Phases 417

hmmer, wavelet soplex, wavelet povray, wavelet

CoV = 0.02 CoV = 0.03 CoV = 0.04
0.95
0.96 0.95
L2 hitrate

0.90
0.94
0.85
0.85
0.92

0.90 0.75 0.80

1 2 1 2 1 2

libquantum, wavelet gcc, wavelet milc, wavelet

CoV = 0.05 CoV = 0.07 CoV = 0.09
0.95 0.80
L2 hitrate

0.85 0.80
0.60
0.75 0.60 0.40
0.65
0.40 0.20
1 2 1 2 1 2

specrand, wavelet astar, wavelet bzip2, wavelet

CoV = 0.10 CoV = 0.13 CoV = 0.19
1.00
0.90
0.85 0.80
0.80
L2 hitrate

0.60
0.70 0.75
0.40
0.60 0.65
0.20
0.50
0.55 0.00
1 2 1 2 1 2

mcf, wavelet gobmk, wavelet lbm, wavelet

CoV = 0.21 CoV = 0.52 CoV = 1.38

0.60 0.60
L2 hitrate

0.02
0.50 0.40
0.40 0.01
0.20
0.30
0.00 0.00
1 2 1 2 1 2
Cluster Cluster Cluster

Fig. 1 CoV metric

clustered both the scaled matrix and the left-top matrix element (Haar wavelet[0]
[0]). For the Setvectors, I clustered the magnitude of the Setvectors; for the Stack
Reuse Distances, I clustered the Stack Distance Vector. The results shown for the
418 M. Zwick

libquantum, wavelet specrand, wavelet hmmer, wavelet

PoEC = 0.00 PoEC = 0.12 PoEC = 0.19
0.95 0.90
0.96
L2 hitrate

0.80
0.85
0.70 0.94
0.75 0.60
0.92
0.65 0.50
0.90
1 2 1 2 1 2

gobmk, wavelet povray, wavelet gcc, wavelet

PoEC = 0.19 PoEC = 0.22 PoEC = 0.41
0.95
0.60
L2 hitrate

0.90 0.80
0.40

0.20 0.85 0.60

0.00 0.80 0.40

1 2 1 2 1 2

astar, wavelet mcf, wavelet milc, wavelet

PoEC = 0.44 PoEC = 0.56 PoEC = 0.94
0.80
0.85 0.60
L2 hitrate

0.50 0.60
0.75
0.40
0.65 0.40
0.30
0.55 0.20
1 2 1 2 1 2

spolex, wavelet lbm, wavelet bzip2 wavelet

PoEC = 0.97 PoEC = 0.97 PoEC = 1.00
1.00
0.95 0.80
L2 hitrate

0.02
0.60
0.85 0.40
0.01
0.20
0.75 0.00 0.00
1 2 1 2 1 2
Cluster Cluster Cluster

Fig. 2 PoEC metric

SVD originate from the column space of U, that has been calculated from the scaled
matrices described above. Column/Rowspace of V didn’t achieve any better results.
PoEC values near 1 indicate good classification performance, PoEC values near
32 Predicting Memory Phases 419

0 indicate poor classification performance. The benchmark mcf for example shows
superior performance for the Activity Vector method, the wavelet method and the
Setvector approach and poor classification performance for the Stack Distance and
SVD method.
Figure 4 depicts the mean classification performance of each method averaged
over the ten SPEC2006 benchmarks. The Setvector approach outperforms all other
methods and achieves about 6% better classification performance than the next best
method, the Activity Vectors method, while the Stack Reuse Distance method shows
about 18% inferior performance than the Setvector method. While the wavelet
based method still shows fair results (about 75% of the performance of the Setvec-
tor method), the Singular Value Decomposition of the scaled matrix has apparantely
not been a good idea at all.

1.0

0.8

0.6
PoEC

0.4

0.2

0.0
r

er

cf

ilc

ay
re
ta

gc

bm
ip

m
lb
m

vr
as

64
bz

hm
go

po
h2

Activity Vectors |Activity Vectors| Haar wavelet Haar wavelet[0][0]

|Setvectors| Stack Reuse Dist. SVD

Fig. 3 Accuracy of phase classification techniques measured by the PoEC metric

Activity Vectors 94%

|Activity Vectors| 86%

Haar wavelet 75%

Haar wavelet [0][0] 73%

|Setvectors| 100%

Stack Reuse Dist. 82%

SVD 9%

0.0 0.2 0.4 0.6 PoEC

Fig. 4 Mean PoEC classification performance

420 M. Zwick

5.2 Computational Performance

The computational performance of the mentioned methods cannot easily be com-

pared in this chapter, because the methods have been implemented in different
programming languages. The Setvector and Activity Vector methods have com-
pletely been implemented in slow Ruby 1.8.6, the wavelet and SVD methods have
been implemented in a mixture of Ruby and R and the Stack Reuse Distances
method has completely been implemented in C for performance reasons.
The calculation of the scaled matrices used by the wavelet method has been
timed for 14 ms per simulated instruction. The calculation of wavelet coefficients in
Ruby was about 15 ns, the calculation of the k-means algorithm in R was about
2.5 ns, including data exchange over the Ruby-R-bridge in both cases.
The Setvector method suffered – as the wavelet method did – from the slow
Ruby interpreter and took about 3.97 ms per instruction. The magnitude calculation
and the clustering algorithm for the Setvectors finished in less than 0.5 ns. The
better clustering performance originates from using scalars (vector magnitues)
rather than vectors of dimensionality 16 16 ¼ 256, as it has been done in the
wavelet technique.
The by far worst computational performance was demonstrated by the Stack
Reuse Distance method, that also took about 3.97 ms of time per simulated instruc-
tion, but this time the method was implemented in C-code to finish the simulation in
just some days instead of months.
Comparing the Ruby wavelet implementation (about 14.0175 ms per simulated
instruction) to the execution time measued by [5] (about 10 ns per simulated
instruction) allows the Setvector method to be estimated for about 5–10 ns per
instruction when implemented on a C code basis; i.e. the Setvector method has
about the same magnitude of calculation complexity as the other mentioned meth-
ods, except the Stack Reuse Distance method. All values have been averaged over
several executions or files.

6 Conclusion

Within this chapter, I introduced a new method for phase classification called
Setvectors. The method is similar to the Activity Vector method proposed in [11],
but it differs in the way the vectors are obtained. While Settle et al. just count
accesses to Super Sets, the Setvector method calculates the number of accesses to
a set that reference a different key. I showed that the proposed Setvector method
outperforms state-of-the-art methods with respect to classification accuracy by
approximately 6–25%, having about the same computational complexity. As a
second contribution, I introduced the PoEC metric that can be used to objectively
evaluate phase classification methods in a more intuitive way than the known
metrics CoV and sweighted. Although I proved the better performance of the PoEC
32 Predicting Memory Phases 421

method compared to the CoV/sweighted method just qualitatively “by inspection”,

it obviously is a more reasonable approach.

References

1. K. Beyls, E.H. D’Hollander, Reuse distance as a metric for cache behavior, in Proceedings of
the IASTED Conference on Parallel and Distributed Computing and Systems, 2001
2. A.S. Dhodapkar, J.E. Smith, Managing multi-configuration hardware via dynamic working set
analysis, in International Symposium on Computer Architecture (ISCA’02), 2002
3. E. Duesterwald, C. Cascaval, S. Dwarkadas, Characterizing and predicting program behavior
and its variability, in 12th International Conference on Parallel Architectures and Compila-
tion Techniques (PACT’03), 2003
4. C. Fang, S. Carr, S. Önder, Z. Wang, Reuse-distance-based miss-rate prediction on a per
instruction basis, in Proceedings of the 2004 Workshop on Memory System Performance, 2004
5. T. Huffmire, T. Sherwood, Wavelet-based phase classification, in Parallel Architectures and
Compilation Techniques (PACT’06), 2006
6. C. Jacobs, A. Finkelstein, D. Salesin, Fast multiresolution image querying, in Proceedings of
the 22nd Annual Conference on Computer Graphics and Interactive Techniques, 1995
7. J. Lau, J. Sampson, E. Perelman, G. Hamerly, B. Calder, The strong correlation between code
signatures and performance. International Symposium on Performance Analysis of Systems
and Software, 2005
8. J. Lau, S. Schoenmackers, B. Calder, Structures for phase classification. International Sym-
posium on Performance Analysis of Systems and Software, 2004
9. C.-K. Luk, R. Cohn, R. Muth, H. Patil, A. Klausner, G. Lowney, S. Wallace, V.J. Reddi,
K. Hazelwood, Pin: building customized program analysis tools with dynamic instrumenta-
tion. Programming Language Design and Implementation, 2005
10. J. MacQueen, Some methods for classification and analysis of multivariate obervations, in
Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability,
vol. 1, pp. 281–297, 1967
11. Settle, A., Kihm, J.L., Janiszewski, A., and Connors, D. A. (2004). Architectural Support for
Enhanced SMT Job Scheduling, in Proceedings of the 13th International Conference of
Parallel Architectures and Compilation Techniques
12. T. Sherwood, E. Perelman, G. Hamerly, B. Calder, Automatically characterizing large scale
program behavior, in ASPLOS-X: Proceedings of the 10th International Conference on
Architectural Support for Programming Languages and Operating Systems, 2002
13. E. Stollnitz, T. DeRose, D. Salesin, Wavelets for computer graphics: a primer. IEEE Com-
puter Graphics and Applications, 1995
14. M. Zwick, M. Durkovic, F. Obermeier, W. Bamberger, K. Diepold, MCCCSim – a highly
configurable multi core cache contention simulator. Technical Report – Technische
Universit€ at M€
unchen, https://mediatum2.ub.tum.de/doc/802638/802638.pdf, 2009
Chapter 33
Information Security Enhancement to
Public–Key Cryptosystem Through Magic
Squares

Gopinath Ganapathy and K. Mani

Abstract The efficiency of a cryptographic algorithm is based on its time taken

for encryption and decryption and the way it produces different cipher text from
a plain text. The RSA, the widely used public key algorithm and other public
key algorithms may not guarantee that the cipher text is fully secured. As an
alternative approach to handling ASCII value of the text in the cryptosystem, a
magic square of order 16 equivalent to ASCII set generated from two different
approaches is thought of in this work. It attempts to enhance the efficiency by
providing add-on security to the cryptosystem. Here, encryption/decryption is
based on numerals generated by magic square rather than ASCII values. This
proposed work provides another layer of security to any public key algorithms.
Since, this model is acting as a wrapper to a public-key algorithm, it ensures that
the security is enhanced. Further, this approach is experimented in a simulated
environment with 2, 4, and 8 processor model using Maui scheduler which is
based on back filling philosophy.

1 Introduction

Today security is an important thing that we need to send data from one location to
another safely. As there is strong security, there is a great hacker and spire at the
other side. Therefore, many models and systems are looking for optimization of the
electronic connected-world. Cryptography accepts the challenges and plays the
vital role of the modern secure communication world. It is the study of mathemati-
cal techniques related to aspects of information security such as confidentiality,

G. Ganapathy (*)
Department of Computer Science, University of Bharthidasan, 620 024, India
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 423

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_33, # Springer ScienceþBusiness Media B.V. 2010
424 G. Ganapathy and K. Mani

data integrity, entity authentication, and data origin authentication. Cryptographic

algorithms are divided into public-key and secret-key algorithms. In public-key
algorithms both public and private keys are used, with the private key computed
from the public key. Secret-key algorithms rely on secure distribution and manage-
ment of the session key, which is used for encrypting and decrypting all messages.
Though, public-key encryption is slower than symmetric-key encryption, it is
used for bulk-data encryption. Hence, in practice cryptosystems are a mixture of
both [1, 2].
There are two basic approaches used to speed up the cryptographic transforma-
tions. The first approach is to design faster (symmetric or asymmetric) crypto-
graphic algorithms. This approach is not available most of the time. The speed
of cryptographic algorithm is typically determined by the number of rounds (in
private-key) or by the size of messages (in public-key case). The second approach is
the parallel cryptographic system. The main idea is to take a large message block,
divide it into blocks of equal sizes and each block can be assigned to one processor
[3]. To perform the operations in parallel, effective scheduling is very important so
that it can reduce the waiting time of message for processing. The back filling is
one such approach.
The security of many cryptographic systems depends upon the generation of
unpredictable components such as the key stream in the one-time pad, the secret
key in the DES algorithms, the prime p, and q in the RSA encryption etc. In all these
cases, the quantities generated must be sufficient in size and the random in the sense
that the probability of any particular value being selected must be sufficiently
small. However, RSA is not semantically secure or secure against chosen cipher
text attacks even if all parameters are chosen in such a way that it is infeasible to
compute the secret key d from the public key (n, e), choosing p, q are very large etc.
Even if the above said parameters are taken carefully, none of the computational
problems are fully secured enough [4]. Because to encrypt the plaintext characters,
their ASCII values are taken and if a character occurs in several places in a plaintext
there is a possibility of same the cipher text is produced. To overcome the problem,
this paper attempts to develop a method with different doubly even magic squares of
order 16 by two different approaches and each magic square is considered as one
ASCII table. Thus, instead of taking ASCII values for the characters to encrypt,
preferably different numerals representing the position of ASCII values are taken
from magic square and encryption is performed using RSA cryptosystem.

2 Methodology

The working methodology of the proposed add-on security model is discussed

stepwise.
l Construct different doubly even magic square of order 16 as far as possible and
each magic square corresponds to one ASCII set.
33 Information Security Enhancement to Public–Key Cryptosystem 425

l To encrypt the character, use the ASCII value of the character to determine the
numeral in the magic squareby considering the position in it. Let NP and NC
denote the numeral of the plaintext and cipher text respectively. Based on NPand
NCvalues, all plaintext and cipher text characters are encrypted and decrypted
respectively using RSA algorithm.
l To speed up the operations, perform them in parallel in a simulated environment.
For that, use Maui scheduler with back filling philosophy.
The methodology of Add-on Security Model is shown in Fig. 1

2.1 Magic Squares and Their Construction

Definition 1: A magic square of order n is an arrangement of integers in an n * n

magic square such that the sums of all the elements in every row, column and along
the two main diagonals are equal.
A normal magic square contains the integers from 1 to n2. It exists for all orders
n  1 except n ¼ 2, although the case n ¼ 1 is trivial as it consists of a single cell
containing the number. The constant sum in every row, column and diagonal is

a
Clear Text
ASCII
value of the Clear Text

Determine the
Starting No., Generate
numeral value from Numeral
Seed No., Doubly Even the Magic Square Encryption
Cipher Text
and Magic that corresponds to Algorithm
Magic Square the position of
Square Sum ASCII value

b
Determine
the position ASCII
Decryption Numeral of the value of
Cipher Text the Clear Text
Algorithm numeral
value in the Clear
Magic text
Square

Fig. 1 Add-on wrapper Model (a) encryption process (b) decryption process
426 G. Ganapathy and K. Mani

called the magic constant or magic sum, M [5]. The magic constant of a normal
magic square depends only on n and has the value

nðn2 þ 1Þ
MðnÞ ¼ (1)
2

For normal magic squares of order n ¼ 3, 4, 5, 6, 7, 8, . . . the magic constants are

15, 34, 65, 111, 175, 265, . . . respectively. Though there is only one magic square of
order 3 apart from trivial notations and reflections, the number of squares per order
quickly sky rockets. There are 880 distinct order 4 squares, and 275, 305, 234
distinct order 5 magic squares [6]. Magic squares can be classified into three types:
odd, doubly even (n divisible by four) and singly even [7] (n even but not divisible
by four). This paper focuses only on doubly even magic square implementation and
their usefulness for public-key cryptosystem.
While constructing the doubly even magic squares of order 16 based on different
views and fundamental properties of magic square the following notations are used.
MS: Magic Square
n: Order of MS where n ¼ 4m, where
m ¼ 1, 2, 3 and 4
MSn: MS of order n
MSB4: Base MS of order 4
MSstart: Starting number of MS
MSTnsum: Total sum of MS of order n
MSDnsum: Diagonal sum of MS of order n
T_No_MS: Total Number of MS
S1 S2 S3 S4: First, Second, Third and Fourth digit of 4 digits seed number. Each
digit has a value from 0 to 7. If S1 ¼ S2 ¼ S3 ¼ S4 ¼ 0, i.e. 0000 then it is MSB4.
The values in the MSB4 are filled as shown in Table 1. Call it as MS4_fill_order
(Minstart, Maxstart)
where the numerals with positive sign represent value to be added with MSstart
and the numerals with negative sign represent value to be subtracted from
MST2sum.
MS4_min_start_value: Minimum starting value of MS
MS4_fill_order (MSstart, MST4sum): function used to fill the values in MS the
order is shown in Table 1.
MSn_base i: ith nth order base MS
MS4_sub i: ith 4th order MS obtained from MSB4 by using suitable transforma-
tion based on Si, where i¼ 1, 2, 3, and 4

Table 1 Magic square filling order

4 MSstart 8 þ12
10 þ14 6 þ2
þ8 12 þ4 MST2sum
þ6 2 þ10 14
33 Information Security Enhancement to Public–Key Cryptosystem 427

MS4_prop_sub i: ith 4th order property based MS obtained from MSB4 by using
the fundamental property of MS4magic square based on Spi, where pi¼ 1, 2, 3, and 4
Prop: properties of fundamentalMS4 with four character values
jj: concatenation

2.2 Construction of Doubly Even Magic Square based

on different Views of fundamental Magic Square

In this work, to generate the doubly even magic square, any seed number, starting
number, and magic sum may be used and the numbers generated will not be in
consecutive order. The algorithm 2.2.1 proposed by Gopinath Ganapathy, and Mani
[8] starts with building 4  4 magic square. Incrementally 8  8 and 16  16 magic
squares are built using 4  4 magic squares as building blocks.

2.2.1 Algorithm

Input: 4 digit Seed number, Starting Number and

Magic Square Sum
Output: Doubly Even Magic Square of order 16.
1. Read MSstart; Si, i ¼ 1, 2, 3, 4 ;MST16sum and T_No_MS
2. MST4sum← d MST16sum/ 4e; 3. MST8sum← dMST16sum/ 2e
4. Count ← 1
5. While Count  T_No_MS do // to generate MS16
5. a. prev← 1;
5. b. for j ← 1 to 4 // to generate four MS8
5.b.1 for i ← 1 to 4 // to generate four MS4
5.b.1.1 if ((j¼2)&&(i ¼ 3)) jj ((j ¼ 3) &&(i ¼2))
MS4_min_start_value
← MS4startþ 16 *(iprev-1) þ1
else
MS4_min_start_value
← MSstartþ 16* (prev-1)
end if
5.b.1.2 MS4_max_start_value
← MST4sum– 16* (prev- 1) - MSstart
5.b.1.3 MS4_base i ← call MS4_ fill_order
(MS4_min_start_value, MS4_max_ start_value)
5.b.1.4 Case Si in 0,1,2,3,4,5,6,7
0: MS4_sub i ← rotate MS4_base i by 0
1: MS4_sub i ← rotate MS4 _base i by 90
2: MS4_sub i ← rotate MS4_base i by 180
428 G. Ganapathy and K. Mani

3: MS4_sub i ← rotate MS4_base i by 270

4: MS4_sub i ← rotate MS4_base i by 180 about a
vertical axis
5: MS4_sub i ← rotate MS4_base i by 90 through
anticlock wise
6: MS4_sub i ← rotate MS4_base i by 180 about
horizontal axis
7: MS4_sub i ← rotate MS4_base i by 180 about the
main diagonal
end case Si
5.b.1.5 prev← prevþ1
end for i
5.b.2 MS8_sub j ←MS4_sub 1jj MS4_sub 2 jj MS4_sub 3 jj MS4_sub 4
5.b.3 case j in 2, 3, 4
2: S1← S2; S2← S3; S3← S4; S4← S1
3: S1← S3; S2← S4; S3← S1; S4← S2
4: S1← S4; S2← S1; S3← S2; S4← S3
end case j
5.b.4 Si← S1jj S2jj S3jj S4, i ← 1, 2, 3, 4
end for j
5.c MS16_count ← MS8_sub1jj MS8_sub2 jjMS8_sub3jjMS8_sub4
5.d Si←(Siþ rnd(10)) mod 8, i ¼ 1,2,3,4
5.e count ← count þ 1
end while count

2.2.2 Example

Based on the Algorithm 2.2.1 the first four MS4_subi, i ¼ 1, 2, 3 and 4 are generated
and they are shown in Tables 2 – 5.
Other magic squares MS8_sub2, MS8_sub3 and MS4_sub4 are generated in
this manner and they are concatenated so that they form MS16 which is shown
in Table 6.
Similarly other MS16 magic squares are generated by using suitable trans-
formations of the seed number. In this paper, only four MS16 magic squares are
generated.

2.3 Construction of Doubly Even Magic Square of Order 16

based on the Properties of 4  4 Magic Square

In this paper, Section 2.1focuses on construction of a 16  16 magic square from

sixteen 4  4 different views of fundamental magic squares. To obtain the
different views of magic squares, each one can be rotated or reflected so that it
can generate eight different views of a 4  4 fundamental magic square. Thus, any
33 Information Security Enhancement to Public–Key Cryptosystem 429

Table 2 MS4_sub1 with (4,1539)

1,535 4 1,531 16
1,529 18 1,533 6
12 1,527 8 1,539
10 1,537 14 1,525

Table 3 MS4_sub2 with (20, 1,523)

(a) MS4_sub2 with (20, 1,523)
1,519 20 1,515 32
1,513 34 1,517 22
28 1,511 24 1,523
26 1,521 30 1,509
(b) 90 Rotation of Table 3a
26 28 1,513 1,519
1,521 1,511 34 20
30 24 1,517 1,515
1,509 1,523 22 32

Table 4 MS4_sub3 with (36, 1,507)

1,503 36 1,499 48
1,497 50 1,501 38
44 1,495 40 1,507
42 1,505 46 1,493

Table 5 MS4_sub4 with (52, 1,491)

1,487 52 1,483 64
1,481 66 1,485 54
60 1,479 56 1,491
58 1,489 62 1,477

square is one of a set of eight that are equivalent under rotation or reflection. In
order to generate different views of a magic square, the general format shown in
Table 7 is used.

2.3.1 Sub-types of 4  4 Magic Square

l Normal
A magic square is said to be normal in which every row, column, and the main
and secondary diagonals sum to the magic sum. That is B þ C þ N þ O ¼ E þ H þ
I þ L ¼ F þ G þ J þ K ¼ A þ D þ M þ P ¼ MST4sum
 430

Table 6 First MS16 using starting no. 4, seed no. 0100, and magic sum 12,345
1535 4 1531 16 26 28 1513 1519 74 76 1465 1471 1455 85 1451 96
1529 18 1533 6 1521 1511 34 20 1473 1463 82 68 1449 98 1454 86
12 1527 8 1539 30 24 1517 1515 78 72 1469 1467 93 1447 88 1459
10 1537 14 1525 1509 1523 22 32 1461 1475 70 80 90 1457 94 1446
1503 36 1499 48 1487 52 1483 64 1439 101 1435 112 1423 116 1419 128
1497 50 1501 38 1481 66 1485 54 1433 114 1438 102 1417 130 1421 118
44 1495 40 1507 60 1479 56 1491 109 1431 104 1443 124 1415 120 1427
42 1505 46 1493 58 1489 62 1477 106 1441 110 1430 122 1425 126 1413
1407 132 1403 144 1391 149 1387 160 1343 196 1339 208 1327 212 1323 224
1401 146 1405 134 1385 162 1390 150 1337 210 1341 198 1321 226 1325 214
140 1399 136 1411 157 1383 152 1395 204 1335 200 1347 220 1319 216 1331
138 1409 142 1397 154 1393 158 1382 202 1345 206 1333 218 1329 222 1317
1375 165 1371 176 186 188 1353 1359 1311 228 1307 240 250 252 1289 1295
1369 178 1374 166 1361 1351 194 180 1305 242 1309 230 1297 1287 258 244
173 1367 168 1379 190 184 1357 1355 236 1303 232 1315 254 248 1293 1291
170 1377 174 1366 1349 1363 182 192 234 1313 238 1301 1285 1299 246 256
G. Ganapathy and K. Mani
33 Information Security Enhancement to Public–Key Cryptosystem 431

Table 7 General format of a fundamental 4  4 magic square

A B C D
E F G H
I J K L
M N O P

l Pan-Magic
It has the additional property that the broken diagonals also sum to the magic sum.
That is B þ G þ L þ M ¼ D þ E þ J þ O ¼ C þ F þ I þ P ¼ A þ H þ K þ N ¼ C þ
H þ I þ N ¼ B þ E þ L þ O ¼ MST4sum. The magic squares depicted shown in
Tables 2 through 5 are normal and Pan Magic.
l Associative
A magic square is an associative, it has the additional property that the eight
symmetrically opposite the centre of the square shown in Table 2 are AP, BO, CN,
DM, EL, FK, GJ and HI having the magic sum magic constant MST4sum.
l Quadrant Associative
It has the property that all diagonally opposite pairs of numbers within the four
main quadrants sum to MST4sum. That is, the diagonally opposite pairs of numbers
within the four quadrants AF, BE, CH, DG, IN, JM, KP and LO are having the
magic sum MST4sum.
l Distributive
It has the property that each of the four integers in the set of numbers (1, 2, 3, 4),
and (5, 6, 7, 8) and (9, 10, 11, 12), and (13, 14, 15, 16) are located in a row and
column where none of the other three numbers in the set are located [9, 10].

2.3.2 Construction of Doubly Even Magic Square of Order 16

Based on the Properties

In this work, to generate the doubly even magic square based on the properties
the algorithm 2.3.3 is used similar to algorithm 2.2.1 with a slight modification.
Apart from any starting number, and magic sum, the fundamental magic square
seed number 0000, and the four character value for the Prop variable are given as
input. The algorithm starts with by constructing 4  4 fundamental magic square
and there are sixteen 4  4 fundamental magic squares are generated. In each
magic square, the properties of fundamental magic square are applied based on
the cyclic left shift operation in the value of Prop variable. The algorithm is
shown in 2.3.3.
432 G. Ganapathy and K. Mani

2.3.3 Algorithm

Input: 4 digit seed number, starting Number, magic square property, and magic
squares sum
Output: Doubly Even Magic Square of order 16.
To generate 4  4 fundamental magic squares, the algorithm uses the steps from
1 to 5.b.1.3 illustrated in Algorithm 2.2.1. To generate further, the following steps
are used.
5.b.1.4 Case Spi in P. A, Q, D
‘P’: MS4_prop_sub i ← Pan-magic of MS4_base i
‘A’: MS4_prop_sub i ← Associativec of MS4_base i
‘Q’: MS4_prop_sub i ← Quadrant Associative of
MS4_base i
‘D’: MS4_prop_sub i ← Distributive of MS4_base i
end Case
5.b.1.5 prev← prevþ1
end for i
5.b.2 MS8_prop_sub j ←MS4_prop_sub 1jj MS4_prop_sub 2
jj MS4_prop_sub 3 jj MS4_prop_sub 4
5.b.3 case j in 2, 3, 4
2: Sp1← Sp2; Sp2← Sp3; Sp3← Sp4; Sp4← Sp1
3: Sp1← Sp3; Sp2← Sp4; Sp3← Sp1; Sp4← Sp2
4: Sp1← Sp4; Sp2← Sp1; Sp3← Sp2; Sp4← Sp3
end case j
5.b.4 Spi← Sp1jj Sp2jj Sp3jj Sp4, i ← 1, 2, 3, 4
end for j
5.c MSP16_count ←
MS8_prop_sub1jjMS8_prop_sub2jjMS8_prop_sub3jjMS8_prop_sub4
5.d Si←(Siþ rnd(10)) mod 8, i ¼ 1, 2, 3, 4
5.e count ← count þ 1
end while count

2.3.4 Magic Square Construction Based on Properties-Example

To generate a 16  16 doubly even magic square equivalent to ASCII set based on

the properties of 4  4 magic square, already generated the fundamental magic
squares MS4_sub1, MS4_sub2, MS4_sub3 and MS4_sub4 are considered. Let, the
value for seed number, starting number, magic sum and Prop are given 0000, 4,
12345 and ‘PAQD’ respectively as input. The Tables 2 – 5 are considered and the 4
 4 magic squares based on the properties are generated as follows.
Since the first value of ‘Prop’ variable is P i.e. pan-magic, apply this property to
the MS4_sub1. The resulting MS4_prop_sub1 is shown in Table 8. Based on the
second value ‘A’ i.e associative, the MS4_prop_sub2 is generated from MS4_sub2
and it is shown in Table 9.
33 Information Security Enhancement to Public–Key Cryptosystem 433

Since the third character is ‘Q’ i.e. quadrant associative, the MS4_prop_sub3 is
generated from MS4_sub3 and it is shown in Table 10.
Now, the fourth letter is ‘D’ i.e. distributive the MS4_prop_sub4 is generated
from MS4_sub4 and it is shown in Table 11. Thus, to form MS8_prop_sub1
concatenate all four MS4_prop_subi, I ¼ 1, 2, 3 and 4 (shown in Tables 8 through
11) and the first quadrant 8  8 magic square is shown in Table 12.
The second quadrant MS8_prop_sub2 magic square is generated by transform-
ing the Seed No. as Sp1← Sp2; Sp2← Sp3; Sp3← Sp4; Sp4← Sp1 and the ‘Prop’ value
is changed to ‘AQDP’, the second quadrant 8  8 magic square is generated from
the next four fundamental 4  4 magic squares shown in Table 13.

Table 8 MS4_prop_sub1 with (4, 1,539)

1,535 4 1,531 16
1,529 18 1,533 6
12 1,527 8 1,539
10 1,537 14 1,525

Table 9 MS4_prop_sub2 with (20, 1,523)

1,523 1,503 22 32
428 26 1,513 1,519
24 30 1,517 1,515
1,511 1,521 34 20

Table 10 MS4_sub3 with (36, 1,507)

1,507 42 1,493 44
1,501 36 1,499 50
38 1,503 48 1,497
40 1,505 40 1,495

Table 11 MS4_sub4 with (52, 1,491)

52 1,489 62 1,483
66 1,479 56 1,485
1,481 60 1,491 54
1,487 58 1,477 64

Table 12 MS8_prop_sub1 with (4, 1,539)

MS4_prop_sub1 MS4_prop_sub2
MS4_prop_sub3 MS4_prop_sub4
434 G. Ganapathy and K. Mani

The third quadrant MS8_prop_sub3 is obtained by transforming the seed number

as Sp1← Sp3; Sp2← Sp4; Sp3← Sp1; Sp4← Sp2 and the ‘Prop’ value is changed to
‘QDPA’. By applying the properties, the resultant magic square is shown is Table 14.
Similarly, the fourth quadrant MS8_prop_sub4 is generated in this manner by
shifting ‘QDPA’ one position towards left, the resultant magic square is shown in
Table 15.
In order to obtain first MS16_prop magic square, concatenate all four
MS8_prop_subi(i ¼ 1, 2, 3, and 4) magic squares and it is shown Table 16.
Similarly, the other MS16_prop magic squares are generated by using suitable
transformations of the seed number.

Table 13 MS8_prop_sub2 with (68, 1,475)

1471 68 1467 80 1456 84 1451 96
1465 82 1469 70 1449 98 1454 86
76 1463 72 1475 92 1448 88 1459
74 1473 78 1461 90 1457 94 1446
1439 101 1435 112 1423 116 1419 128
1433 114 1438 102 1417 130 1421 118
109 1431 104 1443 124 1415 120 1427
106 1441 110 1430 122 1425 126 1413

Table 14 MS8_prop_sub3 with (132, 1,411)

1411 138 1397 140 149 1393 158 1387
1405 132 1403 146 162 1383 152 1390
134 1407 144 1401 1385 157 1395 150
136 1409 142 1399 1391 154 1382 160
1375 165 1371 176 1363 1349 182 192
1369 178 1374 166 188 186 1353 1359
173 1367 168 1379 184 190 1357 1355
170 1377 174 1366 1351 1361 194 180

Table 15 MS8_prop_sub4 with (196, 1,347)

196 1345 206 1339 1327 212 1323 224
210 1335 200 1341 1321 226 1325 214
1337 204 1347 198 220 1319 216 1331
1343 202 1333 208 218 1329 222 1317
1315 1301 230 240 1291 1285 256 254
236 234 1305 1311 258 252 1289 1287
232 238 1309 1307 244 250 1295 1297
1303 1313 242 228 1293 1299 246 248

Table 16 First MS16_prop magic square using starting

no. 4, seed no. 0100, ‘PAQD’, and magic sum 12345
MS8_prop_sub1 MS8_prop_sub2
MS8_prop_sub3 MS8_prop_sub4
33 Information Security Enhancement to Public–Key Cryptosystem 435

3 Encryption/Decryption of Plain Text Using RSA

Cryptosystem with Magic Square

To show the relevance of this work to the security of public-key encryption

schemes, a public-key cryptosystem RSA is taken. RSA was proposed by Rivest
et al. To encrypt a plaintext M the user computes C ¼ Me mod n and decryption is
done by calculating M ¼ Cd mod n. In order to thwart currently known attacks, the
modulus n and thus M and C should have a length of 512–1,024 bits in this paper.

3.1 Wrapper Implementation-Example

In order to get a proper understanding of the subject matter of this paper, let p ¼ 11,
q ¼ 17 and e ¼ 7, then n ¼ 11(17) ¼ 187, (p  1)(q  1) ¼ 10(16) ¼ 160. Now
d ¼ 23. To encrypt, C ¼ M7mod 187 and to decrypt M ¼ C23mod 187. Suppose the
message is to be encrypted is “BABA”. The ASCII values for A and B are 65 and 66
respectively. To encrypt B, the numerals which occur at 66th position in first
(Table 6) and third MS16 (not shown here) are taken because B occurs in first
and third position in the clear text. Similarly, to encrypt A, the numerals at 65th
position in second and fourth MS16 (not shown here) are taken. Thus NP(A) ¼ 42
and 48, NP(B) ¼ 36 and 44. Hence NC(B) ¼ 367mod 187 ¼ 9, NC(A) ¼ 427mod 187
¼ 15, NC(B) ¼ 447mod 187 ¼ 22, NC(A) ¼ 487mod 187 ¼ 157. Thus, for same A
and B which occur more than once, different cipher texts are produced. Similar
encryption and decryption operations can be performed by considering the numer-
als occur at the positions of 66 and 65 for the characters B and A respectively from
MS16_prop magic square.

4 Parallel Cryptography

To speed up cryptographic transformation, the parallel cryptography is used.

Effective scheduling is important to improve the performance of crypto system in
parallel. The scheduling algorithms are divided into two classes: time sharing and
space sharing. Backfilling is the space sharing optimization technique. Maui is a
job scheduler specifically designed to optimize system utilization in policy
driver, heterogeneous high performance cluster (HPC) environment. The philoso-
phy behind it is essentially schedule jobs in priority order, and then backfill in the
holes. Maui has a two-phase scheduling algorithm. In the first phase, the high
priority jobs are scheduled using advanced reservation. A backfill algorithm is
used in the second phase to schedule low-priority jobs between previously selected
jobs [11]. In our study the Maui Scheduler with back filling scheduling technique is
used to calculate the encryption/decryption time of given message in simulated
environment.
436 G. Ganapathy and K. Mani

Table 17 Encryption and decryption time using RSA on a pentium processor based on properties
of magic squares
File Size No. of processors
(MB) 1 2 4 8
E D T E D T E D T E D T
(ms) (ms) (ms) (ms) (ms) (ms) (ms) (ms) (ms) (ms) (ms) (ms)
1 362 405 767 560 607 1,167 502 515 1,017 450 447 897
2 700 707 1,407 737 756 1,493 590 599 1,189 514 530 1,044
3 1,332 1,370 2,702 1,073 1,097 2,170 753 672 1,425 610 605 1,215
4 2,601 2,618 5,219 1,700 1,741 3,441 1,084 1,100 2,184 765 776 1,541
E - encryption time, D - decryption time, ms - milliseconds

5 Experimental Result

The methodology proposed is implemented in visual Cþþ version 6.0. The time
taken for encryption and decryption of various file sized message in simulated
parallel environment using RSA public-key crypto system based on the properties
of magic squares is shown in Table 17. The simulation scenario configured in our
implementation consisting of 2, 4, and 8 processors.
From Table 17, we observe that as the file size is increased in double, encryption
and decryption time is also increased in double for single processor. Moreover, the
time taken for encryption and decryption is almost same. Parallel encryption and
decryption have more effect if the file size is increased.

6 Conclusion

An alternative approach to existing ASCII based cryptosystem a number based on

magic squares generated by two different schemes is thought of and implemented.
It is noted that that the numerals generated using different views of magic square for
encryption may be obtained by the eavesdropper (but it is very difficult) because of
some kind of relationship exists among the numerals. But, in the case of magic
squares generated using the properties, it is very difficult to obtain the numerals by
him because of the numbers do not have any kind of relationship and hence it is very
difficult to obtain such numerals for encryption. Thus, the second approach is better
than the first one. This methodology will add-on one more layer of security, it adds
numerals for the text even before feeding into public-key algorithm for encryption.
This add-on security is built using magic square position equivalent to the character
to be encrypted.
Further efficiency of cryptographic operation depends on performing them in
parallel. Simulation using different number of processors for encryption/decryption
has shown that the time taken for decryption is approximately 1.2 times larger than
the corresponding time for encryption.
33 Information Security Enhancement to Public–Key Cryptosystem 437

References

1. A. Hamalainenn, M. Tommiska, J. Skytta, 6.78 Gigabits per Second Implementation of the

IDEA Cryptographic Algorithm LNCS 2438 (Springer-Verlag, Berlin, 2002), pp. 760–769
2. T. Wollinger, J. Guajardo, C. Paar, Cryptography in embedded systems: an overview, in
Proceedings of the Embedded World 2003 Exhibition and Conference, 18–20 Feb, 2003, pp.
735–744
3. J. Piepryk, D. Pointcheval, Parallel authentication and public-key encryption, The Eigth
Australasian Conference on Information Security and Privacy (Springer-Verlag, July 2003),
pp. 383–401
4. A.J. Menezes, P.C. Van Oorschot, S. Vanstone, Handbook of Applied Cryptography, (CRC
Press, Boca Ration, FL, 1997)
5. Wikipedia (Feb 2007) http://en.wikipedia.org/wiki/Magicsquare, pp. 1–3
6. A. Rogers, P. Loly, The Inertial Properties of Magic Squares and Cube, Nov 2004, pp. 1–3.
7. C.A. Pickover, The Zen of Magic Squares, Circles and Stars (Universities Press, Hyderabad,
India, 2002)
8. G. Ganapathy, K. Mani, Add-On Security Model for Public-KeyCryptosystem Based on Magic
Square Implementation, WCECS, (20–22 Oct 2009, San Francisco, CA), p. 317–321.
9. G.A. Foster, St. Joseph’s Healthcare Hamilton, Hamilton, et al. (Global Forum 2009) SAS®-
Magic Squares
10. J.P. Robertson, Theory of 4  4 Magic Squares (17 Jan 2009)
11. Y. Zhang, H. Franke, J.E. Moreira, A. Sivasubramanian, An Integrated Approach to Parallel
Scheduling using Gang-Scheduling, Backfilling and Migration (Springer-Verlag, JSSPP, UK,
2006), pp. 133–158
Chapter 34
Resource Allocation for Grid Applications: An
Economy Model

G. Murugesan and C. Chellappan

Abstract With the emergence of grid environments featuring dynamic resources

and varying user profiles, there is an increasing need to develop reliable tools that
can effectively coordinate the requirements of an application with available com-
puting resources. The ability to predict the behavior of complex aggregated systems
under dynamically changing workloads is particularly desirable, leading to effec-
tive resource usage and optimization of networked systems. To explore these issues,
some economic/market based models have been introduced in the literature, where
users, external schedulers, and local schedulers negotiate to optimize their objec-
tives. In this chapter, we have proposed a mathematical model for Grid resource
allocation with aims to minimize the cost of the grid users when multiple grid users
are trying to utilize the grid system at the same time.

1 Introduction

Grid computing is the sharing, selection and aggregation of resources such as

supercomputers, mainframes, storage systems, data sources which are geographi-
cally dispersed and the individual resources can vary widely in capacity, capability
and availability. In Grid computing the user (customer) can rent the resource
residing on the grid for executing their computationally intensive applications
instead of purchasing their own special expensive resources. The definition of
Grid is taken from the electrical power Grid that provides pervasive access to
power without caring where and how the power was produced. In the same way,
the computational Grid should provide pervasive access to high and computing
resources, without consideration for where and how the jobs were executed.

G. Murugesan (*)
Department of Computer Science and Engineering, Anna University, Chennai, 600 025,
Tamil Nadu, India
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 439

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_34, # Springer ScienceþBusiness Media B.V. 2010
440 G. Murugesan and C. Chellappan

The major classifications of Grids are computational grid, scavenged grid and data
grid. The first one is mainly concerned with computing power, the second classifi-
cation focuses on how the unused resources can be utilized and the last one is
concerned with data access capability.
The most challenging task in grid computing is the allocation of resources with
respect to the user application requirement; i.e. mapping of jobs to various
resources. This is a known NP-complete problem. For example mapping of 10
jobs on to 10 resources produce 1010 possible mappings. This is because every job
can be mapped to any of the resources. Another complexity in the resource alloca-
tion process is the lack of availability of information about the status of resources.
Much of the work was done on finding an optimal allocation of resources. Some of
the work used conventional strategies that were concerned with the overall system
performance but did not consider economics (prices) for allocating jobs to the
resources. On the other hand, a large number of projects use economic strategies
for allocating jobs. This work also follows the economic strategies for allocating
jobs to the resources. In our scenario, there are three classes of entities: Users,
Scheduler, and Resources. There are n users (also called as sources) S1, S2, . . ., Sn,
a Scheduler and m number of Resources (also called as Processors) R1, R2, . . ., Rm.
Each user i has Wi jobs and all the jobs are divisible jobs.
The users have jobs that need to be executed. The users send the jobs to the
scheduler which in turn identifies the resources with respect to the job requirement.
The resources represent the computing power of the grid and are where the users’
jobs get executed. Resource providers make a profit by selling computing power.
Users are the entities who can only send the job to the scheduler. Scheduler acts as a
mediator between the user and the resources. Finally resources are the entities that
just receive the jobs and process them.

2 Grid Economy Model

While comparing the grid system with the system around us, it is found that the Grid
system has high similarity with the market. The market is a decentralized, competi-
tive and dynamic system where the consumer and the producer have their own
objectives. Based on this observation, economic models are introduced to the Grid
in order to optimize the resource management and scheduling problems. The basic
components of the market model are the producers, consumers and commodities,
and which is analogous to the resource owners, resource users and various comput-
ing resources in the Grid environment. The decision making parameter in the
market models are prices and quality of commodities. For example, a consumer
usually wants to get better services (e.g., a smaller makespan) with a budget as
small as possible, while a provider usually wants to get a higher resource utilization
to raise its profits.
The use of economic methods in Grid scheduling involves interacting processes
between resource providers and users, analogous to various market behaviors, such
34 Resource Allocation for Grid Applications: An Economy Model 441

as bargain, bid, auction and so on. Buyya et al. discusses some economic models
that can be applied to the Grid world, including the Commodity Market Model,
Tender/contract-net Model, and Auction Model.
Due to the introduction of economic model in Grid computing, new research
opportunities have arisen. Because the economic cost and profit are considered by
Grid users and resource providers respectively, new objective functions and sched-
uling algorithms optimizing them are proposed. There are several algorithms called
deadline and budget constrained (DBC) scheduling algorithms are presented which
consider the cost and makespan of a job simultaneously. These algorithms imple-
ment different strategies. For example, guarantee the deadline and minimize the
cost or guarantee the budget and minimize the completion time. The difficulties to
optimize these two parameters in an algorithm lie in the fact that the units for cost
and time are different, and these two goals usually have conflicts (for example, high
performance resources are usually expensive). Economic models are not only
useful when economic cost/profit is explicitly involved, but also to construct new
scheduling methods with traditional objectives.
Economic methods for scheduling problems are very interesting because of their
successes in our daily lives. Most of the models can only support relatively simple
Grid scheduling problems such as independent tasks. For more complex applica-
tions, such as those consisting of dependent tasks and requiring cross-site coopera-
tion, more sophisticated economic models might be needed. With our knowledge
most of the economic models are focused on feasible solutions with respect to the
deadline and budget constraints. In our work we initially focuses on independent
workload model and later focusing on dependent workload model to achieve both
feasible and optimal allocation with linear programming approach.

3 Resource Management Challenges

The Grid environment contains heterogeneous resources, local management sys-

tems (single system image OS, queuing systems, etc.) and policies, and applications
(scientific, engineering, and commercial) with varied requirements (CPU, I/O,
memory, and/or network intensive). The producer (also called resource owners)
and consumers (who are the users) have different goals, objectives, strategies, and
demand patterns. Most importantly, both resources and end-users are geographi-
cally distributed with different time zones. A number of approaches for resource
management architectures have been proposed and the prominent ones are: cen-
tralized, decentralized, and hierarchical.
In managing the complexities present in large-scale Grid-like systems, tradi-
tional approaches are not suitable as they attempt to optimize system-wide mea-
sures of performance. Traditional approaches use centralized policies that need
complete state information and a common resource management policy, or decen-
tralized consensus based policy. Due to the complexity in constructing successful
Grid environments, it is impossible to define an acceptable system-wide performance
442 G. Murugesan and C. Chellappan

matrix and common fabric management policy. Therefore, hierarchical and decen-
tralized approaches are suitable for Grid resource and operational management.
Within these approaches, there exist different economic models for management
and regulation of supply-and-demand for resources.

4 Resource Allocation Model

A generic grid computing system infrastructure considered here comprises a net-

work of supercomputers and/or a cluster of computers connected by local area
networks, as shown in Fig. 1. We consider the problem of scheduling large-volume
loads (divisible loads) within a cluster system, which is a part of grid infrastructure.
We envisage this cluster system as a cluster node comprising a set of computing
nodes. Communication is assumed to be predominant between such cluster nodes
and is assumed to be negligible within a cluster node. The underlying computing
system within a cluster can be modeled as a fully connected bipartite graph
comprising sources, which have computationally intensive loads to be processed
(very many operations are performed on them) and computing elements called
sinks, for processing loads (data). To organize and coordinate distributed resources
participating in the grid computing environment, we utilize the resource model
which contains more than one source and a resource, based on which the uniform
and scalable and allocating computational resources are identified.
In a dynamic Grid environment, nodes may join or leave frequently, and nodes
status may change dynamically. In our model, we are considering the environment
as static; i.e. once the node joins the scheduling process, it has to be active
throughout the processing schedule. But in divisible load theory all the resources
that we propose unequal division of load from a resource with respect to the
processor capacity and the availability and the divisible load is assign to a set of
processors participate in the scheduling process.
We assume a star like structure with a scheduler as the central component; acts
as a controller to the model. The right of the scheduler; a set of processors

S1 P1

S2 P2

S3 scheduler P3

Fig. 1 Resource allocation Sm Pn

model
34 Resource Allocation for Grid Applications: An Economy Model 443

(processing elements) to execute the jobs assigned by the scheduler. The other side
of the scheduler are a set of sources; grid users. The scheduler collects the load from
all the sources and distribute to a set of processors p1, p2, . . ., pn. Initially all the
loads are held by the scheduler. All the communications are starts from the
scheduler and there is no direct communication between the processors. For
simplicity here we are assuming that the scheduler is only communicating with
processors to distribute the loads and get back the result.
Here we are considering the communication delay (transmitting time) to send a
portion of load to a processor and the result returning time is negligible. The
processor starts processing only after receiving entire workload which is assigned
to them. Also we assume that processors having independent communication
hardware which allows for simultaneous communications and computations on
the previously received loads. Additional constraints may be imposed on the
processors to the existence of other more urgent computations or maintenance
periods the availability of processor pi may be restricted to some interval [ri, di].
Where ri be the release time of the processor pi and di be the deadline of the
processor pi. By such a restriction we mean that computations may take place only
in the interval [ri, di]. A message with the load may arrive or start arriving before ri.
We assume that computations start immediately after the later of the two events: ri
or the load arrival. The computation time must be fit between the later of the above
two events. Also the loads are received from different sources with different
capacity and each load become divisible without any restriction. Each load received
from different sources is divided into a set of different tasks (portion of workload).
Maximum of one task may be assigned or allotted to a processor. The load portion
depends upon the capacity of the processor. Suppose, if there are m number of
sources {S1, S2, . . ., Sm}and n number of processors {P1, P2, . . ., Pn}and the
workloads from each sources become L ¼ {L1, L2, . . ., Lm}. Where L1 be the
total workload received from the source S1 and so on. Each workload Li may be
divided into T1, T2, . . .. The load L can be reordered by the scheduler to achieve
good performance of the computation. The scheduler splits the workload into tasks
and sends them to processors to perform the specified process. Only a set of
processors may be used to perform the workload of a source.
We denote that aij be the size of the task assigned to the processor pj. It is
expressed in load units (e.g. in bytes). aij ¼ 0 implies that pj is not in the set of
processor selected to perform the process of ith source workload. The total work-
load from a source is the sum of the sizes of the workload parts. i.e. Saij ¼ Li. Not
only pj is selected by the scheduler, but also the sequence of activating the
processors in pj and the division of load Li into chunks aij. Here we consider the
processors as unrelated processors; its communication links and start-up time are
specific for the task. Similarly the processor computing rates depend on the
processor and the task.
Let zj be the communication rate of the link to processor pj perceived by task Tj.
Transferring the portion of workload unit aij to pj takes zjaij time units. Let tj be the
processing rate of processor pj perceived by task Tj. To process aij portion of load
by processor pj takes tjaij time units. Let cj the processing cost to process the potion
444 G. Murugesan and C. Chellappan

of workload Tj. The total processing cost to process a portion of workload aij by the
processor pj becomes cjaij cost units.
In this work we analyze the complexity of scheduling the divisible loads L1,
L2, . . . , Lm of sizes T1, T2, . . . on n parallel processors P1, P2, . . . , Pn which are
interconnected together. We assume that the processors have sufficient memory
buffers to store the received portion of the workloads and computations. All the
processor will start processing immediately after receiving their entire portion of
workloads. One processor can process more than one source’s portion of workload.
By constructing a schedule the scheduler decides on the sequence of the tasks,
the set of processors assigned to each portion of workloads, the sequence of
processor activation and the size of the load parts. Our objective is to minimize
the usage of the grid user cost. We assumed that there is no separate start-up time
for individual processors and there is no fixed cost to utilize the processors. All the
processors are dedicated processors. But practically it is not possible, to simplify
our model as well as reduce the number of variables and constraints. The following
notations are used to formulate the mathematical model.

5 Design of Economy Model

To design the mathematical model we have used the following notations/variables

are used.
cj – Amount to spend to utilize jth processor
aij – Portion of workload from ith source to jth processor
xij – Binary variable
di – Deadline to complete the ith source job
bi – Budget allotted for the ith source job
sj – Schedule period of jth processor
tj – Time required to perform operation on one unit of job by jth processor
oi – Total workload of ith source
zj – Time taken to transfer a unit of workload to jth processor
sj – Scheduled time ith source

PMinimize
P
cj aij xij
i j

Subject to
PP PP
1. zj aij xij þ tj aij xij  di
Pi Pj Pi P
j
2. zj aij xij þ tj aij xij  si
Pi Pj i j
3. cj aij xij  bi
Pi j
4. aij ¼ oi
j
34 Resource Allocation for Grid Applications: An Economy Model 445
P
5. xij ¼ 1
i
6. 0
aij
7. xij ¼ f0; 1g
8. sj  0
9. zj ¼ 0
The objective is to minimize the total cost of the grid user those who are
assigning job to the grid system. The Eq. (1) represent the cost of all jobs that are
being assigned to a resource is the objective function. The Eqs. (1) through (7)
specifies the constraints used the mathematical model. The constraints (1) represent
the deadline associated with each sources. Constraints (2) match the workload
within the availability of resources. Constraints (3) are the budget for each source’s
workload. Constraint (4) represents the total workload of each sources involved in
the scheduling process. Constraints (5) makes sure that a portion of workload is
assigned to only one resource; i.e., there is no overlap between processing of
workloads. Constraint (6) makes sure that a portion of workload divided from the
total workload becomes a whole number. Constraint (7) is to set the binary value
either 0 or 1. The constraint (8) is a non-negativity constraints and the constraint (9)
we are assuming that there is no communication delay. But practically it is not
possible, for the simplicity of the model we assumed zj ¼ 0.

6 Experimental Results

Let us assume that the Grid system consists of five processors (resources) namely
P1, P2, P3, P4, and P5 with three sources S1, S2, and S3 are trying to utilize the grid
system to execute their workloads.
Table 1 shows the processors involved in the Grid system, the processing
capacity of each processor per unit workloads, the processing cost of each processor

Table 1 Processor capacity

Processors Processing time per Processing cost ($) Available
unit Workload (h) time (hrs)
P1 2 4 30
P2 3 3 50
P3 4 2 50
P4 3 3 50
P5 2 5 50

Table 2 Details of workloads

Sources Work loads (kb) Budget ($) Deadline (hrs)
S1 25 60 60
S2 30 100 70
S3 40 120 130
446 G. Murugesan and C. Chellappan

Table 3 Processor allotments

Source Allotted processor Workload allotted (kb) Time taken to complete (hrs)
S1 P1 15 57
P2 7
P5 3
S2 P3 3 68
P4 2
P5 25
S3 P2 16 124
P3 6
P4 16
P5 2

to execute a unit workload and the available time of each processor. Table 2 shows
the details of the different sources which are trying to utilize the grid system, total
workload of each source and the budget allotted to each of the sources to complete
their workloads and the expected time to complete the process of each workload.
Using the details given in the table we have formed the mathematical model and
solved the equation using LINGO package. After execution of the mathematical
model, the maximum cost for processing all the three source workloads are Rs. 351.
Table 3 shows the details of workload allotment to the processors involved in the
process. From the table it is clear that the total workload of S1 is divided into three
parts, the total workload of S2 is divided into three parts and the total workload of S3
is divided into four parts and allotted into processors. Also it shows that the
completion time of each source’s workloads.

7 Related Works

Till date several grid scheduling algorithms have been proposed to optimize the
overall grid system performance. The study of managing resources in the Grid
environment started from 1960s. The economic problem [1] results from having
different ways for using the available resource, so as how to decide what the best
way to use them is. Utilization of Grid must be cheaper for the Grid user than
purchasing their own resources [2] and must satisfy their requirements. On the other
hand, resource providers must know if it is worth to provide their resources for
usage by the Grid users. Also the pricing of resources should not be per time unit or
slot (e.g. cost per minutes) [3]. Because it leads to big difference in speeds of
processors, so the price per unit time of a processor might be cheaper, but the user
must pay large amount of money due to slow processing resources. Moreover the
users have to know how many time units they need to finish their jobs. Thus the cost
of Grid user must be determined based on the tasks the resource is processing.
Deadline scheduling algorithm [4] is one of the algorithms which follow the
economic strategy. The aim of this algorithm is to decrease the number of jobs that
do not meet their deadlines. The resources are priced according to their perfor-
mance. This algorithm also has a facility of fallback mechanism; which mean that it
34 Resource Allocation for Grid Applications: An Economy Model 447

can inform the grid user to resubmit the jobs again and the jobs which do not met the
deadline of the available resources. Nimrod/G [5, 6] includes four scheduling
algorithms which are cost, time, conservative time and cost-time. Cost scheduling
algorithm try to decrease the amount of money paid for executing the jobs with
respect to the deadline. Time scheduling algorithms attempt to minimize the time
required to complete the jobs with respect to their budget allotment. The conserva-
tive time scheduling algorithm aims to execute the jobs within the stipulated budget
and the deadlines. The cost-time scheduling algorithm works similar to cost sched-
uling algorithm except that when there are two or more resources with the same
price, it employs time scheduling algorithm. It does not deal with co-allocation.
The effective workload allocation model with single source has been proposed [7]
for data grid system. Market-based resource allocation for grid computing (Gomaluch
and Schroeder) supports time and space shared allocations. Furthermore it supports
co-allocation. It is supposed that resources have background load which change with
time and that has the highest priority for execution, so they are not fully dedicated to
the grid. Gridway [8] is an agent based scheduling system. It aims to minimize the
execution time, total cost and the performance cost ratio of the submitted job. Market
economy based resource allocation in Grids [9] is an auction based user scheduling
policies for selecting resources were proposed. GridIs [10] a P2P decentralized
framework for economic scheduling using tender model. The author tries to perform
the process without considering the deadline and the algorithm is implemented with
the help of a variable called conservative degree, its value between 0 and 1.
The time and cost trade-off has been proposed [11] with two meta-scheduling
heuristics algorithms, that minimize and manage the execution cost and time of user
applications. Also they have presented a cost metric to manage the trade-off
between the execution cost and time. Compute power market [12] is architecture
responsible for managing grid resources, and mapping jobs to suitable resources
according to the utility functions used. Parallel virtual machine [13, 14] enables the
computational resources to be used as if they are a single high performance
machine. It supports both execution on single and multiple resources by splitting
the task into subtasks. Grid scheduling by using mathematical model was proposed
[15] with equal portion of load to all the processors; i.e., the entire workload
received from a source is equally divided and a portion of load is assigned to a
processor with the help of random numbers to divide the entire workload from a
source. The work closest to ours is [6] where the authors proposed the algorithm
that claims to meet the budget and deadline constraints of jobs. However, the
algorithm proposed is ad-hoc and does not have an explicit objective.

8 Conclusion

In this chapter, we have proposed scheduling strategies for processing multiple

divisible loads on grid systems. In real life the loads will come dynamically and the
resources are also in dynamic in nature. But here we have considered both loads and
448 G. Murugesan and C. Chellappan

the resources are in static manner. This work can be extended for dynamic nature
also. We are in the process of dynamic arrival of loads and the resources. Also we
have assumed that there is sufficient buffer space available in all the resources. The
experimental result demonstrates the usefulness of this strategy. We need a system
to find out the execution time of a task and also the cost of usage of a processor/
resource. The execution time entirely depends upon the processor speed.

References

1. J. Nakai, Pricing computing resources: reading between the lines and beyond. Technical
Report, National Aeronautics and Space Administration, 2002
2. L. He, X. Sun, G.V. Laszewski, A qos guided scheduling algorithm for grid scheduling, 2003
3. L. He, Designing economic-based distributed resource and task allocation mechanism for self-
interested agents in computational grids, in Proceeding of GRACEHOPPER, 2004
4. A. Takefusa, S. Matsuoka, H. Casanova, F. Berman, A study of deadline scheduling for client-
server system on the computational grid, in HPDC’01: Proceedings of the 10th IEEE
International Symposium on High Performance Distributed Computing, pp. 406–415, 2001
5. R. Buyya, Economic-based distributed resource management and scheduling for grid computing.
PhD Thesis, Monash University, Melbourne, Australia, 2002
6. R. Buyya, D. Abramson, J. Giddy, Nimrod/g: An architecture for a resource management and
scheduling system in a global computational grid, HPC ASIA’2000, 2000
7. M. Abdullah, M. Othman, H. Ibrahim, S. Subramaniam, Load allocation model for scheduling
divisible data grid applications. J. Comput. Sci. 5(10), 760–763 (2009)
8. R.A. Moreno, A.B. Alonso-Conde, Job scheduling and resource management technique in
economic grid environments. European Across Grids Conference, pp. 25–32, 2003
9. S.R. Reddy Market economy based resource allocation in grids. Master’s Thesis, Indian
Institute of technology, Kharagpur, India. 2006
10. L. Xiao, Y. Zhu, L.M. Ni and Z. Xu, Gridis: An incentive-based grid scheduling, in IPDPS’05:
Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium,
Washington, DC, 2005, IEEE Computer Society
11. S.K. Garg, R. Buyya, H.J. Siegel, Scheduling parallel applications on utility grids: time and
cost trade-off management, in Proceedings of the Thirty-Second Australasian Computer
Science Conference (ACSC2009), Wellington, Australia, Conferences in Research and Prac-
tice in Information Technology (CRPIT), vol. 91
12. R. Buyya, S. Vazhkudai, Computer power market: towards a market oriented grid. CCGRID,
pp. 574–581, 2001
13. A. Beguelin, J. Dongarra, A. Geist, V. Sunderam, Visualization and debugging in a heteroge-
neous environment. Computer 26(6), 88–95 (1993)
14. V.P.V.M. Sunderam, A framework for parallel distributed computing. Concurrency Prac. Exp.
2(4), 315–340 (1990)
15. G. Murugesan, C. Chellappan, An optimal allocation of loads from multiple sources for Grid
Scheduling, ICETIC’2009, International Conference on Emerging Trends in Computing,
2009
16. R. Buyya, M. Murshed, D. Abramson, A deadline and budget constrained cost-time optimi-
zation algorithm for scheduling task farming applications on global grids, Technical Report
CSSE-2002/109, Monash University, Melbourne, Australia, 2002
17. J. Gomaluch, M. Schroeder, Market-based resource allocation for grid computing: A model
and simulation, Middleware workshops. pp. 211–218
18. G. Murugesan, C. Chellappan, An economic allocation of resources for multiple grid applica-
tions, in WCECS2009, Proceedings of the World Congress on Engineering and Computer
Science, vol I, San Francisco, CA, 2009
34 Resource Allocation for Grid Applications: An Economy Model 449

19. M. Othman, M. Abdullah, H. Ibrahim, S. Subramaniam, Adaptive divisible load model for
scheduling data-intensive grid applications: computational science, LNCS, 4487, 246–253
(2007)
20. S. Viswanathan, B. Veeravalli, T.G. Robertazzi, Resource-aware distributed scheduling
strategies for large-scale computational cluster/grid systems. IEEE Trans. Parallel Distributed
Syst.18: 1450–1461, 2007
21. Q. Xiao, Design and analysis of a load balancing strategy in data grids, Future Generation
Comput. Syst. 16, 132–137, 2007
Chapter 35
A Free and Didactic Implementation
of the SEND Protocol for IPv6

Say Chiu and Eric Gamess

Abstract IPv6 adds many improvements to IPv4 in areas such as address space,
built-in security, quality of service, routing and network auto-configuration. IPv6
nodes use the Neighbor Discovery (ND) protocol to discover other nodes on the
link, to determine their link-layer addresses, to find routers, to detect duplicate
address, and to maintain reachability information about the paths to active neigh-
bors. ND is vulnerable to various attacks when it is not secured. The original
specifications of ND called for the use of IPsec as a security mechanism to protect
ND messages. However, its use is impractical due to the very large number of
manually configured security associations needed for protecting ND. For this
reason, the Secure Neighbor Discovery Protocol (SEND) was proposed. In this
work, we present Easy-SEND, an open source implementation of SEND that can be
used in production environment or as a didactic application for the teaching and
learning of the SEND protocol. Easy-SEND is easy to install and use, and it has an
event logger that can help network administrators to troubleshoot problems or
students in their studies. It also includes a tool to generate and verify Cryptographi-
cally Generated Addresses (CGA) that are used with SEND.

1 Introduction

IPv6 [1–3] (Internet Protocol version 6) is a solution to the problem of the shortage
of public IPv4 addresses that faces Internet. IPv6 adds many improvements to IPv4
in areas such as quality of service, routing and network auto-configuration. Even if
IPv6 has been around for more than 10 years now, there is a lack of IPv6 network

S. Chiu (*)
Universidad Central de Venezuela, Facultad de Ciencias, Escuela de Computación, Paseo Los
Ilustres, Urb. Valle Abajo, Caracas 1020, Venezuela
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 451

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_35, # Springer ScienceþBusiness Media B.V. 2010
452 S. Chiu and E. Gamess

specialists in Venezuela and around the world. Therefore, the training of IPv6
specialists has become an important issue. In the undergraduate program of Com-
puter Science at Central University of Venezuela (in Spanish: Universidad Central
de Venezuela), some courses have been upgraded or added to the curriculum to face
the problem. For example, Advanced Network Protocols (in Spanish: Protocolos
Avanzados en Redes) is a new course that was introduced to the curriculum of the
undergraduate program of Computer Science in 2005. Its objectives include the
understanding of IPv6 standards, such as the ND [4] (Neighbor Discovery) protocol
and the SEND [5] (Secure Neighbor Discovery) protocol.
Since ND [4] is supported by all the actual modern operating systems, a variety
of laboratories are done in the course (Advanced Network Protocols) to strengthen
student’s knowledge about this important protocol. However, we have been facing
the lack of support for SEND [4] (as stated is Section 5) from manufacturers and it
has been almost impossible for us to do laboratories to clarify the complex
procedures involved in SEND. Therefore, we decided to develop a new application
(Easy-SEND) from scratch that implements the SEND protocol, with good support
for the teaching and learning process. Its main goal is to be used as a didactic
application in advanced courses related to networks at Central University of
Venezuela. Additionally, Easy-SEND can be used in production networks where
security is important as a replacement of the ND protocol.
The rest of this research is organized as follows: An overview of ND is presented
in Section 2. Vulnerability issues for ND are discussed in Section 3. SEND is
presented in Section 4. Related works are viewed in Section 5. Easy-SEND is
introduced and justified in Section 6. Conclusions and future work are discussed in
Section 7.

2 Neighbor Discovery Protocol Overview

The ND [4] (Neighbor Discovery) protocol solves a set of problems related to the
interaction between nodes attached to the same link. It defines mechanisms to
solution each of the following problems:
l Router discovery: During router discovery, a host discovers the local routers on
an attached link. This process is equivalent to ICMPv4 router discovery [6].
l Prefix discovery: How hosts discover the set of network prefixes that define
which destinations are on-link for an attached link.
l Parameter discovery: Through this process, nodes learn some operating para-
meters such as the link MTU (Maximum Transmission Unit) or the default hop-
limit value to place in outgoing packets.
l Address auto-configuration: How nodes automatically configure addresses
for an interface in either the presence or absence of an address configuration
server, such as a DHCPv6 [7] (Dynamic Host Configuration Protocol for IPv6)
server.
35 A Free and Didactic Implementation of the SEND Protocol for IPv6 453

l Address resolution: Is a process by which nodes determine the link-layer address

of an on-link destination (neighbor) given only the destination’s IPv6 address. It
is equivalent to ARP (Address Resolution Protocol) in IPv4.
l Next-hop determination: How nodes determine the IPv6 address of the neighbor
to which a packet must be forwarded, based on the destination address. The
next-hop address is either the destination address or the address of an on-link
router.
l Neighbor Unreachability Detection (NUD): Is a process by which nodes deter-
mine that a neighbor is no longer reachable. For neighbors used as routers,
alternate default routers can be tried. For both routers and hosts, address resolu-
tion can be performed again.
l Duplicate Address Detection (DAD): During duplicate address detection, a node
determines that an address it wishes to use is not already in use by another node.
This process is equivalent to using gratuitous ARP frames in IPv4.
l Redirect: How a router informs a host of a better first-hop node to reach a
particular destination. It is equivalent to the use of the ICMPv4 Redirect
message.
The ND protocol has five messages or PDUs (Protocol Data Units) provided by
ICMPv6 [8] (Internet Control Message Protocol version 6), an updated version of
ICMPv4 [9] (Internet Control Message Protocol version 4). These messages are:
RS (Router Solicitation), RA (Router Advertisement), NS (Neighbor Solicitation),
NA (Neighbor Advertisement), and Redirect. RS are sent by IPv6 hosts to discover
neighboring routers on an attached link. RA are sent by IPv6 routers periodically
(unsolicited multicast router advertisements) or in response to a RS message
(solicited router advertisements). NS are sent by IPv6 nodes to resolve a neighbor’s
IPv6 address to its link-layer address (MAC address) or to verify if an IPv6 node is
still reachable (NUD). NA are sent by IPv6 nodes in response to a NS message or to
propagate a link-layer address change. Redirect messages are sent by IPv6 routers
to inform hosts of a better first-hop for a destination.
ND messages consist of an ND message header, composed of an ICMPv6
header, some ND message-specific data, and zero or more ND options. Fig. 1
shows the format of an ND message.

Neighbor Discovery Message

IPv6 Header ND Message-

ICMPv6 ND Message
Next Header = 58 specific
Header Options
(ICMPv6) data

Neighbor Discovery
Message Header

Fig. 1 Format of an ND message

454 S. Chiu and E. Gamess

Table 1 Options for Neighbor Discovery

Type Option name Description
1 Source link-layer Included in NS, RS and RA messages to indicate the link-layer
address address of the sender of the packet
2 Target link-layer It indicates the link-layer address of the target. It is used in NA
address and Redirect messages
3 Prefix information Appear in RA messages to provide hosts with on-link prefixes
and prefixes for address auto-configuration.
4 Redirected header It is used in Redirect messages and contains all or part of the
packet that is being redirected
5 Maximum Transmission This option is used in RA messages to ensure that all nodes on a
Unit (MTU) link use the same MTU value in those cases where the link MTU
is not well known

ND message options provide additional information, such as link-layer addresses,

on-link network prefixes, on-link MTU information, redirection data, mobility infor-
mation, and specific routes. These options are formatted in type-length-value (TLV)
format. Table 1 shows the main options of the ND protocol with a brief description.

3 Vulnerabilities of the Neighbor Discovery Protocol

ND can suffer different attacks that can be classified in three types as follow:
l Impersonation/Spoofing: class of attacks in which a malicious node successfully
masquerades as another by falsifying data and thereby gaining an illegitimate
advantage. This type of attacks is easily made since there is no control over
MAC addresses and they can be change with little effort.
l DoS (Denial of Service): type of attacks in which a malicious node prevents
communication between the node under attack and all other nodes.
l Redirection: class of attacks in which a malicious node redirects packets away
from the legitimate receiver to another node on the link.
According to [10], the ND protocol suffers frequent attacks that include, but are
not limited to: Neighbor Solicitation Spoofing, Neighbor Advertisement Spoofing,
Neighbor Unreachability Detection Failure, Duplicate Address Detection DoS
Attack, Malicious Last Hop Router, Spoofed Redirect Message, Bogus On-Link
Prefix, Bogus Address Configuration Prefix, Parameter Spoofing, Replay Attacks,
and Neighbor Discovery DoS Attack.

4 Secure Neighbor Discovery Protocol

As stated in Section 3, ND is insecure and susceptible to malicious interference. To

counter the threats of the ND protocol, IPsec [11, 12] (Internet Protocol Security)
can be used. One of the main objectives of ND is the auto-configuration of hosts
35 A Free and Didactic Implementation of the SEND Protocol for IPv6 455

with no human intervention. With IPsec, the key exchange can be done automati-
cally when hosts already have a valid IPv6 address. Without an initial IPv6 address,
the usage of IPsec can be quite tedious and impracticable. For these reasons, the
IETF (Internet Engineering Task Force) developed the SEND [5] (Secure Neighbor
Discovery) protocol.
SEND is a security extension of the ND protocol. It introduces new messages
(CPS and CPA) and options (CGA, Timestamp, Nonce, RSA Signature, Trust
Anchor and Certificate) to the ND protocol, as well as defense mechanisms
against attacks on integrity and identity. The main components used in the SEND
protocol are:
l Certification paths: anchored on trusted parties certify the authority of routers. A
host must be configured with a trust anchor to which the router has a certification
path before the host can adopt the router as its default router. CPS (Certification
Path Solicitation) and CPA (Certification Path Advertisement) messages are
used to discover a certification path to the trust anchor.
l Cryptographically Generated Addresses (CGA): are used to make sure that the
sender of a Neighbor Discovery message is the owner of the claimed address. A
public-private key pair is generated by all nodes before they can claim an
address. A new ND option, the CGA option, is used to carry the public key
and associated parameters.
l RSA Signature option: is used to protect all messages relating to Neighbor and
Router Discovery.
l Timestamp option: Provides replay protection against unsolicited advertise-
ments and redirects.
l Nonce option: Ensures that an advertisement is a fresh response to a solicitation
sent earlier by the node.
Similarly to the ND protocol, SEND uses ICMPv6 [8] messages. Two new
messages were added for the ADD (Authorization Delegation Discovery) process.
Table 2 shows the value of the Type field from the ICMPv6 message and a brief
description.
SEND defines two mechanisms (CGA and ADD) presented in the following
sections for securing the ND protocol.

Table 2 ICMPv6 messages for SEND

ICMPv6 SEND message Description
type
148 Certification Path Sent by hosts during the ADD process to request a
Solicitation (CPS) certification path between a router and one of the host’s
trust anchors
149 Certification Path The CPA message is sent in reply to the CPS message and
Advertisement (CPA) contains the router certificate
456 S. Chiu and E. Gamess

sec sec sec U/L I/G

64 65 66 67 68 69 70 71 72 ... 127
Network Prefix Interface Identifier

Hash
Function

Public Key Auxiliary Parameters

Fig. 2 Generation process of a CGA

4.1 Cryptographically Generated Address

CGAs [13] (Cryptographically Generated Addresses) are IPv6 addresses generated

with a hash function, from a public key and auxiliary parameters. This provides a
method for securely associating a cryptographic public key with an IPv6 address in
the SEND protocol. The node generating a CGA address must first obtain a RSA
(Rivest, Shamir, and Adelman) key pair (SEND uses an RSA public/private key
pair). The node then computes the interface identifier part (which is the rightmost
64 bits) and appends the result to the prefix to form the CGA address (see Fig. 2).

4.2 Authorization Delegation Discovery

The ADD (Authorization Delegation Discovery) is used to certify the authority of

routers via a trust anchor. A “trust anchor” is an authoritative source that can be
trusted by hosts, and which is used to certify the authority of routers: “certification
paths” can be established from the trust anchors to enable a certified router to
perform certification of another router; before any host can adopt a router as its
default router, the host must be configured with a trust anchor that can certify the
router via a certification path. Hence, the certificate format mandated by the SEND
protocol enables the hosts to learn a certification path of a router, and to validate the
authenticity of a router.

5 Related Works

At the present time, manufacturers of network devices and developers of operating

systems do not include functionalities of the SEND protocol in their products. The
unique implementation of SEND had been developed by members of the working
group that did its specification [5], especially James Kempf who works at DoCoMo
35 A Free and Didactic Implementation of the SEND Protocol for IPv6 457

USA Labs. DoCoMo’s Open Source SEND Project is an implementation of SEND

developed by DoCoMo USA Labs that works in the operating system user-space.
It can be installed in Linux (kernel 2.6) or FreeBSD 5.4.
The SEND implementation from DoCoMo USA Labs has a limited debugging
mode where users can see information related to the internal states and the opera-
tions executed by the application. It also includes two tools (cgatool and ipexttool)
to generate keys, CGA addresses and the extension for IPv6 addresses required in
X.509 certificates [14]. The application allows two configuration modes: basic and
advanced modes. In the basic mode, users can create a SEND environment for hosts
only, without on-link router discovery. In the advanced mode, hosts can discover
on-link routers by using certification path and their correspondent trust anchor.
The SEND application from DoCoMo USA Labs has some drawbacks. It is not
able to manage address collisions during the DAD (Duplicate Address Detection)
process, due to the difficulties carried by its integration to the kernel. Additionally,
users must have good knowledge in the specification of firewall rules (ip6tables),
since they are essential for the configuration of the application. Also, DoCoMo
USA Labs has announced a few weeks ago that it will no longer maintain the SEND
project1. Hence, the source code is no longer available to download in the web site
and support has been canceled.
JSend2 is announced in SourceForge as a free and open source implementation
of the SEND protocol in Java. Its main objective is to be used for learning and
testing purpose. For now, this project has not released any files.

6 A Didactic Implementation of the SEND Protocol

Easy-SEND is a free and open source application developed in Java that implements
the SEND [5] (Secure Neighbor Discovery) protocol. It also includes a tool (CGA-
Gen) to generate and verify CGAs [13] (Cryptographically Generated Addresses).
We developed the application in the Linux user-space that does not require
modification at the kernel level. Easy-SEND works as a firewall between the
network interface card and the IPv6 stack, by using the ip6tables filtering rules.
The actual version is limited to the creation of a secure environment for IPv6 hosts;
that is, hosts are not able to participate in the Router Discovery process.
The architecture and the functionality of Easy-SEND are summarized in Fig. 3.
libipq3 provides a mechanism for passing packets out of the stack for queuing to
user-space, then receiving these packets back into the kernel with a verdict

1
http://www.docomolabs-usa.com/lab_opensource.html
2
http://sourceforge.net/projects/jsend
3
http://www.netfilter.org/projects
458 S. Chiu and E. Gamess

specifying what to do with the packets (such as acceptation or rejection). These

packets may also be modified in the user-space prior to reinjection back into the
kernel.

User space
User
Process

libipq

2
ip6tables

ip_queue

3
Netfilter
rules Packets out
no
Packets in drop?
1
kernel

Fig. 3 Easy-SEND architecture

Initially, ND messages are selected based on ip6tables rules (see Fig. 4) and sent
to ip_queue, a queue managed by libipq. From the first to the fifth line of Fig. 4,
ip6tables rules capture RS (type 133), RA (type 134), NS (type 135), NA (type
136), and Redirect (type 137) messages, respectively. Library libipq allows the
manipulation of ND packets (aggregation and verification of options) based on the
SEND specifications before the determination of the verdict (acceptation or rejec-
tion of the packets). Messages marked for rejection will be dropped. Message
marked for acceptation will be further processed.

ip6tables -A INPUT -p icmpv6 -j QUEUE --icmpv6-type "133"

ip6tables -A INPUT -p icmpv6 -j QUEUE --icmpv6-type "134"
ip6tables -A INPUT -p icmpv6 -j QUEUE --icmpv6-type "135"
ip6tables -A INPUT -p icmpv6 -j QUEUE --icmpv6-type "136"
ip6tables -A INPUT -p icmpv6 -j QUEUE --icmpv6-type "137"

Fig. 4 Capture rules for ND messages with ip6tables

Since library libipq was written in C, we used a wrapper, called Virtual Services
IPQ4 or VServ IPQ for short, which allows Java users to call the libipq native
functions from Java applications.

4
http://www.savarese.org/software/vserv-ipq
35 A Free and Didactic Implementation of the SEND Protocol for IPv6 459

To develop the application, we used the waterfall methodology [15, 16]. It is a

sequential methodology, in which development is seen as flowing steadily downwards
(like a waterfall) through the phases of Conception, Requirement, Analysis, Design,
Coding, and Testing. The application’s general structure consists of a set of classes
that interacts with each others. Fig. 5 shows the principal classes of Easy-SEND.

Fig. 5 Diagram of principal classes

Class SENDOPT is an abstract class where get and set methods are defined to
access and modify SEND options. Class CGA allows the generation and mani-
pulation of CGA addresses, as well as the validation of CGA options included
in ND messages. Methods of the RSASignature class permit the creation and
verification of digital signatures. TimeStamp and Nonce are classes with methods
to produce and handle timestamps for the protection against replay attacks. Class
Monitor is for the capture and manipulation of packets that arrive to the ip6tables
queue. The rules for the reception and transmission of packets are also defined in
class Monitor. NDMessage is a class for the aggregation (outgoing packets) and
removal (incoming packets) of SEND options in messages. Its methods automati-
cally update the checksum field. Class KeyManagement allows users to generate,
load and save RSA keys.
To validate our implementation of the SEND protocol, we make several test
experiments. For reasons of space, in this work we will only briefly describe two of
our experiments. In the first experiment, we setup a test-bed (see Fig. 6) where we
460 S. Chiu and E. Gamess

CGA: 2004::34cc:6bba:5947:d208 PC1

eth0

eth0

PC2 CGA: 2004::3869:5ddb:cdba:f44c

Fig. 6 Test-bed topology

Fig. 7 Ping between two PCs that run Easy-SEND

connected two PCs in an Ethernet LAN. Each PC had 512 MB of RAM, and we
installed Ubuntu 8.04 LTS (a community developed, Linux-based operating sys-
tem), Java 2SE 1.6.0_07 and Easy-SEND.
Fig. 7 shows the ND messages sent by PC1 and PC2 during the execution of a
ping command from PC1 to PC2. We can observe that Easy-SEND modifies the
original messages (NS and NA) to add SEND options (CGA, Timestamp, Nonce,
and RSA Signature). We used Wireshark [17] (a popular network analyzer formerly
known as Ethereal) to analyze the SEND protocol.
In the second experiment, we made interoperability tests between Easy-SEND and
the SEND implementation of DoCoMo USA labs. We made a test-bed similar to the
one shown in Fig. 6, but we installed the SEND implementation of DoCoMo USA
Labs in PC1 and Easy-SEND in PC2. We made different communication tests (FTP,
SSH, PING, TFTP, etc.) between the two PCs and had a perfect interoperability.
35 A Free and Didactic Implementation of the SEND Protocol for IPv6 461

Easy-SEND has a powerful event logger to record all the SEND events for
troubleshooting and learning purposes. It is based on Apache log4j5, a Java-based
logging framework. Our event logger has three messages levels (INFO, DEBUG,
and ERROR) that can be activated. INFO messages show information related to the
states and processes that occur in the application. DEBUG messages allow users to
get a hexadecimal dump of the messages sent or received by Easy-SEND. ERROR
messages are shown when abnormal situations occur.
Typically, logs can be displayed in real-time in a Unix shell. It is also possible to
redirect the logs to text files for later analysis. Our event logger can show many
information to users helping them in the troubleshooting or learning process. Fig. 8

[06:53:07,577] INFO Monitor - NetfilterPacket Hook Local Input

[06:53:07,578] INFO Monitor - Incoming Interface: eth0
[06:53:07,579] INFO Monitor - Verifying SEND options
[06:53:07,579] INFO Monitor - ICMPv6 Type: <- Neighbor Solicitation
[06:53:07,579] INFO CGA - Verifying CGA
[06:53:07,580] INFO CGA - Verifying address: fe80:0:0:0:2c8e:4f5b:84b:80c1

[06:53:07,580] DEBUG CGA - CGA Parameters:

1b be bd c4 5f 26 a4 c0 66 6b eb 46 73 44 fe 32
fe 80 00 00 00 00 00 00 00 30 81 9f 30 0d 06 09
2a 86 48 86 f7 0d 01 01 01 05 00 03 81 8d 00 30
81 89 02 81 81 00 b0 89 3b 51 88 d3 f2 3c 9b a0
98 4f 99 01 b4 5f de 81 89 b3 d6 ba 96 ff a7 05
b3 49 4c 28 ce 60 2e ba e7 f6 76 f4 c2 87 f9 51
36 be 08 12 ab c2 e5 00 50 8a b5 be fb 04 79 99
79 b4 ee 88 5b 45 b7 f1 88 a9 e2 e4 52 3a 75 9b
ea 94 87 38 56 49 8d 63 66 b6 78 1e d0 41 4a f5
25 e9 cf 69 50 9b f2 b9 68 91 a9 8c e6 8a 6f 61
8f 45 8d 40 6a fc a4 26 ae 1e a8 7d 8b 48 88 c2
20 b3 c3 e3 93 2b 02 03 01 00 01

[06:53:07,581] INFO CGA - Verification of CGA was successful!

[06:53:07,581] INFO TimeStamp - Verifying Timestamp
[06:53:07,581] INFO TimestampCache - 1.223292216E9 > 1.223292209E9 ?
[06:53:07,581] INFO RSASignature - Verifying KeyHash
[06:53:07,582] INFO RSASignature - Keyhash was successfully verified
[06:53:07,582] INFO RSASignature - Verifying digital signature
[06:53:07,585] INFO RSASignature - Digital signature successfully verified

[06:53:07,585] DEBUG Monitor - Modified Packet

60 00 00 00 00 20 3a ff fe 80 00 00 00 00 00 00
2c 8e 4f 5b 08 4b 80 c1 fe 80 00 00 00 00 00 00
20 ee 39 bb 94 e7 35 6a 87 00 7b a6 00 00 00 00
fe 80 00 00 00 00 00 00 20 ee 39 bb 94 e7 35 6a
01 01 00 0c 29 c4 87 be

[06:53:07,585] INFO Monitor - Reinjected!!!

Fig. 8 Example of logs

5
http://logging.apache.org/log4j/1.2/index.html
462 S. Chiu and E. Gamess

shows the trace generated by packet 9 (a Neighbor Solicitation message) for the
capture of Fig. 7.
At the present time, Easy-SEND is the unique free and open source implementa-
tion of the SEND protocol that is active, since DoCoMo USA Labs is no longer
supporting its implementation. The source code of Easy-SEND can be downloaded
from SourceForge (a famous web-based source code repository) at http://sourceforge.
net/projects/easy-send To facilitate the learning and teaching process, a virtual
appliance is also available. A virtual appliance is a minimalist virtual machine
image designed to run under some sort of virtualization technology and aimed
to eliminate the installation, configuration and maintenance costs. Our virtual
appliance can be run with VMWare Player.6

7 Conclusions and Future Work

Security aspects of IPv6 networks are not fundamentally different from IPv4 net-
works, therefore, many of the possible attacks to IPv4 protocols can also be done in
IPv6 protocols, and the ND protocol is not the exception. SEND has been developed
to minimize the attacks that can be done to IPv6 nodes during their auto-configura-
tion process. Unlike IPsec, SEND is more easy and practicable to be implemented.
Note that [18] proposes a new protocol called TRDP (Trusted Router Discovery
Protocol) to secure the router discovery process. Compared with the ADD (Autho-
rization Delegation Discovery) process introduced in the SEND protocol, TRDP
obviates the burdensome work for a host to parse the lengthy certification path,
improves efficiency of network communication between the router and hosts during
the router authentication process, and also reduces the exposure to attacks on both
hosts and the access router.
At Central University of Venezuela, SEND is an important topic in the syllabus
of the Advanced Network Protocols course. Instructors had a hard time teaching
SEND, so some laboratories were introduced to clarify the complex processes
associated to SEND. Initially, the SEND implementation of DoCoMo USA Labs
was used, but this implementation has been discarded since its debugging system
has some weaknesses and DoCoMo USA Labs announced that it will no longer
maintain the project. Easy-SEND was developed to face this situation.
Easy-SEND is a free and open source implementation of the SEND protocol
developed under the GNU General Public License in Java. It has a powerful event
logger with three levels of messages. It main goal is to be used for teaching and
learning purpose, but can also be used in a production environment. Since it is open
source, its source code can be enhanced or adapted by other researchers for specific
needs.

6
http://www.vmware.com/products/player
35 A Free and Didactic Implementation of the SEND Protocol for IPv6 463

At the present time, Easy-SEND is the unique open source project that imple-
ments the SEND protocol that is active. We plan to further develop it. Our goal is to
offer the application to other national and international universities to be used in
their curriculum for teaching the SEND protocol. As future work, we plan to
integrate to our application the process of Router Discovery with the ADD trust
mechanism, and extension fields to the CGA parameter data structure to support
different hash functions [19]. Also, the implementation of Easy-SEND in the Linux
user-space can not be considered as a final solution, so we will adapt our application
to the kernel to optimize the aggregation and the verification of SEND message
options. Finally, we want to port our application to Microsoft Windows operating
systems.

References

1. J. Davies, Understanding IPv6, 2nd edn. (Microsoft Press, Jan 2008)

2. S. Deering, R. Hinden, Internet Protocol, Version 6 (IPv6) Specification. RFC 2460. Dec 1998
3. S. Hagen, IPv6 Essentials, 2nd edn. (O’Reilly, Jan 2008)
4. T. Narten, E. Nordmark, W. Simpson, H. Soliman, Neighbor Discovery for IP Version 6
(IPv6). RFC 4861. Sep 2007
5. J. Arkko, J. Kempf, B. Zill, P. Nikander, Secure Neighbor Discovery (SEND). RFC 3971. Mar
2005
6. S. Deering, ICMP Router Discovery Messages. RFC 1256. Sep 1991
7. R. Droms, J. Bound, B. Volz, T. Lemon, C. Perkins, M. Carney, Dynamic Host Configuration
Protocol for IPv6 (DHCPv6). RFC 3315. Jul 2003
8. A. Conta, S. Deering, M. Gupta, Internet Control Message Protocol (ICMPv6) for the Internet
Protocol Version 6 (IPv6) Specification. RFC 4443. Mar 2006
9. A. Farrel, The Internet and Its Protocols: A Comparative Approach, 1st edn. (Morgan
Kaufmann, May 2004)
10. P. Nikander, J. Kempf, E. Nordmark, IPv6 Neighbor Discovery (ND) Trust Models and
Threats. RFC 3756. May 2004
11. S. Kent, IP Authentication Header. RFC 4302. Dec 2005
12. S. Kent, IP Encapsulating Security Payload (ESP). RFC 4303. Dec 2005
13. T. Aura, Cryptographically Generated Addresses (CGA). RFC 3972. Mar 2005
14. R. Housley, W. Polk, W. Ford, D. Solo, Internet X.509 Public Key Infrastructure Certificate
and Certificate Revocation List (CRL) Profile. RFC 3280. Apr 2002
15. J. Charvat, Project Management Methodologies: Selecting, Implementing, and Supporting
Methodologies and Processes for Projects (John Wiley & Sons. Feb 2003)
16. W. Royce, Managing the development of large software systems: concepts and techniques,
in Proceedings of the 9th International Conference on Software Engineering, Monterey,
California, United States. 30 Mar–2 Apr 1987
17. A. Orebaugh, G. Ramirez, J. Beale, Wireshark & Ethereal Network Protocol Analyzer Toolkit.
(Syngress, Feb 2007)
18. J. Zhang, J. Liu, Z. Xu, J. Li, X. Ye, TRDP: a trusted router discovery protocol, in Proceeding
of the 7th International Symposium on Communications and Information Technologies,
Sydney, Australia. 17–19 Oct 2007
19. M. Bagnulo, J. Arkko, Cryptographically Generated Addresses (CGA) Extension Field
Format. RFC 4581. Oct 2006
Chapter 36
A Survey of Network Benchmark Tools

Karima Velásquez and Eric Gamess

Abstract Nowadays, there are a wide variety of network benchmark tools, giving
researchers and network administrators many options to work with. However, this
variety tends to hinder the selection process of the appropriate tool. Furthermore,
sometimes users are forced to try several tools in order to find one that calculates a
desired gauge, so they have to learn how to manipulate different tools and how
to interpret the obtained results. This research offers a compilation of network
benchmark tools currently used, with the purpose of guiding the selection of one
tool over the others, by outlining their main features, strengths and weaknesses.

1 Introduction

It is common that researchers and/or network administrators have to measure

different network indicators. Several tools are available for this purpose, but not
all of them have the same characteristics and evaluate the same performance
parameters. Sometimes, researchers are obliged to use a number of these tools
and complete several experiments in order to find significant results. According to
the test to be performed, and to the network or device under test, one benchmark
tool can be more appropriated than another. This can complicate and delay the
evaluation process, since researchers would have to learn to use different tools. For
example, Gamess and Velásquez [1] had to use two different performance tools and
write their own benchmarks in order to evaluate the forwarding performance on
different operating systems.

K. Velásquez (*)
Universidad Central de Venezuela, Facultad de Ciencias, Escuela de Computación, Paseo Los
Ilustres, Urb. Valle Abajo, Caracas 1020, Venezuela
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 465

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_36, # Springer ScienceþBusiness Media B.V. 2010
466 K. Velásquez and E. Gamess

This research presents an up-to-date survey of network benchmark tools cur-

rently used, to help researchers and network administrators to make a right decision
about the more appropriate tool for their needs. Additionally, this work outlines
desirable features for benchmark tools to be developed in the future.
This work is organized as follow. In Section 2, we discuss previous works related
to our study. Section 3 contains a deep study and description of seven popular
benchmark tools. Section 4 shows a comparative analysis of the seven tools.
Conclusions and future work are discussed in Section 5.

2 Related Works

The interest of network benchmark tools has been very important in the Internet
community, to the point that Parker and Schmechel [2] wrote an RFC (Request For
Comment) to present some of them, when RFCs are usually used to specify
protocols. Botta et al. [3] present an exhaustive state-of-the-art survey on available
bandwidth estimation tools. They divided the tools in three categories: (1) end-to-
end capacity estimation tools, (2) available bandwidth estimation tools, and (3)
TCP throughput and bulk transfer capacity measurement tools. They also introduce
a tool called BET (Bandwidth Estimation Tool) and they compare it with other
performance tools, in terms of accuracy and total time used to complete the
measurement process. Ubik and Král [4] summarize their experience with band-
width estimation tools; they focus on finding the size and location of bottlenecks.
They present a classification of end-to-end bandwidth estimation tools based on a
number of properties, which include determining bandwidth (installed bandwidth)
vs. throughput (available bandwidth), sender only software vs. sender and receiver
software, etc. They also describe properties for a few selected tools, present the
results of measurement with one tool (pathload) on a high-speed scenario and
results of combined measurements with several tools over a real fast long-distance
network. Gupta et al. [5] perform an experimental comparison study of both passive
and active bandwidth estimation tools for 802.11-based wireless networks, and
conclude that for wireless networks a passive technique provides greater accuracy.
Strauss et al. [6] describe Spruce, a simple tool for measuring available bandwidth,
and then compare it with other existing tools over many different Internet paths.
Their study is based on accuracy, failure patterns, probe overhead, and implemen-
tation issues. Montesino [7] presents a comparative analysis of state-of-the-art
active bandwidth estimation techniques and tools. He offers a short overview of a
set of tests performed over different conditions and scenarios, which were done to
assess the performance of active bandwidth estimation tools. Mingzhe et al. [8]
present WBest, a wireless bandwidth estimation tool designed to accurately esti-
mate the available bandwidth in IEEE 802.11 networks, explaining that most of the
existing tools were not designed for wireless networks. They define the algorithm
employed by WBest and present a set of experiments and analysis of the results.
They also compare their tool with other tools like pathchirp [9] and pathload [10].
36 A Survey of Network Benchmark Tools 467

3 Network Benchmark Tools

For our study, we selected seven different throughput and bulk capacity measure-
ment tools: Netperf, D-ITG, NetStress, MGEN, LANforge, WlanTV, and TTCP. We
chose these tools for their popularity and also because they are active projects.
Many performance tools available in Internet have no updated versions and seem to
be dead. These seven tools also offer different features, so they provided us with a
much wider range for our analysis.
Our goal is to present a clear description of the available tools that can help users
choose one that is best suited for a specific experiment. Additionally, we gathered
some desirable features that a throughput and bulk transfer capacity measurement
tool should have since our goal is to develop a new network benchmark tool from
scratch.
During our analysis, each necessary step, questions that arose, the efforts it took
to find the answers to them, features that were helpful or confusing, and the required
time were carefully recorded. Progress and experiences were discussed among the
authors.
According to [11], most of the tools included in our study belong to the group of
Throughput and Bulk Transfer Capacity (BTC) measurement tools, which are
benchmark tools that use large TCP and/or UDP transfers to measure the available
throughput in an end-to-end path. The experiences with each of the tools are
described below. For tools that require a Unix environment, tests were conducted
under Linux Debian 5.0 (Lenny) with kernel version 2.6.26. For Windows, tests
were performed under Windows XP Professional SP3.

3.1 Netperf

Netperf1 is an open source benchmark tool that can be used to measure various
aspects of networking performance. Its primary focus is on bulk data transfer and
request/response performance using either TCP or UDP and the BSD socket
interface.
Netperf is designed around the basic client-server model. There are two execu-
tables (netperf and netserver). The netserver program can be invoked by inetd (the
system daemon), or can be run as a standalone daemon. In the first method, users
must have administrator privileges; the second method implies that users must
remember to run the program explicitly. Unless users want to change the default
port (option -p) or want to work with IPv6 (option -6) instead of IPv4, the server is
invoked without options.

1
http://www.netperf.org
468 K. Velásquez and E. Gamess

3.1.1 Installation and Documentation

We installed Netperf using Aptitude, a Debian package management utility. In

order to do so, we had to edit the /etc/apt/sources.list file (root privileges are
required), to add the non-free branch indicator to the listed mirrors, so the package
could be found. Installation from source files is also possible when the binary
distribution is not available. For our Debian version, the compilation of the source
code was quite easy.
Documentation can be found at the official website. It includes detailed infor-
mation on how to install the tool using the source files and a general description of
the tool. It also describes how different tests can be done, including bulk data
transfer and request/response performance experiments. Finally, it has a section
with examples and some troubleshooting recommendations.
The information reported in the man pages was not accurate. In fact, most
options described were inconsistent to their actual use in the tool. We suspect that
the man pages have not been upgraded with the new releases of the tool and
correspond to an old release. So, we recommend to follow the indications from
the tool’s help (invoked with the -h option) instead of using the man pages.

3.1.2 Testing and Usability

Netperf can be executed from a console using the command netperf -H

<hostNamejIPAddress>, specifying the remote host name or IP address. When
netperf is executed, a control connection is established to the remote system
(netserver) to pass test configuration parameters and results to and from the remote
system. The control connection is a TCP connection using BSD sockets.
The duration of the test can be configured. Final reports include sender and
receiver buffer socket size in bytes, sent message size in bytes, duration of the test
in seconds, and throughput. Results are reported as a table. It is possible to change
units for the output, but they have to be set at the beginning of the test.

3.1.3 Supported Platforms and Protocols

Netperf is only supported by Unix platforms. It is possible to use IPv4 or IPv6 as the
network layer protocol, just by specifying the option -4 or -6 respectively. It allows
the usage of both TCP and UDP as the transport layer protocol.

3.2 D-itg

D-ITG2 (Distributed Internet Traffic Generator) is an open source packet

level traffic generation platform that produces IPv4/IPv6 traffic for variable

2
http://www.grid.unina.it/software/ITG/index.php
36 A Survey of Network Benchmark Tools 469

inter-departure time and packet size. It was conceived to be used as a distributed

performance tool, able to measure one-way-delay (OWD), round-trip-time (RTT),
packet loss rate, jitter and throughput [12].
D-ITG follows the client-server model. There are four basic executables that
form the platform components: ITGSend, ITGRecv, ITGLog, and ITGDec. ITGSend
acts as the client, and can generate several data flows simultaneously as specified by
the input file (configuration file). ITGRecv acts as the server and can receive several
data flows from different clients simultaneously. ITGLog is the log server of the
platform, and receives information from ITGSend and ITGRecv. ITGDec is a utility
that analyzes the results of the experiments. Additionally there are two more
executables, ITGPlot and ITGapi. ITGPlot is an Octave3 based tool to plot the
data contained in the log files obtained with ITGDec (such as delay.dat, bitrate.dat,
jitter.dat and packetloss.dat), that contain an average delay, bit rate, jitter and
packet loss rate, respectively, which are calculated each millisecond and reported
in their respective files. ITGapi is a Cþþ API that enables the remote control of
traffic generation.
When D-ITG is executed, a control connection is established to the remote
system, to pass test configuration parameters such as network protocol to use.

3.2.1 Installation and Documentation

We downloaded the D-ITG-2.3-Familiar.zip file for our Linux version, which

contains the precompiled binaries to execute this tool. This version did not work
on our system. Then we tried the D-ITG-2.6.1d.zip file, which contains the source
files for compilation. Following the compilation instructions included in this distri-
bution, we used the make command but were not able to compile the source files.
We presume incompatibility issues between the tool and our Linux version. Finally,
we tried on an older version of Debian, with 2.6.24 kernel; this time we could
compile the source files without any problems, and transferred the resulting binary
files to our original system, where they were executed correctly.
We downloaded the D-ITG-2.6.1d-WINbinaryIPv4.zip and D-ITG-2.6.1d-
WINbinaryIPv6.zip files for Windows. They contain the binary files ready for
execution. No installation process is required.
Documentation can be found on the official website. It includes a manual (which
can also be accessed from the tool with -h option) along with several examples of use.

3.2.2 Testing and Usability

D-ITG is able to measure one-way-delay, round-trip-time, packet loss rate, jitter

and throughput.

3
http://www.octave.org
470 K. Velásquez and E. Gamess

This tool can work in two modes: single mode and script mode. The single mode
allows users to generate a single flow of packets from a client (ITGSend) to the
server (ITGRecv). The script mode enables ITGSend to simultaneously generate
several flows from a single client. Each flow is managed by a single thread, and an
additional thread acts as master and coordinates the other threads. To generate n
flows, the script file (input file) must contain n lines, each of which is used to specify
the characteristics of one flow.
Execution is started via console (no GUI is offered) using different options
provided by the tool, which can be consulted with the -h option. Some configurable
parameters include the TTL (Time to Live), inter-departure time, payload size and
protocol type.

3.2.3 Supported Platforms and Protocols

D-ITG is available for Unix platforms as well as Windows platforms. IPv4 and IPv6
are both supported as network layer protocols; in Unix, to select one or the other
users must specify the right type of IP address (IPv4 address or IPv6 address) with
the –a option in ITGSend. In Windows there are two different binary files, one for
IPv4 support and another one for IPv6 support.
By default, UDP is used as the transport layer protocol, but it is possible to use
TCP and even ICMP. Upper layer protocols supported include Telnet, DNS, and
RTP for VoIP applications.

3.3 NetStress

NetStress4 is a simple benchmark tool used to measure network performance for

both wired and wireless networks. It implements bulk data transfer using TCP.
Network performance is reported in terms of throughput. Once again, the client-
server model is used during the testing process.

3.3.1 Installation and Documentation

NetStress installation is very easy. Once the setup file was downloaded, we only had
to execute it and follow the installation assistant’s instructions.
On the official website there is a link to the NetStress help file that contains a
brief description of the tool and its purpose, system requirements, and descriptions
on how to run the server and the client. Additionally, it includes a section

4
http://www.performancewifi.net/performance-wifi/main/netstress.htm
36 A Survey of Network Benchmark Tools 471

explaining how to interpret the results. This help can also be accessed within the
application, in the Help menu choosing the Contents option.

3.3.2 Testing and Usability

NetStress only measures the maximum network throughput. It provides a GUI

where users can select different options. Once started, a mode must be selected:
server or client. The server listens for network packets, whereas the client transmits
network packets. The tool reports the amount of bytes sent, bytes received, and the
throughput, in text mode and graphically.
There is no way to specify the duration of the experiment or the amount of data
to transfer, so users must close the application to stop the experiment.

3.3.3 Supported Platforms and Protocols

NetStress was developed for Windows platforms. It employs IPv4 at the network
layer and TCP at the transport layer. It does not offer support for other protocols.

3.4 MGEN

The Multi-Generator5 (MGEN) is an open source benchmark developed by the

PROTocol Engineering Advanced Networking (PROTEAN) research group at the
Naval Research Laboratory (NRL). MGEN provides the ability to perform IP
network performance tests and measurements using UDP/IP traffic. It supports
both, unicast and multicast traffic generation. It follows the client-server model,
using the same program for both ends.
Currently, two different versions (3.X and 4.X) are available, which are not
interoperable. Some versions of MGEN (such as 3.X) have a graphical user inter-
face. MGEN version 4.X must be launched currently from the command-line.

3.4.1 Installation and Documentation

Two packages are available for Linux platforms: source code and a precompiled
version. We installed the source files to guarantee system compatibility. We started
to uncompress the tar ball and used the command make –f Makefile.linux, since
there are several makefiles for Unix systems and one must be chosen according to

5
http://cs.itd.nrl.navy.mil/work/mgen
472 K. Velásquez and E. Gamess

the target system. For Windows platforms there is no need for installation, since the
ZIP file includes an executable file ready to use.
The distribution contains an HTML file with extensive documentation. It
explains the tool’s usage and the different options. It also offers several examples
on how to use the tool to evaluate performance.

3.4.2 Testing and Usability

Input files (configuration files) can be used to drive the generated loading patterns
over the time. A command-line option is also available.
Input files can be used to specify the traffic patterns of unicast and/or multicast
UDP/IP applications. Defined data flows can follow periodic (CBR), Poisson, and
burst patterns. These data flows can be modified during the experiment, since the
input file allows users to change a given data flow at specific times. Some fields of
the IP header can be set. When a multicast flow is defined, users can specify the
TTL (Time to Live) value. For both unicast and multicast flows, the ToS (Type of
Service) value can also be specified. For IPv6 flows, a Flow Label value can be
defined. For multicast flows, users can control when to join or leave the multicast
group, indicating the IP address of the multicast group to join or leave and the time
to do so.
Results are shown on standard output, or they can be redirected to a logfile for
later analysis. These results are only the report of the packets exchanged; no
additional information is given.
To obtain statistics, users must record the results in a logfile, which can be later
used as an input for the trpr6 (Trace Plot Real-time) program. trpr analyzes the
output of MGEN and creates an output that is suitable for plotting. It also supports a
range of functionalities for specific uses of the gnuplot7 graphing program. Impor-
tant results, such as the throughput, delivery latency, loss rate, message reordering,
and multicast join/leave latency can be calculated from the information in the
output logfile. However, it is left to users to do this calculation.

3.4.3 Supported Platforms and Protocols

The updated version of the MGEN toolset, MGEN 4.0, provides support for Win32
platforms in addition to a broad base of Unix-based platforms, including MacOS X.
Several enhancements are planned, including support for TCP, since MGEN
currently only supports UDP. IPv4 and IPv6 can be used as the network layer
protocol.

6
http://pf.itd.nrl.navy.mil/protools/trpr.html
7
http://www.gnuplot.info
36 A Survey of Network Benchmark Tools 473

3.5 LANforge

Candela Technologies’ LANforge8 consists of two tools: LANforge-FIRE configu-

ration, used for traffic generation that simulates the edge of the tested network; and
LANforge-ICE feature set that is used to simulate the core of a network.
LANforge is a proprietary tool that can be purchased at its website. Candela
Technologies offers several options of purchase. To obtain this tool we subscribed
and created an account which enabled us to download a trial version of the tool.

3.5.1 Installation and Documentation

To install LANforge on a Linux platform, the LANforgeServer-X.Y.Z_Linux-x64.

tar.gz file is needed (where X.Y.Z stands for the version, in our case it was version
5.0.9). Once uncompressed, we executed the install.bash file. A home directory for
the tool is required. By default, /home/lanforge is used, but users can change it by
executing install.bash –d <directory>. Once executed install.bash, we had to run
the lfconfig script (with the –cwd options) from the tool’s home directory. Both
scripts are self explanatory, and give users enough feedback. For the GUI (Graphic
User Interface), we had to install the LANforgeGUI_X.Y.Z_Linux.tar.bz2 package,
and execute the lfgui_install.bash file.
For the installation on the Windows platform, we downloaded two files:
LANforge-Server-X.Y.Z-Installer.exe and LANforge-GUI-X.Y.Z-Installer.exe.
Then, we just followed the installation assistant’s instructions. The installation
process is easy and fast.
A complete documentation is available at the website, including a user’s guide,
an installation guide, and an upgrade guide for every package. The installation
guide includes a step-by-step description of the installation process; as well as a
troubleshooting section. Also, tutorials are available on the website. These include
explicit examples on how to use the tool for different scenarios, along with detailed
screenshots to describe the configuration and testing process.

3.5.2 Testing and Usability

LANforge is a tool that allows users to simulate networks and perform tests over
them. It includes a GUI that makes the testing process easier. There are several tests
that can be performed with this tool, including configuration of virtual routers (only
for the Linux version). Some of the results that can be obtained with this tool are:
bytes received and transmitted, packets received and transmitted, bits per second
(bps) received and transmitted, collisions and errors; these results are shown in text
and graphically, in real time.

8
http://www.candelatech.com/lanforge_v3/datasheet.html
474 K. Velásquez and E. Gamess

Users can configure many aspects of the test, including the number of packets to
send, interface type of the endpoint to simulate (ISDN, T1, modem, etc), and even
custom payloads can be defined. LANforge also incorporates a VoIP Call Generator
which currently supports H.323 and SIP (Session Initiated Protocol). The voice
payload is transmitted with RTP (Real-time Transport Protocol) which runs over
UDP. RTCP (Real-time Transport Control Protocol) is used for latency and other
accounting. Jitter buffers are used to smooth out network jitter inherent in RTP
traffic.
WAN links can also be simulated and tested. Various characteristics can be
added to the traffic flowing though them, including maximum bandwidth, latency,
jitter, dropped packets, duplicated packets, bit and byte errors, etc. Results are
reported as text and graphically, using a vertical bar graph.
The LANforge GUI has many tabs grouping different tests, and it can be
confusing trying to use it because of the many options offered. However, there is
a detailed description of each tab at the GUI User Guide that can be found at the
website. This guide shows several examples and screenshots that clarify the pro-
cess.

3.5.3 Supported Platforms and Protocols

LANforge is available for Linux, Windows, and Solaris platforms (the Solaris
platform was not included in this research). Supported protocols include, but are
not limited to, raw Ethernet, MPLS, IPv4, IPv6, ICMP, OSPF, BGP, UDP, TCP,
FTP, HTTP, and HTTPS.

3.6 WLAN Traffic Visualizer

WLAN Traffic Visualizer9 (WlanTV for short) provides measurements of traffic load
and visualization of sequence of frames in IEEE 802.11 WLANs [4]. WlanTV is
published under the GPL terms. It is developed in Java and both source and JAR
distributions are available for download at the website.

3.6.1 Installation and Documentation

WlanTV requires Wireshark10 (Windows) or TShark (Linux), and JRE 1.6 (Java
Runtime Environment) or later to build and run the program. Only standard Java
packets are needed. It uses TShark to parse the log files and to perform live captures.

9
http://sourceforge.net/projects/wlantv
10
http://www.wireshark.org
36 A Survey of Network Benchmark Tools 475

We installed WlanTV from the JAR distribution and used the command java –jar
wlantv.jar, which opens a GUI from which users can run the experiments.
Documentation is not very extensive. The distribution includes a very small
README file with a few instructions about system prerequisites and not many
instructions about installation and execution. However, an example of capture is
available for download at the website.

3.6.2 Testing and Usability

This tool does not use the client-server model. Only one computer is needed to
conduct the experiments. It sniffs the traffic of an IEEE 802.11 network and reports
the results of that capture. For the experiments, users can employ a file from a
previous capture, or they can start a new live capture. In the later option, users must
stop the live capture in order to observe the results; which include the frame count,
byte count, capture duration, transmission rate during the capture, detailed infor-
mation for each packet (protocols, length, addresses, etc.), and bandwidth distribu-
tion. Results are shown in text mode and graphically.

3.6.3 Supported Platforms and Protocols

WlanTV is available for Linux and Windows platforms. This tool reports statistics
for 802.11 protocols.

3.7 TTCP

TTCP11 (Test TCP) is a command-line sockets-based benchmark tool for measuring

TCP and UDP performance between two systems. TTCP is designed around
the client-server model. There is only one executable, ttcp, and users must specify
which end will be transmitting (option -t) and which end will be receiving
(option -r) the data during the experiment.

3.7.1 Installation and Documentation

To install TTCP, gcc12 is needed to compile the file ttcp.c and generate the
executable ttcp, which includes both the sender and the receiver. Documentation
is offered via man pages, and also by invoking the tool without options.

11
ftp://ftp.eenet.ee/pub/FreeBSD/distfiles/ttcp
12
http://gcc.gnu.org
476 K. Velásquez and E. Gamess

3.7.2 Testing and Usability

Users must start the receiver and then start the transmitter, both via command-line.
The transmitting side sends a specified number of packets (option –n, 2048 by
default) to the receiving side. At the end of the test, the two sides display the
number of bytes transmitted and the time elapsed for the packets to travel from one
end to the other.
Results are shown on standard output. These results include the throughput and
an average of system calls. It is possible to change the units for the output, but they
have to be set at the beginning of the test.

3.7.3 Supported Platforms and Protocols

TTCP is only supported on Unix platforms. However, some ports for Windows are
available (e.g. PCATTCP). IPv4 is used as the network layer protocol. It allows the
usage of both TCP and UDP as the transport layer protocol.

4 Comparative Analysis

Table 1 presents the main characteristics of the reviewed tools. We considered

some interesting features such as: date of the last release, A/P, privileges, supported
platforms, protocols, reported results, and clock synchronization.
The date of the last release is important since it can indicate how much time the
project has been inactive. The A/P category indicates if the tool is Active (A) or
Passive (P). Active tools will affect the normal traffic in the network, by injecting
their own packets; while passive tools only capture the traffic that passes through
the interface. The Privileges row shows information about the user permission
required to install or run the application. We limited the study of supported plat-
forms to Linux and Windows. For the protocols, we focused on network and
transport layer protocols. For the Reported results we include the delay, jitter and
throughput. In the row labeled Sync required, we indicate if clock synchronization
is required between computers that run any of the processes involved in the tests.
We used ‘–’ to denote features that are not available, not supported, nor documen-
ted by some tools.
According to Table 1, some projects are more active than others. It seems that
MGEN has been inactive for a while, since its last version was released more than
5 years ago, which may be a reason for its lack of support of IPv6. TTCP seems to
be dead, however NTTCP (New Test TCP) and ETTCP (Enhanced Test TCP) are
recent forks of TTCP. Another important issue is the cost of the tools. Most of the
evaluated tools are free, except LANforge. LANforge price starts at $1999 (USD);
this could be an impediment for some researchers or network administrators. It is
also important to consider if the tool is active or passive. Among the studied tools,
 36

Table 1 Tool’s main features

Feature Netperf D-ITG NetStress MGEN LANforge WlanTV TTCP
Evaluated version 2.4.4 2.6.1d 1.0.8350 4.2b6 5.0.9 1.3.0 –
Last release 10/2007 09/2008 09/2008 10/2005 04/2009 12/2008 –
Free Yes Yes Yes Yes No Yes Yes
Open source Yes Yes No Yes No Yes Yes
A/P Active Active Active Active Active Passive Active
Privileges User User User User User Root User
A Survey of Network Benchmark Tools

Supported platforms Linux Linux, Windows Windows Linux, Windows Linux, Windows Linux, Windows Linux
Network protocols IPv4, IPv6 IPv4, IPv6, ICMPv4, IPv4 IPv4, IPv6 IPv4, IPv6, ICMPv4, IPv4, IPv6 IPv4
ICMPv6 ICMPv6
Transport protocols TCP, UDP TCP, UDP TCP UDP TCP, UDP TCP, UDP TCP, UDP
Reported results Throughput Delay, jitter Throughput Delay and Delay, jitter and Throughput Throughput
and throughput throughput throughput
User interface Console Console GUI Console GUI GUI Console
Sync required No No No No No – No
477
478 K. Velásquez and E. Gamess

only WlanTV does not introduce additional overhead to the network by injecting
traffic; however this also implies that this tool is not able to measure the throughput
and only reports results of the transmission rate for the captured packets during
the test.
We also noticed that WlanTV users must have root privileges for execution since
the tool requires the network interface card to be set in promiscuous mode. Also, we
noticed that most of the studied tools already have IPv6 support. Network through-
put is the most common reported result; followed by delay and jitter.
In relation to the user interface, only a few tools have a GUI (NetStress,
LANforge, and WlanTV). To ease the testing process, some tools (D-ITG and
MGEN) use an input file that contains the parameters and the streams description
for the tests, which is more flexible than the traditional command line arguments.
One last issue of great importance is the clock synchronization, especially when
measuring delays. Manual clock synchronization is not recommended due to of its
poor accuracy. One solution is to use NTP (Network Time Protocol); this requires
additional knowledge and configuration.

5 Conclusions and Future Work

In this work, we presented the results of our analysis in which we evaluated

and compared several current network benchmark tools. We studied seven tools:
Netperf, D-ITG, NetStress, MGEN, LANforge, WlanTV and TTCP. The evaluation
was categorized in: installation and documentation, testing and usability, and
supported platforms and protocols. Table 1 summarizes the results of our study.
We consider that none of the seven tools studied is complete, so in general, users
must install several of these tools to perform basic performance tests. Netperf is a
simple tool and users can easily learn how to use it; however, the results of the tests
are limited to throughput. D-ITG is a more powerful tool and reports more results,
such as jitter, delay, and throughput; however, the times reported by D-ITG for the
delay are not realistic. In our experiments, we had delays that were over an hour.
We believe it was not an installation problem, since there are some test results
reported on the D-ITG manual with similar times. Furthermore, users must face
some problems during installation in some Debian versions. NetStress is an easy
tool to work with that offers a GUI; but has poor protocol support, and tests cannot
be neither parameterized nor stopped. MGEN allows the parameterization of tests,
and also offers a complete documentation; however, it is not that simple to use,
because of the many options it offers. Moreover, it does not show results directly,
and users must calculate them from the information of the exchanged packets.
LANforge provides support for many different protocols and tests; it is easy to
install and offers good documentation. WlanTV is a passive tool that offers relevant
indicators of performance for traffic captured on WiFi networks. TTCP is an easy
tool to work with that reports the throughput of a network; but offers just a few
36 A Survey of Network Benchmark Tools 479

parameters to customize the experiment (buffer size and number) as well as a poor
protocol support.
For future work, we plan to design and implement our own network benchmark
tool that will incorporate the strengths of the evaluated tools, but eliminates their
weaknesses, following the client-server model. Our tool will be compliant with
RFCs 1242 [13], 2544 [14] and 5180 [15], where a set of recommendations are
suggested to develop benchmarks. One of our main interests is to offer support for
IPv6 protocol. Also, it will be distributed under the GNU General Public License,
allowing users to modify the source code for any particular requirement.
Among the features we plan to include in our tool is the ability to measure the
packet loss rate, throughput and RTT for different protocols (e.g. TCP, UDP, ICMP,
and IP). Tests also may be parameterized, allowing users to define custom traffic
generation that follow different models (random variables), including CBR, burst
traffic, Poisson, Exponential and Pareto. Also, users should be able to specify
values for some header fields, such as the TTL and ToS in IPv4, and the Traffic
Class, Hop Limit and Flow Label in IPv6.
As shown in our analysis, most of the tools do not have specific support for WiFi.
For IEEE 802.11, only end-to-end performance evaluation is offered. Users have no
way to obtain performance result (bandwidth, delay, packet loss rate) between a
mobile station and an access point (or wireless router). So we also plan to adapt the
tool that we will develop for use in access points. Since the shipped firmware of
access points (or wireless routers) does not allow users to customize them, several
free Linux-based firmwares (DD-WRT13, OpenWrt14, HyperWRT15, Tomato16)
were developed by the Internet community. We evaluated them and we decided
to use DD-WRT mega version 2.24, for its flexibility and because we were able to
successfully install it on an ASUS WL-500W wireless router. A cross compiler suite
(also called toolchain) is a set of tools that includes a compiler capable of creating
executable code for a platform other than the one on which the compiler is run.
Cross compiler tools are used to generate executables for embedded system or small
devices upon which it is not feasible to do the compiling process, because of the
limited resources (memory, disk space, computational power, etc). From the source
codes of different GNU packages (binutils, gcc, gþþ, glibc and uClibc), we created
the cross compiler suite that generates the appropriate executable code for our
ASUS WL-500W wireless router (MIPS ELF based wireless router). uClibc17 is a C
library for developing applications for embedded Linux systems. It is much smaller
than the GNU C Library (glibc), but nearly all applications supported by glibc also
work perfectly with uClibc as well. To test our cross compiler suite, we developed
several small applications for our wireless router and we observed that the code

13
http://www.dd-wrt.com
14
http://www.openwrt.org
15
http://sourceforge.net/projects/hyperwrt
16
http://www.polarcloud.com/tomato
17
http://www.uclibc.org
480 K. Velásquez and E. Gamess

generated by the cross compiler is much smaller when uClibc is used rather than
when glibc is used. This result is important due to the limited capacity of flash
memory of the wireless router, so we will use uClibc when possible, and glibc when
using advanced library functions not supported by uClibc.

Acknowledgment We would like to thank Candela Technologies for the trial version of their tool
(LANforge), which allowed us to include it in our study.

References

1. E. Gamess, K. Velásquez, IPv4 and IPv6 Forwarding Performance Evaluation on Different

Operating Systems, in XXXIV Conferencia Latinoamericana de Informática (CLEI 2008),
Sept 2008
2. S. Parker, C. Schmechel, Some Testing Tools for TCP Implementors. RFC 2398. Aug 1998
3. A. Botta, A. Pescapé, G. Ventre, On the Performance of Bandwidth Estimation Tools, in
Proceedings of the 2005 Systems Communications (ICW’05), Aug 2005
4. S. Ubik, A. Král, End-to-End Bandwidth Estimation Tools. CESNET Technical Report. 25/
2003. Nov 2003
5. D. Gupta, D. Wu, P. Mohapatra, C. Chuah, Experimental Comparison of Bandwidth Estima-
tion Tools for Wireless Mesh Networks. IEEE INFOCOM mini-conference, Apr 2009
6. J. Strauss, D. Katabi, F. Kaashoek, A Measurement Study of Available Bandwidth Estimation
Tools. IMC2003. Oct 2003
7. F. Montesino, Assessing active bandwidth estimation tools in high performance networks.
TERENA networking conference, June 2004
8. L. Mingzhe, M. Claypool, R. Kinicki, WBest: a bandwidth estimation tool for IEEE 802.11
wireless networks, in Proceedings of 33rd IEEE Conference on Local Computer Networks
(LCN), Oct 2008
9. V. Ribeiro, R. Riedi, R. Baraniuk, J. Navratil, L. Cottrell, pathChirp: efficient available
bandwidth estimation for network paths, in Proceedings of Passive and Active Measurements
(PAM) Workshop, Apr 2003
10. M. Jain, C. Dovrolis, Pathload: a measurement tool for end-to-end available bandwidth, in
Proceedings of Passive and Active Measurements (PAM) Workshop, Mar 2002
11. R. Prasad, M. Murray, C. Dovrolis, K. Claffy, Bandwidth estimation: metrics, measurement
techniques, and tools. IEEE Netw. 17(6), Nov-Dec 2003, 27–35 (Apr 2004)
12. A. Botta, A. Dainotti, A. Pescapé, Multi-Protocol and Multi-Platform Traffic Generation and
Measurement. INFOCOM 2007 Demo Session. May 2007
13. S. Bradner, Benchmark Terminology for Network Interconnection Devices. RFC 1242. July
1991
14. S. Bradner, J. McQuaid, Benchmarking Methodology for Network Interconnect Devices. RFC
2544. Mar 1999
15. C. Popoviciu, A. Hamza, G. Van de Velde, D. Dugatkin, IPv6 Benchmarking Methodology for
Network Interconnect Devices. RFC 5180. May 2008
Chapter 37
Hybrid Stock Investment Strategy Decision
Support System
Integration of Data Mining, Artificial Intelligence and
Decision Support

Chee Yong Chai and Shakirah Mohd Taib

Abstract This chapter discusses the continuous effort to explore stock price and
trend prediction from finance perspective as well as from the integration of three
major research areas namely data mining, artificial intelligence and decision support.
These areas have been explored to design a hybrid stock price prediction model with
relevant techniques into the stock price analysis and prediction activities.

1 Introduction

Throughout the centuries and in various countries, many people had tried to predict
the movement of share prices and beat the market but no one can really accurately
predict the movement of a particular share prices for company listed in the stock
exchange. No one knows the futures; hence no one knows what will happen to a
company in the future where such uncertainty caused the movement of the share price
unpredictable. The advances and great movements in computing give implications to
finance and economics today. New area of research known as financial engineering or
computational finance is highly mathematical and cross-disciplinary field which
relies on finance, mathematics and programming. Thus it enables financial analyst
to analyze data efficiently. Much work can be done using mathematical model in
spreadsheet program but the more demanding work needs one to write a computer
program. This chapter presents the study on designing a decision model for invest-
ment strategy by utilizing financial methods and computation techniques.
The movement of the share price is unpredictable. There was a study done to
estimate and predict the movement of share prices by using statistics, math and

C.Y. Chai (*)

Department of Engineering, Development Division, Petronas Carigali Sdn. Bhd., Tower 1,
Petronas Twin Towers, 50088 Kuala Lumpur, Malaysia
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 481

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_37, # Springer ScienceþBusiness Media B.V. 2010
482 C.Y. Chai and S.M. Taib

forecasting methods based on the historical share price data and this named as
Technical Analysis (TA). According to Rockefeller [1], TA has been around for
100 years and before the existence of computers, it was done only by professionals
who had access to real time price on the stock exchange.
There are tools and software created to automatically illustrates and generate
various graphs and perform various complex calculations in the TA. These tools
eliminate the haste of Financial Analyst in performing TA manually by paper
and pen in the past. With computer capabilities, now investors also perform TA
themselves with the help of appropriate software that help them in making their
trading and investment decision.
Other than assisting and making the TA process more efficient and effective,
IT field also brought a totally new perspective and thinking into stock analysis
through the researches and studies on the implementation of AI and DM concepts in
the stock analysis. For example, there were studies conducted on the combining the
Candlestick analysis (one form of the TA) with the Genetic Algorithms in stock
analysis [2]. In addition, Quah [3] stated that with the significant advancement in
the field of Artificial Neural Network (ANN), the generalization ability of the ANN
is able to create an effective and efficient tool for stock selection.

1.1 High Risk Investment

Investors face the highest risks compared to other form of financial investments
such as bond, Treasury bill and fixed deposit when they invest in the stock market.
Stock price fluctuates almost every second and most people see the share price
movement as unpredictable or in other term “random walk”. Some investors
involve in this activity mostly based on speculation where their aim is more to
obtaining capital gain rather than earning dividend as their investment return.
Investing based on speculation sometimes caused many individual investors had
their ‘hand burn’ where they lose their initial investment. Kamich [4] defined that
stock market is a private or public market for the trading of company stocks at an
agreed price. One of the main objectives of the company to be listed in the stock
market is to raise capital to fund future expansion.
If the investors able to purchase the correct share when the price at its lowest
range and it can be sold at a better price later, they will be able to earn big chunk of
money from the stock market. Of course, this is not easy to be done as no one able to
predict the future, as well as the movement of the share price accurately.

2 Finance Theories and Analysis in Stock Price Prediction

The attempt of analyzing the stock performance using financial theories can be
traced back up to late 1800s, where the oldest approach to common stock selection
and price prediction, TA is created and used [5]. Edward and Magee [6], stated that
37 Hybrid Stock Investment Strategy Decision Support System 483

throughout the years of stock market study, two distinct schools of thought has
arisen each with radically different methods of obtaining answers on what share to
buy and when to buy and sell the share. These two different schools of thought in
stock analysis are Fundamental Analysis (FA) and TA.
Jones [5] defined FA as a method of security valuation which involves analyzing
basic financial variables such as sales, profit margin, depreciation, tax rate, sources
of financing, asset utilization and other factors. Edward and Magee [6] notified that
FA relies heavily on statistics and people who performing FA will be looking
through the auditor’s reports, the profit-and-loss statement, balance sheet, dividend
records, and policies of the companies whose shares is under their observation.
They will also analyze business activities and daily news to estimate the company’s
future business condition. The investors who use FA will purchase stocks that are
viewed as underpriced by the analysts who believe that in the future, the share price
will increase to the level that it should be as computed in FA.
Another theory, TA is defined as the study of how securities prices behave
and how to exploit that information to make money while avoiding losses [1]. TA
focuses on the share price to assess and evaluate the demand and the supply for the
shares based on the market price itself and do not listen to chatter about securities.
Jones [5] claimed that technical analyst believes that the market itself is its own best
source of data. It is believed that all the investors’ reactions towards all the
information regarding the security already embedded in the share price.
According to Rockefeller [1], TA works because people constantly repeat
behavior under similar circumstances and TA is a forecasting method where it
uses past and current behavior (price) to predict future behavior (price). Concurrent
to this behavior repetition concept, Edward and Magee [6] noted that share price
move in trends and they tend to continue until something happens that will change
the supply and demand balance. As a conclusion, TA charts and analyzes the
historical share prices to reveal the pattern, formation and investor’s behavior and
interpret it to predict the possible future share price trend.

3 Data Mining (DM) and Artificial Intelligence (AI)

Generally, prediction of stock is a very difficult task as it behaves like ‘random

walk’ process and the prediction might be run off due to some unexpected news that
have direct impacts on the respective company. The obvious complexity of the
problem paved the way for the importance of intelligent prediction paradigms [7].
One of the most popular researches in this field is on the usage and implementation
of ANN in the stock market analysis. Kamruzzaman et al. [8] said that stock
analysis has been one of the most important applications of neural networks in
finance where numerous researches had been done on ANN in stock market index
prediction, stock performance/selection prediction, and stock risk prediction.
Quah [3] further explained that ANN generalization ability is able to infer the
characteristics of performing stocks from the historical pattern. ANN is used to
484 C.Y. Chai and S.M. Taib

discover the non-linear relationship between financial variables and stock price
which cannot be done by the computerized rule-based expert system. In stock
prediction using ANN, the input will be historical data such financial variables
and the output of it will be the performance of stock.
TA is about find the pattern in the historical data to predict the future price
movement, therefore researchers found DM is a suitable technique to reveal the
hidden pattern in the historical data and predict the price trend. Both Y.Y. Shi and
Z.K. Shi [9] has conducted a study on clustering technique (one of DM techniques)
in stock prediction. In their studies, they identified that clustering technique is one
of the efficient methods for stock investment analysis and they also noticed the
weakness of this technique in stock prediction where constructed cluster might not
completely reflecting the characteristics of the stock. In the study, the idea of rules
mining of stock market using the statistical pattern recognition is proposed in order
to overcome the weakness in clustering techniques. From techno-index in TA, Shi
et al. [9] developed an algorithm of fuzzy kern clustering and tested this idea on
Shanghai and Shenzhen Stock Market historical data since 1997. The result of the
study clearly indicates that statistic rules in essential trends of the stock market and
implied rules can be recognized to a certain degree through the construction of
proper clustering algorithm [9].
The authors’ proposed model for stock price prediction is based on Decision
Support System (DSS) model by Turban et al. [10]. AI and DM are included in one
of the subsystem which is Knowledge based subsystem.

4 DSS Model for Stock Investment Strategy

The authors have designed DSS model and architecture with some minor modifica-
tion for stock investment strategy. This model can be used to adapt AI and DM
elements into the system and utilizing TA and FA methods. The flow of the system
is shown in Fig. 1.

5 Architecture of Stock Investment Strategy Decision

Support System

Based on the architecture that shown in Fig. 2, the flow of the system activities can
be divided into two which reflect the two main processes of the system. The two
main processes of the system are analyzing the raw financial data from TA concept
(TA process) and analyzing the financial data from FA concept (AI process) with
both processes have the same aim of predicting stock price and trends. The first flow
of activities will be started from database where it provides raw financial data to
DM component in knowledge-based subsystem, the output from DM component
37 Hybrid Stock Investment Strategy Decision Support System 485

Technical Analysis (TA)

Evaluate three separate
components of TA

Get the historical

information showing the Amend the values of the
relative contribution of three components
each component in
making the final decision

Amend the values of all

Evaluate the fourth
components
component (Risk)

Financial Analysis (FA) Arrive at a predicted

share price

Fig. 1 The flow of Stock Investment Strategy Decision Support System

Model Management Subsystem

Technical
Recommendation Pattern and price
Analysis
Component prediction
Component

Artificial
Intelligent
Data Mining
Component
Component
(Fundamental
Analysis)

Knowledge-Based Subsystem

Raw Financial
Data

Fig. 2 The architecture of Stock Investment Strategy Decision Support System

486 C.Y. Chai and S.M. Taib

together with raw financial data from database will be sent to TA component
in model management subsystem for further processing and the output of TA
component will be the input for Recommendation Component. The second flow
of activities will be from raw financial data to AI component in knowledge-based
subsystem. Finally the output from AI component will be sent to Recommendation
Component in model management subsystem.
TA component focuses on illustrating historical stock price data into the form of
graph as well as other TA indicators to interpret investors and market sentiment
in the effort of predicting the stock price movement and trend. Among the most
popular charting techniques that being used in TA are Line Graph, Open-High-
Low-Close chart and Japanese Candlestick chart. The mechanism proposed here is
the system that able to indentify and recognized pattern directly from the raw
financial data using rules generated from the underlying reasoning and interpreta-
tion in various TA techniques. In achieving this, TA component of the system
will use rule-based TA model that is constructed from underlying rules in various
TA charting techniques.

5.1 DM Component

Rockefeller [1] highlighted that the same TA method would not work in the same
stock analysis all the time. The same TA method that proven working on a
particular stock does not necessary mean the same method will also work in
other stocks. Therefore, instead of identifying and interpreting patterns from the
raw financial based on the rules generated from one TA method only, the system
will identify the patterns based on underlying rules in various TA methods. This
will reduce the dependency on a single TA method in analyzing a particular stock
as well as provide better reasoning and analysis capability in estimating future
trend of the stock as we compare the same pattern using different rules from
different techniques which give higher accuracy of the prediction. This system
will use the association rules mining to determine the accuracy and effectiveness
of TA methods in predicting price movement and trend for a particular stock.
Therefore, the flow of activities and information in analyzing the stock in TA
concept will start from database and end at recommendation component as
illustrated in Fig. 3
The first phase of TA process will be the financial raw data being mined using
the association rules mining. With association rules mining, all the underlying rules
that reflecting the interpretation of TA in stock price and trend prediction will be
mined by calculating their respective support and confidence values for a particular
stock using the historical price of the stock. By using association rules mining, the
system will be able to identify which association rules that actually work on that
particular stock based on the historical data stored in the database based on the
confidence and support. Those confidence and support value together with raw
financial data will be the input for TA component.
37 Hybrid Stock Investment Strategy Decision Support System 487

Fig. 3 Activities and Raw Financial Data

Information Flow for TA
process

Data Mining Component

Association Rules Mining

TA Component
Rule-based TA Model

Recommendation Component

In the rule-based TA model, the recent historical stock price movement and its
related information will be analyzed by all association rules to identify pattern. If
recent historical price information triggered more than one association rules, the
output from the association rules will be taken into consideration based on weight
that are determined based on the support and confidence level. If the triggered
association rules have a confidence level that is more than 0.50 and it’s the highest
among other triggered association rules, that association rules will have the highest
weight. On the other hand, if an association rule has the confidence level below than
0.50, it means that although the rule has been triggered in the past, the actual result
is different from the expected result. The rule has wrongly predicted the result more
than 50% based on historical data. This indicates that the association rule does not
work well with the stock, hence this association rule will receive lower weight.
Besides that, if an association rule is triggered has the support value of 0, it means
the rule never being triggered before in the past and the system will assign a new
confidence value of 0.5 for the associate rule because since the rule never triggered
before, hence the rule will have a 50% chance that it will predict correctly. If all the
triggered associate rules have confidence level below than 0.50 the system will notify
the user that prediction might have higher risk to be incorrect and the system would
still provide the stock trend prediction. If from all association rules, the predicted
stock trend is a mixture of up and down, the system will compute the likelihood of the
expected trend to be happening based on the expected trend that has the highest
accumulated weight which this process will be done by the Inference Engine in the
TA Component.

5.2 TA Component

There are six main sub-components included in TA component as shown in Fig. 4.

488 C.Y. Chai and S.M. Taib

Raw Financial
Data

Stock Trend Stock Price Stock Trading Stock Volatility

Evaluation Evaluation Signal Evaluation
Component Component Component Component

Output from
Back-test
Data Mining
Component
Component

Inference
Engine

Fig. 4 Processes in TA Component

5.2.1 Stock Trend Evaluation Component

The first process would be in the Stock Trend Evaluation Component where the
latest two weeks historical data is used to sun through the underlying rules derived
from the Open-High-Low-Close chart and Japanese Candlestick chart. This process
will detect the pattern, gauge the investor’s sentiment on the stock and discover the
current trend of a particular stock. This component has the most association rules
compared to other components. The output for this process will be the trend of the
particular stock in the near future which will be uptrend, downtrend, indecision or
no trend. Example of association rules from this component illustrated in Fig. 5.

5.2.2 Stock Price Evaluation Component

This component will evaluate either the stock has been overvalued or undervalued
using the Williams %R Indicator and Relative Strength Index. Williams %R
indicator is developed with the aim to assess either the particular stock is being
overbought or oversold. There will be two outputs from two indicators which
the result might be contradicted with one to another. This contradiction issue will
be sorted out in back-test component. The output for this process is the evaluation
37 Hybrid Stock Investment Strategy Decision Support System 489

Rule 1:- If the "high" is higher than previous day’s "high" for 3 consecutive day AND
"low" is higher than previous day’s "low" for 3 consecutive days AND
"close" is higher than previous day’s "close" for 3 consecutive days THEN
the stock is in uptrend.

Fig. 5 Sample of association rule from Stock Trend Evaluation Component

Fig. 6 Sample of association Williams %R Indicator

rule from Stock Price Rule 1:- If the today’s Williams %R is above –20%
Evaluation Component THEN the stock is overvalued.

Relative Strength Index

Rule 1:- If the today’s RSI is above 70%
THEN the stock is overvalued.

on the stock either it’s overvalued, undervalued or reasonable. A sample of associ-

ation rules used in this component is shown in Fig. 6

5.2.3 Stock Trading Signal Component

Another component is used to identify any buying or selling signal generated based
on TA for that particular stock. There will be six TA methods being used to trace the
trading signals which are Weighted Multiple moving Average, Momentum, Engulfing,
Doji, Long Shadow and Gap. Unlike ‘Stock Price Evaluation Component’,
association rules in ‘Stock Trading Signal Component’ are not being triggered
all the time. There is high tendency that none of the association rules in this
component is being triggered most of the time because trend reversal itself does
not happen all the time and this component focuses on tracing trend reversal
that leads to the generation of trading signal. The output of this component will
be trading signal and forecasted stock trend. Trading signal is not generated
everyday for a particular stock and the same goes to the forecasted stock trend.
Some of the association rules from various indicators used in this component is
illustrated in Fig. 7.

5.2.4 Stock Volatility Component

The forth component in the TA component is used to evaluate the risks and the
noise level of a particular stock. Volatility is a measure of price variation which is
directly linked to risk level and noise level. Stock with high volatility means trading
is riskier but has more profit potential, while low volatility means less immediate
risk. Therefore, this component serves to evaluate the stock’s volatility level and
provides a warning if the volatility level is high because high volatility will cause
the system’s forecast and prediction less accurate due to the high noise level in the
stock movement.
490 C.Y. Chai and S.M. Taib

Weighted Multiple Moving Average

Rule 1:- If 5-days Moving Average decline and moves towards 10-days Moving Average
AND the trend is uptrend
THEN the uptrend will continue.

10-days Momentum Indicator

Rule 1:- If today’s momentum higher than yesterday’s momentum
AND yesterday’s momentum is above 100
AND last 2 day’s momentum is below 100
THEN buying signal is generated.

Fig. 7 Sample of association rule from Stock Trading Signal Component

There are two indicators embedded in this component which are Average True
Range and Average Volume Indicator. The Average True Range is a useful measure
of volatility. The increase in Average True Range signifies that highs and lows
are getting farther apart which means volatility is rising and risk has increased. The
Average Volume Indicator is basically an indicator that shows the average of the
stock’s volume over a period. Volume is an important confirming indicator for any
forecasts especially trading signal where any trading signal generated without
accompanied by signification increase in volume is normally a false trading signal.
The output from this component will be the percentage of changes in the today’s
Average True Range as compared to the result from the previous day and the
percentage of changes in volume as compared to 10-days average volume.

5.2.5 Back-Test Component

Back-test component is one of the main parts in the TA component. Output from
the various components explained earlier might have contradicted forecasts that
even occurred within a particular component itself. Back-test component will
weight each output from all of the TA indicators based on the performance of
those indicators for that particular stock in the past where indicators that performed
the best in the past will be given the highest weight. This component receives an
important input from the DM component which is the support and confidence values
of each of the association rules used in the three components explained earlier based
on the historical stock price. All the outputs from Stock Trend Evaluation Compo-
nent, Stock Price Evaluation Component and Stock Trading Signal Component
together with their respective weights are sent to Inference Engine.

5.2.6 Inference Engine

Inference Engine will produce the final decision on the various different forecasts
and predictions generated by other respective components. Inference engine works
37 Hybrid Stock Investment Strategy Decision Support System 491

based on all the forecasts and their respective weights. If there are contradict
forecasts incurred, those forecasts with higher weight will be chosen as the
decision. All of the outputs from four other components will be included because
each of the indicators in all the components are complementary to each other.
Verification of forecast done by indicators is very important to reduce the unprofit-
able trade resulted from false trading signal generated by a particular indicator. A
final recommendation will be generated by Inference Engine that includes Trading
Signal (Trading Signal is not generated everyday), Forecasted Stock Trend, Stock
Price Evaluation and risks level. Provided together is the possibility level for the
generated trading signal, forecasted stock trend and stock price evaluation which is
derived from the confidence level from each indicators.

5.2.7 AI Component

TA techniques exclusively cannot recognize the non-linear relationship between

financial variables and stock price. Therefore, AI component that particularly using
ANN will be included in the system. This ANN is based on the concepts in FA.
With the inputs of ANN which are extracted from raw financial data, the system
would be able to discover the non-linear relationship between financial variables
and stock price. Furthermore, by integrating FA concept into ANN, the system is
able to compare and back-test the stock price and trend prediction done by TA to
achieve higher prediction accuracy than only use TA or FA in ANN alone.
The inputs for the ANN will be from company and economic factors. Inputs
from company factor will be Return on Equity, Historical P/E Ratio, Prospective
P/E ratio, Cash flow yield, Payout Ratio, Operating Margin, Current Ratio and Debt
to Capital at book. As for economic factor, the inputs will be Inflation, Interest rate,
GDP Growth. Hence, the output from the ANN in the AI component will be the
forecasted stock price in the next 4 month. The forecasted stock price will be in the
format of percentage of increase or decrease such as less than 5%, between 5% and
10% or above 10% from the base stock price which is the price of the day the
forecast is computed by ANN.

5.2.8 Recommendation Component

Both outputs from TA component and AI component are the inputs for the Recom-
mendation component. Forecasted Stock Trend, Stock Price Evaluation, risks level
and Trading Signal are the inputs obtained from TA component while 4-month
target stock market price is the input obtained from AI component. The main reason
that the output from AI does not undergo the back-test component like in TA
component is because the ANN itself has been trained using historical financial
data before producing the 4 month target stock price prediction. This component
task is to assess whether the 4 month target stock price is achievable or not based on
the predicted near future stock trend, evaluation of the current stock price as well as
492 C.Y. Chai and S.M. Taib

trading signals. Hence, it actually integrates both long term and short term forecast
for the stock and present it to the user. The final outputs from this component will be
the predicted near future stock trend to enable user to know the current and possible
trend of the stock with the historical accuracy that the prediction accurate, stock
price evaluation on either the stock price is reasonable, undervalued or overvalued
and with the historical accuracy, risks level to alert users on the possibility that the
system’s prediction is wrong due to high level of volatility, trading signal (buy or
sell signal) if there are trading signal generated for that stock and with historical
accuracy and finally the 4 month targeted stock price.
The recommendation will be provided together with all the outputs given by
various TA indicators and their historical accuracy in presenting the fine-grained
information to the users for their use in making trading decision making concurrent
with the role of the system as a Decision Support System. The system does not
dictate, but recommend. Any final decision is based on the user’s judgment and
evaluation themselves.

6 Conclusion

A study on the FA and TA has been done in order to understand how stock price
prediction works in Finance world as well as to create the TA Model and ANN
based on FA. The further study on DSS model and architecture is done to ensure
Stock Investment Strategy Decision Support System follows the actual basic
framework and model of a DSS. The combination of AI and DM is applied in
this hybrid model particularly in the knowledge-based subsystem to generate better
analysis of the historical data. The output from the system will be the result analysis
and recommendation to the user. The analysis will be generated by the TA compo-
nent and AI component where it will be presented to the users to aid them in their
trading decision making with the presentation of this fine-grained information.
However, it will only present the information and outputs derived from various
TA and FA techniques used in the system. The recommendation component will
judge, evaluate and provide trading recommendation based on the same outputs that
are presented to the user in the analysis.

References

1. B.Rockefeller, TA for Dummies (Wiley Publishing, New York, 2004)

2. P.Belford, Candlestick Stock Analysis with Genetic Algorithms (ACM Digital Library, 2006)
3. T.S.Quah, Improving Returns on Stock Investment through Neural Network Selection
(Nanyang Technical University, Singapore, 2006)
4. B.M.Kamich, How TA Works (New York Institute of Finance, New York, 2003)
5. C.P.Jones, Investments, 14th edn. (John Wiley & Sons, 2007)
6. R.D.Edward, J. Magee, TA of Stock Trends, 8th edn. (St. Lucie Press, FL, 2001)
37 Hybrid Stock Investment Strategy Decision Support System 493

7. A.Abraham, B. Nath, P.K. Mahanti, Hybrid Intelligent Systems for Stock Market Analysis
(Springer-Verlag Berlin Heidelberg, 2001)
8. J.Kamruzzaman, R.Begg, R. Sarker, Artificial Neural Networks in Finance and
Manufacturing (Idea Group Publishing, Hershey, PA, 2006)
9. Y.Shi, Z. Shi, Clustering Based Stocks Recognition (ACM Digital Library, 2006)
10. E.Turban, J.E.Aronson, T.P.Liang, R. Sharda, Decision Support System and Intelligence
Systems (Pearson Education, Upper Saddle River, NJ, 2007)
Chapter 38
Towards Performance Analysis of Ad hoc
Multimedia Network

Kamal Sharma, Hemant Sharma, and Dr. A.K. Ramani

Abstract A primary challenge in mobile ad hoc network scenario is the dynamic

differentiation of provided levels of Quality of Service (QoS) depending on client
characteristics and current resource availability. In this context the paper focuses on
characterizing the average end-to-end delay and maximum achievable per-node
throughput for In-vehicle ad hoc multimedia network with stationary and mobile
nodes. This work considers an approximation for the expected packet delay based
on the assumption that the traffic at each link acts as an independent M/D/1 queue.
The chapter establishes the network model, throughput model and delay model for
ad hoc multimedia network. It further presents analysis and evaluation of the delay
model based on an experimental multimedia communication scenario.

During the last years, the evolutionary development of in-car electronic systems has
led to a significant increase of the number of connecting cables within a car. To
reduce the amount of cabling and to simplify the networking of dedicated devices,
currently appropriate wired bus systems are being considered. These systems are
related with high costs and effort regarding the installation of cables and accessory
components. Thus, wireless systems are a flexible and very advanced alterative to
wired connections. However, the realization of a fully wireless car bus system is
still far away. Though, cost-effective wireless subsystems which could extend or
partly replace wired bus systems are already nowadays conceivable. A promising
technology in this context is specified by the latest Bluetooth standard [1].
In the wireless networks, Bluetooth is one of the communication technologies
emerging as the de facto standard for “last-meter” connectivity to the traditional
fixed network infrastructure [2]. Bluetooth-enabled portable devices can interconnect
to form a particular incarnation of Personal Area Networks (PAN) called piconet,
which consists of one master and up to seven slaves. The master device has direct

K. Sharma (*)
Institue of Computer Science and Information Technology, Devi Ahilya University, Khandwa
Road, Indore, India
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 495

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_38, # Springer ScienceþBusiness Media B.V. 2010
496 K. Sharma et al.

visibility of all slaves in its piconet and can handle three types of communication
services, which in Bluetooth terminology are called logical transports: unicast packet-
oriented Asynchronous Connection-oriented Logical transports (ACL), broadcast
packet-oriented Active Slave Broadcast (ASB), and circuit-oriented Synchronous
Connection Oriented (SCO). Bluetooth logical transports have very different quality
characteristics, e.g., in terms of reliability and maximum throughput. It is also vital to
predict actual performance of Bluetooth systems so that it can be applied to the design
of a system.
The use of Bluetooth has been restricted to (safety) uncritical applications in
and around the car, e.g. communication and infotainment applications or applica-
tions regarding the exchange of pure car-specific information (e.g. control and
status data of the body electronics). Inside a car, Bluetooth allows the wireless
connection of a variety of car-embedded electronic devices such as control panels,
small displays or headsets with other electronic in-car systems. But beyond repla-
cing wires between in-car devices, the scope of the application of Bluetooth in
automotive environments is to ensure wireless interoperability between mobile
devices (e.g. PDA, notebook, mobile phone) and car-embedded devices (e.g.
navigation system, car radio, car-embedded phone). The main principle behind is
to allow devices to co-operate and share resources, and to control mobile devices
by using the car embedded user interface, or the other way round, to access car-
embedded systems by mobile devices. In this way, the functionality of an in-car
electronics environment can be greatly extended, and beyond that, the personaliza-
tion of the car and its services can be realized.
In addition, current implementations of the Bluetooth software stack do not
allow applications to exploit the limited performance functions included in the
specification in a portable way. The result is that the development of Bluetooth
operations in multimedia ad hoc applications currently depends on specific imple-
mentation details of the target Bluetooth hardware/software platform. This rele-
vantly complicates service design and implementation, limits the portability of
developed applications, and calls for the adequate modeling of performance para-
meters corresponding to potential ad hoc applications and services.
This chapter discusses the model and analysis for common ad hoc network
performance attributes for pervasive in-vehicle multimedia services for packet
based data communication.

1 In-Vehicle Multimedia Network

An ad hoc multimedia network in a car and participating devices are shown in

Fig. 1. The diagram presents an ad hoc network containing a smart phone, an
infotainment system, and rear seat entertainment system. The devices are equipped
with Bluetooth module and are capable of establishing a Bluetooth connection. A
piconet can be established between Infotainment system and iPhone when the
driver or any other occupant of the vehicle pairs the phone. Similarly a piconet
could be established between Infotainment System and Rear Seat Entertainment
38 Towards Performance Analysis of Ad hoc Multimedia Network 497

Infotainment System Rear Seat

iPhone Entertainment
piconet

piconet

Rear

Rear
Central Unit

Rear

Rear

Fig. 1 Typical pervasive vehicle network with multimedia devices

system. The piconet shall enable sharing of iPhone contents or access to contents
from internet, if appropriate application framework is available at the infotainment
system.
The network contains a pure multimedia part, represented by Multimedia Bus,
and a control and external communication part, represented by Networking for
Automotive. Multimedia Bus represents the wired multimedia resources. External
multimedia resources can form an ad hoc network via the communication interfaces
of the vehicle network.
Bluetooth technology is based on a master-slave concept where the master
device controls data transmissions through a polling procedure. The master is
defined as the device that initiates the connection. A collection of slave devices
associated with a single master device is referred to as a piconet.
The master dictates packet transmissions within a piconet according to a time-
slot process. The channel is divided into time slots that are numbered according to
an internal clock running on the master. A time division duplex (TDD) scheme is
used where the master and slaves alternatively transmit packets, where even
numbered time slots are reserved for master-slave transmissions, while odd num-
bered time slots are reserved for slave-master transmissions.
Figure 2 provides an overview of the multimedia PAN inside the car with the
role of different nodes. The piconet here consists of a master, the infotainment
device, and two slaves, the smart phone (iPhone) and rear-seat entertainment
device. As soon as the infotainment system is up and running, its Bluetooth module
is ready to pair with other available device. Pairing with iPhone or with rear-seat
entertainment unit or with both establishes the network.
498 K. Sharma et al.

Master

Infotainment System

iPhone

Rear Seat Entertainment

Slave 2
Slave 1

Fig. 2 In-car ad hoc multimedia PAN

1.1 System Architecture

The network organizes infotainment application contents into communication

channels to facilitate identification and the communication of intended data
streams. Therefore, every application subscribes to the channels that it is interested
in, and his corresponding device node will try to retrieve any contents belonging to
those channels. The framework follows the approach used in Internet-based pod-
casting protocols and structures channels into different media streams. To make
efficient use of contacts with a small duration, the streams are further divided into
data packets, transport-level data units of a size that can typically be downloaded in
an individual node encounter. The data packets are further organized into protocol
data units (PDU), the atomic transport unit of the network.
The system architecture for ad hoc multimedia network is illustrated in Fig. 3.
The transport layer acts directly on top of the link-layer without any routing layer.
To distribute contents among the communicating nodes, the framework does not
rely on any explicit multi-hop routing scheme. Instead of explicitly routing the data
38 Towards Performance Analysis of Ad hoc Multimedia Network 499

Application Layer

Media Stream
Content
Application Synchronization Information Video
Service Interface MP3

Transport Layer

Media Stream
Node Discovery Application Transport Packets
Transport Handler Services

Communication Paired
[Broadcast / Multicast] Connections PDU Size

Data Link Layer

Fig. 3 Overview of ad hoc multimedia network

through specific nodes, it relies on a receiver-driven application-level dissemination

model, where contents are routed implicitly as nodes retrieve contents that they
request from neighboring nodes. It distinguishes between an application layer, a
transport layer, and the data link layer. In a first level of aggregation, the applica-
tions organize their data contents in media stream channels.
Below the application layer, the transport layer organizes the data into data
streams. Streams are smaller data units that should be able to communicate over
short contacts. The use of smaller file blocks is also supported by the idea of
integrating forward error correction, for instance the use of fountain codes, to
speed up and secure the data transfer, specifically when packets are received
unordered. The streams packets themselves are then again cut into smaller parts
to optimize the interaction with the data link layer; i.e., the size is set to the PDU
size of the data link layer.
The proposed system is designed to work on any MAC architecture, however,
to be effective even in the presence of short contact durations, short setup times
and high data rates are important for achieving high application communication
throughput.
500 K. Sharma et al.

1.2 Application Scenarios

The ad hoc multimedia network consisting of Infotainment system and the slave
devices shall provide following services:
l Access to audio contents from iPhone to Infotainment system that could be
played and hearable on the vehicle’s sound system
l Access to video contents from iPhone to rear-seat entertainment unit via the
piconet master.
l Access to internet from rear-seat unit using iPhone via the piconet master
l Applications based on Information contents received by the iPhone to help safe
driving, such as weather information or traffic information

2 Performance Modelling

The main parameters that can influence the performance of the network described
in previous section are, basically,
l The number of application channels
l The offered traffic per application
l The application profile, defined by the objective parameters (thresholds) and
weights
The performance indicators that should be considered for this network are: the
packet delay, the packet error rate, the latency, and the throughput. Packet
transmission and reception are synchronized by the system clock (625 per s slot
interval). This paper focuses on throughput and delay performance in this point-
to-multipoint communication environment. We try to simplify and abstract the
system as much as possible while preserving the essence of the network which
affects the actual performance of the system.

2.1 Network Model

The model considered here is that of a wireless ad hoc network with nodes assumed
either fixed or mobile. The network consists of a normalized unit area torus contain-
ing n nodes [9]. For the case of fixed nodes, the position of node i is given by Xi. A
node i is capable of transmitting at a given transmission rate of W bits/s to j if,






Xk  Xj
 ð1 þ DÞ
Xi  Xj
;

Where D > 0, such that the node Xk will not impede Xi and Xj communication.
This is called the protocol model [9].
38 Towards Performance Analysis of Ad hoc Multimedia Network 501

For the case of mobile nodes, the position of node i at any time is now a function
of time. A successful transmission between nodes i and j is governed again by above
equation, where the positions of the nodes are time dependent. Time is slotted to
simplify the analysis. Also, at each time step, a scheduler decides which nodes are
sources, relays, or destinations, in such a manner that the association pair (source-
destination) does not change with time.
The multimedia network presented in previous section resembles to multihop
wireless ad hoc network, and therefore, can be modeled as a queuing network as
shown in Fig. 4a. The stations of the queuing network correspond to the nodes of the

a
p51(n)

p41(n)

4
2

p21(n) p14(n)

p12(n) p15(n)
p31(n)

5
3
p13(n)

Representation of wireless ad hoc network as a queuing network

Filter for absorbing 1-p(n) λi

FCFS
packets destined Σ Queue
Xi
to the nodes

λ
p(n)

Packets Generated
Packets Absorbed

Representation of Node

Fig. 4 Network model for ad hoc multimedia network

502 K. Sharma et al.

wireless network. The forwarding probabilities in the queuing network, denoted by

pij, correspond to the probability that a packet that is transmitted by node i enters the
queue of node j. Figure 4b shows a representation of a node in the ad hoc network as
a station in the queuing network.
The delay in a wireless network equals the sum of queuing and transmission
delays at source and intermediate nodes. We will use the queuing network model,
shown in figures above, in order to mathematically analyze the packet delay.

2.2 Packet Delay Model

In order to model the delay, we consider one to five ad hoc multimedia applications
with ACL connections in the piconet. In this scenario, packet types are adaptively
selected among DM 1, DM3 and DM5 according to the amount of data waiting for
transmission in buffers. We consider only DM packets assuming that data packets
are vulnerable to bit errors due to error-prone wireless channel characteristics. To
ensure reliable data transmission, FEC (Forward Error Correction) and ARQ
(Automatic Repeat Request) [1] are employed. We simulated simple wireless
channel model, where BER (Bit Error Rate) is constant in a Simulation.
The evaluation of the probability of the queuing delay exceeding a threshold
Dth, the delay threshold, starts from the expression of the delay

X
Pq X
Nactive ðiÞ1
D ¼ Lij
i¼0 j¼0

where pq denotes the number of packets in the queue of the desired application,
Nactive(i) is the number of active applications during the transmission of packet
i, and lij is the length of the packet transmitted by application j during the ith round.
The limiting effect to the delay, for very high values, is determined by the
control imposed to the incoming traffic when the buffer occupancy gets close to
its limits.

2.3 Throughput Model

Taking into account that the main purpose of this ad hoc network is to provide
access to multimedia contents, the throughput model has to reflect the character-
istics of the network traffic. Throughput performance is evaluated when there are
one or more slave device applications, for which an ACL or SCO connection is
established.
It is assumed that packets are always transmitted and received in the cor-
responding slots. In other words, the master device polls a slave device with an
38 Towards Performance Analysis of Ad hoc Multimedia Network 503

ACL connection in a round-robin fashion and it transmits a packet to the polled

slave. When the slave device is polled by the master, the slave sends a packet to the
master in the consecutive TX slot of the reception of the packet from the application
at master. Various packet types can be used for ACL connections when there are no
SCO connections in a piconet. Only DHl or DM1 packets are assumed to be
transmitted if at least one ACL connection is present in a piconet. In order to see
the maximum possible performance, asymmetric TX is considered, where different
packet types are used for down (master to slave) and up (slave to master) stream
traffic.
A buffer analysis can be conducted for the network throughput resorting to the
classical queuing theory [2] since the arrival process is modelled by an Interrupted
Poisson Process (IPP). The number of packets in the queue is well approximated by
a geometric random variable with parameter (1  r), where r is the activity
coefficient [3] given by the ratio between the arrival rate, given by the sum of the
arrival rates of the Nactive active applications in the piconet and the service rate m:

ðNactive Þl
r¼
m

To take into account the effects of a finite buffer, a feedback is considered,

blocking the incoming traffic whenever the buffer occupancy exceeds 90% of its
capacity. In the simulations, the buffer capacity is considered at 256 Kb.

3 Performance Evaluation

The performance of network has been evaluated when the applications on piconet
nodes and transferring audio or data packets. We present here the analysis based on
data packets.

3.1 Simulation Setup

The number of applications in the modelled network is relatively small and they are
assumed to transfer data traffic between the nodes in the piconet. For the data traffic,
this combination is expected to result in burst traffic streams in the network. The
voice traffic was simulated using a measured trace of PCM audio stream, but not
considered here.
Transmitted packets may be lost due to bit errors and is modeled with a constant
loss probability. All lost ACL packets are resent according to the Bluetooth ARQ
scheme. In all the simulations presented herein, the packet loss probability was set
to 103.
504 K. Sharma et al.

3.2 Delay Analysis

Average packet delay, time between arrival of a new packet (data) into a buffer for
transmission and reception of the packet at a receiver, is measured for each of traffic
load as the number of ACL connections changes.
Figure 5 presents the packet latency as a function of the offered traffic per
application for the network, where the number of applications is limited to 50,
the objective thresholds are packet error ERRth ¼ 103 and delay threshold Dth ¼
10 ms, and the average duration of the vertical handover procedure is assumed to be
100 ms.
The weights reflect the relative importance of each objective, according to the
type of service. The objective values are used as threshold to determine a perfor-
mance dissatisfaction value U on the basis of the weights assigned to the profile
parameters.
For low traffic and a profile where packet error EER has higher priority, the delay
can even exceed that of the abstract Bluetooth system; otherwise, the delay lies
between the performance of the ad hoc network and abstract Bluetooth network.

3.3 Throughput Analysis

The throughput performance of the system is considered without SCO connections

as the number of ACL connections increases. In terms of throughput, in Fig. 6, the

102
Bluetooth
Uth =.2 WD =.3 WEER =.3
Uth =.5 WD =.3 WEER =.3
101 Uth =.3 WD =.5 WEER =.5
Packet delay D [s]

100

10–1

10–2 EERth = 1e -3
Dth = 10 ms

10–3
105 106
Packet traffic [bit / s]

Fig. 5 Packet delay as function of buffered packet traffic

38 Towards Performance Analysis of Ad hoc Multimedia Network 505

106

EERth =1e-3
Dth = 10 ms

105
Throughput [bit /s]

104

Uth = .2 WD = .3 WPER = .3
Uth = .5 WD = .3 WPER = .3
Uth = .3 WD = .5 WPER = .5
103
105 106
Packet traffic [bit / s]

Fig. 6 Throughput as function of buffered packet traffic

available bandwidth is shown as a function of the offered traffic per application,

comparing the three scenarios, with packet EERth ¼ 103 and Dth ¼10 ms. Again,
the number of applications in the network is considered to be around 5 and other
parameters have the same values as that for packet delay measurement.
Figure 6 below indicates an improvement in the throughput can be obtained by
choosing a high weight for the packet error EER, thus forcing a control on the
reliability of the transmission.
On the other hand, if the focus is on the delay, the performance in terms of
available throughput drops to low values in correspondence to a high offered traffic
load. Moreover, note that, in this case, a control on the packet error rate and delay is
performed.

4 Summary

This chapter presented the network model, delay model and throughput model for
ad hoc In-car multimedia network. We have modelled and evaluated the perfor-
mance of the multimedia piconet in Bluetooth supporting point-to-multipoint
communications. Our performance models provide a performance metric when
ACL slave applications are supported. An approximation for the packet error
probabilities has been used and the influence of the main parameters on the network
performance has been studied. In order to guarantee low delay requirements, the
number of application running simultaneously may need to be limited.
506 K. Sharma et al.

These results can serve as valuable inputs to the optimal design of Bluetooth
based in-car multimedia applications or interfaces with special throughput and
delay requirements.

References

1. Bluetooth 3.0 Specifications, www.bluetooth.com

2. L. Kleinrock, Queuing Systems (Wiley, 1975)
3. G. Lin, G. Noubir, R. Rajaraman, Mobility models for ad hoc network simulation, in Proceed-
ings of the IEEE INFOCOM Conf., Mar 2004, pp. 454–463
4. Q. Zhang, W. Zhu, Y.-q. Zhang, End-to-end QoS for video delivery over wireless internet.
Proc. IEEE 03 93(1), 123–134 (2005)
5. K.K. Sharma, H. Sharma, Dr. A.K. Ramani, mCAR: software framework architecture for
in-vehicle pervasive multimedia services. Proc. IMECS I, 1100–1105 (2009)
6. K.K.[AU3] Sharma, H. Sharma, Dr. A.K. Ramani, Performance analysis of ad hoc multimedia
services in automotive environment, in Proceedings of WCECS, 2009, pp. 218–222
7. X. Masip-Bruin, M. Yannuzzi, J. Domingo-Pascual, A. Fonte, M. Curado, E. Monteiro,
F. Kuipers, P. Van Mieghem, S. Avallone, G. Ventre, P. Aranda-Guti´errez, M. Hollick,
R. Steinmetz, L. Iannone, K. Salamatian, Research challenges in QoS routing. Comput.
Commun. 29, 563–581 (2006)
8. X. Zhou, J. Wei, C.-Z. Xu, Quality-of-service differentiation on the internet: A taxonomy.
J. Netw. Comput. Appl. 30(1), 354–383 (2007)
9. J. Li, C. Blake, D.S.D. Couto, H.I. Lee, R. Morris. Capacity of ad hoc wireless networks, in
MobiCom ’01: Proceedings of the 7th annual international conference on Mobile computing
and networking (ACM Press, New York, 2001), pp. 61–69
Chapter 39
Towards the Performance Optimization
of Public-key Algorithms Using Fuzzy
Modular Arithematic and Addition Chain

Gopinath Ganapathy and K. Mani

Abstract In most of the public-key cryptosystems like RSA, ElGamal, etc.;

modular exponentiation plays a vital role for performing encryption/decryption
operations. In other public-key cryptosystems like ECC, scalar point multiplication,
kP where k is an arbitrary integer in the range 1 < k < ord(P) and P is a point in
the elliptic curve is the central operation. In cryptographic algorithms, exponent
is always an integer and can be performed faster than the traditional square and
multiply method by iteratively reducing the small gain may be made if the numbers
of multiplications are organized properly. For that some integer can be represented
in the form of sum of squares and based on the sum of squares larger exponent
e can be reduced into smaller one. Then, the addition chain is used to minimize
the number of multiplications in the smaller exponent to speed up the operations.
Similarly, in the case of ECC to speed up kP fuzzy modular arithmetic is considered.

1 Introduction

Data security and cryptographic techniques are essential for safety relevant appli-
cations. Cryptography keeps the message communication secure so that eavesdrop-
pers cannot decipher the transmitted message. It provides the various security
services like confidentiality, integrity, authentication, and non-repudiation. There
are several cryptographic algorithms available for both symmetric and public-key
cryptosystem [1]. Slow running cryptographic algorithms cause customer dissatis-
faction and inconvenience. On the other hand, fast running encryption algorithms

G. Ganapathy (*)
Professor and Head, Department of Computer Science, University of Bharthidasan, Trichy,
620 024, India
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 507

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_39, # Springer ScienceþBusiness Media B.V. 2010
508 G. Ganapathy and K. Mani

lead high speed. To achieve this, fast exponentiation methods and high speed
custom hardware devices are needed.
Public-key cryptosystems like RSA, ElGamal often involve raising large
elements of some group fields to large powers. The performance and practicality
of such cryptosystems is primarily determined by the implementation efficiency of
the modular exponentiation. Fast exponentiation is becoming increasingly impor-
tant with the widening use of encryption. In order to improve the time requirements
of the encryption process, minimizing the number of modular multiplication is
essential. For example, to compute C ¼ xe mod n, based on paper-and-pencil
method, it requires (e  1) modular multiplication of x. i.e. x–x2–x3–xe1–xe.
Several authors have proposed the fast modular exponentiation like left-to-right,
right-to-left, multi-exponentiation by Interleaved exponentiation, sliding window
and window NAF exponential methods etc. [2]. But in this paper, in order to reduce
larger exponent into smaller an alternative approach based on sum of squares is
thought of because an integer is represented as several forms of sum of squares
Then, the addition chain is used to reduce the number of multiplications.
ECC was proposed in 1985 by Neal Koblitz and Victor Miller. It is an alternative
to the established cryptosystems like RSA, ElGamal, and Rabin etc. It guarantees all
the security services with the shorter keys. The use of shorter length implies less
space for key storage, less arithmetic cost and time saving when keys are transmitted.
These characteristics make ECC the best choice to provide security in wireless
networks. ECC has increased as evidenced by its inclusion in standards by in credited
standards organizations such as American National Standards Institute (ANSI),
Institute of Electrical and Electronic Engineers (IEEE), International Standards
Organization (ISO), and National Institute of Standards and Technology (NIST).
The security of ECC is based on the discrete logarithm problem over the points on an
elliptic curve. But, it is more difficult problem than the prime number problem of
RSA algorithm [3]. These two problems are closely related to the key length of
cryptosystems. If the security problem is more difficult, then smaller key length can
be used with sufficient security. This smaller key length makes ECC suitable for
practical applications such as embedded systems and wireless applications [4].
A vast amount of research has been conducted on secure and efficient imple-
mentation since then. All focus on scalar point multiplication that is kP, where k is a
scalar and P is a point on elliptic curve. Efficient hardware and software implemen-
tation of scalar point multiplication is the main research topic on ECC in recent
years. To encrypt and decrypt the message using any public key cryptosystem
like ECC, modular arithmetic plays a vital role and it is the key operation in modern
cryptology. Even though, many algorithms have been developed for faster imple-
mentation, none of the work reveals the use of fuzzy modular arithmetic to speed up
the encryption and decryption operations on ECC.
Fuzzy modular arithmetic was proposed by Wael Adi. The key idea used in
fuzzy modular arithmetic is not to compute the result exactly as in the traditional
modular arithmetic because the traditional modular arithmetic uses division
operation for modulo reduction m. Thus, instead of the full reduction, a pseudo
fuzzy randomized partial reduction is performed.
Towards the Performance Optimization of Public-key Algorithms 509

The rest of the paper is organized as follows. A brief introduction to sum of

squares, addition chain, Fermat’s Theorem and EC is provided in Section 2. A short
description of fuzzy modular arithmetic is presented in Section 3. Section 4 describes
the applications of sum of squares, and addition chain to reduce the number of
multiplication in modular exponentiation. Section 5 presents the application of
fuzzy modular arithmetic, and Fermat’s theorem to speed up the operation of scalar
multiplication in ECC. Finally, we draw our conclusion in Section 6.

2 Concept of Sum of Squares, Addition Chain, Elliptic Curve,

and Fermat Theorem

In this section, we present the representation of some integers based on the sum of
two, three, and four squares, the concept of addition chain which is used to find the
minimal number of multiplications required for an exponent, the mathematical
preliminaries of elliptic curve, addition of points in elliptic curve, and Fermat’s
theorem to find the modular inverse of an integer.

2.1 Sum of Squares

In this section, we explore integers that may (or may not) be expressed as a sum of
two, three and four squares [5–7].

2.1.1 Sum of Two Squares

The Theorems 1 and 2, and Corollaries 1 and 2 are useful to represent an integer in
the form of sum of two squares.
l Theorem 1 (Fermat)
An odd prime p is expressible as a sum of two squares if and only if p  1(mod 4)
l Corollary 1
Any prime p is of the form 4k þ 1 can be represented uniquely (aside from the
order of the summands) as a sum of two squares.
l Theorem 2
Let the positive integer n be written as n ¼ N2 m, where m is square-free. Then
n can be represented as the sum of two squares if and only if m contains no prime
factor of the form 4k þ 3.
510 G. Ganapathy and K. Mani

l Corollary 2
A positive integer n is representable as a sum of two squares if and only if each
of its prime factors of the form 4k + 3 occurs to an even power.
l Theorem 3
Any positive integer n can be represented as the difference of two squares if and
only if n is not of the form 4k + 2.
l Example
13 ¼ 22 þ 32 (We do not consider 13 ¼ 22 þ 32 as distinct from 13 ¼ 32 þ 22).
In this case, the total number of pairs (x, y) of integers such that x2 þ y2 ¼ n is 8. For
n ¼ 13, we get the pairs (2, 3) and (3, 2). This result is due to Fermat.

2.1.2 Sum of Three Squares

l Theorem 4
The Diophantine equation x12 þ x22 þ x32 ¼ n where n is a positive integer has
integral solutions if and only if n is not of the form 4k (8m þ 7).
l Theorem 5
No positive integer of the form 4n (8m þ 7) can be represented as the sum of
three squares.
l Example
The integer 14, 33, and 67 can be represented as sum of three squares.
14 ¼ 32 þ 22 þ 12
33 ¼ 52 þ 22 þ 22
67 ¼ 72 þ 32 þ 32

2.1.3 Sum of Four Squares

l Theorem 6 (Lagrange)
Any positive integer n can be written as the sum of four squares, some of which may
be zero.
l Theorem 7
Any prime p of the form 4m þ 3 can be expressed as the sum of four squares.
l Lemma 1 (Euler)
If the integers m and n are each the some of four squares, then m  n is likewise
so representable.
Towards the Performance Optimization of Public-key Algorithms 511

l Example
To represent the integer 459 as sum of four squares, we have
459 ¼ 32.3.17 (Euler Identity)
¼ 32(12 þ 12 þ 12 þ 02)(42 þ 12 þ 02 þ 02)
¼ 32[(4 þ 1 þ 0 þ 0)2 þ (1  4 þ 0  0)2 þ (0  0  4 þ 0)2 þ (0 þ 0  1  0)2]
¼ 32[52 þ 32 þ 42 þ 12]
459 ¼ 152 þ 92 þ 122 þ 32

2.1.4 Integers Which Cannot Be Represented as Sum of Squares

l If n  3 (mod 4), then n cannot be written as sum of two squares. More generally,
if n contains a prime factor p  3 (mod 4) raised to an odd power, then n cannot
be written as a sum of two squares.
l Example
78 ¼ 2  3  13 is such a number.
For the general case, when n is an arbitrary positive integer, a formula for the
number of pairs (x, y) of integers such that x2 þ y2 ¼ n was first found by Jacobi. It
is expressed in terms of the number of divisors of n, thus: Let the number of divisors
of n of the type 1 (mod 4) be d1(n), and let the number of divisors of n of the type 3
(mod 4) be d3(n). Then the number of pairs (x, y) of integers such that x2 þ y2 ¼ n is
equal to 4[d1(n)  d3(n)].
l Example:
Take n ¼ 65. The divisors of 65 are 1, 5, 13, 65. These are all of the form 1
(mod 4), so d1(n) ¼ 4, d3(n) ¼ 0, 4 [d1(n)  d3(n)] ¼ 16. The integer pairs
(x, y) for which x2 þ y2 ¼ 65 are (1, 8), (8, 1), (4, 7), (7, 4),
and these are indeed 16 in number. Or take n ¼ 39. The divisors of 39 are 1, 3,
13, 39, so that d1(n) ¼ 2, d3(n) ¼ 2, d1(n)  d3(n) ¼ 0. And indeed there are no
pairs (x, y) of integers for which x2 þ y2 ¼ 39.

2.2 Addition Chain

To compute xe by multiplication, the most economical way is to reduce into

addition, since the exponents are additive. Formally, an addition chain can be
defined as follows.

2.2.1 Definition (Addition Chain)

An addition chain for e is a sequence of integers

1 ¼ a0; a1 ; a2 ; . . . ; an ¼ e (1)
512 G. Ganapathy and K. Mani

with the property that

ai ¼ aj þ ak; for some k  j < i for all i ¼ 1; 2; . . . ; r:

without loss of generality that an addition chain is “ascending”

1 ¼ a0; a1 ; a2 ; . . . ; an ¼ e (2)

For if any two a’s are equal, one of them may be dropped; and arrange the
sequence (1) into ascending order and remove terms >e without destroying the
addition chain property (2). The shortest length, r, for which there exists an addition
chain for e is denoted by l(r). Clearly, the problem of finding the shortest ones
becomes more and more complicated as r grows larger [8].

2.3 Elliptic Curve

This section focus on the concept of elliptic curve, addition of two points in elliptic
curve, and scalar point multiplication.

2.3.1 Definition (Elliptic Curve over GF(p))

Let p be a prime greater than 3 and let a and b be two integers such that 4a3 + 27b2 ¼ 6 0
(mod p). An elliptic curve E over the finite field Fp is the set of points (x, y) E Fp  Fp
satisfying the Weierstrass equation

E : y2 ¼ x3 þ ax þ b (3)

together with point at infinity O. The inverse of the point P ¼ (x1, y1) is P ¼ (x1,
y1). The sum P + Q of the points P ¼ (x1, y1) and Q ¼ P ¼ (x2, y2) (where P, Q 6¼
O and P ¼ Q) is the point R ¼ (x3, y3) where

ð y2  y1 Þ
l¼ ; x3 ¼ l2  x1  x2 ; y3 ¼ ðx1  x3 Þl  y1 : (4)
ð x2  x1 Þ

For P ¼ Q, the doubling formulae are

3x21 þ a
l¼ ; x3 ¼ l2  2x1 ; y3 ¼ ðx1  x3 Þl  y1 : (5)
2y1

The point at infinity O plays a rule similar to that of the number 0 in normal
addition. Thus, P þ Q ¼ P, and P þ (P) ¼ O for all points P. The point on E
together with the operation of “addition” forms an abelian group [9]. Since the point
Towards the Performance Optimization of Public-key Algorithms 513

Table 1 Double-and-add algorithm

Algorithm 1 Elliptic Curve point Multiplication: Binary Method
P j
l1
Input: A point P an l-bit integer k ¼ kj 2 ;kj 2 f0; 1g
Output: Q ¼ [k]P j¼0

1. Q ! P
2. for j from l–1 to 0 by 1 do:
3. Q ! [2] Q
4. If kj ¼ 1 then
Q P+Q
5. end if
6. end for

or scalar multiplication is the main operation on ECC, the ‘kP’ multiplication is

based on point doubling and point addition and it can be calculated by using double-
and-add algorithm, which is shown in Table 1 [10].

2.3.2 Modular Multiplication Inversion

It is done according to Fermat’s theorem, a–1 ¼ ap–2 mod p, if gcd (a, p) ¼ 1. In this
paper, we are interested the E defined over GF (p), p is a prime. Thus, Fermat’s
theorem is used to find multiplicative inverse modulo p. The multiplication inver-
sion can be performed by modular exponentiation of a by p  2 and it can be
realized by using the square and multiply algorithm [11].

3 Fuzzy Modular Arithmetic

The concept of fuzzy modular arithmetic for cryptographic schemes was proposed
by Wael Adi. As illustrated by Wael Adi [12, 13], in partial modulo reduction
instead of performing division, the division is replaced by repeated subtraction. For
that some random multiples of m is used. Since the operation used is not exact but
involves some random multiples of m from the value to be reduced modulo m, and
hence this approach is called fuzzy. The result of the process is not the rest of the
division usually computed, but another value and it can be used for calculation. The
numerical example of serial fuzzy modular multiplication is shown in Fig. 2.
The modulus m ¼ 13, a multiple of m i.e., m ¼ m·t ¼ 13·2 ¼ 26 is subtracted
each time. The 2’s complement of m is . . .1100110.
In Fig. 1, the conventional multiplication of two binary integers, say A ¼ 35, and
B ¼ 33 with m ¼ 13 is shown and the result is 11 (mod 13). Since, fuzzy modular
multiplication mainly focuses on repeated subtraction instead of division for modulo
reduction m, in this work m ¼ 2  13 ¼ 26 is considered and the process is shown in
Fig. 2. It is noted that subtraction is performed whenever the bit position in B is zero.
514 G. Ganapathy and K. Mani

Fig. 1 Conventional
multiplication A = (35)2 = × 100011 35 ×
B = (33)2 = 100001 33
100011 1155 = 11 (mod 13)
000000
000000
000000
000000
100011
10010000011

Fig. 2 Numerical example

for fuzzy modular A = (35)2 = 100011 × 35 ×
multiplication B = (33)2 = 100001 33
000000100011 + 35
111111111100110 – 2 × 26
1111111111100110 – 4 × 26
1111111111100110 – 8 × 26
11111111100110 16 × 26
00000100011 32 × 35
....00000000101110111 375 = 11 (mod 13)

4 Applications of Sum of Squares, and Addition Chain

in Reducing the Number of Multiplication in Modular
Exponentiation

In many public-key cryptosystems such as RSA, the modular exponentiation xe

mod n, where x, e, and n are positive integer is a corner stone operation. It is
performed using successive modular multiplications. Several authors have pro-
posed the methods to reduce the total number of multiplication. Nadia Nedjah
et al proposed an idea to obtain addition chains with a minimal number of multipli-
cation and hence implementing efficiently the exponentiation operation using
genetic algorithms. To find optimal addition chains a genetic algorithm approach
was illustrated by Nareli Cruz-Cortes et al. [14]. But, in this work, we proposed an
algorithm to reduce the higher power of an exponent into smaller one using the sum
of squares. The proposed pesudocode is shown in Section 4.1

4.1 Pseudocode

The pseudocode illustrates the representation of an integer as sum of squares and

the number multiplications required for an exponent.
Towards the Performance Optimization of Public-key Algorithms 515

1. Read the exponent e.

2. Check, whether e is prime or not.
a. If e is a prime and e  1 mod 4 then express e as sum of two squares.
b. If e is a prime and e  3 mod 4, then if e is of the form 4k þ 3 it cannot be
expressed as sum of squares.
c. If e is not a prime and one of the prime factor say p is p  (3 mod 4), express
the factor as sum of squares.
d. If e is not a prime and if e is not of the form 4k (8m þ 7),based on Diophantine
equation, express e ¼ x12 þ x22 þ x32 as sum of three squares.
e. If e not a prime and of the form 4m þ 3, express e as the sum of four squares
e ¼ x12 þ x22 þ x 32 þ x 42.
Once the exponent is expressed as the sum of squares, use the addition chain to
find the minimal number of multiplication if it is possible.

4.2 Example

In order to understand the pseudocode, the example illustrated in various cases are
useful.
l Case 1: When e is a prime and e  1 mod 4
Suppose e is 29, then e  1 mod 4. Using sum of two squares, say
e ¼ 52 þ 22; x29 ¼ x25 þ 4 ¼ (x4)7x
we first evaluate y ¼ x4
Then, we form y1¼ y2 ¼ (x 4)2; y2 ¼ y12
x29 ¼ y * y1 * y2 * x
Thus, the whole process takes 7 multiplication, which is equal to shorter addition
chain as illustrated by Nareli Cruz-Cortes et al.
l Case 2: when e is a prime and e  3 (mod 4)
Suppose e is 11, and e is in the form of 4k þ 3, where k ¼ 1. It cannot be
expressed as sum of squares.
Thus, we form x11 ¼ x(x5)2
We first evaluate y ¼ x(x4) ¼ x(x2)2;
Then, we form x11 ¼ y2 * x which requires 5 multiplication

l Case 3: when e is not a prime and of the prime factor is of the form, p  3 (mod 4)
Suppose e is 78. Then, the factors of 78 are 2, 3, and 13. One of the prime factor,
say p ¼ 3 is of the form p  3 (mod 4). Thus, 78 cannot be expressed as a sum of
two squares. In this case, we have to consider a1l factors of 78.
516 G. Ganapathy and K. Mani

Table 2 Minimum number of multiplication for some exponent e

Exponent (e) Shortest addition chain l(r) No. multiplication
11 1–2–3–5–10– 11 5
15 1–2–3–5–10–15 5
27 1–2–3–6–12–15–27 6
29 1–2–3–5–7–14–28-29 7
78 1–2–3–6–12–15–27-39-78 8

Thus, 78 ¼ 2  3  13
x78 ¼ ((x2)3)13
We first evaluate y1¼ x * x
y2 ¼ y1 * y1 * y1
x78 ¼ y213 ¼ y212y2
y3 ¼ y2 * y2
y4 ¼ y 36
x
78
¼ y4 * y2
Thus, the exponent 78 requires 8 multiplications.
l Case 4: when e is expressed as sum of three squares
Suppose e ¼ 27, then 27 ¼ 12 þ 12 þ 52
x27 ¼ x2(x5)5
Thus, the exponent 27 requires 6 multiplication
l Case 5: when e is expressed as sum of four squares
Suppose e ¼ 15, then 15 ¼ 12 þ 12 þ 22 þ32
x15 ¼ x2 · x4 · x9
¼ x2(x4)3x
¼ x x2(x4)3
Thus, the exponent 15 requires 5 multiplications.
Similar calculations can be performed other exponents also. The addition chain
for the exponents 11, 15, 27, 29, and 78 are shown in the Table 2.

5 Implementation of ECC Using Fuzzy Modular Arithmetic

The execution of ECC scheme is mostly dominated by the efficient implementation

of finite field arithmetic and point multiplication kP. To compute kP, several
methods such as binary method, binary NAF method, windowed NAF etc. have
been proposed [15]. To implement these methods, efficient hardwired architecture
and a platform that support the efficient computation of such methods are needed.
Previous hardware work includes: the first ASIC implementation with Motorola
M68008 microcomputer, reconfigurable finite-field multiplier, parallelized field
Towards the Performance Optimization of Public-key Algorithms 517

Table 3 Addition of two points in EC using fuzzy modular arithmetic and Fermat’s theorem
Algorithm 2 EC point addition and doubling using fuzzy modular arithmetic
Input: p1¼ (x1,y1), p2 ¼ (x2,y2) and m Input: p1¼ (x1,y1), a and m
Output: p1 þ p2 ¼ p3 ¼ (x3,y3) Output: 2 p1 ¼ p3 ¼ (x3,y3)
1. Tl1 y2  y1 1. Tl1 3x21 þ a
2. Tl2 x2  x1 2. Tl2 2 y1
3. Tlm2 Tl2 m2 mod m 3. Tlm2 Tl2 m2 mod m
(Fermat’s Theorem) (Fermat’s Theorem)
4. T1
lm2 Tlm2  z1 m(z1m  Tlm2 ) 4. T1
lm2 Tlm2  z1 m(z1 m  Tlm2 )
5. l Tl1 T1 lm2 5. l Tl1 T1 lm2
6. Tx3 l2  x1  x2 6. Tx3 l2  2 x1
7. If Tx3 > m then 7. If Tx3 > m then
8. x3 Tx3  z2 m (z2 m  Tx3 ) 8. x3 Tx3  z2 m (z2 m  Tx3 )
Else Else
9. x3 Tx3 9. x3 Tx3
10. end if 10. end if
11. Ty3 (x1  x3) l  y1 11. Ty3 (x1  x3) l  y1
12. If Ty3 > m then 12. If Ty3 > m then
13. y3 Ty3  z3 m 13. y3 Ty3  z3 m
else else
14. y3 Ty3 14. y3 Ty3
15. end if 15. end if

Use Fuzzy Modular Arithmetic to

Key Generation speed up addition and scalar point
multiplication, and inversion
Select of elliptic curve points
Ep(a, b)
Select p, q
n = p×q
Select e

Public Key Private Key

(e,n,Ep) (d,n,Ep)

Clear text P
(M) Cipher text C = (x, y) C
Imbedd M in Ep and
select the corresponding = Pe mod n P = Cd mod n
point say P = (x, y). = e × P mod n = d × C mod

P
M Reverse of imbedding P into M

Fig. 3 ECC with fuzzy modular arithmetic model

518 G. Ganapathy and K. Mani

multiplier etc. [16]. The main objective of this work is to perform modular opera-
tions without trail division because the hardware implementation of trial division is
an expensive. Thus, this work emphasizes fuzzy modular arithmetic as illustrated
by W. Adi no division is required and division is repeatedly replaced by subtraction
and hence the hardware implementation is relatively easy. Moreover, to perform
kP, repeated addition is used. To add two points of EC, in both formula modular
multiplication inversions is involved and it takes considerable amount of time. To
avoid it, Fermat’s Theorem is used in this paper which is illustrated in Section 2.3.2.
The algorithm which describes the EC point addition and doubling using Fuzzy
Modular Arithmetic and Fermat’s theorem is shown in Table 3 [17].
The proposed methodology to speed up the modular arithmetic is shown in
Fig. 3.

6 Conclusion

A straight forward, cheap algorithm for reducing the number of multiplications

normally performed in an exponentiation has been described and it can be used for
any public-key cryptosystem. Further, to increase the speed of kP, which is the key
operations in performing the encryption and decryption process of ECC, fuzzy
modular arithmetic and Fermat’s theorem are used. In fuzzy modular arithmetic to
perform modular operation, repeated subtraction is used instead of division and
hence hardware complexity is relatively reduced than division operation. To per-
form inversion operation with respect to modulus, Fermat’s theorem is used which
is further simplified by repeated multiplication.

References

1. A. Menezes, P. Van Oorschot, S. Vanstone, Handbook of Applied Cryptography, (CRC Press,

Boca Raton, FL, 1997)
2. B. Moller, Improved techniques for fast exponentiation, Information Security and Cryptology –
ICISC 2002 (Springer-Verlag, LNCS 2587), p. 298–312
3. R. Cheung, W. Telle, P. Cheung, Customizable elliptic curve cryptosystems. IEEE TVLSI, 13(a),
1048–1049 (Sept 2005)
4. D. Hankerson, L. Herandez, A. Menezes, Software Implementation of Elliptic Curve Cryp-
tography over Binary Fields (CHES, LNCS 1965, Springer-Verlag, Berlin Heidelberg, 2000),
pp. 1–24
5. J.A. Anderson, J.M. Bell, Number Theory with Applications (Prentice Hall, Englewood.
Cliffs, NJ, 1997)
6. M.D. Burton, Elementary Number Theory, 6th edn. (Tata Mc-Graw-Hill, New Delhi, 2007)
7. N. Koblitz, A Course in Number Theory and Cryptography, 2nd edn. (Springer-Verlag, Berlin,
1994)
8. D. Knuth, The Art of Computer Programming – Semi Numerical Algorithms, vol. 2, 3rd edn.
(Addison-Wesley, 1998)
Towards the Performance Optimization of Public-key Algorithms 519

9. N. Koblitz, Elliptic curve cryptosystems. Math. Comput. 48(177), 203–209 (Nov 1982)
10. J. Lutz, A. Hasan, High performance FPGA based elliptic curve cryptographic co-processor.
ITCC 04, international conference on information technology coding and computing, vol. 2,
2004, p. 486
11. I. Blake, G. Seroussi, N.P. Smart, Elliptic Curves in Cryptography, (Ser. London Math. Soc.
Lecture Note Series, Cambridge University Press, New York, 1999)
12. W. Adi, Fuzzy modular arithmetic for cryptographic schemes with applications for
mobile security, IEEE, Institute of Technology Technical University of Braunschweig,
Braunschweig, 2000, p. 263–265
13. A. Hanoun, W. Adi, F. Mayer-Lindenberg, B. Soundan, Fuzzy modular multiplication archi-
tecture and low complexity IPR-protection for FPGA technology. IEEE, 2006, p. 325–326
14. N. Cruz-Cortes et al., Finding Optimal Addition Chain Using a Genetic Algorithm Approach,
LNCS, vol. 3801 (Spring Berlin/Heidberg, 2005), pp. 208–215
15. N. Koblitz, Elliptic curve cryptosystems. Math. Comput. 48(177), 203–209 (Nov 1982)
16. Y.-J. Choi, M.-S. Kim, H.-R. Lee, H.-W. Kim, Implementation and analysis of Elliptic Curve
Cryptosystems over Polynomial basis and ONB. PWAEST, 10, 130–134 (Dec 2005)
17. G. Ganapathy, K. Mani, Maximization of Speed in Elliptic Curve Cryptography Using Fuzzy
Modular Arithmetic over a Microcontroller based Environment, Lecture Notes in Engineering
and Computer Science WCECS (San Francisco, CA, 20–22 Oct 2009), pp. 328–332
Chapter 40
RBDT-1 Method: Combining Rules and Decision
Tree Capabilities

Amany Abdelhalim and Issa Traore

Abstract Most of the methods that generate decision trees for a specific problem
use examples of data instances in the decision tree generation process. This chapter
proposes a method called “RBDT-1” - rule based decision tree - for learning a
decision tree from a set of decision rules that cover the data instances rather than
from the data instances themselves. RBDT-1 method uses a set of declarative rules
as an input for generating a decision tree. The method’s goal is to create on-demand
a short and accurate decision tree from a stable or dynamically changing set of
rules. We conduct a comparative study of RBDT-1 with existing decision tree
methods based on different problems. The outcome of the study shows that in
terms of tree complexity (number of nodes and leaves in the decision tree) RBDT-1
compares favorably to AQDT-1, AQDT-2 which are methods that create decision
trees from rules. RBDT-1 compares favorably also to ID3 which is a famous method
that generates decision trees from data examples. Experiments show that the
classification accuracies of the different decision trees produced by the different
methods under comparison are equal.

1 Introduction

Decision Tree is one of the most popular classification algorithms used in data
mining and machine learning to create knowledge structures that guide the decision
making process.
The most common methods for creating decision trees are those that create
decision trees from a set of examples (data records). We refer to these methods as
data-based decision tree methods.

A. Abdelhalim (*)
Department of Electrical and Computer Engineering, University of Victoria, 3055 STN CSC,
Victoria, B.C.,V8W 3P6, Canada
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 521

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_40, # Springer ScienceþBusiness Media B.V. 2010
522 A. Abdelhalim and I. Traore

On the other hand, to our knowledge there are only few approaches that create
decision trees from rules proposed in the literature which we refer to as rule-based
decision tree methods.
A decision tree can be an effective tool for guiding a decision process as long as
no changes occur in the dataset used to create the decision tree. Thus, for the data-
based decision tree methods once there is a significant change in the data, restruc-
turing the decision tree becomes a desirable task. However, it is difficult to
manipulate or restructure decision trees. This is because a decision tree is a
procedural knowledge representation, which imposes an evaluation order on the
attributes. In contrast, rule-based decision tree methods handle manipulations in the
data through the rules induced from the data not the decision tree itself. A declara-
tive representation, such as a set of decision rules is much easier to modify and
adapt to different situations than a procedural one. This easiness is due to the
absence of constraints on the order of evaluating the rules [1].
On the other hand, in order to be able to make a decision for some situation we
need to decide the best order in which tests should be evaluated in those rules. In
that case a decision structure (e.g. decision tree) will be created from the rules. So,
rule-based decision tree methods combine the best of both worlds. On one hand
they easily allow changes to the data (when needed) by modifying the rules rather
than the decision tree itself. On the other hand they take advantage of the structure
of the decision tree to organize the rules in a concise and efficient way required to
take the best decision. So knowledge can be stored in a declarative rule form and
then be transformed (on the fly) into a decision tree only when needed for a decision
making situation [1].
In addition to that, generating a decision structure from decision rules can
potentially be performed faster than generating it from training examples because
the number of decision rules per decision class is usually much smaller than the
number of training examples per class. Thus, this process could be done on
demand without any noticeable delay [2, 3]. Data-based decision tree methods
require examining the complete tree to extract information about any single
classification. Otherwise, with rule-based decision tree methods, extracting infor-
mation about any single classification can be done directly from the declarative
rules themselves [4].
Although rule-based decision tree methods create decision trees from rules, they
could be used also to create decision trees from examples by considering each
example as a rule. Data-based decision tree methods create decision trees from data
only. Thus, to generate a decision tree for problems where rules are provided, e.g.
by an expert, and no data is available, rule-based decision tree methods are the only
applicable solution.
This chapter presents a new rule-based decision tree method called RBDT-1.
To generate a decision tree, the RBDT-1 method uses in sequence three differ-
ent criteria to determine the fit (best) attribute for each node of the tree, referred
to as the attribute effectiveness (AE), the attribute autonomy (AA), and the
minimum value distribution (MVD). In this chapter, the RBDT-1 method is
compared to the AQDT-1 and AQDT-2 methods which are rule-based decision
40 RBDT-1 Method: Combining Rules and Decision Tree Capabilities 523

tree methods, along with ID3 which is one of the most famous data-based
decision tree methods.

2 Related Work

There are few published works on creating decision structures from declarative
rules.
The AQDT-1 method introduced in [1] is the first approach proposed in the
literature to create a decision tree from decision rules. The AQDT-1 method uses
four criteria to select the fit attribute that will be placed at each node of the tree.
Those criteria are the cost1, the disjointness, the dominance, and the extent, which
are applied in the same specified order in the method’s default setting.
The AQDT-2 method introduced in [4] is a variant of AQDT-1. AQDT-2 uses five
criteria in selecting the fit attribute for each node of the tree. Those criteria are the
cost1, disjointness, information importance, value distribution, and dominance,
which are applied in the same specified order in the method’s default setting. In
both the AQDT-1 & 2 methods, the order of each criterion expresses its level of
importance in deciding which attribute will be selected for a node in the decision
tree. Another point is that the calculation of the second criterion – the information
importance – in AQDT-2 method depends on the training examples as well as the
rules, which contradicts the method’s fundamental idea of being a rule-based decision
tree method. AQDT-2 requires both the examples and the rules to calculate the
information importance at certain nodes where the first criterion – Disjointness – is
not enough in choosing the fit attribute. AQDT-2 being both dependent on the
examples as well as the rules increases the running time of the algorithm remark-
ably in large datasets especially those with large number of attributes.
In contrast the RBDT-1 method, proposed in this work, depends only on the rules
induced from the examples, and does not require the presence of the examples
themselves.
Akiba et al. [5] proposed a rule-based decision tree method for learning a single
decision tree that approximates the classification decision of a majority voting
classifier. In their proposed method, if-then rules are generated from each classifier
(a C4.5 based decision tree) and then a single decision tree is learned from these
rules. Since the final learning result is represented as a single decision tree,
problems of intelligibility and classification speed and storage consumption are
improved. The procedure that they follow in selecting the best attribute at each node
of the tree is based on the C4.5 method which is a data-based decision tree method.
The input to the proposed method requires both the real examples used to create the

1
In the default setting, the cost equals 1 for all the attributes. Thus, the disjointness criterion is
treated as the first criterion of the AQDT-1 and AQDT-2 methods in the decision tree building
experiments throughout this paper.
524 A. Abdelhalim and I. Traore

classifiers (decision trees) and the rules extracted from the classifiers, which are
used to create a set of training examples to be used in the method.
In [6], the authors proposed a method called Associative Classification Tree
(ACT) for building a decision tree from association rules rather than from data.
They proposed two splitting algorithms for choosing attributes in the ACT one
based on the confidence gain criterion and one based on the entropy gain criterion.
The attribute selection process at each node in both splitting algorithms relies on
both the existence of rules and the data itself as well. Unlike our proposed method
RBDT-1, ACT is not capable of building a decision tree from the rules in the absence
of data, or from data (considering them as rules) in the absence of rules.

3 Rule Generation and Notations

In this section, we present the notations used to describe the rules used by our
method. We also present the methods used to generate the rules that will serve as
input to the rule-based decision tree methods in our experiments.

3.1 Notations

In order for our proposed method to be capable of generating a decision tree for a
given dataset, it has to be presented with a set of rules that cover the dataset. The
rules will be used as input to RBDT-1 which will produce a decision tree as an
output. The rules can either be provided up front, for instance, by an expert or can
be generated algorithmically.
Let a1 ; :::; an denote the attributes characterizing the data under consideration,
and let D1 ; :::; Dn denote the corresponding domains, respectively (i.e. Di represents
the set of values for attribute ai ). Let c1 ; :::; cm represent the decision classes
associated with the dataset.

3.2 Rule Generation Method

The RBDT-1 takes rules as input and produces a decision tree as output. Thus, to
illustrate our approach and compare it to existing similar approaches, we use in
this chapter two options for generating a set of disjoint rules for each dataset. The
first option is based on extracting rules from a decision tree generated by the ID3
method, where we convert each branch – from the root to a leaf – of the decision
tree to an if–then rule whose condition part is a pure conjunction. This scenario
will ensure that we will have a collection of disjoint rules. We refer to these rules
as ID3-based rules.
40 RBDT-1 Method: Combining Rules and Decision Tree Capabilities 525

Using ID3-based rules will give us an opportunity to illustrate our method’s

capability of producing a smaller tree while using the same rules extracted from a
decision tree generated by a data-based decision tree method without reducing the
tree’s accuracy.
The second rule generation option consists of using an AQ-type rule induction
program. We use the AQ19 program for creating logically disjoint rules which we
refer to as AQ-based rules.

4 RBDT-1 Method

In this section, we describe the RBDT-1 method by first outlining the format of the
input rules used by the method and the attribute selection criteria of the method.
Then we summarize the main steps of the underlying decision tree building process
and present the technique used to prune the rules.

4.1 Attribute Selection Criteria

The RBDT-1 method applies three criteria on the attributes to select the fittest attribute
that will be assigned to each node of the decision tree. These criteria are the Attribute
Effectiveness, the Attribute Autonomy, and the Minimum Value Distribution.

4.1.1 Attribute Effectiveness (AE)

AE is the first criterion to be examined for the attributes. It prefers an attribute

which has the most influence in determining the decision classes. In other words, it
prefers the attribute that has the least number of “don’t care” values for the class
decisions in the rules, as this indicates its high relevance for discriminating among
rule sets of given decision classes. On the other hand, an attribute which is omitted
from all the rules (i.e. has a “don’t care” value) for a certain class decision does not
contribute in producing that corresponding decision. Choosing attributes based on
this criterion maximizes the chances of reaching leaf nodes faster which on its turn
minimizes the branching process and leads to producing a smaller tree.
Using the notation provided above (see Section 3), letVij denote the set of values
for attribute aj involved in the rules in Ri , which denote the set of rules associated
with decision class ci , 1  i  m. Let DC denote the ‘don’t care’ value, we calculate
Cij ðDCÞ as shown in (1):


1 if DC 2 Vij
Cij ðDCÞ ¼ (1)
0 otherwise
526 A. Abdelhalim and I. Traore

Given an attribute aj , where 1  j  n, the corresponding attribute effectiveness

is given in (2).

P
m
m Cij ðDCÞ
i¼1
AEðaj Þ ¼ (2)
m
(where m is the total number of different classes in the set of rules).
The attribute with the highest AE is selected as the fit attribute.

4.1.2 Attribute Autonomy (AA)

AA is the second criterion to be examined for the attributes. This criterion is

examined when the highest AE score is obtained by more than one attribute. This
criterion prefers the attribute that will decrease the number of subsequent nodes
required ahead in the branch before reaching a leaf node. Thus, it selects the
attribute that is less dependent on the other attributes in deciding on the decision
classes. We calculate the attribute autonomy for each attribute and the one with the
highest score will be selected as the fit attribute.
For the sake of simplicity, let us assume that the set of attributes that achieved
the highest AE score are a1 ; :::; as ; 2  s  n. Let vj1 ; :::; vjpj denote the set of
possible values for attribute aj including the “don’t care”, and Rji denote the rule
subset consisting of the rules that have aj appearing with the value vji , where
1  j  s and 1  i  pj . Note that Rji will include the rules that have don’t care
values for aj as well.
The AA criterion is computed in terms of the Attribute Disjointness Score
(ADS), which was introduced by Imam and Michalski [1]. For each rule subset
Rji , let MaxADSji denote the maximum ADS value and let ADS_Listji denote a list
that contains the ADS score for each attribute ak , where 1  k  s; k 6¼ j.
According to [1], given an attribute aj and two decision classes ci and ck (where
1  i; k  m; 1  j  s), the degree of disjointness between the rule set for ci and
the rule set for cj with respect to attribute aj is defined as shown in (3):
8
>
> 0 if Vij  Vkj
< 1 if V  V
ij kj
ADSðAj ; Ci ; Ck Þ ¼ (3)
>
> 2 if V ij \ V kj 6¼ ð or Vij or Vkj Þ
:
3 if Vij \ Vkj ¼ 
 
The Attribute Disjointness of the attribute aj , ADS aj score, is the summation of
the degrees of class disjointness ADSðaj ; ci ; ck Þ given in (4):
X
m X  
ADSðaj Þ ¼ ADS aj ; ci ; ck (4)
i¼1 1ks
i6¼k
40 RBDT-1 Method: Combining Rules and Decision Tree Capabilities 527

Thus, the number of ADS_List that will be created for each attribute aj as well as
the number of MaxADS values that are calculated will be equal to pj . The MaxADSji
value as defined by Wojtusiak [8] is 3  m  ðm  1Þ where m is the total number
of classes in Rji . We introduce the AA as a new criterion for attribute aj as given
in (5):

1
AAðaj Þ ¼ (5)
P
pj
AAðaj ; iÞ
i¼1

where AAðaj ; iÞ is defined as shown in (6):

8
>
> 0 if MaxADSji ¼ 0
> 1 if
>
>
> !!
>
>
  < MaxADS 6¼ 0 ^ ðs ¼ 2Þ _
AA aj ; i ¼ ji
ð9l : MaxADSji ¼ ADS Listji ½lÞ (6)
>
> " #
>
>
>
> P
s1
>
> 1 þ ðs  1Þ  MaxADSji  ADS Listji ½l otherwise
:
l¼1;l6¼j

The AA for each of the attributes is calculated using the above formula and the
attribute with the highest AA score is selected as the fit attribute. According to the
above formula, AAðaj ; iÞ, equals zero when the class decision for the rule subset
examined corresponds to one class, in that case MaxADS¼ 0, which indicates that
a leaf node is reached (best case for a branch). AAðaj ; iÞ equals 1 when s equals
2 or when one of the attributes in the ADS_list has an ADS score equal to MaxADS
value (second best case). The second best case indicates that only one extra node
will be required to reach a leaf node. Otherwise AAðaj ; iÞ will be equal to 1 þ (the
difference between the ADS scores of the attributes in the ADS_list and the
MaxADS value) which indicates that more than one node will be required until
reaching a leaf node.

4.1.3 Minimum Value Distribution (MVD)

The MVD criterion is concerned with the number of values that an attribute has in
the current rules. When the highest AA score is obtained by more than one attribute,
this criterion selects the attribute with the minimum number of values in the current
rules. MVD criterion minimizes the size of the tree because the fewer the number of
values of the attributes the fewer the number of branches involved and consequently
the smaller the tree will become [1]. For the sake of simplicity, let us assume that
the set of attributes that achieved the highest AA score are a1 ; . . . ; aq ; 2  q  s.
Given an attribute aj (where 1  j  q), we compute corresponding MVD value as
shown in (7).
528 A. Abdelhalim and I. Traore

 
 [ 
 
MVDðAj Þ ¼  Vij  (7)
1im 

(where |X| denote the cardinality of set X).

When the lowest MVD score is obtained by more than one attribute, any of these
attributes can be selected randomly as the fit attribute. In our experiments in case
where more than two attributes have the lowest MVD score we take the first
attribute.

4.2 Building the Decision Tree

In the decision tree building process, we select the fit attribute that will be assigned
to each node from the current set of rules CR based on the attribute selection criteria
outlined in the previous section. CR is a subset of the decision rules that satisfy the
combination of attribute values assigned to the path from the root to the current
node. CR will correspond to the whole set of rules at the root node.
From each node a number of branches are pulled out according to the total
number of values available for the corresponding attribute in CR.
Each branch is associated with a reduced set of rules RR which is a subset of CR
that satisfies the value of the corresponding attribute. If RR is empty, then a single
node will be returned with the value of the most frequent class found in the whole
set of rules. Otherwise, if all the rules in RR assigned to the branch belong to the
same decision class, a leaf node will be created and assigned a value of that decision
class. The process continues until each branch from the root node is terminated with
a leaf node and no more further branching is required.

5 Illustration of the RBDT-1 Method

In this section the RBDT-1 method is illustrated by a dataset named the weekend
problem which is a publicly available dataset.

5.1 Illustration

The Weekend problem is a dataset that consists of ten data records obtained from
[12]. For this example, we used the AQ19 rule induction program to induce the rule
set shown in Table 1 which will serve as the input to our proposed method RBDT-1.
AQ19 was used with the mode of generating disjoint rules and with producing
a complete set of rules without truncating any of the rules. The corresponding
40 RBDT-1 Method: Combining Rules and Decision Tree Capabilities 529

Table 1 The weekend rule set induced by AQ19

Rule# Description
1: Cinema Parents-visiting¼“yes” & weather ¼ “don’t care” & Money ¼“rich”
2: Tennis Parents-visiting¼”“no” & weather ¼“sunny” & Money ¼“don’t care”
3: Shopping Parents-visiting¼“no” & weather ¼“windy” & Money ¼“rich”
4: Cinema Parents-visiting¼“no” & weather ¼“windy” & Money ¼“poor”
5: Stay-in Parents-visiting¼“no” & weather ¼“rainy” & Money ¼“poor”

Fig. 1 The decision tree parents-visiting

generated by RBDT-1 for the yes
weekend problem -> cinema
no
weather
windy
money
poor
-> cinema
rich
-> shopping
sunny
-> tennis
rainy
-> stay-in

decision tree created by the proposed RBDT-1 method for the weekend problem is
shown in Fig. 1. It consists of three nodes and five leaves with 100% classification
accuracy for the data.
Figure 2 depicts the tree created by AQDT-1 & 2 and the ID3 methods consisting
of five nodes and seven leaves, which is bigger than the decision tree created by the
RBDT-1 method. Both trees have the same classification accuracy as RBDT-1. Due
to space limitations, we are unable to include here all of our work. We refer the
reader to our extended technical report on RBDT-1 [13] for more details on the
decision tree generation and discussion on tree sizes.

6 Experiments

In this section, we present an evaluation of our proposed method by comparing it

with the AQDT-1 & 2 and the ID3 methods based on 16 public datasets. Our
evaluation consisted of comparing the decision trees produced by the methods for
each dataset in terms of tree complexity (number of nodes and leaves) and accu-
racy. Other than the weekend dataset, all the datasets were obtained from the UCI
machine learning repository [14].
We conducted two experiments for comparing RBDT-1 with AQDT-1 & 2 and
the ID3; the results of these experiments are summarized in Tables 2 and 3. The
530 A. Abdelhalim and I. Traore

Fig. 2 The decision tree weather

generated by AQDT-1, windy
AQDT-2 & ID3 for the money
weekend problem poor
-> cinema
rich
parents-visiting
yes
-> cinema
no
-> shopping
sunny
parents-visiting
yes
-> cinema
no
-> tennis
rainy
money
poor
-> stay-in
rich
-> cinema

Table 2 Comparision of tree complexities of the RBDT-1, AQDT-1, AQDT-2 & ID3 methods
using ID3-based rules
Dataset Method Dataset Method
Weekend RBDT-1 MONK’s 1 RBDT-1
Lenses RBDT-1, ID3 MONK’s 2 RBDT-1, AQDT-1 & 2
Chess ¼ MONK’s 3 ¼
Car RBDT-1, ID3 Zoo ¼
Shuttle-L-C ¼ Nursery RBDT-1
Connect-4 RBDT-1 Balance RBDT-1, ID3

Table 3 Comparison of tree complexities of the RBDT-1, AQDT-1, AQDT-2 & ID3 methods
using AQ-based rules
Dataset Method Dataset Method
Weekend RBDT-1 Monk’s2 RBDT-1
Lenses RBDT-1, ID3 Monk’s3 ¼
Zoo RBDT-1 Chess ¼
Car RBDT-1 Balance RBDT-1, ID3
Monk’s1 RBDT-1 Shuttle-L-C RBDT-1, AQDT-1 & 2

results in Table 2 are based on comparing the methods using ID3-based rules as
input for the rule-based methods. The ID3-based rules were extracted from the
decision tree generated by the ID3 method from the whole set of examples of each
dataset used. Thus, they cover the whole set of examples (100% coverage). The size
of the extracted ID3-based rules is equal to the number of leaves in the ID3 tree.
40 RBDT-1 Method: Combining Rules and Decision Tree Capabilities 531

In Table 3, the experiment was conducted using AQ-based rules. We used

AQ-based rules generated by the AQ19 rule induction program with 100% correct
recognition on the datasets used as input for the rule-based methods under compar-
ison. No pruning is applied by any method in both comparisons presented in
Tables 2 and 3.
In the following tables, the name of the method that produced a less complex tree
appears in the table in the method column; the “¼” symbol indicates that the same
tree was obtained by all methods under comparison. All four methods run under the
assumption that they will produce a complete and consistent decision tree yielding
100% correct recognition on the training examples. For 7 of the datasets listed in
Table 2, RBDT-1 produced a smaller tree than that produced by AQDT-1 & 2 with
an average of 274 nodes less while producing the same tree for the rest of the
datasets. On the other hand, RBDT-1 produced a smaller tree with 5 of these
datasets compared to ID3 with an average of 142.6 nodes less while producing
the same tree for the rest of the datasets. The decision tree classification accuracies
of all four methods were equal.
Based on the results of the comparison in Table 3, the RBDT-1 method per-
formed better than the AQDT-1 & 2 and the ID3 methods in most cases in terms of
tree complexity by an average of 33.1 nodes less than AQDT-1 & 2 and by an
average of 88.5 nodes less than the ID3 method. The decision tree classification
accuracies of all four methods were equal.

7 Conclusions

The RBDT-1 method proposed in this work allows generating a decision tree from a
set of rules rather than from the whole set of examples. Following this methodo-
logy, knowledge can be stored in a declarative rule form and transformed into a
decision structure when it is needed for decision making. Generating a decision
structure from decision rules can potentially be performed much faster than by
generating it from training examples.
In our experiments, our proposed method performed better than the other three
methods under comparison in most cases in terms of tree complexity and achieves
at least the same level of accuracy.

References

1. I. F. Imam, R.S. Michalski, Should decision trees be learned from examples of from decision
rules? Source Lecture Notes in Computer Science, in Proceedings of the 7th International
Symposium on Methodologies, vol. 689. (Springer, Berlin/Heidelberg, 1993), pp. 395–404
2. J.R. Quinlan, Discovering rules by induction from large collections of examples, in Expert
Systems in the Microelectronic Age, ed. by D. Michie (Edinburgh University Press, Scotland,
1979), pp. 168–201
532 A. Abdelhalim and I. Traore

3. I. H. Witten, B.A. MacDonald, Using concept learning for knowledge acquisition, Int. J. Man
Mach. Stud. 27(4), 349–370 (1988)
4. R.S. Michalski, I.F. Imam, Learning problem-oriented decision structures from decision rules:
the AQDT-2 system, Lecture Notes in Artificial Intelligence, in Proceedings of 8th Interna-
tional Symposium Methodologies for Intelligent Systems, vol. 869 (Springer Verlag, Heidel-
berg, 1994), pp. 416–426
5. Y. Akiba, S. Kaneda, H. Almuallim, Turning majority voting classifiers into a single decision
tree, in Proceedings of the 10th IEEE International Conference on Tools with Artificial
Intelligence, 1998, pp. 224–230
6. Y. Chen, L.T. Hung, Using decision trees to summarize associative classification rules. Expert
Syst. Appl. Pergamon Press, Inc. Publisher, 2009, 36(2), 2338–2351
7. R.S. Michalski, K. Kaufman, The aq19 system for machine learning and pattern discovery: a
general description and user’s guide, Reports of the Machine Learning and Inference Labora-
tory, MLI 01-2, George Mason University, Fairfax, VA, 2001
8. J. Wojtusiak, AQ21 user’s guide. Reports of the Machine Learning and Inference Laboratory,
MLI 04-5, George Mason University, 2004
9. R.S. Michalski, I.F. Imam, On learning decision structures. fundamenta informaticae 31(1),:
49–64 (1997)
10. R.S. Michalski, I. Mozetic, J. Hong, N. Lavrac, The multi-purpose incremental learning
system AQ15 and its testing application to three medical domains, in Proceedings of AAAI-
86, Philadelphia, PA, 1986, pp. 1041–1045
11. F. Bergadano, S. Matwin, R.S. Michalski, J. Zhang, Learning two-tiered descriptions of
flexible concepts: the POSEIDON system. Mach. Learning 8(1), pp. 5–43 (1992)
12. A. Abdelhalim, I. Traore, The RBDT-1 method for rule-based decision tree generation,
Technical report #ECE-09-1, ECE Department, University of Victoria, PO Box 3055, STN
CSC, Victoria, BC, Canada, July 2009
13. A. Asuncion, D.J. Newman, UCI machine learning repository [http://www.ics.uci.edu/mlearn/
MLRepository.html]. University of California, School of Information and Computer Science,
Irvine, CA, 2007
Chapter 41
Computational and Theoretical Concepts
for Regulating Stem Cells Using Viral and
Physical Methods

Aya Sedky Adly, Olfat A. Diab Kandil, M. Shaarawy Ibrahim, Mahmoud

Sedky Adly, Ahmad Sedky Adly, and Afnan Sedky Adly

Abstract Regenerative medicine is the application of tissue, sciences, engineering,

computations, related biological, and biochemical principles that restore the struc-
ture and function of damaged tissues and organs. This new field encompasses many
novel approaches to treatment of disease and restoration of biological function.
Scientists are one-step closer to create a gene therapy/stem cell combination to
combat genetic diseases. This research may lead to not only curing the disease, but
also repairing the damage left behind. However, the development of gene therapy
vectors with sufficient targeting ability, efficiency, and safety must be achieved
before gene therapy can be routinely used in human. Delivery systems based on
both viruses and non-viral systems are being explored, characterized, and used for
in vitro and in vivo gene delivery. Although advances in gene transfer technology
have been made, an ideal system has not yet been constructed. The development of
scalable computer systems constitutes one-step toward understanding dynamics
and potential of this process. Therefore, the primary goal of this work is to develop
a computer model that will support investigations of both viral and non-viral
methods for gene transfer on regenerative tissues including genetically modified
stem cells. Different simulation scenarios were implemented, and their results were
encouraging compared to ex-vivo experiments, where, the error rate lies in the
range of acceptable values in this domain of application.

1 Introduction

The transfer of genes into cells has played a decisive role in our knowledge of
modern cell biology and associated biotechnology especially in the field of regen-
erative medicine. It is an exciting new area of medicine that has attracted much

A.S. Adly (*)

Faculty of Computers and Information, Computer Science department, Helwan University, Cairo,
Egypt
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 533

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_41, # Springer ScienceþBusiness Media B.V. 2010
534 A.S. Adly et al.

interest since the first submission of clinical trials. Preliminary results were very
encouraging and prompted many investigators and researchers. However, the
ability of stem cells to differentiate into specific cell types holds immense potential
for therapeutic use in gene therapy. Realization of this potential depends on
efficient and optimized protocols for genetic manipulation of stem cells. It is widely
recognized that gain/loss of function approaches using gene therapy are essential
and valuable for realizing and understanding specific functions in studies involving
stem cells.
Integrating all these aspects into a successful therapy is an exceedingly complex
process that requires expertise from many disciplines, including molecular and cell
biology, genetics and virology, in addition to bioinformatics, bioprocess
manufacturing capability and clinical laboratory infrastructure. Gene therapy deli-
vers genetic material to the cell to generate a therapeutic effect by correcting an
existing abnormality or providing cells with a new function. Researchers have
successfully induced differentiated cells that can behave like pluripotent stem
cells. These cells are artificially derived from a non-pluripotent cells, typically
adult somatic cells, by inserting certain genes. Induced Pluripotent Stem Cells are
believed to be identical to natural pluripotent stem cells in many respects, but the
full extent of their relation is still being assessed [1–4]. This new field holds the
realistic promise of regenerating damaged tissues, and empowers scientists to grow
them in vitro. This is a completely new concept compared to the traditional
medicine [5–7]. Whereas modeling of gene delivery systems for regulating stem
cell lineage systems is critical for the development of an integrated attempt to
develop ideas in a systematic manner, it has not been a research area that has
received the desired mount of attention. Over the last few years or so, there has been
a noticeable change in this respect, and there is now a growing awareness of
the need to use models and computer simulation to understand its processes and
behaviors.

2 Methods Used in Gene Therapy

Vectors are agents used to carry foreign genes into host cells that will produce the
protein encoded by those genes. They contain an origin of replication, which
enables the vector, together with the foreign DNA fragment inserted into it, to
replicate. The development of efficient model for vector systems is crucial in
supporting success of gene therapy [8–10]. The ideal vector for gene therapy
would introduce a transcriptionally regulated gene into all organs or tissues relevant
to the genetic disorder by a single safe application. However, it rapidly became
obvious that there is no universally applicable ideal vector for gene transfer; each
has its own advantages and disadvantages, see Table 1 [11–14]. In general, a vector
used for gene delivery would have at least the following characteristics: (i) speci-
ficity for the targeted cells; (ii) resistance to metabolic degradation and/or attack by
the immune system; (iii) safety, i.e., minimal side effects; and (iv) an ability to
41 Computational and Theoretical Concepts for Regulating Stem Cells 535

Table 1 Commonly used viral and non-viral vectors

Retrovirus: Viral vector (can carry up to 8,000 base pairs)
Target Dividing cells (possible to target specific types by engineering proteins
on virus surface)
Activation Its RNA travels to the cell’s nucleus to be converted to DNA before genes
are activated
Integration The DNA made from the virus’ RNA template integrates into the host cell’s
genome in random locations, then duplicated when the cell divides
Side effects There is the chance that it will disrupt another gene. Can cause an immune
response, which can be reduced by removing proteins on the surface of
virus that trigger it
Letivirus: Viral vector (can carry up to 8,000 base pairs)
Target Dividing cells and non-dividing cells (possible to target specific cell types by
engineering proteins on the virus surface)
Activation Via infectious particles. Infecting neighboring cells via direct contact with
the host cells
Integration Long-term stable gene expression
Side effects Possible proviral insertional mutagenesis in target cells.
Adenovirus: Viral vector (can carry up to 7,500 base pairs)
Target Dividing cells and non-dividing cells (possible to target specific cell types by
engineering proteins on the virus surface)
Activation Its DNA travels to the cell’s nucleus, where its genes are activated
Integration Its DNA will not integrate, and after a week or two, the cell will discard it
Side effects Can cause an immune response in the patient, which can be reduced by
removing proteins on the surface of the virus that trigger the response
Adeno-associated virus: viral vector (can carry up to 5,000 base pairs)
Target Dividing and non-dividing cells but need assistance of a ‘helper’ virus to
replicate inside cells (possible to target specific cell types by engineering
proteins on the virus surface)
Activation Its DNA travels to the cell’s nucleus, where its genes are activated
Integration Its DNA will integrate into the host cell genome. 95% of the time, it will
integrate into a specific region on Chromosome 19, greatly reducing the
chance other genes disruptions
Side effects These viruses typically will not cause an immune response in the patient
Herpes simplex virus: viral vector (can carry up to 20,000 base pairs)
Target Cells of the nervous system
Activation Its DNA travels to the cell’s nucleus, where its genes are activated
Integration It will not integrate, but can stay in the cell’s nucleus for a long time,
as a separate circular piece of DNA that replicates when the cell divides.
It will not disrupt other genes
Side effects Can cause an immune response in the patient, which can be reduced
by removing proteins on the surface of the virus that trigger the response
Liposome: non-viral vector (there is no maximum length)
Target Not specific for any cell type (enter cells far less effectively than viruses)
Activation The plasmid DNA is transported to nucleus, where its genes will
be activated
Side effects It will not generate an immune response, but some types can be toxic
Naked DNA: non-viral vector (there is no maximum length)
Target Not specific for any cell type (enter cells far less effectively than viruses)
Activation Unless it is specifically engineered to do so, plasmid DNA will not integrate
Side effects It will not generate an immune response, and it is generally not toxic
536 A.S. Adly et al.

express, in an appropriately regulated fashion, the therapeutic gene for as long as

required. In general terms, gene delivery methods can be sub-divided into two
categories: (a) the use of biological vectors and (b) techniques employing either
chemical or physical approaches [11, 13, 15].
Genes may be introduced into cells by transfection or transduction. Transfection
utilizes non-viral methods as chemical or physical methods to introduce new genes
into cells. Usually, small molecules, such as liposomes, are employed to facilitate
the entry of DNA encoding the gene of interest into the cells. Brief electric shocks
are additionally used to facilitate the process. These non-viral methods have a
number of advantages [16, 17]: they are safer, cell independent, and there is no
limitation to the size of the genes that can be delivered. In addition, gene transfec-
tion is a fundamental technology indispensable to the further research development
of basic biology and medicine regarding stem cells. However, despite the fact that it
is widely used to transfer genes into living cells, the steps that limit DNA transfer
remain to be determined. On the other hand, transduction utilizes viral vectors for
DNA transfer. Engineered viruses are very efficient and can be used to introduce
almost any genetic information into cells. However, there are usually limitations in
the size of the introduced gene. Additionally, some viruses only infect specific cells
effectively. In most cases, a genetic material carried by viral vectors is stably
integrated into the host cell genome [18, 19].
The development of efficient computer model for such systems is crucial in
supporting the success of gene therapy. A computational approach towards under-
standing these processes therefore has two major objectives. The first objective is to
provide a qualitative and quantitative explanatory model describing the inner work-
ings of cell regulation. Computer simulations can help linking observations to
hypotheses by reproducing the expected behavior from a theoretical model. The
second objective is concerned with predictive models capable of extrapolating from
a given state of the cell or its components to subsequent states.

3 Proposed Model

Models are used particularly in the natural sciences and engineering disciplines
such as physics, biology, and electrical engineering and can be understood as a
representation of the essential aspects of a certain system, and description of its
behavior in order to presents its knowledge in usable form. This work applies the
idea of intelligent agents where its structure is defined as a specific environment in
which agents are situated. The Capabilities of agents include both ability to adapt
and ability to learn.
Adaptation implies sensing the environment and reconfiguring in response. This
can be achieved through the choice of alternative problem solving rules or algorithms,
or through the discovery of problem solving strategies. Learning may proceed through
observing actual data, and then it implies a capability of introspection and analysis of
behavior and success. Alternatively, learning may proceed by example and
41 Computational and Theoretical Concepts for Regulating Stem Cells 537

generalization, and then it implies a capacity to abstract and generalize. We could

group agents used in this work into five classes based on their degree of perceived
intelligence and capability: Simple reflex agent, Model-based reflex agent, Goal-
based agent, Utility-based agent, and learning agent Figs. 1 through 5 [20].

Simple reflex agent Precepts

Sensors

What is the

Environment
world like now?

Condition / action rules Action to be

(if / then) done

Actions
Actuators

Fig. 1 Block diagram of a simple reflex agent. This type acts only on the basis of the current
percept. The agent function is based on the condition-action rule

Model-based reflex agent Precepts

Sensors

State

How the world evolves What is the

Environment

world like now?

What my actions do

Condition / action rules Action to be

(if / then) done

Actions
Actuators

Fig. 2 Block diagram of a model-based reflex agent. This type can handle partially observable
environments by keeping track of the current state of its world
538 A.S. Adly et al.

Goal-based agent Precepts

Sensors

State
What is the
How the world evolves
world like now?

Environment
What my actions do
What will happen if
I do this action?

Action to be
Goals
done
Actions
Actuators

Fig. 3 Block diagram of a goal-based agent. These agents are model-based agents, which store
information regarding desirable situations

Utility-based agent Precepts

Sensors
State

How the world evolves What is the

world like now?
What my actions do
Environment

What will happen if

I do this action?

Utility How happy I will be in such a state

Goals Action to be done

Actions
Actuators

Fig. 4 Block diagram of a utility-based agent. It is possible to define a measure of how desirable a
particular state is. This measure can be obtained through the use of a utility function which maps a
state to a measure of the utility of the state

The motion for all agents is performed in a stepwise fashion over the space
defined by a regular 2-dimensional grid. Within the grid each object or molecule
occupies a square. The system development will be considered at discrete time
41 Computational and Theoretical Concepts for Regulating Stem Cells 539

Learning agent
Precepts
Performance Standard Sensors

Critic

Environment
Changes
Learning Performance
element element
Knowledge

Problem
Experiments Actions
Generator
Actuators

Fig. 5 Block diagram of a learning agent: Learning has an advantage that it allows the agents to
initially operate in unknown environments and to be-come more competent than its initial
knowledge alone might allow

steps tm ¼ m * Dt, (m ¼ 0, 1, 2, . . .) with a fixed length Dt. In general, models are

constructed to enable reasoning within an idealized logical framework about such
processes and are an important component of scientific theories. Idealized here
means that the model may make explicit assumptions that are known to be incom-
plete in some detail. Such assumptions may be justified on the grounds that they
simplify the model while, at the same time, allowing the production of acceptably
accurate solutions. We could summarize some of the essential points that were
taken into consideration:
l Rule-based agents are the basic entities in the system whose:
Behavior is non-deterministic.
Interactions and responding abilities are governed by a set of rules that define
their behavior.
Rules will often contradict and there must be some mechanism for selecting over
them.
Not aware of the system macro-view although having the capability of respond-
ing to local environmental factors as well as external dynamics and signaling
under various conditions.
Not able to know the state and current behavior of every other agent.
l Only four common types of viral vectors and two of non-viral vectors are
supported: Retroviruses, Adenoviruses, Adeno-associated viruses, Herpes sim-
plex viruses, Liposomes, and Naked DNA.
l Each vector have six main properties that affect its behavior and uses: Type,
Target, Max carrying length, Activation, Integration, Side effects.
540 A.S. Adly et al.

l A virus has a glycoprotein spikes on its surface that helps it to enter or exit the
cell. The H spikes help a virus enter a cell, and the N spikes help it leave.
l After virus transcription, viral proteins are made in the cytoplasm using the host
cell’s ribosomes. Some of these viral proteins move to the nucleus to aid in viral
replication, others move to aid in virus assembly.
l Probability of infection is highly dependent on both viral propagation and
concentration.
l Applying electric pulses to a cell, a process called electroporation, create
transient pores in the plasma membrane that allow transport of traditionally
nonpermeant molecules into cells via both diffusion and electrophoresis.
l In order to successfully deliver non-permeant molecules into cells they must be
located within a critical distance during electroporation in order to reach the
plasma membrane before the pores are closed.
l Delivering non-permeant molecules into cells using electric pulsation depends on
several factors including the field magnitude, pulse duration, pulse interval, pore
size, molecules concentration, cell size, cell type and its cellular response.
l Cells differentiation and regeneration are dynamic phenomena, which are highly
related to time factors, interaction dependent and appear to be stochastic.
l Cell differentiation follows a hierarchical organization.
l Cells which have oscillatory chemical reactions within, could differentiate
through interaction with other cells.
l Differentiated state of a cell is preserved by the cell division and transmitted to
its offspring.
l During cell division, each offspring has two choices, either to preserve its parent
state or to change it, and the probability of each choice seems to depend also on
both its stability and differentiability.
l A kind of memory is formed through the transfer of initial conditions (e.g.
chemical compositions) to offspring.
l Cells with high level of stemness could lead to several distinct types of tissues,
these abilities decrease with lower levels.
l Stem cells normally reside in a specific micro-environmental niche. These cells
can migrate from one crypt to another within their system.
l Maturity is the capability level of a cell for being able to accomplish tasks and
functions and usually when it increases the stemness level decreases.
l The cell society as a whole is stabilized through cell-to-cell interactions e.g.
when the number of one type of cells is decreased by external removal, the
distribution is recovered by further differentiations, replications, renewal, and
regeneration.
l Each cell has its own internal dynamics and genetic switching according to its
state and condition.
l Interactions among cells are given by the diffusive transport of chemicals
between each cell and its neighboring according to their environment.
l Within each cell, a network of biochemical reactions also includes reactions
associated with genetic expressions, signaling, pathways, chemical concentra-
tions, and so on.
41 Computational and Theoretical Concepts for Regulating Stem Cells 541

l A cellular state is represented by both its internal and external dynamics.

l Cell volume is treated as a dynamical variable that increases as a result of
transportation of chemicals from the environment, which leads to the cell
division when it is larger than a given threshold.
l Here the environment doesn’t mean external environment for individual organ-
ism, but is intended as interstitial environment of each cell.
l The concentrations of chemicals in the daughter cells are almost equal to the
concentrations of the mother cell.
l The rates of chemicals transported into a cell are proportional to differences of
chemicals concentrations between cell inside and outside.
l Some chemicals can penetrate the membrane and others cannot. The behavior of
dynamics depends on their number and ability. This penetration ability should be
different for different cell states.
l Each cell takes penetrable chemicals from the medium, while the reactions in the
cell usually transform them to impenetrable chemicals.
l The nature of internal dynamics depends on the choice of the reaction network,
and on the number of paths in the reaction matrix. When the number of paths is
small, cellular dynamics may falls into a steady state without oscillation. On the
other hand, when it is large, many chemicals generate each other. Then chemical
concentrations take constant values, which are often almost equal. Any way it is
not easy to estimate the number of paths in real biochemical data, although they
may suggest the medium number of paths.
l Number of penetrating chemicals is another control parameter for capacity of
oscillation or differentiation. When it is small, the rate of reaction networks,
which show oscillatory dynamics, is small. On the other hand, when it is too
large, it would be difficult to have oscillatory dynamics.
l The oscillatory dynamics is rather common as the number of auto-catalytic paths
is increased.
l Tiny changes during division amplify the cell instability. When it exceeds some
threshold, the differentiations start. Then the emergence of another cell type
stabilizes the dynamics of each cell again.
l Each internal dynamics of a certain type of cells is gradually modified with the
change of distribution of other cells.
l There is a higher level dynamics, which controls the rate of cell division and
differentiation according to the number of each cell type. These dynamics should
be stochastic, and robust against external perturbations.

4 Simulation Results

Stem cell lines vary in their characteristics and behavior not only because they are
derived from genetically outbred populations, but also because they may undergo
progressive adaptation upon long-term culture in vitro.
542 A.S. Adly et al.

Hence, the task is to model the dynamic processes that drive and control the
cellular attributes. Clearly, it is presently impossible to describe these processes in
any reasonable detail. Therefore, it will be necessary to propose a simplified basic
scheme of the cellular dynamics.
Therefore, different simulation scenarios were observed, each with different
results, and different error rates lies in the range of acceptable values in this domain
of application. We present the results for some of these simulations along with
comparisons to experimental results.
To validate utility and performance we attempt to simulate and observe 16
fundamentally different but important cellular phenomena for which experimental
measurements were available or for which results are well characterized. These
included: (1) Intracellular Signaling; (2) Signaling Ligands; (3) Transcription
Factors; (4) Cell Surface Receptors; (5) Cell Cycle; (6) Apoptosis; (7) DNA Repair;
(8) Protein Synthesis; (9) Protein Folding; (10) Protein Processing; (11) RNA
Binding; (12) Metabolism; (13) Transport; (14) Chromatin Regulators; (15) Cyto-
skeleton; (16) Adhesion. These processes span the range of most kinds of cellular
behavior, from simple chemical kinetics to more complex metabolism and genetic
circuits. For a number of these phenomena we investigate its utility in monitoring or
simulating the effects of finite numbers and stochasticity on certain processes.
For observing non-viral transfection effect we attempt to simulate the process of
electroporation of Chinese hamster ovary cells (CHO), which are spherical in
suspension with a mean diameter of 13.5F1.4 Am, that isn’t significantly different
from control cells (P ¼ 0.01) [21, 22]. After the aphidicolin (a reversible inhibitor of
nuclear DNA replication, used for blocking the cell cycle at early S phase) treatment,
cells synchronized in G2/M phase significantly increased in size, their diameter
reaching 17.1F1.55 Am. Indeed, in aphidicolin synchronized cells, synthesis of
proteins and lipids has been shown to be still present while cell cycle was blocked.
These syntheses led to an increase of cell volume, which is in agreement with
previous experimental values investigated by Muriel, et al. [21], and
Sukhorukov, et al. [22]. However, cells were pulsed under optimum conditions for
gene transfer. Ten pulses, lasting 5 ms, were applied. Under those conditions, 63% of
control cells were permeabilized at 0.8 kV/cm and 81% at 0.9 kV/cm (Fig. 6a). The
associated fluorescence intensity, related to the number of molecules incorporated
into the electropermeabilized cells (Fig. 6b), increased with an increase in the
electric field intensity. The amount of electroloaded molecules in permeabilized
cell increased with the electric field intensity, in agreement with previous experi-
mental results [21, 23]. The percentage of permeabilized cells and the permeabiliza-
tion efficiency for cells synchronized in G1 phase were slightly smaller than for
control cells (Fig. 6a, b). Cells synchronized in G2/M phase exhibited a one and a
half fold increase in the percentage of permeabilization at 0.8 kV/cm and a slight
increase at 0.9 kV/cm (Fig. 6a). This synchronization effect was more pronounced
when fluorescence intensities were compared. Electropermeabilized cells in G2/M
phase exhibited a two-fold increase in fluorescence intensity at 0.8 and 0.9 kV/cm
(Fig. 6b). However, this increase had to be associated with the fact that in G2/M
phase, cells had 4N instead of 2N chromosomes and therefore two times more sites
41 Computational and Theoretical Concepts for Regulating Stem Cells 543

a b
500
100

Fluorescence (a.u.)
Permeabilization %

80 400

60 300

40 200

20 100

0 0
0.8 0.9 0.8 0.9
E (kv / cm) E (kv / cm)
Control Cells Control Cells
Aphidicolin synchronized cells Aphidicolin synchronized cells
Sodium butyrate synchronized cells Sodium butyrate synchronized cells

Fig. 6 Simulation results: (a) Percentage of permeabilized cells. Cells were submitted to 10
pulses of 5 ms duration in pulsing buffer having propidium iodide. Permeabilization was deter-
mined based on total number of cells. (b) Associated mean fluorescence level of the permeabilized
cells. “Values represent means”

a b
Transduction Effieciency %

100
100
Transduction Effieciency %

90
80
70 80
60
50 60
40 ± ±
30 % 40 %
20
10
0 20
Cycle 1 Cycle 2 Cycle 3
0
Number of cycles Static Centrifugation

Fig. 7 (a) Effect of transduction cycle number on efficiency of transduction, error rate ¼ 11.64%.
(b) Static transduction versus centrifugational transduction, error rate ¼ 5.22%

for PI interaction [21, 24]. These simulation values were in agreement with previous
experimental values investigated by Muriel, et al. [21].
For observing viral transduction effect we attempt to simulate the process of self-
renewal and differentiation of Human Fetal Mesenchymal Stem Cells (hfMSCs)
after transduction with onco-retroviral and lentiviral vectors. Retroviral and lenti-
viral approaches offered the initial methodology that launched the field, and estab-
lished the technological basis of nuclear reprogramming with rapid confirmation
across integrating vector systems [25, 26]. However a static transduction of three
samples of first trimester fetal blood-derived hfMSCs from 9+1, 9+6, and 10+6 weeks
of gestation revealed a mean efficiency of 20.3%  7.2%, 38.6%  6.4%, and 47%
 3.2% after 1, 2, and 3 cycles of transduction, respectively with an error rate ¼
11.64% (Fig. 6c, 7, 8 and 9). Simulating transduction under centrifugational forces
544 A.S. Adly et al.

a b

Mean fluorescence intensity (MFI)

100 2000
Transduction Effieciency %

90 1800
80 1600
70 1400
60 1200
50 1000
40 800
30 600
20 400
10 200
0 0
3 6 11 22 3 6 11 22
Multiplicity of Infection Multiplicity of Infection

Fig. 8 (a) Percentage of lentiviral-transduced cells with varying MOI, error ¼ 5.4%. (b) Mean
fluorescence intensity with varying MOI after lentiviral transduction, error ¼ 9.8%

Cytoskeleton Adhesion Intracellular

5% 6%
Signaling
Chromatin 15%
Regulators
Signaling
3%
Ligands
Transport
4%
6%
Metabolism Transcription
13% Factors
13%

Cell Surface
RNA Binding Receptors
4% 9%

Cell Cycle
Protein 3%
Processing Protein Apoptosis
8% Folding Protein DNA Repair 1%
4% Synthesis 3%
5%

Fig. 9 A pie chart depicting the simulated average results of distribution within the HSC profile
for gene products with known or putative functions

also demonstrated roughly a twofold increase in transduction efficiency (n ¼ 3) from

37.8%  2.6% to 77.3%  6.8% after 2 cycles (paired t-test; p ¼ 0.002) with an error
rate ¼ 5.22% (Fig. 6d). And using a lentiviral vector agent, transduction efficiencies
were of 98.9%  1.1% having error rate ¼ 5.4% with MFIs of 2,117  633 were
achieved with the same three samples and with error rate ¼ 9.8% (Fig. 6e,f). This
was achieved after a single cycle at MOI of 11. Raising viral titers was also
detrimental to the hfMSCs, with massive cell death observed after MOI of >22.
41 Computational and Theoretical Concepts for Regulating Stem Cells 545

These simulation values were in agreement with previous experimental values

investigated by Chan, et al. [27].
Generally, for observing the basic cellular phenomena we attempt to simulate
the genome wide transcriptional changes during early stages of hematopoietic
differentiation (Fig. 6g) and the simulated values were in agreement with previous
experimental values investigated by Ivanova, Dimos et al. [28].

References

1. A.L. David, D. Peebles, Gene therapy for the fetus: is there a future? Best Pract. Res. Clin.
Obstet. Gynaecol. 22(1), 203–218 (2008)
2. M. Baker, (2007-12-06). Adult cells reprogrammed to pluripotency, without tumors. Nature
Reports Stem Cells. Retrieved on 2007-12-11
3. R.S. Oliveri. Epigenetic dedifferentiation of somatic cells into pluripotency: cellular alchemy
in the age of regenerative medicine? Regen. Med. 2(5), 795–816 (Sept 2007)
4. P. Collas, C.K. Taranger, Toward reprogramming cells to pluripotency. Ernst Schering Res
Found Workshop. 2006;(60):47–67
5. J. Yu, J. Domen, D. Panchision, T.P. Zwaka, M.L. Rohrbaugh, et al. Regenerative medicine,
NIH Report, 2006
6. A. Atala, Recent developments in tissue engineering and regenerative medicine. Curr. Opin.
Pediatr. 18(2), 167–171, 2006
7. Y. Junying, J.A. Thomson, Regenerative Medicine, 2006
8. M. Colavito, M.A. Palladino, Gene Therapy, 1st edn. (Benjamin Cummings, Aug 2006)
ISBN-10: 0805338195
9. D.V. Schaffer, W. Zhou, Gene Therapy and Gene Delivery Systems (Springer, Feb 2006)
ISBN: 3540284044
10. A. Battler, J. Leor, Stem Cell and Gene-Based Therapy: Frontiers in Regenerative Medicine
(Springer, June 2007) ISBN: 184628676X
11. J. Phillips, C. Gersbach, A. Garcia, Virus-based gene therapy strategies for bone regeneration.
Biomaterials 28, 211–229 (2007)
12. Y. Yi, S.H. Hahm, KH. Lee, Retroviral gene therapy: safety issues and possible solutions.
Curr. Gene Ther.5, 25–35 (2005)
13. L.F. Maxfield, C.D. Fraize, J.M. Coffin, Relationship between retroviral DNA-integration-site
selection and host cell transcription. Proc. Natl. Acad. Sci. USA 102, 1436–41 (2005)
14. W. Weber, M. Fussenegger Pharmacologic transgene control systems for gene therapy. J Gene
Med 8(5), 535–556 (2006)
15. S. Mehier-Humbert, R.H. Guy, Physical methods for gene transfer: improving the kinetics of
gene delivery into cells. Adv. Drug Deliv. Rev. 57 733–753 (2005)
16. C. Capaccio, N.S. Stoykov, R. Sundararajan, D.A. Dean, Quantifying induced electric field
strengths during gene transfer to the intact rat vasculature. J. Electrostat. 67 652–662, (2009)
17. J.P. Roberts, Gene therapy’s fall and rise (again). Scientist 18(18), 22–24 (2004)
18. U. Bastolla, M. Porto, H.E. Roman, M. Vendruscolo, Structural Approaches to Sequence
Evolution: Molecules, Networks, Populations, 1 edn. (Springer, Jan 2007) ISBN-10:
3540353054
19. A.J. Nair, Introduction to Biotechnology and Genetic Engineering. Firewall/isp (May 2008)
ISBN: 8131803392
20. S.J. Russell, P. Norvig, Artificial Intelligence: A Modern Approach, 2nd edn. (Prentice Hall,
Upper Saddle River, NJ, 2003) ISBN: 0-13-790395-2
546 A.S. Adly et al.

21. M. Golzio, J. Teissie, M.-P. Rols, Cell synchronization effect on mammalian cell permeabi-
lization and gene delivery by electric field. Biochimica et Biophysica Acta 1563(2002),
23–28, (2002)
22. V. Sukhorukov, C. Djuzenova, W. Arnold, U. Zimmermann, J. Membr. Biol. 142 77–92
(1994)
23. M.P. Rols, J. Teissie, Biophys. J. 58 1089–1098 (1990)
24. H.A. Crissman, J.A. Steinkamp, Eur. J. Histochem. 37 (1993) 129–138, 1993
25. T.J. Nelson, A. Terzic Induced pluripotent stem cells: reprogrammed without a trace. Regen.
Med. (2009). doi:10.2217/rme.09.16
26. C.A. Sommer, M. Stadtfeld, G.J. Murphy et al. Induced pluripotent stem cell generation using
a single lentiviral stem cell cassette. Stem Cells (2009). doi:10.1634/stemcells.2008-1075
27. J. Chan, K. O’Donoghue, J. de la Fuente et al. Human fetal mesenchymal stem cells as
vehicles for gene delivery. Stem Cells 23, 93–102 (2005)
28. N.B. Ivanova, J.T. Dimos, C. Schaniel et al. A stem cell molecular signature (Sept 2002).
doi:10.1126/science.1073823
Chapter 42
DFA, a Biomedical Checking Tool for the Heart
Control System

Toru Yazawa, Yukio Shimoda, and Tomoo Katsuyama

Abstract We made our own detrended fluctuation analysis (DFA) program. We

applied it to the cardio-vascular system for checking characteristics of the heartbeat
of various individuals. Healthy subjects showed a normal scaling exponent, which
is near 1.0. This is in agreement with the original report by Peng et al. published in
1990s. We found that the healthy exponents span from 0.9 to 1.19 in our temporary
guideline based on our own experiments. In the present study, we investigated the
persons who have an extra-systole heartbeat, so called as premature ventricular
contractions (PVCs), and revealed that their arrhythmic heartbeat exhibited a low
scaling exponent approaching to 0.7 with no exceptions. In addition, alternans,
which is the heartbeats in period-2 rhythms, and which is also called as the
harbinger of death, exhibited a low scaling exponent like the PVCs. We may
conclude that if it would be possible to make a device that equips a DFA program,
it might be useful to check the heart condition, and contribute not only in statistical
physics but also in biomedical engineering and clinical practice fields; especially
for health check. The device is applicable for people who are spending an ordinary
life, before they get seriously heart sick.

1 Introduction

The healthy heart muscles can contract periodically and rhythmically, so the heart
can supply the blood to the entire body through the cardio-vascular system. Since
significant decrease or complete stop of the blood flow caused by heart’s illness
can immediately be lethal, the sick hearts are evidently facing death at every turn
more than the healthy heart. Regardless of advancement of medical technology,

T. Yazawa (*)
Department of Biological Science, Tokyo Metropolitan University, 1-1 Minami-Ohsawa,
Hachioji, 192-0397 Tokyo, Japan
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 547

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_42, # Springer ScienceþBusiness Media B.V. 2010
548 T. Yazawa et al.

cardio-vascular failure especially life threatening attacks are still taking place
unpredictably in normally looking subjects. Gradually changing activity of cellu-
lar and tissue systems must be involved. Therefore, detection of “hidden” early
symptom of normally looking subjects is one of the main goals. For understand-
ing hidden information in the control system, nonlinear analysis has been pro-
posed to be an applicable method since decades ago. Indeed, the DFA is a method
that is potentially useful to diagnose a heart condition [1]. However, it has not
been appeared as a practical tool for the use in a bio-medical engineering and
clinical practice field. We have made our own DFA program (by K. Tanaka) [2].
We here demonstrate that our method can be useful to calculate the scaling
exponent for persons, who have an extra-systolic heartbeat, for persons who
have alternans-heartbeats, and for other types of heartbeats, including normal,
healthy hearts.

2 Methods

2.1 Finger Blood-Pressure Pulse

Heartbeats have been recorded by a conventional electrophysiological method. For

sensing the heartbeats, we performed it either electrically (EKGs) or mechanically
(blood pressure pulses) in our series of studies. From human subjects we mostly
used the finger pulse recording with a Piezo-crystal mechano-electric sensor,
connected to a Power Lab System (AD Instruments, Australia). All data were
registered at 1 kHz in rate of sampling.

2.2 DFA Methods

We made our own programs for measuring beat-to-beat intervals (by Tanaka) and
for calculating the approximate scaling exponent. The DFA program (by Tanaka)
can instantly calculate the scaling exponents if a time series data exist.
After recording EKG or finger pressure pulses, we first made the beat-to-beat
interval time series. Preferable data length could be from some hundreds to about
2,000 heartbeats. From the time series we calculated the approximate scaling
exponent. For preparing figures in this report, we presented (1) heartbeat-data by
plotting heart rates against time, which is the representation of the original data
used for the DFA, (2) he approximate scaling exponents calculated by the DFA, and
(3) a graph with one or sometime three least mean square regression lines. From the
graph, one can read an approximate slope to which the exponent corresponds
directly. While beating hearts show a fluctuating pattern, we hardly predict the
scaling exponent by eye-observation of such dynamic patterns visualized as the
42 DFA, a Biomedical Checking Tool for the Heart Control System 549

time series. Only a DFA program can do that. The scaling exponent reflects hidden
information carried by quantitative measures in a dynamically changing interval of
heartbeat and velocity of blood flow under the functional control of the autonomic
nervous system. An EKG-chart data cannot tell us what the scaling exponent is.
Mathematical procedures of DFA and explanation of biomedical meaning have
been documented elsewhere [1, 3].

2.3 Volunteers and Ethics

Subjects were selected from colleagues of university laboratories including our-

selves and companies, individuals of which visited to our exhibition place and
asked us to record and test his/her heartbeats at the Innovation Japan 2006, 2007,
2008, and 2009, for example, NOMS Co. Ltd., Nihon-Koden Co. Ltd. and Toshiba
Co. Ltd, across over 100 companies. All subjects were treated under the ethical
control regulation of our University Committee.

3 Results

3.1 Extra-Systole

The subject, whose heart had an arrhythmia due to the ventricular extra-systole,
exhibited an abnormal record such as shown in Fig. 1. This is so called as premature
ventricular contractions (PVCs). We here show one example obtained from a
female subject age 58. We have over 20 subjects who exhibited the ventricular
extra-systole. All of them, and importantly with no exceptions, showed the same
results; that is, a lower exponent.
In the ventricle extra-systole, ventricle muscle generally tried to generate its own
rhythm before normal rhythmic excitation arrived to the ventricle. Therefore a

Fig. 1 Typical extra-systole.

Recorded from a finger by a
piezo-crystal pulse-recording
device. We consider a small
beat is one heartbeat (see
asterisks). 58 years female
550 T. Yazawa et al.

lower/smaller pulse, comparing to a normal pulse, was generated like a small pulse
in Fig. 1. But this PVC is considered by physicians as a not so serious problem of
the heart, if this happens less frequently. Thus, it is “benign” arrhythmia, though we
cannot agree with the idea, because of DFA results as shown below. If it may
happen a lot, like more than 10 times per 1 min, physicians may take great care of it
finally.
As shown in Fig. 2, this person’s extra-systole appeared less frequently, the
count of which was 12 within the span of the record. The period length of this
recording was for about 5,800 heartbeats, which corresponded to about 1.5 h.
Therefore, this person’s heart is “normal” in terms of a physician’s guideline.
However, we found out that our DFA’s results was that the subject’s heart shown
in Figs.1 and 2 exhibited low exponents, about 0.7. Followings are brief description
how we calculate the scaling exponents.
For the computation of the scaling exponents, at a final step, we need to calculate
the slope in the graph of Box-size vs. Variance as shown in Fig. 3. Here we used
Box-size across a range 100–1,000. Before this final step, there are procedures that
are much complex. In the mid-way of the DFA, the least-mean square methods
were involved which are linear-, quadratic-, cubic-, and bi-quadratic-fitting meth-
ods. Details of those “mid-way” procedures are not shown here. The graph Fig. 3 is
only the final step. The scaling exponents, being made after a measurement of such
a least-mean square fitting functions are indicated as inset. When we repeated
linear-, quadratic-, cubic-, and bi-quadratic fitting, we are finally able to obtain a
steady result of calculation. For example, there is no big difference between B and C

Fig. 2 Time series data of a

long term recording. Twelve
PVCs (premature ventricular
contractions) are recorded. 58
years female shown in Fig. 1

Fig. 3 An example DFA.

The scaling exponent for this
heart is 0.65 or 0.67. 58 years
female
42 DFA, a Biomedical Checking Tool for the Heart Control System 551

in Fig. 3. Consequently, we learn that the result B (and C) is the scaling exponent.
From this calculation, an approximate scaling exponent was finalized for this
subject. The figure indicates that the scaling exponent of this person is way down
below from the normal value (i.e., 1.0). We determined that the scaling exponent
for this heart is about 0.7.
This (0.7) is not normal. We were intrigued her heartbeats and she kindly
allowed us to measure finger pulses multiple times over the course of 6 years.
Figure 4 shows the result of our “automatic calculation” program. The data length
was about 7,000 beats. The scaling exponents shown here is 0.91 and 0.65. The
former value corresponds to 30–90 box-size and the latter to 70–240 box-size. In a
short window size, i.e., 30–90, her heart looks fine but in long window size, 70–240,
which is similar range to those shown in Fig. 3, we obtained a low scaling exponent,
0.65. Her body system is not perfect. Indeed, we never see the perfect 1/f rhythm in
her data.
The ‘skipping heartbeat’ makes the subject feel very uncomfortable. She
described that she was not feeling very well, even though the physician says it is
okay. It might be not very good psychologically. Needless to say, ventricular
tachycardia is a serious cause for a life threatening cardiac dysfunction. Impor-
tantly, this tachycardia is derived from randomly repetitive, very fast beatings, in
other words, a series of extra-systoles. Although the causes of the extra-systole have

Fig. 4 An example of automatic computation of the DFA. The same subject shown in Figs. 1
through 3. (a) Finger-pulse test, “*” An asterisk shows PVC. (b) About 7,000 heartbeats.
Sporadically appearing extra-systoles can be seen by downward swings. (c) DFA results at
different box size. (d) Slopes corresponding to the scaling exponents in (c). 58 years female
552 T. Yazawa et al.

not been fully understood physiologically by now, we confirmed by DFA that at

least one reason of this person’s extra-systole is due to the autonomic nervous
system dysfunction, because, surprisingly enough, this person’s extra-systole
totally disappeared later. After 2 years, she nicely explained that her (58 years)
environments had improved, in other words, the psychological stress of her life had
been decreased during the years. As a result, we luckily discovered that the reason
why the person’s scaling exponent was pushed up as close as to 0.9 (data not
shown).

3.2 Alternans with Low Exponent

We have examined DFA of an Alternans person (Fig. 5). The Alternans was
documented in 1872 by Traube. However, the alternans did not get particular
attention from doctors until recently, indeed before alternans was noticed as
harbinger of death. We finally found that alternans heartbeat always exhibited a
low scaling exponent when we measure heartbeats of animal models (crustacean
animals such as crabs and lobsters, n ¼ 13) and humans (n ¼ 8, five Japanese and
three Americans).

3.3 Extraordinary High Exponent

We have met a person who had a high scaling exponent, which was 1.5 at 30–60
box-size (Fig. 6). This man was tall but not fat, and apparently healthy. The time
series look like as being normal (Fig. 9b). However, his heart rate at rest was
relatively high (HR ¼ 70 per min, Fig. 9b) comparing to that of normal man (HR ¼
60 per min, Fig. 9a). We wondered why. We asked him a possible explanation, and
finally found out that he has two more brothers, who both had already died because

Fig. 5 Alternans shows very

low scaling exponents. Heart
rate trace clearly shows an
abnormality of its rhythm.
The scaling exponents were
0.5–0.7 at different box size.
58 years male. We have four
human alternans subjects and
many model animals too. All
of them exhibited a low
scaling exponent like this
example
42 DFA, a Biomedical Checking Tool for the Heart Control System 553

Fig. 6 A DFA example for a person who has a high scaling exponent. 59 years male

Fig. 7 A typical normal DFA

for a middle aged subject.
56 years female

of detail-unknown heart problem. This high exponent might be some sign for a
certain heart or heart-control system problems, although we have no idea for the
reason. We consider that cardiac physicians and geneticists may identify the reason,
why the family has a potential risk, if they study his family.

3.4 Normal Exponent

The DFA revealed that middle age people often show a normal exponent or a little
lower exponent (Fig. 7). As shown in Fig. 7, the scaling exponents were unstable at
a different box size. We would say that this person’s heart could be normal, but
having a stressful life, because the DFA showed roughly 0.77 when measuring the
entire range of the box size. If a person is in perfect health, the person would have
an exponent of 1.0 at entire range of the box size.
554 T. Yazawa et al.

Fig. 8 EKG obtained at

hospital health check.
54 years male

Fig. 9 Time series data of two persons. (a) Normal size healthy mail, age 53. This heartbeat looks
normal. This person’s EKG during this health check (Fig. 8) showed no problem with his heart, but
his scaling exponent was 0.8, slightly below than normal (1.0). The detailed examination of his
pulse record finally revealed that he had abnormal heartbeats. (b) A tall, close to 2 m high, healthy
mail, age 59. This time the series again looked kind of normal. But this person had a scaling
exponent of 1.5 (Fig. 6). Both (a) and (b) were recorded simultaneously at rest. Note that heart rate
of (b) is greater than that of (a) in average, though (b) is greatly taller than (a). Generally, big body
mass lowers heart rate but here not the case

3.5 DFA Is Beneficial

For the medical examination of EKG, usually five or six pulses has been recorded
as shown in Fig. 8. In this record, the each wave form looks normal and also it
shows that there are no Q-T elongation and no S-T elevation. However, the DFA
revealed, that his heart exhibits a low scaling exponent, 0.8 (see Fig. 9a). So we
performed a detailed examination of his pulse record for a period of 2 h. And
finally we discovered that he had an abnormal heartbeat, such as an extra-systole
and even tachycardia too, although those events occurred rarely (data not shown).
The findings imply that his cardio-vascular control system must have some
problems. Even though we cannot tell what the physiological mechanism is at
the moment, DFA can find out the some physiological problem on cardio-vascular
system.

4 Discussion

To perform DFA we normally need a pretty long term recording of heartbeats [1].
On the other hand, for an ordinal EKG done for a hospital health check, it would
take a very short period of time. Indeed, a typical EKG needs only 6 beats or so, as
42 DFA, a Biomedical Checking Tool for the Heart Control System 555

shown in Fig. 8. Therefore, EKG recordings at medical health check unfortunately

are useless for the DFA due to the short period length of recording.
Other important recording condition is the rate of digital sampling. Our DFA
method needs the heartbeat recording at 1 kHz sampling rate, i.e., 1,000 dot per
second. This notion was derived from our own previous research. According to our
own cardio-vascular experiments on crabs and lobsters and much older great
heritages of the crustacean heart research [4, 5, 11], we have had experienced that
cardio-regulatory nerve impulses exhibits fluctuation of about 1 ms in time scale
[6, 7]. Since the autonomic nerve activities directly affect the fluctuation of the rate
of heartbeat [8], we must observe the fluctuation of the heartbeat in the time scale of
1 ms at a lowest limit. Therefore, we considered that the heartbeat fluctuation must
be observed with a time scale in millisecond order, although the size of data
increases dramatically comparing to the way of ordinal EKG recording (normally
4 ms). This notion which is derived from our crustacean studies must be applicable
to the human heart too, because life on earth uses identical DNA-genetic control
systems as the conventional way in the development of structure and function of the
cardio-vascular system [9–11]. Unfortunately, the conventional Holter EKG
recording system has not equipped 1 ms resolution in time scale (about 4–8 ms).
So, we needed to record human heartbeat by our own machine. This is the reason
why we preferred our own system to collect data instead of using the existing data
base. Thus, we are always performing recording at a rate of 1 ms (1 kHz sampling)
for the rate of data sampling.
Using the DFA, and using the principles of 1/f scaling in the dynamics of
the cardio-vascular system [12], we have demonstrated the scaling exponents
reflect the balance of the control system, that is, the autonomic nervous system.
In other words, the DFA can determine whether the entire body system is function-
ing in sick or in healthy, simply by the scaling exponent. If we notice that the
scaling exponents are lower, we must assume that the subjects are not normal and
the subjects might be exhibiting abnormal “stressful” heartbeats.
Empirical evidences illustrated the dynamics of adaptive heart movement
behavior. The dynamic behavior results from in a range of multi-elementally
actions in the heart and nerve cells. Those are including molecules contributing
for heart muscle contraction, molecules for regulating ionic flow of the myocar-
dium, molecules for generating pace-making signals for contraction, molecules for
neuro-muscular transmission, and so on. We expect that future research will
identify parameters that determine the interaction between a heart function (e.g.,
generating rhythmically from inside the heart) and control processes (e.g., generat-
ing rhythmically from inside the brain). Moreover, we must consider the interaction
between the heart and external factors in the environment (e.g., circadian rhythms)
and biological, behavioral and medical factors (e.g., stress, fighting, and disease).
However, we must pay attention to the fact that all of such elements are nonlinearly
connected. That means that the best way to identify parameters is the way of
complex dynamics analysis performing together with traditional physiological
methods. Therefore, collaborating intimately together of biology and physics is
the evocative solution in the future study.
556 T. Yazawa et al.

We might conclude that the DFA has benefits. As a tool for diagnosing the
condition of the cardio-vascular system, DFA can find out something more than
those tests, which are not be able to be checked out by a simple traditional EKG.

Acknowledgment We are very grateful to all volunteers, for allowing us to test their heartbeats by
finger pulse tests. This research was supported by the grants 2008-JST04-065, H20DG407 NOMS,
and the grant TMU-San-Gaku-Ko-H20Project-D56I0106.

References

1. C-K. Peng, S. Havlin, HE. Stanley, AL. Goldberger, Quantification of scaling exponents and
crossover phenomena in nonstationary heartbeat time series. Chaos 5, 82–87 (1995)
2. T. Yazawa, K. Tanaka, T. Katsuyama, Neurodynamical control of the heart of healthy and
dying crustacean animals. Proceedings CCCT05, vol. 1 (Austin, TX, USA, July 24–27 2005),
pp. 367–372
3. T. Katsuyama, T. Yazawa, K. Kiyono, K. Tanaka, M. Otokawa, Scaling analysis of heart-
interval fluctuation in the in-situ and in-vivo heart of spiny lobster. Panulirus japonicus, Bull.
Univ. Housei Tama 18, 97–108 (2003) (Japanese)
4. J.S. Alexandrowicz, The innervation of the heart of the crustacea. I. Decapoda, Quat. J.
Microsc. Sci. 75, 181–249 (1932)
5. D.M. Maynard, Circulation and heart function. The Physiology of Crustacea, vol. 1
(Academic, New York, 1961), pp. 161–226
6. T. Yazawa, K. Kuwasawa, Intrinsic and extrinsic neural and neurohymoral control of the
decapod heart. Experientia 48, 834–840 (1992)
7. T. Yazawa, T. Katsuyama, Spontaneous and repetitive cardiac slowdown in the freely moving
spiny lobster, Panulirus japonicus. J. Comp. Physiol. A 87, 817–824 (2001)
8. J.P. Saul, Beat-to-beat variations of heart rate reflect modulation of cardiac autonomic
outflow. News Physiol. Sci. 5, 32–27 (1990)
9. W.J. Gehring, Master Control Genes in Development and Evolution: The Homeobox Story
(Yale University Press, New Haven, 1998)
10. I. Sabirzhanova, B. Sabirzhanov, J. Bjordahl, J. Brandt, P.Y. Jay, T.G. Clark, Activation of
tolloid-like 1 gene expression by the cardiac specific homeobox gene Nkx2-5. Dev. Growth
Differ. 51, 403–410 (2009)
11. J.L. Wilkens, Evolution of the cardiovascular system in Crustacea. Am. Zool. 39, 199–214
(1999)
12. M. Kobayashi, T. Musha, 1/f fluctuation of heartbeat period. IEEE Trans. Biomed. Eng. 29,
456–457 (1982)
Chapter 43
Generalizations in Mathematical Epidemiology
Using Computer Algebra and Intuitive Mechanized
Reasoning

Davinson Castaño Cano

Abstract We are concerned by imminent future problems caused by biological

dangers, here we think of a way to solve them. One of them is analyzing endemic
models, for this we make a study supported by Computer Algebra Systems (CAS)
and Mechanized Reasoning (MR). Also we show the advantages of the use
of “CAS” and “MR” to obtain in that case, an epidemic threshold theorem. We
prove a previously obtained theorem for SnIR endemic model. Moreover using
“CASþMR” we obtain a new epidemic threshold theorem for the SnImR epidemic
model and for the staged progressive SImR model. Finally we discuss the relevance
of the theorems and some future applications.

1 Introduction

At the moment, we are at the edge of a possible biological problem. Some people
say that the nineteenth century was the century of chemistry, the twentieth was the
century of physics, and they say that the twenty-first will be the century of biology.
If we think, the advances in the biological field in the recent years have been
incredible, and like the physics and its atomic bomb, with biology could create
global epidemics diseases. Also the climate change could produce a new virus
better than the existing virus, creating an atmosphere of panic. For these reasons
and others, we think in a solution using mathematical models with computer
algebra and mechanized reasoning. Specifically we consider the SIR (Suscepti-
ble-Infective-Removed) model, with differential susceptibility and multiple kinds
of infected individuals. The objective is to derive three epidemic threshold

D.C. Cano
Logic and Computation Group, Engineering Physics Program, EAFIT University, Carrera 49 N
7 Sur, 50, Medellı́n, Suramérica, Colombia
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 557

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_43, # Springer ScienceþBusiness Media B.V. 2010
558 D.C. Cano

theorems by using the algorithm MKNW given in [1] and a little bit of mechanized
reasoning.
Briefly the MKNW runs on: Initially we have a system of ordinary non-lineal
differential equations S, whose coefficients are polynomial. We start setting all
derivates to zero for finding equilibrium; we solve the system finding the equilib-
rium point T. Then we compute the Jacobian Jb for the system S and replace T in S.
We compute the eigenvalues for Jb; from the eigenvalues we obtain the stability
conditions when each eigenvalue is less than zero. Finally we obtain the reproduc-
tive number for the system S in the particular cases. Using deductive reasoning we
obtain some theorems based on the particular cases.
The MKNW algorithm is not sufficient to prove the threshold theorems that will
be considered here and for this reason, it is necessary to use some form of
mechanized reasoning, specifically some strategy of mechanized induction.
The threshold theorem that we probe in Section 2 was originally presented in [2]
using only pen and paper and human intelligence. A first contribution of this paper
is a mechanized derivation of such theorem using CAS.
The threshold theorem to be proved in Section 3 is original and some particular
cases of this theorem were previously considered via CAS in [3, 4] and without
CAS in [5].
The threshold theorem to be proved in Section 4 is original and similar models
were before considered without CAS in [6].

2 CA And MR Applied to the SNIR Epidemic Model

We introduce the system for the SnIR epidemic model, which has n groups of
susceptible individuals and which is described by the next system of Eq. [2]:

d
Xi ðtÞ ¼ mðpi X0  Xi ðtÞÞ  li Xi ðtÞ
dt
d Xn
YðtÞ ¼ lk Xk ðtÞ  ðm þ g þ dÞ YðtÞ
dt k¼1
d
ZðtÞ ¼ gYðtÞ  ðm þ eÞ ZðtÞ (1)
dt

we define the rate of infection as:

li ¼ ai bZYðtÞ (2)

and we define pi as follow:

X
n
pi ¼ 1 (3)
i¼1
43 Generalizations in Mathematical Epidemiology 559

This is a system with (n þ 2) equations and each previous constant is defined like
it is shown:
m is the natural death rate.
g is the rate at which infectives are removed or become immune.
d is the disease-induced mortality rate for the infectives.
e is the disease-induced mortality rate for removed individuals.
ai is the susceptibility of susceptible individuals.
b is the infectious rate of infected individuals.
Z is the average number of contacts per individual.
Each function or group is defined as follow:
Xi(t) are the n groups of susceptible in the time equal t
Y(t) is the group of infectives in the time equal t.
Z(t) is the group of removed in the time equal t.

2.1 The Standard SIR Model

As a particular case we analyze the standard SIR model [7] which has just one
group of susceptible and is described in the next equation system:

d
X1 ðtÞ ¼ mðp1 X0  X1 ðtÞÞ  l1 X1 ðtÞ
dt
d
YðtÞ ¼ l1 X1 ðtÞ  ðm þ g þ dÞYðtÞ
dt
d
ZðtÞ ¼ gYðtÞ  ðm þ eÞZðtÞ (4)
dt

In the infection-free equilibrium there is no variation in time. So the derivates are

canceled and the solution for the previous system, it’s given by:

X1 ¼ p 1 X0 ; Y ¼ 0 (5)

After, we generate the Jacobian matrix for the equations system:

" #
m  a1 ZbY a1 ZX1 b
; (6)
a1 ZbY a1 ZX1 b  m  g  d

we substitute the solution (5) in the Jacobian:

" #
m a1 Zp1 X0 b
(7)
0 a1 Zp1 X0 b  m  g  d
560 D.C. Cano

Now, we find the eigenvalues for the previous matrix:

m; a1 Zp1 X0 b  m  g  d (8)

and its corresponding stability condition is:

a1 Zp1 X0 b  m  g  d < 0 (9)

this can be rewritten as:

a1 Zp1 X0 b
<1 (10)
mþgþd

using the next expression:

R0 < 1 (11)

Finally we find the basic reproduction number which it represents the condition
of equilibrium:

a1 Zp1 X0 b
R0 ¼ (12)
mþgþb

2.2 The S2IR Model

As another particular case we analyze the S2IR model where there are two groups of
susceptible and the equations for this system are:

d
X1 ðtÞ ¼ mðp1 X0  X1 ðtÞÞ  l1 X1 ðtÞ
dt
d
X2 ðtÞ ¼ mðp2 X0  X2 ðtÞÞ  l2 X2 ðtÞ
dt
d
YðtÞ ¼ l1 X1 ðtÞ þ l2 X2 ðtÞ  YðtÞm  gYðtÞ  YðtÞd
dt
d
ZðtÞ ¼ gYðtÞ  ðm þ eÞZðtÞ (13)
dt

Initially we find the infection-free equilibrium solution for the previous system:

X1 ¼ p 1 X0 ; X2 ¼ p 2 X0 ; Y ¼ 0 (14)
43 Generalizations in Mathematical Epidemiology 561

Equally we generate the Jacobian matrix for the equations system and substitut-
ing the infection-free equilibrium point in the Jacobian:
2 3
m 0 a1 Zp1 X0 b
6 7
4 0 m a2 Zp2 X0 b 5 (15)
0 0 a1 Zp1 X0 b þ a2 Zp2 X0 b  m  g  d

we find the eigenvalues for the previous Jacobian,

m; m; a1 Zp1 X0 b þ a2 Zp2 X0 b  m  g  d (16)

and the corresponding stability condition is:

a1 Zp1 X0 b þ a2 Zp2 X0 b  m  g  d < 0 (17)

this can be rewritten as:

a1 Zp1 X0 b þ a2 Zp2 X0 b
<1 (18)
mþgþd
also it can be written using (11) as:
a1 Zp1 X0 b þ a2 Zp2 X0 b
R0 ¼ (19)
mþgþd
this is the basic reproductive number for S2IR model.

2.3 The S3IR, The S4IR and S5IR Models

Here we show the S3IR, the S4IR and S5IR models where there are three, four and
five groups of susceptibles, respectively. With these models we do the same
process, and we only show the basic reproductive number.

a1 Zp1 X0 b þ a2 Zp2 X0 b þ a3 bZp3 X0

R0 ¼ (20)
mþgþd

this is the basic reproductive number for S3IR model.

a1 Zp1 X0 b þ a2 Zp2 X0 b þ a3 bZp3 X0 þ a4 bZp4 X0

R0 ¼ (21)
mþgþd

this is the basic reproductive number for S4IR model.

X0 bZða1 p1 þ a2 p2 þ a3 p3 þ a4 p4 þ a5 p5 Þ
R0 ¼ (22)
mþgþd

and this is the basic reproductive number for S5IR model.

562 D.C. Cano

2.4 The SnIR Model

Theorem. For the equations system given by (1). The infection-free equilibrium is
locally stable if R0 < 1, and is unstable if R0 > 1, where:

 
P
n
X0 bZ ai pi
i¼1
R0 ¼ (23)
mþgþd

We can probe the theorem looking the inequalities corresponding to stability

conditions for each system previously considered, we have listed in the Fig. 1.
Using mechanized induction we obtain the general expression for the stability
conditions for a system with (n + 2) equations.

n 
P
X0 bZ ai pi
i¼1
<1 (24)
mþgþd

We can represent into a schematic diagram, the deductive reasoning using

“MR”.
Here we have an idea for the MR, it finds the similar components in each item
and it has a viewer or a detector that find the sequential form for the dissimilar parts.
It is just an idea, we believe this system have to be improved by the scientific
community.

Fig. 1 Inductive mechanized reasoning for the SnIR Model

43 Generalizations in Mathematical Epidemiology 563

3 CA and MR Applied to the SNIMR Epidemic Model

The SnImR epidemic model is made by a group of equations which has n groups of
susceptible individuals and m groups of infected people, this system is illustrated in
the next equations:
!
d Xm
Xi ðtÞ ¼ mðpi X0  Xi ðtÞÞ  ai Z bj Yj ðtÞ Xi ðtÞ
dt i¼1
!
d Xn
Yj ðtÞ ¼ Zbj Yj ðtÞ ai Xi ðtÞ  ðm þ g þ dÞYj ðtÞ
dt i¼1
d
ZðtÞ ¼ gYj ðtÞ  ðm þ eÞZðtÞ (25)
dt

pi is defined in (3). This is a system with (n + m + 1) equations and each constant

were defined in the last model.

3.1 The SI2R Model

Like in all cases, we analyze a particular case with one group of susceptibles and
two groups of infectives. The following equations describe this case:
!
d X2
X1 ðtÞ ¼ mðp1 X0  X1 ðtÞÞ  a1 Z bj Yj ðtÞ X1 ðtÞ
dt j¼1
!
d X1
Y1 ðtÞ ¼ Zb1 Y1 ðtÞ bi Xi ðtÞ  ðm þ g þ dÞY1 ðtÞ
dt i¼1
!
d X1
Y2 ðtÞ ¼ Zb2 Y2 ðtÞ ai Xi ðtÞ  ðm þ g þ dÞY2 ðtÞ
dt i¼1
d
ZðtÞ ¼ gYj ðtÞ  ðm þ eÞZðtÞ (26)
dt

Solving the system for the infection-free equilibrium, we find:

mp1 X0
Y1 ¼ 0; Y2 ¼ 0; X1 ¼ ! (27)
P
2
m þ a1 Z bj Yj
j¼1

We generate a Jacobian as in the others cases. After, we substitute the infection-

free equilibrium point and we find the eigenvalues for the system,
564 D.C. Cano

m; a1 Zp1 X0 b1  m  g  d; a1 Zp1 X0 b2  m  g  d (28)

The stability conditions shall satisfy,

a1 Zp1 X0 b2  m  g  d < 0
a1 Zp1 X0 b1  m  g  d < 0 (29)

Moreover the inequalities in (29) can be written like the others cases as:

a1 Zp1 X0 b2
R0;2 ¼
mþgþd
a1 Zp1 X0 b1
R0;1 ¼ (30)
mþgþd

Here, we have the two basic reproductive numbers for SI2R model.

3.2 The S2I2R Model

We analyze a particular case where we have two groups of susceptible and two
groups of infective, in addition we solve the system for the infection-free equilib-
rium and we find:

mp1 X0 mp2 X0
Y1 ¼ 0; Y2 ¼ 0; X1 ¼ ! ; X2 ¼ ! (31)
P
2 P
2
m þ a1 Z bj Yj m þ a2 Z bj Yj
j¼1 j¼1

Also we generate a Jacobian; we substitute the infection-free equilibrium point

and find the eigenvalues for this system:

m; m; a1 Zp1 X0 b1 þ a2 Zp2 X0 b1  m  g  d; a1 Zp1 X0 b2

(32)
þa2 Zp2 X0 b2  m  g  d

In like manner that we obtained the last reproductive numbers, we obtain the two
basic reproductive numbers for S2I2R model:

a1 Zp1 X0 b2 þ a2 Zp2 X0 b2
R0;2 ¼
mþgþd
a1 Zp1 X0 b1 þ a2 Zp2 X0 b1
R0;1 ¼ (33)
mþgþd
43 Generalizations in Mathematical Epidemiology 565

3.3 The SnImR Model

Theorem. For the equations system given by (25). The infection-free equilibriums
are locally stable if the reproductive numbers of infection R0,j < 1, and is unstable if
R0,j > 1, with j from 1 to m, where:
n 
P
X0 Z ai pi bj
i¼1
R0;j ¼ (34)
mþgþd
To prove the theorem, we used mechanized induction starting from the particular
results previously obtained in (30) and (33). Finally, we find the general solution for
the basic reproductive numbers for the SnImR model according with (Fig. 2).

4 CA and MR Applied to the Staged Progressive

SIMR Epidemic Model

At this moment, we are concerned in the analysis of the staged progressive SImR
epidemic model. This has m groups of infected; in this case the infection is staged
and progressive, which is described by next system:

d
XðtÞ ¼ mðx0  XðtÞÞ  lXðtÞ
dt
d
Y1 ðtÞ ¼ lXðtÞ  ðm þ g1 þ d1 ÞY1 ðtÞ
dt
d
Yj ðtÞ ¼ Yj1 Yj1 ðtÞ  ðm þ gj þ dj ÞYj ðtÞ
dt
d
ZðtÞ ¼ gm Ym ðtÞ  ðm þ eÞZðtÞ (35)
dt

Fig. 2 Inductive mechanized reasoning for the SnImR Model

566 D.C. Cano

with the restriction for j is: 2  j  m. And where the rate of infection is:
!
X
m
l¼a bj Yj ðtÞ Z (36)
j¼1

This is a system with (m þ 2) equations and all the constants were described
before.

4.1 The Staged Progressive SI2R Model

Now, we analyze a particular case with one group of susceptible and two groups of
infectives. The following equations describe this case:

d
XðtÞ ¼ mðx0  XðtÞÞ  lXðtÞ
dt
d
Y1 ðtÞ ¼ lXðtÞ  ðm þ g1 þ d1 ÞY1 ðtÞ
dt
d
Y2 ðtÞ ¼ l1 Y1 ðtÞ  ðm þ g2 þ d2 ÞY2 ðtÞ
dt
d
ZðtÞ ¼ l2 Y2 ðtÞ  ðm þ eÞZðtÞ (37)
dt

Solving the system for the infection-free equilibrium, we find:

mx0
X¼ ; Y1 ¼ 0; Y2 ¼ 0 (38)
m þ aZb1 Y1 þ aZb2 Y2

We generate a Jacobian coming off the equations system and substituting the
infection-free equilibrium point:
 
 m aZx0 b1  aZx0 b2
 
 
 0 aZx0 b1  m  g1  d1 aZx0 b2  (39)
 
 0 g m  g  d2 
1 2

Now we obtain the characteristic polynomial:

ðl þ mÞðl2 þ 2lm þ lg2 þ ld2  aZx0 b1 l  aZx0 b1 m

 aZx0 b1 g2  aZx0 b1 d2 þ m2 þ mg2 þ md2 þ g1 l
þ g1 m þ g1 g2 þ g1 d2 þ d1 l þ d1 g2 þ d1 d2
 aZx0 b2 g1 (40)
43 Generalizations in Mathematical Epidemiology 567

From this polynomial, we obtain the basic reproduction number using the Routh-
Hurwitz theorem:
x0 aZðb1 d2 þ b1 m þ b1 g2 þ b2 g1 Þ
R0 ¼ (41)
ðm þ g2 þ d2 Þðm þ g1 þ d1 Þ

4.2 The Staged Progressive SI3R Model

We analyze a particular case with three groups of infectives. We solve this system
in the same way that we did in the Section 4.1. For this model the basic reproduction
number that we find is:
R0 ¼ ðx0 aZðb1 d2 m þ b1 d2 l3 þ b1 ml3 þ b1 g2 m þ b1 g2 g3
þb1 g2 d3 þ b1 m2 þ b1 d2 d3 þ g1 b2 m þ g1 b2 g3 þ g1 b2 d3
þg1 b3 g2 þ b1 md3 ÞÞ=ððm þ g3 þ d3 Þðm þ g2 þ d2 Þðm
þg1 þ d1 ÞÞ (42)

4.3 Staged Progressive SImR Model

Theorem. For the equations system given by (35). The infection-free equilibrium is
locally stable if the reproductive numbers of infection R0 < 1, and is unstable if R0
> 1, with j from 1 to m, where:
!
Pm Q
j1 Q
m
x0 aZ bj gk ðm þ gl þ dl Þ
j1 k¼1 l¼jþ1
R0 ¼ Q
m (43)
ðm þ gl þ dl Þ
l¼1

To prove the theorem, we used mechanized induction starting from the particu-
lar results previously obtained. To conclude, we find the general solution for the
basic reproductive numbers for the Staged Progressive SImR model according with
(Fig. 3).

Fig. 3 Inductive mechanized reasoning for the staged progressive SImR Model
568 D.C. Cano

5 Conclusions

We finally obtain three theorems which can help us to demonstrate that CASþMR
are important tools for solving problems in every situation that mathematics could
model. The theorems are useful to make strategies to fight against epidemic
diseases in the future biological dangers.
Due to use CAS, in our case “Maple 11”, we can proceed to solve the mathe-
matical problem and we can obtain results very fast that without them could take us
too much time.
The use of CAS+MR can help in teaching and learning the mathematics to
engineering, whose don’t have time and need to give quickly solutions. It can be
implemented in engineer programs.
Through the MR we already found the general forms for the infection free
equilibrium, we can see the importance of developing software it can do the
mechanized reasoning automatically.

Acknowledgments The authors would like to thank Prof. Dr. Eng. Andrés Sicard and Prof. Mario
Elkin Vélez for keeping company in this work in the Logic and Computation group, Prof. Eng.
Maurio Arroyave for supporting this work from the engineering physics program. They are also
very grateful for Prof. Dr. Eng. Félix Londoño for his help in the World Congress on Engineering
and Computer Sciences and the Student Organization and Ms. Angela Echeverri for the economi-
cal assistance in the trip to the congress. This research was support by the Logic and Computation
group at EAFIT University.

References

1. C.W. Brown, M.E. Kahoui, D. Novotni, A. Weber, Algorithmic methods for investigating
equilibria in epidemic modeling. J. Symb. Comput. 41, 1157–1173 (2006)
2. J.M. Hyman, J. Li, Differential susceptibility epidemic models. J. Math. Biol. 50, 626–644
(2005)
3. J.A.M. Taborda, Epidemic thresholds via computer algebra, MSV 2008, pp. 178–181, http://
dblp.uni-trier.de/rec/bibtex/conf/msv/Taborda08
4. D. Hincapié et al., Epidemic thresholds in SIR and SIIR models applying an algorithmic
method, Lectures Notes Bioinformatics, 5354, 119–130, 2008
5. S.B. Hsua, Y.H. Hsiehb, On the role of asymptomatic infection in transmission dyanamics of
infectious diseases, Bull. Math. Biol. 70, 134–155 (2008)
6. J.M. Hyman, J. Li, An intuitive formulation for the reproductive number for spread of diseases
in heterogeneous populations. Math. Biosci. 167, 65–86 (2000)
7. N.T.J. Bailey, The Mathematical Theory of Epidemics, 1957
Chapter 44
Review of Daily Physical Activity Monitoring
System Based on Single Triaxial Accelerometer
and Portable Data Measurement Unit

Mihee Lee, Jungchae Kim, Sun Ha Jee, and Sun Kook Yoo

Abstract Main objective of this pilot study was to present a method to convenient
monitoring of detailed ambulatory movements in daily life, by use of a portable
measurement device employing single tri-axial accelerometer. In addition, the
purpose of this review article is to provide researchers with a guide to understanding
some commonly-used accelerometers in physical activity assessment. Specially, we
implemented a small-size wearable data storing system in real time that we used
Micro SD-Memory card for convenient and long period habitual physical activity
monitoring during daily life. Activity recognition on these features was performed
using Fuzzy c means classification algorithm recognized standing, sitting, lying,
walking and running with 99.5% accuracy. This study was pilot test for our devel-
oped system’s feasibilities. Further application of the present technique may be
helpful in the health promotion of both young and elderly.

1 Introduction

Over the past two decades, a striking increase in the number of people with the
metabolic syndrome worldwide has taken place. This increase is associated with the
global epidemic of obesity and diabetes. With the elevated risk not only of diabetes
but also of cardiovascular disease from the metabolic syndrome, there is urgent
need for strategies to prevent the emerging global epidemic [1, 2]. Although the
metabolic syndrome appears to be more common in people who are genetically
susceptible, acquired underlying risk factors-being overweight or obese, physical
inactivity, and an atherogenic diet-commonly elicit clinical manifestations [3].

M. Lee (*)
Graduate Programs of Biomedical Engineering, Yonsei University, 262 Seongsanno,
Seodaemoon-Ku, Seoul 120-749, Korea
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 569

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_44, # Springer ScienceþBusiness Media B.V. 2010
570 M. Lee et al.

Current guidelines recommend practical, regular, and moderate regimens of

physical activity (eq, 30 min moderate-intensity exercise daily) [4]. Regular and
sustained physical activity will improve all risk factors of the metabolic syndrome.
Sedentary activities in leisure time should be replaced by more active behavior such
as brisk walking, jogging, swimming, biking, golfing, and team sports. Combina-
tion of weight loss and exercise to reduce the incidence of type 2 diabetes in patients
with glucose intolerance should not be dismissed [5].
A variety of methods exist to quantify levels of habitual physical activity
during daily life, including objective measures such as heart rate, one- and
three-dimensional accelerometer, and pedometer, as well as subjective recall
questionnaires like the International Physical Activity Questionnaire and physical
activity logbooks [7, 8]. Yet all possess some important limitations. Heart rate
monitors have been widely used to quantify physiological stress, but their efficacy
at low intensities has been questioned due to the potential interference of environ-
mental conditions and emotional stress [9]. A wide range of self-report activity
questionnaires exist that are well suited to large surveillance studies but are limited
due to their reliance on subjective recall. Pedometers are an inexpensive form of
body motion sensor, yet many fail to measure slow walking speeds or upper body
movements, and most are unable to log data to determine changes in exercise
intensity [10]. The most common accelerometers used in human activity research
measure accelerations either in a vertical plane (uni-axial), or in three planes (tri-
axial), with excellent data-logging abilities [1, 10].
Main objective of this study was to present a method to convenient monitoring of
detailed ambulatory movements in daily life, by use of a portable measurement device
employing single tri-axial accelerometer. In addition, the purpose of this review article
is to provide researchers with a guide to understanding some commonly-used accel-
erometers in physical activity assessment. The literature related to physical activity
measurement is reviewed in this paper. For organizational purposes, the literature is
presented under the following topics: (a) Measurement of Physical Activity, (b)
Behavioral Observation, (c) Pedometers, and (d) Accelerometers.

1.1 Measurement of Physical Activity

Direct measurements of physical activity include behavioral observation and

motion sensors (pedometers and accelerometers). An estimate of energy expendi-
ture can be extrapolated from these direct measures of physical activity based on
previously validated equations. The ability to examine health in relation to physical
activity requires an accurate, precise and reproducible measurement of physical
activity. Physical activity measurement tools should provide an objective measure
of physical activity, be utilized under normal daily living conditions, have the
capacity to record data over a prolonged period of time, and produce minimal
discomfort to subjects [11]. Consideration should be given to the number of
subjects to be monitored, and size and cost of the measurement tool, and the activity
being monitored.
44 Review of Daily Physical Activity Monitoring System 571

1.2 Behavioral Observation

Behavioral observation is a seldom utilized, time intensive technique that has

primarily be applied when monitoring children and adolescents. This technique
requires skilled observers who monitor individuals for a set period of time in an
effort to code their behavior to describe overall physical activity patterns and
estimate energy expenditure based on published MET values [12].

1.3 Pedometers

The concept of a pedometer was first introduced by Leonardo Da Vinci 500 years
ago. Pedometers are inexpensive tools that measure the number of steps taken by
responding to the vertical acceleration of the trunk. Mechanical pedometers have a
level arm is triggered when a step is taken and a ratchet within the pedometer that
rotates to record the movement [13]. Electronic pedometers have a spring sus-
pended pendulum arm that moves up and down thereby opening and closing an
electrical circuit [7]. Output from the pedometer consists of number of steps taken,
distance walked (with stride information), and estimated number of calories
expended (with body weight considered) [13]. However, it is assumed that a person
expends a constant amount of energy per step regardless of the speed (e.g. Yamax
pedometer, 0.55 kcal · kg1 · step1) [7].
Vertical acceleration of the hip during walking varies from 0.5 to 8 m · s2,
however some pedometers do not register movement below 2.5 m · s2 or distin-
guish vertical accelerations above a certain threshold, and therefore are inaccurate
at slow or very fast walking speeds and during running [7, 14].
Limitations of the pedometer include insensitivity to: (1) static or sedentary acti-
vity, (2) isometric activity, (3) movement occurring with the arms, and (4) slow or fast
walking velocities [7, 13, 14]. In addition, they lack internal clocks and data storage
capability which makes an analysis of overall physical activity patterns difficult [7].
Pedometers are primarily utilized in health promotion and maintenance and may
provide valuable feedback to a participant concerning their level of activity. Advan-
tages of utilizing a pedometer include low cost, size, and compliance of subjects.

1.4 Accelerometers

Accelerometer is based on the assumption that limb movement and body accelera-
tion are theoretically proportional to the muscular forces responsible for the accel-
erations, and therefore energy expenditure may be estimated by quantifying these
accelerations [11, 13]. Interest in monitoring the acceleration of the body started in
the 1950s when Broubha and Smith (1958) suggested an association between the
integral of vertical acceleration versus time and energy expenditure [15].
572 M. Lee et al.

Acceleration is the change in velocity over time, and the degree of acceleration
detected provides an index of movement intensity. Accelerations increase from the
head to the feet and are generally the greatest in the vertical direction. Running
produces the greatest vertical direction accelerations (8.1–12 g) at the ankle and up
to 5 g at the back [16]. Bouten et al. (1996) recommends that accelerometers should
be able to measure accelerations up to 12 g for measurements occurring in daily
living activities and exercise. Processing and filtering is required to identify and
include accelerations that are outside the normal acceleration range of human
movement.
Low pass filters remove high frequency signals that may occur as a result
of external vibrations [17]. Factors which can influence accelerometer output
include the acceleration due to body movement and external vibrations (loose
straps, transportation) [18]. Most accelerometers are oriented to monitor the vertical
plane such that linear acceleration with movement is not recorded [19].

1.4.1 Types of Accelerometer

Accelerometers can be classified based on the process by which acceleration is

measured and recorded (piezoelectric versus piezoresistive) or the number of planes
of movement it measures. Piezoresistive accelerometers utilize a spring element with
strain sensitive gages connected via a Wheatstone bridge. The deflection of the spring
element during acceleration causes a deformation of the gages, producing an electri-
cal output in response to the change in resistance [13]. Piezoelectric accelerometers
consist of a piezoelectric element and a seismic mass. When movement occurs, the
sensor undergoes acceleration which causes deformation of the piezoelectric element
(e.g. bending). Conformational changes in the element produce a voltage signal that
is proportional to stress, or acceleration of the limb. When the body part to which the
accelerometer is attached accelerates a charge is generated that is proportional to the
force exerted by the subject. An acceleration-deceleration curve wave is created and
the area under the wave is summed to yield the final count value [13, 19]. Ceramic
transducers are more reliable than spring mechanisms, sensitive at very low frequen-
cies, and require less power; however, aging of the ceramic may result in a loss of
piezoelectric capabilities causing a decreased sensitivity to movement [20].
Accelerometers can be uni-axial (one plane of movement), tri-axial (three planes of
movement) or “omni-directional”, monitoring acceleration in every plane other than
that perpendicular to the point of reference. The uni-axial accelerometer is most
sensitive to bending in one direction and is typically utilized to monitor acceleration
in the vertical plane. An omni-directional accelerometer is most sensitive in one
direction (vertical), but has the ability to quantify deformations in other planes (hori-
zontal and lateral). Tri-axial accelerometers quantify acceleration in the vertical, ante-
rior-posterior and medial-lateral plane and provide a measure of counts in each plane as
well as an overall vector magnitude [13, 19]. There appears little advantage to using the
triaxial accelerometer versus a single axis when the monitored activity is primarily
walking [21]. However, activity energy expenditure during sedentary activities,
44 Review of Daily Physical Activity Monitoring System 573

as determined by a tri-axial accelerometer, correlates better with measured energy

expenditure when compared to a uni-axial accelerometer [22].

1.4.2 Placement of Monitoring Sensors

Accelerometers are worn while the participant performs different activities and the
counts from the monitor are regressed against a criterion measure of energy
expenditure (e.g. whole room calorimetry, indirect calorimetry, doubly labeled
water, or heart rate). Output from the monitors can be utilized to assess frequency,
duration and intensity of physical activity. Energy expenditure is predicted by
converting monitor counts into a unit of energy expenditure (MET, kcal · min1,
or kcal · kg · min1).
The characteristics of the population being studied and/or the specific activity may
affect the accuracy of the motion sensors [23]. The output of the accelerometer will
depend on the placement site (e.g. hip, wrist, ankle), the orientation relative to the
subject, the posture of the subject, and the activity being performed. Accelerometer
output is influenced by the placement of the monitors as body parts are differentially
active depending upon the activity. Placing the accelerometer on the limbs while
monitoring sedentary and daily living activities may produce more accurate results
than hip placement as the limbs are likely to reflect activities which occur in these
circumstances. The primary acceleration which occurs during walking or running is
in the vertical direction, therefore placing the accelerometer at a site which reflects
acceleration at the center of the body mass (e.g. hip) is likely to be more accurate than
the limbs. The velocity of movement can also have an effect on the relationship
between accelerometer counts and energy expenditure.

1.4.3 Limitations for Physical Activity Monitoring

The disadvantages of utilizing the accelerometer to estimate energy expenditure

include subject compliance, altered activity patterns, and a cost of approximately
$400 per monitor [13, 24]. One limitation of use is the inability to account for
increased energy cost of walking uphill or cost of carrying a load. Accelerometers
do not respond to activities where there is minimal movement at the body’s center
of gravity (e.g. rowing and cycling) and they are not responsive to isometric work as
even thought the limbs are not moving, energy is being expended [25, 26].
However, Westerterp (1999) suggests that the effect of static exercise on the
total level of physical activity is negligible. Furthermore, estimates of energy
expenditure are derived from equations that were developed within specific popula-
tions performing primarily walking activities in the laboratory. Consequently,
energy estimation equations may not be applicable to the daily living activities
performed within the general population. The primary advantages of utilizing an
accelerometer include the ability to respond to both frequency and intensity of
574 M. Lee et al.

movement, its relatively low cost, ability to store large amounts of data over
extended periods of time, small size and minimal participant discomfort [27].
These findings indicate that differences between uniaxial, triaxial, and omnidi-
rectional accelerometry devices are minimal and all three provide reasonable
measures of physical activity during treadmill walking. Furthermore, the inclusion
of stride length in accelerometry based energy expenditure prediction equations
significantly improved energy expenditure estimates.
The purpose of this study was to present a method to convenient monitoring of
detailed ambulatory movements in daily life, by use of a portable measurement
device employing single tri-axial accelerometer. Specially, we implemented a
small-size wearable data store system in real time that we used Micro SD-Memory
card (Secure Digital memory card) for convenient and long period habitual physical
activity monitoring during daily life. In this work, the performance of activity
recognition algorithms under conditions akin to those found in real-world settings
is assessed. Activity recognition results are based on acceleration data collected
from single tri-axial accelerometer placed on subjects’ waist under semi-naturalistic
conditions for pilot test.

2 Material and Method

2.1 Portable Data Measurement Unit

Our long-term aim in developing an accelerometry monitoring system was to

develop a practical system that could be used to monitor and assess physical
activities in free-living subjects. Therefore, we developed a wearable device con-
sisted of Micro SD-Memory card connector (AUTF-08WP01, AUSTONE Elec-
tronics, Inc., USA) with mini USB socket (5P, SMT type, SHIH HAN CO., LTD.,
Russia). Measured acceleration signal was stored on micro SD-Memory card or
transmitted wirelessly using Zigbee-compatible 2.4G bandwidth for wireless com-
munication, and CC2420 (Chipcon Co. Ltd., Norway) with a simple interface
circuit around the chip. In addition, a ceramic chip antenna TI-AN048 (SMD
type, Texas Instruments Co. USA) was applied for stable wireless transmission.
For this we used ADXL330 (Analog Devices, Inc., USA), an acceleration sensor
that is composed of a single chip and can detect tri-axis acceleration information,
and measured acceleration information of axis X, Y and Z according to the subject’s
posture and activity. Using the implemented system, we measured change in
acceleration signal according to the change of activity pattern. Li-Ionic batteries,
micro processor units and micro SD memory card, as shown in Fig. 1. This equip-
ment was small (60  40  20 [mm]) and light enough to carry without and
restriction. Sampling frequency was 100 Hz. Data was downloaded via USB, and
processed offline by a PC. The equipment was designed to be attached on the waist
(see Fig. 1).
44 Review of Daily Physical Activity Monitoring System 575

One Chip Mini USB

Filter
Tri-axis MCU Interface
&
Accelerometer (Artmega 128)
Amp
(X,Y,Z) Zigbee
(CC 2420)

Li Ionic Vcc(+3.3 V)
AU6337
Battery (SD card
controller)

Charge
Controller Micro SD
Memory card

A / C Adapter (or Portable Device

USB)

Subject PC

Fig. 1 Architecture of the wearable measurement device

Table 1 Participant characteristics

Case Gender (M/F) Age (year) Height (cm) Weight (kg) BMI (kg w m2)
A M 26 182 76 22.9
B M 28 179 85 26.5
C M 28 177 69 22.0
D F 24 167 55 19.7
E F 33 160 55 21.5

2.2 Physical Activity Data Collection

The data for these experiments were gathered in an unsupervised pilot study in which
healthy young (age 24–33) subjects performed a variety of activities in the three times
on outdoor conditions. Characteristics of the participants are presented in Table 1.
We put the acceleration measuring sensor system on the left waist of the sub-
jects, and measured change in acceleration signal according to change in ambula-
tory movement and physical activities. In the experiment performed in this
research, the change of acceleration was measured while the subject was repeating
postures such as standing, sitting, lying, walking and running. A representatives set
of routine data is shown in Fig. 2.
576 M. Lee et al.

Fig. 2 Representative data from the daily routine for each of the three axes of the tri-axial device

While the movements and postures contained within the routine are by no means a
complete set of all possible activities that a given person might perform, they do form
a basic set of simple activities which form an underlying structure to a person’s daily
life, and are likely to provide a great deal of information in terms of the person’s
balance, gait and activity levels if they can be accurately identified.

2.3 Feature Extraction

Features were computed on 512 sample windows of acceleration data with 256
samples overlapping between consecutive windows. At a sampling frequency of
100 Hz, each window represents 5.2 s. Maximum acceleration, mean and standard
deviation of acceleration channels were computed over sliding windows with 50%
overlap has demonstrated success in past works. The 512 sample window size
enabled fast computation of FFTs used for some of the features. The DC feature for
normalization is the mean acceleration value of the signal over the window. Use of
mean of maximum acceleration features has been shown to result in accurate
recognition of certain postures and activities.

3 Results

This section describes the experiments and experimental results of the human
posture recognition system. In the pilot test, a subject continuous posture change
including standing, sitting, lying, walking and running. In the experiment, each
posture was recognized third.
44 Review of Daily Physical Activity Monitoring System 577

Table 2 Clustering results Parameters and Jaccard Purity Efficiency

of different posture in a real posture score
continuous motion
Standing 0.99 0.99 1
Sitting 1 1 1
Lying 1 1 1
Walking 0.99 1 0.99
Running 1 1 1
Average 0.98 0.99 0.99

Fig. 3 Activity recognition result using Fuzzy c means classification algorithm

Mean and standard deviation of acceleration and correlation features were

extracted from acceleration data. Activity recognition on these features was per-
formed using Fuzzy c means classification algorithm recognized standing, sitting,
lying, walking and running with 99.5% accuracy as shown in Table 2 and Fig. 3.

4 Discussion and Conclusion

This paper proposes an ambulatory movement’s recognition system in daily life. A

portable acceleration sensor module has been designed and implemented to measure
human body motion. A small portable device utilizing single tri-axis accelerometer
578 M. Lee et al.

was developed, which detects features of ambulatory movements including vertical

position shifts. The classification method based on Fuzzy c means classification
algorithm, recognition accuracy of over 99% on a five activities (standing, sitting,
lying, walking and running). These results are competitive with prior activity recog-
nition results that only used laboratory data. However, several limitations are also
observed for the system. Firstly, collected data was from younger (age 24–33)
subjects. Secondly, single accelerometer of placed on body waist typically do not
measure ascending and descending stairs walking.
Accelerometers are preferable to detect frequency and intensity of vibrational
human motion [28]. Many studies have demonstrated the usefulness of accelerometer
for the evaluation of physical activity, mostly focusing on the detection of level
walking or active/rest discrimination [29–31].
This study was pilot test for our developed system’s feasibilities. Further appli-
cation of the present technique may be helpful in the health promotion of both
young and elderly, and in the management of obese, diabetic, hyperlipidemic and
cardiac patients. Efforts are being directed to make the device smaller and allow
data collection for longer time periods. Implementation of real-time processing
firmware and encapsulation of the hardware are our future studies.

Acknowledgments This study was supported by a grant of the Seoul R&BD Program, Republic
of Korea (10526) and the Ministry of Knowledge Economy (MKE) and Korea Industrial
Technology Foundation (KOTEF) through the Human Resource Training Project for Strategic
Technology.

References

1. P. Zimmet, K.G. Alberti, J. Shaw, Global and societal implications of the diabetes epidemic.
Nature 414, 782–787 (2001)
2. S.M. Grundy, B. Hansen, S.C. Smith Jr, J.I. Cleeman, R.A. Kahn, Clinical management of
metabolic syndrome: report of the American Heart Association/National Heart, Lung, and
Blood Institute/American Diabetes Association conference on scientific issues related to
management. Circulation 109, 551–556 (2004)
3. R.H. Eckel, S.M. Grundy, P.Z. Zimmet, The metabolic syndrome. Lancet 365, 1415–28
4. P.D. Thompson, D. Buchner, I.L. Pina, et al., Exercise and physical activity in the prevention
and treatment of atherosclerotic cardiovascular disease: a statement from the Council on
Clinical Cardiology (Subcommittee on Exercise, Rehabilitation, and Prevention) and the
Council on Nutrition, Physical Activity, and Metabolism (Subcommittee on Physical
Activity). Circulation 107, 3109–3116 (2003)
5. J. Tuomilehto, J. Lindstrom, J.G. Eriksson, et al., Prevention of type 2 diabetes mellitus by
changes in lifestyle among subjects with impaired glucose tolerance. N. Engl. J. Med. 344,
1343–1350 (2001)
6. Y. Ohtaki, M. Susumago, A. Suzuki, et al., Automatic classification of ambulatory movements
and evaluation of energy consumptions utilizing accelerometers and a barometer. Microsyst.
Technol. 11, 1034–1040 (2005)
7. D.R. Bassett Jr. Validity and reliability issues in objective monitoring of physical activity.
Res. Q Exerc. Sport 71, S30–S36 (2000)
44 Review of Daily Physical Activity Monitoring System 579

8. C.L. Craig, A.L. Marshall, M. Sjostrom, A.E. Bauman, et al., International physical activity
questionnaire: 12-country reliability and validity. Med. Sci. Sports Exerc. 35, 1381–1395
(2003)
9. K.M. Allor, J.M. Pivarnik, Stability and convergent validity of three physical activity assess-
ments. Med. Sci. Sports Exerc. 33, 671–676 (2001)
10. M.J. LaMonte, B.E. Ainsworth, C. Tudor-Locke, Assessment of physical activity and energy
expenditure, in Obesity: Etiology, Assessment, Treatment and Prevention, ed. by R.E. Andersen
(Human Kinetics, Champaign, IL, 2003), pp. 111–117
11. C.V.C. Bouten, W.P.H.G. Verboeket-Van Venne, et al., Daily physical activity assessment:
comparison between movement registration and doubly labeled water. J. Appl. Physiol. 81,
1019–1026 (1996)
12. U.S. Department of Health & Human Services, Physical Activity and Health: A Report of the
Surgeon General (U.S. Department of Health and Human Services, Centers for Disease
Control and Prevention, National Center for Chronic Disease Prevention and Health Promo-
tion, The President’s Council on Physical Fitness and Sports, Atlanta, GA, 1996)
13. P.S. Freedson, K. Miller Objective monitoring of physical activity using motion sensors and
heart rate. Res. Quart. Exerc. Sport 71, 2129 (2000)
14. H.J. Montoye, H.C.G. Kemper, W.H.M. Saris, R.A. Washburn, Measuring Physical Activity
and Energy Expenditure (Human Kinetics, Campaign, IL, 1996)
15. R.K. Dishman, R.A. Washburn, D.A. Schoeller, Measurement of physical activity. Quest. 53,
295–309 (2001)
16. A. Bhattacharya, E.P. McCutcheon, E. Shvartz, J.E. Greenleaf, Body acceleration distribu-
tion and O2 uptake in humans during running and jumping. J. Appl. Physiol. 49, 881–887
(1980)
17. G.J. Welk, Physical Activity Assessments for Health-Related Research (Human Kinetics,
Champaign, IL, 2002)
18. C.V.C. Bouten, A.A.H.J. Sauren, M. Verduin, J.D. Janssen, Effects of placement and orienta-
tion of body-fixed accelerometers on the assessment of energy expenditure during walking.
Med. Biol. Eng. Comput. 35, 50–56 (1997)
19. K.Y. Chen, D.R. Bassett, The technology of accelerometry-based activity monitors: current
and future. Med. Sci. Sport Exerc. 37(11), S490–S500 (2005)
20. E.L. Melanson, P.S. Freedson Physical activity assessment: a review of methods. Crit. Rev.
Food Sci. Nutr. 36, 385–396 (1996)
21. T.G. Ayen, H.J. Montoye Estimation of energy expenditure with a simulated threedimensional
accelerometer. J. Ambul. Monit. 1(4), 293–301 (1988)
22. K.R. Westerterp, Physical activity assessment with accelerometers. Int. J. Obes. 23(Suppl 3),
S45–S49 (1999)
23. B.G. Steele, B. Belza, K. Cain, C. Warms, J. Coopersmith, J. Howard, Bodies in motion:
monitoring daily activity and exercise with motion sensors in people with chronic pulmonary
disease. J. Rehabil. Res. Dev. 40(Suppl 2) 45–58 (2003)
24. M.J. Lamonte, B.E. Ainsworth, Quantifying energy expenditure and physical activity in the
context of dose response. Med. Sci. Sports Exerc. 33, S370–S378 (2001)
25. R.K. Dishman, R.A. Washburn, D.A. Schoeller, Measurement of physical activity. Quest. 53,
295–309 (2001)
26. K.R. Westerterp, Physical activity assessment with accelerometers. Int. J. Obes. 23(Suppl 3),
S45–S49 (1999)
27. M.J. Mathie, A.C.F. Coster, N.H. Lovell, B.G. Celler, Accelerometry: providing an integrated,
practical method for long-term, ambulatory monitoring of human movement. Physiol. Meas.
25, R1–R20 (2004)
28. C.V.C. Bouten, K.T.M. Koekkoek, M. Verduin, R. Kodde, J.D. Janssen, A triaxial acceler-
ometer and portable data processing unit for the assessment of daily physical activity. IEEE
Trans. Biomed. Eng. 44(3):136–147 (1997)
580 M. Lee et al.

29. A.K. Nakahara, E.E. Sabelman, D.L. Jaffe, Development of a second generation wearable
accelerometric motion analysis system. Proceedings of the first joint EMBS/BMES confer-
ence, 1999, p. 630
30. K. Aminian, P. Robert, E.E. Buchser, B. Rutschmann, D. Hayoz, M. Depairon, Physical
activity monitoring based on accelerometry. Med. Biol. Eng. Comput. 37, 304–308 (1999)
31. M.J. Mathie, N.H. Lovell, C.F. Coster, B.G. Celler, Determining activity using a triaxial
accelerometer, in Proceedings of the Second Joint EMBS/BMES Conference, 2002,
pp. 2481–2482
Chapter 45
A Study of the Protein Folding Problem
by a Simulation Model

Omar Gaci

Abstract In this paper, we propose a simulation model to study the protein folding
problem. We describe the main properties of proteins and describe the protein folding
problem according to the existing approaches. Then, we propose to simulate the
folding process when a protein is represented by an amino acid interaction network.
This is a graph whose vertices are the proteins amino acids and whose edges are
the interactions between them. We propose a genetic algorithm of reconstructing
the graph of interactions between secondary structure elements which describe the
structural motifs. The performance of our algorithms is validated experimentally.

1 Introduction

Proteins are biological macromolecules participating in the large majority of

processes which govern organisms. The roles played by proteins are varied and
complex. Certain proteins, called enzymes, act as catalysts and increase several
orders of magnitude, with a remarkable specificity, the speed of multiple chemical
reactions essential to the organism survival. Proteins are also used for storage and
transport of small molecules or ions, control the passage of molecules through the
cell membranes, etc. Hormones, which transmit information and allow the regula-
tion of complex cellular processes, are also proteins.
Genome sequencing projects generate an ever increasing number of protein
sequences. For example, the Human Genome Project has identified over 30,000
genes which may encode about 100,000 proteins. One of the first tasks when
annotating a new genome is to assign functions to the proteins produced by the
genes. To fully understand the biological functions of proteins, the knowledge of
their structure is essential.

O. Gaci
Le Havre University, 25 rue Phillipe Lebon, 76600 Le Havre, France
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 581

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_45, # Springer ScienceþBusiness Media B.V. 2010
582 O. Gaci

In their natural environment, proteins adopt a native compact three dimensional

form. This process is called folding and is not fully understood. The process is a result
of interactions between the protein’s amino acids which form chemical bonds.
In this study, we propose to study the protein folding problem. We describe this
biological process through the historical approaches to solve this problem. Then,
we treat proteins as networks of interacting amino acid pairs [1]. In particular, we
consider the subgraph induced by the set of amino acids participating in the
secondary structure also called Secondary Structure Elements (SSE).We call this
graph SSE interaction network (SSE-IN). We begin by recapitulating relative works
about this kind of study model. Then, we present a genetic algorithm able to
reconstruct the graph whose vertices represent the SSE and edges represent spatial
interactions between them. In other words, this graph is another way to describe
the motifs involved in the protein secondary structures.

2 The Protein Folding Problem

Several tens of thousands of protein sequences are encoded in the human genome.
A protein is comparable to an amino acid chain which folds to adopt its tertiary
structure. Thus, this 3D structure enables a protein to achieve its biological func-
tion. In vivo, each protein must quickly find its native structure, functional, among
innumerable alternative conformations.
The protein 3D structure prediction is one of the most important problems of
bioinformatics and remains however still irresolute in the majority of cases. The
problem is summarized by the following question: being given a protein defined by
its sequence of amino acids, which is its native structure? In other words, we want
to determine the structure whose amino acids are correctly organized in three
dimensions in order to this protein can achieve correctly its biological function.
Unfortunately, the exact answer is not always possible that is why the research-
ers have developed study models to provide a feasible solution for any unknown
sequences. However, models to fold proteins bring back to NP-Hard optimization
problems [2]. Those kinds of models consider a conformational space where the
modeled protein tries to reach its minimum energy level which corresponds to its
native structure.
Therefore, any algorithm of resolution seems improbable and ineffective; the
fact is that in the absolute no study model is yet able to entirely define the general
principles of the protein folding.

2.1 The Levinthal Paradox

The first observation of spontaneous and reversible folding in vitro was carried
out by Anfinsen [3]. He deduced that the native structure of a protein corresponds
45 A Study of the Protein Folding Problem by a Simulation Model 583

to a conformation with a minimal free energy, at least under suitable environ-

mental conditions. But if the protein folding is indeed under thermodynamic
control, a judicious question is to know how a protein can find, in a reasonable
time, its structure of lower energy among an astronomical number of possible
conformations.
As example, a protein of 100 residues can adopt 2100 (1030) distinct conforma-
tions when we suppose that only two possibilities are accessible to each residue. If
the passage from a conformation to another is carried out in 1013 s (which
corresponds to time necessary for a rotation around a connection), this protein
would need at least 1017 s, i.e. approximately three billion years, “to test” all
possible conformations. The proteins however manage to find their native structures
in a lapse of time which is about the millisecond at the second. The apparent
incompatibility between these facts, raised initially by Levinthal [4], was quickly
set up in paradox and made run enormously ink since.
Levinthal gives the solution of its paradox: proteins do not explore the integrality
of their conformational space, and their folding needs to be “guided”, for example,
via the fast formation of certain interactions which would be determining for the
continuation of the process.

2.2 Motivations

To be able to understand how a protein accomplishes its biological function, and to

be able to act on the cellular processes in which the protein intervenes, it is essential
to know its structure. Many protein native structures were determined experimen-
tally – primarily by crystallography with X-rays or by Nuclear Magnetic Resonance
(NMR) – and indexed in a database accessible to all, Protein Data Bank (PDB) [5].
However, the application of these experimental techniques consumes a con-
siderable time [6, 7]. Indeed, the number of protein sequences known [8] is much
more important than the number of solved structures [5], this gap continues to grow
quickly.
The design of methods making it possible to predict the protein structure from
its sequence is a problem whose stakes are major, and which fascine many of
scientists for several decades. Various tracks were followed with an aim of solving
this problem, elementary in theory but extremely complex in practice.

3 Approaches to Study the Protein Folding Problem

The existing models for the protein folding problem study depend, amongst other
things, on the way that the native structure is supposed be reached. Either, a protein
folds following a preferential folding path [9], or a protein folds by searching the
native state among an energetic landscape organized as a funnel.
584 O. Gaci

The first hypothesis implies the existence of preferential paths for the folding
process. In the simplest case, the folding mechanism is comparable to a linear
reaction. Thus, when the steps are enough specifics, only a local region of the
conformational space will be explored. This concept is nowadays obsolete since we
know the existence of parallel folding paths.
The second hypothesis defines the folding by the following way (Dill, 1997):
a parallel flow process of an ensemble of chain molecules; folding is seen as more like
trickle of water down mountainsides of complex shapes, and less like flow through a single
gallery.

In other words, the folding can be described as a set of transitions between

structures whose energies become weaker. It allows guiding the protein by a funnel
effect toward the conformational state whose energy level is the minimum that is
the native conformation. Then, the polypeptide chain explores only a fraction of the
accessible states.
This last hypothesis is the one accepted in this chapter, the protein folding is a
process by which a large number of conformations are accessible and which leads a
sequence into its native structure with the lowest energy level (see Fig. 1).

Fig. 1 Evolution in the description of folding paths in an energy landscape. Top, the protein
folding according to the Anfinsen theory: a protein can adopt a large number of conformations.
Bottom-left, a protein folds by following only one specific path in the conformational space.
Bottom-right, from a denatured conformation, a protein searches its native structure whose energy
level is minimum in a minimum time
45 A Study of the Protein Folding Problem by a Simulation Model 585

3.1 Latest Approach

Many systems, both natural and artificial, can be represented by networks, that is,
by sites or vertices bound by links. The study of these networks is interdisciplinary
because they appear in scientific fields like physics, biology, computer science or
information technology.
These studies are lead with the aim to explain how elements interact with each
other inside the network and what the general laws which govern the observed
network properties are.
From physics and computer science to biology and social sciences, researchers
have found that a broad variety of systems can be represented as networks, and that
there is much to be learned by studying these networks. Indeed, the studies of the
Web [10], of social networks [11] or of metabolic networks [12] contribute to put in
light common non-trivial properties of these networks which have a priori nothing
in common. The ambition is to understand how the large networks are structured,
how they evolve and what are the phenomena acting on their constitution and
formation.
In [13], the authors propose to consider a protein as an interaction network
whose vertices represent the amino acids and an edge describes a specific type of
interaction (which is not the same according to the object of study). Thus, a protein,
molecule composed of atoms becomes a set constituted by individuals (the amino
acids), by interactions (to be defined according to the study) which evolves in a
particular environment (describing the experimental conditions).
The vocabulary evolves but the aim remains the same, we want to better
understand the protein folding process by the way of the modeling. The interaction
network of a protein is initially the one built from the primary structure. The goal is
to predict the graph of the tertiary structure through a discrete simulation process.

3.2 The Amino Acid Interaction Network

The 3D structure of a protein is represented by the coordinates of its atoms. This

information is available in Protein Data Bank (PDB), which regroups all experi-
mentally solved protein structures. Using the coordinates of two atoms, one can
compute the distance between them. We define the distance between two amino
acids as the distance between their Ca atoms. Considering the Ca atom as a “center”
of the amino acid is an approximation, but it works well enough for our purposes.
Let us denote by N the number of amino acids in the protein. A contact map matrix
is a N x N 0-1 matrix, whose element (i, j) is one if there is a contact between amino
acids i and j and zero otherwise. It provides useful information about the protein.
For example, the secondary structure elements can be identified using this matrix.
Indeed, a-helices spread along the main diagonal, while b-sheets appear as bands
parallel or perpendicular to the main diagonal [14]. There are different ways to
586 O. Gaci

define the contact between two amino acids. Our notion is based on spatial
proximity, so that the contact map can consider non-covalent interactions. We
say that two amino acids are in contact if and only if the distance between them
is below a given threshold. A commonly used threshold is 7 Å´ [1] and this is the
value we use.
Consider a graph with N vertices (each vertex corresponds to an amino acid) and
the contact map matrix as incidence matrix. It is called contact map graph. The
contact map graph is an abstract description of the protein structure taking into
account only the interactions between the amino acids.
First, we consider the graph induced by the entire set of amino acids participat-
ing in folded proteins. We call this graph the three dimensional structure elements
interaction network (3DSE-IN), see Fig. 2.
As well, we consider the subgraph induced by the set of amino acids participat-
ing in SSE. We call this graph SSE interaction network (SSE-IN) (see Fig. 2).
In [15] the authors rely on amino acid interaction networks (more precisely they
use SSE-IN) to study some of their properties, in particular concerning the role
played by certain nodes or comparing the graph to general interaction networks
models. Thus, thanks to this point of view the Protein Folding Problem can be
tackle by the graph theory.
To manipulate a SSE-IN or a 3DSE-IN, we need a PDB file which is transformed
by a parser we have developed. This parser generates a new file which is read by the
GraphStream library [16] to display the SSE-IN in two or three dimensions.

4 Folding a Protein in a Topological Space by Bio-Inspired

Methods

In this section, we treat proteins as amino acid interaction networks (see Fig. 2). We
describe a bio-inspired method we use to fold amino acid interaction networks. In
particular, we want to fold a SSE-IN to predict the motifs which describe the
secondary structure.

4.1 Genetic Algorithms

The concept of genetic algorithms has been proposed by John Holland [17] to
describe adaptive systems according to biological process.
The genetic algorithms are inspired from the concept of natural selection pro-
posed by Charles Darwin. The vocabulary employed here is the one relative to the
evolution theory and the genetic. We speak about individuals (potential solutions),
populations, genes (which are the variables), chromosomes, parents, descendants,
reproductions, etc.
45 A Study of the Protein Folding Problem by a Simulation Model 587

Fig. 2 Protein 1DTP SSE-IN (top) and the 1DTP 3DSE-IN (bottom). From a pdb file a parser we
have developed produces a new file which corresponds to the SSE-IN graph displayed by the
GraphStream library [16]

At the beginning, we conserve a population among which it exists a solution

which is not yet optimal. Then, the genetic algorithm make evolves this population
by an iterative process. Certain individuals reproduce themselves, others mute or
disappear and only the well adapted individuals are supposed to survive. The
588 O. Gaci

genetic heritage between generations must help to produce individuals which are
better and better adapted to correspond to the optimal solution.

4.2 Motif Prediction

In previous works [18], we have studied the protein SSE-IN. We have identified
notably some of their properties like the degree distribution or also the way in
which the amino acids interact. These works have allowed us to determine criteria
discriminating the different structural families. We have established a para-
llel between structural families and topological metrics describing the protein
SSE-IN.
Using these results, we have proposed a method to deduce the family of an
unclassified protein based on the topological properties of its SSE-IN, see [19].
Thus, we consider a protein defined by its sequence in which the amino acids
participating in the secondary structure are known. Then, we apply a method able to
associate a family from which we rely to predict the fold shape of the protein. This
work consists in associating the family which is the most compatible to the
unknown sequence. The following step is to fold the unknown sequence SSE-IN
relying on the family topological properties.
To fold a SSE-IN, we rely on the Levinthal hypothesis also called the kinetic
hypothesis. Thus, the folding process is oriented and the proteins don’t explore their
entire conformational space. In this paper, we use the same approach: to fold a SSE-
IN we limit the topological space by associating a structural family to a sequence
[19]. Since the structural motifs which describe a structural family are limited, we
propose a genetic algorithm (GA) to enumerate all possibilities.
In this section, we present a method based on a GA to predict the graph whose
vertices represent the SSE and edges represent spatial interactions between two
amino acids involved in two different SSE, further this graph is called Secondary
Structure Interaction Network (SS-IN) (see Fig. 3).

Fig. 3 2OUF SS-IN (left) and its associated incidence matrix (right). The vertices represent the
different a-helices and an edge exists when two amino acids interact
45 A Study of the Protein Folding Problem by a Simulation Model 589

4.3 Dataset

Thereafter, we use a dataset composed by proteins which have not fold families
in the SCOP v1.73 classification and for which we have associated a family in [19].

4.4 Overall Description

The GA has to predict the adjacency matrix of an unknown sequence when it is

represented by a chromosome. Then, the initial population is composed of proteins
of the associated family with the same number of SSEs. During the genetic process,
genetic operators are applied to create new individuals with new adjacency matri-
ces. We want to predict the studied protein adjacency matrix when only its
chromosome is known.
Here, we represent a protein by an array of alleles. Each allele represents a SSE
notably considering its size that is the number of amino acids which compose it.
The size is normalized contributing to produce genomes whose alleles describe a
value between 0 and 100. Obviously, the position of an allele corresponds to the
SSE position it represents in the sequence. In the same time, for each genome we
associate its SS-IN incidence matrix.
The fitness function we use to evaluate the performance of a chromosome is the
L1 distance between this chromosome and the target sequence.

4.5 Genetic Operators

Our GA uses the common genetic operators and also a specific topological operator.
The crossover operator uses two parents to produce two children. It produces
two new chromosomes and matrices. After generating two random cut positions,
(one applied on chromosomes and another on matrices), we swap respectively the
both chromosome parts and the both matrices parts. This operator can produce
incidence matrices which are not compatible with the structural family, the topo-
logical operator solve this problem.
The mutation operator is used for a small fraction (about 1%) of the generated
children. It modifies the chromosome and the associated matrix. For the chromo-
somes, we define two operators: the two position swapping and the one position
mutation. Concerning the associated matrix, we define four operators: the row
translation, the column translation, the two position swapping and the one position
mutation.
These common operators may produce matrices which describe incoherent
SS-IN compared to the associated sequence fold family. To eliminate the wrong
cases we develop a topological operator.
590 O. Gaci

The topological operator is used to exclude the incompatible children generated

by our GA. The principle is the following; we have deduced a fold family for the
sequence from which we extract an initial population of chromosomes. Thus, we
compute the diameter, the characteristic path length and the mean degree to
evaluate the average topological properties of the family for the particular SSE
number. Then, after the GA generates a new individual by crossover or mutation,
we compare the associated SS-IN matrix with the properties of the initial population
by admitting an error rate up to 20%. If the new individual is not compatible, it is
rejected.

4.6 Algorithm

Starting from an initial population of chromosomes from the associated family, the
population evolves according to the genetic operators. When the global popula-
tion fitness cannot increase between two generations, the process is stopped, see
Algorithm 1.

The genetic process is the following: after the initial population is built, we
extract a fraction of parents according to their fitness and we reproduce them to
produce children. Then, we select the new generation by including the chromo-
somes which are not among the parents plus a fraction of parents plus a fraction of
children. It remains to compute the new generation fitness.
When the algorithm stops, the final population is composed of individuals
close to the target protein in terms of SSE length distribution because of the
choice of our fitness function. As a side effect, their associated matrices are
supposed to be close to the adjacency matrix of the studied protein that we want
to predict.
In order to test the performance of our GA, we pick randomly three chromo-
somes from the final population and we compare their associated matrices to the
45 A Study of the Protein Folding Problem by a Simulation Model 591

sequence SS-IN adjacency matrix. To evaluate the difference between two matri-
ces, we use an error rate defined as the number of wrong elements divided by the
size of the matrix. The dataset we use is composed of 698 proteins belonging to the
All alpha class and 413 proteins belonging to the All beta class. A structural family
has been associated to this dataset in [19].
The average error rate for the All alpha class is 16.7% and for the All beta class
it is 14.3%. The maximum error rate is 25%. As shown in Fig. 4, the error rate
strongly depends on the initial population size. Indeed, when the initial popula-
tion contains sufficient number of individuals, the genetic diversity ensures better
SS-IN prediction. When we have sufficient number of sample proteins from the
associated family, we expect more reliable results. Note for example that when

30
All alpha dataset

20
Error Rate

10

0
[1

[6

[1

[1

[2
-5

-1

1-

6-

1-
]

0]

15

20

[
]

Initial population

30
All beta dataset

20
Error Rate

10

0
[1

[6

[1

[1

[2
-5

-1

1-

6-

1-
]

0]

[
5]

0]

Initial population

Fig. 4 Error rate as a function of the initial population size. When the initial population size is
more than 10, the error rate becomes less than 15%
592 O. Gaci

the initial population contains at least ten individuals, the error rate is always less
than 15%.

5 Conclusions

In this paper, we present a simulation model to study the protein folding problem.
We describe the reasons for which this biological process is a problem not yet
solved.
We summarize relative works about how to fold an amino acid interaction
networks. We need to limit the topological space so that the folding predictions
become more accurate. We propose a genetic algorithm trying to construct the
interaction network of SSEs (SS-IN). The GA starts with a population of real
proteins from the predicted family. To complete the standard crossover and muta-
tion operators, we introduce a topological operator which excludes the individuals
incompatible with the fold family. The GA produces SS-IN with maximum error
rate about 25% in the general case. The performance depends on the number of
available sample proteins from the predicted family, when this number is greater
than 10; the error rate is below 15%.
The characterization we propose constitutes a new approach to the protein
folding problem. Here we propose to fold a protein SSE-IN relying on topological
properties. We use these properties to guide a folding simulation in the topological
pathway from unfolded to folded state.

References

1. A.R. Atilgan, P. Akan, C. Baysal. Small-world communication of residues and significance

for protein dynamics. Biophys. J. 86(1 Pt 1), 85–91 (2004)
2. K.A. Dill, S. Bromberg, K.Z. Yue, K.M. Fiebig, D.P. Yee, P.D. Thomas, H.S. Chan, Principles
of protein folding: a perspective from simple exact models. Protein Sci. 4(4), 561–602 (1995)
3. C.B. Anfinsen, Principles that govern the folding of protein chains. Science 181, 223–230
(1973)
4. C. Levinthal. Are there pathways for protein folding? J. Chim. Phys. 65, 44–45 (1968)
5. H.M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T.N. Bhat, H. Weissig, P.E. Bourne,
I.N. Shindyalov, The protein data bank. Nucleic Acids Res. 28, 235–242 (2000)
6. M. Kataoka, Y. Goto, X-ray solution scattering studies of protein folding. Folding Des. 1,
107–114 (1996)
7. K.W. Plaxco, C.M. Dobson, Time-relaxed biophysical methods in the study of protein folding.
Curr. Opin. Struct. Biol. 6, 630–636 (1996)
8. A. Bairoch, R. Apweiler, The swiss-prot protein sequence database and its supplement trembl.
Nucleic Acids Res. 28, 45–48 (2000)
9. R.L. Baldwin, Why is protein folding so fast? Proc. Natl. Acad. Sci. USA 93, 2627–2628
(1996)
10. A. Broder, R. Kumar, F. Maghoul, P. Raghavan, S. Rajagopalan, R. Stata, A. Tomkins,
J. Wiener. Graph structure in the Web. Comput. Netw. 33(1–6), 309–320, (2000)
45 A Study of the Protein Folding Problem by a Simulation Model 593

11. S. Wasserman, K. Faust, Social network analysis: methods and applications. Structural
Analysis in the Social Sciences, vol. 8 (Cambridge University Press, Cambridge, 1994)
12. H. Jeong, B. Tombor, R. Albert, Z.N. Oltvai, A.-L. Barabàsi, The large-scale organization of
metabolic networks. Nature 407(6804), 651–654 (2000)
13. N.V. Dokholyan, L. Li, F. Ding, E.I. Shakhnovich, Topological determinants of protein
folding. Proc. Natl. Acad. Sci. USA 99(13), 8637–8641 (2002)
14. A. Ghosh, K.V. Brinda, S. Vishveshwara, Dynamics of lysozyme structure network: probing
the process of unfolding. Biophys. J. 92(7), 2523–2535, (2007)
15. U.K. Muppirala, Z. Li, A simple approach for protein structure discrimination based on the
network pattern of conserved hydrophobic residues. Protein Eng. Des. Sel 19(6), 265–275
(2006)
16. A. Dutot, F. Guinand, D. Olivier, Y. Pigné, GraphStream: A Tool for bridging the gap between
Complex Systems and Dynamic Graphs, Proc of EPNACS: Emergent Properties in Natural
and Artificial Complex Systems, Dresden, Germany, 137–143 (2007)
17. J.H. Holland, Adaptation in Natural and Artificial System (MIT Press, Cambridge, MA, 1992)
18. O. Gaci, Building a parallel between structural and topological properties. In Advances in
Computational Biology (Springer, 2010)
19. O. Gaci, Building a topological inference exploiting qualitative criteria. Evol. Boinformatics
(2010)
Chapter 46
Analysing Multiobjective Fitness Function
with Finite State Automata

Nada M.A. Al Salami

Abstract This research analyses and discusses the use of Multiobjective fitness
function to evolve Finite State Automata. Such automata can describe system’s
behavior mathematically in an efficient manner. However system’s behavior must
highly depend on its input-output specifications. Genetic Programming is used, and
the fitness function is built to guide the evolutionary process in two different cases.
First case: Single point fitness function is used where the only focus is on the
correctness of the evolved automata. Second case: multiobjective fitness function is
used since every real-world problem involves simultaneous optimization of several
incommensurable and often competing objectives. Multiobjective optimization is
defined as a problem of finding a Finite State Automata which satisfies: parsimony,
efficiency, and correctness. It has been presented that for large and complex
problems it is necessary to divide them into sub problem(s) and simultaneously
breed both sub-program(s) and a calling program.

1 Introduction

The evolutionary multi-objective optimization (EMOO) is a popular and useful

field of research and developing algorithms to solve many real-life multiobjective
problems. Almost every real-world problem involves simultaneous optimization of

Nada M.A. Al Salami (1971) Assistance Professor in Management Information System Depart-
ment, of Al Zaytoonah University, Amman, Jordan. She interested in the theory of computer
science and evolutionary algorithms

N.M.A.A. Salami
Department of Management Information Systems, Faculty of Economic and Business, Al
Zaytoonah University of Jordan, 130, Amman, 11733, Jordan
e-mail: [email protected]

S.-I. Ao et al. (eds.), Machine Learning and Systems Engineering, 595

Lecture Notes in Electrical Engineering 68,
DOI 10.1007/978-90-481-9419-3_46, # Springer ScienceþBusiness Media B.V. 2010
596 N.M.A.A. Salami

several incommensurable and often competing objectives such as performance and

cost. As a consequence, there is usually not a single optimum but rather a set of
optimal trade-offs, which in turn contains the single-objective optimal. This makes
clear that the optimization of multiple objectives adds a further level of complexity
compared to the single-objective case. Multiopjective optimization defined as a
problem of finding a vector of decision variables which satisfies constraints opti-
mizes a vector function whose elements represent the objective functions. These
function forms a mathematical description of performance criteria which are
usually in conflict with each other. Consider a minimization problem; it tends to
find a set of E for: MinE E Ø F (E), E  Rn, where, E ¼ {E1, E2, E3, . . ., En} is an n-
dimension vector having n decision variables or parameters and Ø defines a feasible
set of E. F ¼ {f1, f2, f3, . . ., fm} is an objective vector with m objective components
to be minimized, which may be competing or noncommensurable to each other [1].
Evolutionary Computing is a research area within Computer Science, which
draws inspiration from the process of natural evolution, application of evolutionary
computing may includes: Bioinformatics, Numerical combinatorial Optimization,
System Modeling and Identifications, Planning and Control, Engineering Design,
Data Mining, Machine Learning, and Artificial Life. Evolutionary computing is
needed for Developing automated problem solvers, where the most powerful
natural problem solvers are human Brain, and evolutionary process. Designing
the problem solvers based on human brain leads to the field of “neurocomputing”.
While the second one leads to evolutionary computing. The algorithms involved in
Evolutionary computing are termed as Evolutionary algorithms (EA). Evolutionary
algorithms (EAs) such as evolution strategies and genetic algorithms have become
the method of choice for optimization problems that are too complex to be solved
using deterministic techniques such as linear programming or gradient (Jacobian)
methods. EAs require little knowledge about the problem being solved, and they are
easy to implement, robust, and inherently parallel. To solve a certain optimization
problem, it is enough to require that one is able to evaluate the objective (cost)
function for a given set of input parameters. Because of their universality, ease of
implementation, and fitness for parallel computing, EAs often take less time to find
the optimal solution than gradient methods. However, most real-world problems
involve simultaneous optimization of several often mutually concurrent objectives.
Multi objective EAs are able to find optimal trade-offs in order to get a set of
solutions that are optimal in an overall sense. In multi objective optimization,
gradient based methods are often impossible to apply. Multiobjective EAs, how-
ever, can always be applied, and they inherit all of the favorable properties from
their single objective relatives. As soon as there are many (possibly conflicting)
objectives to be optimized simultaneously, there is no longer a single optimal
solution but rather a whole set of possible solutions of equivalent quality. Consider,
for example, the design of an automobile. Possible objectives could be: minimize
cost, maximize speed, minimize fuel consumption and maximize luxury. These
goals are clearly conflicting and, therefore, there is no single optimum to be found.
Multiobjective EAs can yield a whole set of potential solutions – which are all
optimal in some sense – and give the engineers the option to assess the trade-offs
46 Analysing Multiobjective Fitness Function with Finite State Automata 597

between different designs. One then could, for example, choose to create three
different cars according to different marketing needs: a slow low-cost model which
consumes least fuel, an intermediate solution, and a luxury sports car where speed is
clearly the prime objective. Evolutionary algorithms are well suited to multi
objective optimization problems as they are fundamentally based on biological
processes which are inherently multiobjective [1–3]. After the first pioneering work
on multi objective evolutionary optimization in the 1980s by Schatner [4, 5],
several different algorithms have been proposed and successfully applied to various
problems, for more details see [6, 7]. In this research, a new multiobjective
evolutionary algorithm is analyzed and discussed; it is developed to deal with
complex system. Since every real life problem are dynamic problem, thus their
behaviors are much complex. Complex systems often include chaotic behavior [8]
(the classic example of chaos theory is “the Butterfly effect”) which is to say that
the dynamics of these systems are nonlinear and difficult to predict over time, even
while the systems themselves are deterministic machines following a strict
sequence of cause and effect. Genetic programming [9–11] suffers from many
serious problems that make it very difficult to predict chaotic behavior. Natural
chaotic systems may be difficult to predict but they will still exhibit structure that is
different than purely random systems [12] for more details see [13–15]. Such
chaotic behavior of real live problem is reduced when the induction process is
based on the meaning of the low-level primitives rather than their structure.
The meaning of a program P is specified by a set of function transformation from
states to states, as given in [12, 16]; hence P effects a transformation on a state
vector X, which consists of an association of the variable manipulated by the
program and their values. A Program P is defined as 9-tuples, called Semantic
Finite State Automata (SFSA): P ¼ (x, X, T, F, Z, I, O, g, Xinitial), where: x is the
set of system variables, X is the set of system states, X¼ {Xinitial, - - - -, Xfinal}, T is
the time scale: T ¼ [0, 1), F is the set of primitive functions, Z is the state transition
function, Z ¼ {(f, X, t): (f, X, t) E F  X  T, z(f, X, t) ¼ (X, t)}, I is the set of
inputs, O is the set of outputs, g is the readout function, and Xinitial is the initial state
of the system: Xinitial E X.
All sets involved in the definition of P are arbitrary, except T, and F. Time scale
T must be some subset of the set [0, 1) of nonnegative integer numbers, while the
set of primitive function F must be a subset of the set CL(FL) of all computable
functions in the language L and sufficient to generate the remainder functions. Two
features characterize state transition function:

Zð; ; tÞ ¼ ðXinitial ; 1Þ if t ¼ 0 (1)

Zðf ; X; tÞ ¼ Zðf ; Zðf ðt  1Þ; X:t  1ÞÞ if t 6¼ 0 (2)

The concepts of reusable parameterized sub-systems is implemented by restrict-

ing the transition functions of the main system, so that it have the ability to call
and pass parameters to one or more such sub-systems. Suppose we have sub-system

P, and main-system P, then they are defined by the following 9-tuples:
598 N.M.A.A. Salami

P (x, X, T, F, Z, 1, 0, Xinitial, g)

P (xXTFZI, O, Xinitial, g)
where

x  x, Xinitial E X, then there exit *f E F, z E Zf, E F, and z EZ, and h is a
function defined over Z with value in X is defined as follows:

h ¼  z ð f ; Xinitial ; 1Þ ¼ Xh; ti (3)

zð f ; X; tÞ ¼ zðh; X; tÞ ¼ Xh;t (4)

*f is a special function we call it sub-SFSA function to distinguish it from other

primitive functions in the set F. Also, we call the sub-system P, sub-SFSA, to
distinguish it from the main SFSA. Formally, a system P is a sub-system of
a system P, iff: x  x, T  T, I  I, O  O, g must be the restriction of g
to O, and F  N, where N is the set of restrictions of F toT. If (f, X, t) is
an element of F   X  T, then there exists f E F, such that the restriction of f to

T is f, and z (f, X, t) is z (f, X, t).

2 Evolutionary Algorithm

On the basis of the theoretical approach sketch in the above section, the evolution-
ary algorithm is defined as 7-tuples: (IOS, S, F, a1, Tmax, b, y), as given in [12, 16].

2.1 Input-Output Specification (IOS)

IOS is establishing the input-output boundaries of the system. It describes the inputs
that the system is designed to handle and the outputs that the system is designed to
produce. An IOS is a 6-tuples: IOS ¼ (T, I, O, Ti, TO, Z). Where T, is the time scale
of IOS, I is the set of inputs, O is a set of outputs, Ti is a set of input trajectories
defined over T, with values in I, TO, is a set of output trajectories defined over T,
with values in O, and Z is a function defined over Ti whose values are subset of TO;
that is, Z matches with each given input trajectories the set of all output trajectories
that might, or could be, or eligible to be produced by some systems as output,
experiencing the given input trajectory. A system P satisfies IOS if there is a state X
of P, and some subset U not empty of the time scale T of P, such that for every input
trajectory g in Ti, there is an output trajectory h in TO matched with g by Z such
that the output trajectory generated by P, started in the state X is:

gðZðf ðgÞ; X; tÞ ¼ ðhðtÞÞ; for every t 2 U (5)

46 Analysing Multiobjective Fitness Function with Finite State Automata 599

2.2 Syntax Term (S)

Refers to the written form of a program as far as possible independent from its
meaning. In particular it concerns the legality or well-formedness of a program
relative to a set of grammatical rules, and parsing algorithms for the discovery
of the grammatical structure of such well-formed programs. S is a set of rules
governing the construction of allowed or legal system forms.

2.3 Primitive Function (F)

Each function must be coupled with its effect on both the state vector X, and the
time scale T of the system. Some primitive functions may serve as primitive
building blocks for more complex functions or even sub-systems.

2.4 Learning Parameter (a1)

Is a positive real number specifying the minimum accepted degree of matching

between an IOS, and the real observed behavior of the system over the time scale,
Tx, of IOS.

2.5 Complexity Parameters (Tmax, b)

Tmax and b parameters are merits of system complexity: size and time, respectively.
It is important to note that there is a fundamental difference between a time scale T
and an execution time of a system. T represents system size, it defines points within
the overall system, whereas, b, is the time required by the machine to complete
system execution, hence it is high sensitive to the machine type.

2.6 System Proof Plan (y)

Prove process should be a part of the statement of system induction problem

especially when the IOS is imprecise or inadequate to generate an accurate system.
We say P is correct iff it computes a certain function f from Xinitial E X properly, that
is if for each Xi E X, P(Xi) is defined, i.e. P does not loop for Xi, and is equal to f(Xi).
Broadly speaking, there have been two main approaches to the problem of
600 N.M.A.A. Salami

developing methods for making programs more reliable: Systematized testing, and
Mathematical proof.
Our works use systematized testing approach as a proof plane. The usual method
for verifying that a program is correct by testing is by choosing a finite sample of
states X1, X2, . . ., Xn and running P on each of them to verify that: P(X1) ¼ f(X1),
P(X2) ¼ f(X2), . . ., P(Xn) ¼ f(Xn). Formally, if testing approach is used for system
verification, a system proof is denoted u ¼ (a2, d), where a2 is a positive real
parameter defining the maximum accepted error from testing process. a2 focus on
the degree of generality, so that a1, and a2, parameters suggest a fundamental
tradeoff between training and generality. On the other hand, d represents a set of test
cases pairs (Oi, Ki), where Ki is a sequence of initial state Xinitial and input Ii.
In addition to using the idea of sub-system functions, i.e., sub-FSA, for complex
software it is better to divide the process of system induction into sub-system(s) and
main-system inductions. Sub-system induction must be accomplished with several
objectives; first one is that a suitable solution to sub-problem must determine a
solution to the next higher level problem. Second is to ensure that the figure of merit
of the sub-system have relationships to the figure of merit of the top-level problem.
The third objective is to ensure specific functional relationships of sub-system proof
plans to the system proof plan of the top-level problems.

3 Evolutionary Process

3.1 Single Objective Evolutionary Process

The basic idea for single objective EAs is to imitate the natural process of biological
evolution. The problem to be solved is therefore described using a SFSA with
certain number of parameters (design variables). One then creates a group of
(n > 0) different parameter vectors and considers it as a population of individuals.
The quantity is called the population size. The quality of a certain SFSA (i.e. an
individual in the population) is expressed in terms of a scalar valued fitness function
(objective function). Depending on whether one wants to minimize or maximize the
objective function, individuals (i.e. parameter vectors) with lower or greater fitness
are considered better, respectively. The algorithm then proceeds to choose the best
individuals out of the population to become the parents of the next generation
(natural selection, survival of the fittest). In this work we always try to minimize
the objective function, i.e. search to find individual with lower fitness value. In this
case, learning parameter a1 must be selected to the minimum accepted degree
of matching between an IOS, and the real observed behavior of the SFSA over the
time scale, TX. Learning parameter a1 is used ultimately to guide the search
process.
Within the context of the suggested mathematical approaches, to automatically
generate a system means search to find an appropriate SFSA satisfying IOS
46 Analysing Multiobjective Fitness Function with Finite State Automata 601

efficiently, with regard to learning and complexity parameters. Then, proof plan u
must be applied to that SFSA for further assessing its performance and correctness.
If that SFSA behaves well with u, it may be considered as a solution or approximate
solution to the problem, else, some or all terms in the statement of the problem of
system induction must be modified, and the process is repeated until a good SFSA is
found, or no further revisions can be made. The search space in Evolutionary
algorithm is the space of all possible computer programs described as 9-tuples
SFSA. Within the context of the suggested mathematical approaches, to automati-
cally generate a system means search to find an appropriate SFSA satisfying IOS
efficiently, with regard to learning and complexity parameters. Then, proof plan u
must be applied to that SFSA for further assessing its performance and correctness.
If that SFSA behaves well with u, it may be considered as a solution or approximate
solution to the problem, else, some or all terms in the statement of the problem of
system induction must be modified, and the process is repeated until a good SFSA is
found, or no further revisions can be made. The search space in Evolutionary
algorithm is the space of all possible computer programs described as 9-tuples
SFSA. The following equation was used to express the quality of a SFSA (i), where
0 < i <¼ n:
 XTx 
0
fitness ðiÞ ¼ a  j¼0
j T ð j Þ  R ð j Þj (6)

Therefore, to generate a new population individuals with smaller fitness value

are chosen and regenerate them to the next population. In addition to reproduction
genetic operation, mutation and crossover genetic operations are also used to
generate new individuals from the old individuals with smaller fitness value.
When the first iteration is completed, then the algorithm loops back to fitness
evaluation phase.

3.2 Multi Objective Evolutionary Process

Two major problems must be addressed when an evolutionary algorithm is applied

to multiobjective optimization. First: how to accomplish fitness assignment and
selection, respectively, in order to guide the search towards the Pareto-optimal set.
Second: how to maintain a diverse population in order to prevent premature
convergence and achieve a well distributed trade-off front. Now, the objective
function should be satisfying the following characteristic [1, 2]:
Complete: so that all pertinent aspects of the decision problem are presented
Operational: in that they can be used in a meaningful manner
Decomposable: if desegregation of objective functions is required
Nonredundant: so that no aspect of the decision problem is considered twice
Minimal: such that there is no other set of objective functions capable of represent-
ing the problem with a smaller number of elements
602 N.M.A.A. Salami

Since there are many objectives to be optimized simultaneously, there is no

longer a single optimal solution but rather a whole set of possible solutions of
equivalent quality. Multi objective EAs can yield a whole set of potential solutions
which are all optimal in some sense, and give the engineers the option to assess the
trade-offs between different designs. One then could, for example, choose to create
three different SFSA according to different needs. Evolutionary algorithms are well
suited to multi objective optimization problems as they are fundamentally based on
biological processes which are inherently multi objective. Multiobjective fitness
measure is adopted to incorporate a combination of three objectives that are:
Correctness: how long the SFSA satisfy IOS as described above.
Parsimony: the length of the evolved SFSA must be smaller as possible i.e.
smallness T. It must contain nonredundant or useless transformations.
Efficiency: although the time required by any machine to complete system execu-
tion is high depended on machine type, however the evolved SFSA must has
smallness b, as possible. Otherwise the complexity of that SFSA is affected.
Accordingly, multiobjective fitness is computed by the following equation:
 XTx 
0
fitnessðiÞ ¼ d1 a  j¼0
j  ð T ð j Þ  ðR ð jÞ Þ j þ d2ðTmax  TiÞ þ d3

 ðb  biÞ (7)

Where: d1, d2, and d3 are weight parameters: (d2, d3) hd1i ¼ 0, bi is the run time
of individual i, Ti is the time scale of the individual i, and Rj is the actual calculated
input trajectory of individual j. Since the Main objective is to satisfy the IOS, the
first weight parameter in Eq. (7), i.e. d1 is always greater than d2 and d3, and grater
than or equal to 1. These values are selected by expert when the algorithm is running
many times. In addition the values of d2 and d3 are selected with respect to different
factors concerning: hardware type, required complexity, problem nature, and other.
In all cases they must be small real numbers greater than or equal to zero. Table 1,
gives different setting for these weight parameters, when the value is equal to 0 it
means this objective is not considered in the evolutionary process. Also if it equal to
1 it means the effect of this objective is not either decreased or decreased. It is clear
that the first setting for the parameters represent single point fitness function, as

Table 1 Weight parameters with different setting

Objective type Objective degree d1 d2 d3
IOS Correctness High 1 0 0
Parsimony High 5 2 0
Efficiency High 5 0 2
Parsimony Moderate 5 2 1
Efficiency Moderate 5 1 2
Parsimony Low 4 0.5 1
Efficiency Low 6 2 0.2
46 Analysing Multiobjective Fitness Function with Finite State Automata 603

given by Eq. (6). In other words, single-objective optimization can be considered a

special case of multiobjective optimization (and not vice versa).
To give rise to the fitness variation in the overall population from one generation
to the next, the fitness of each individual is computed F:
0
00 fitness ðiÞ
fitness ðiÞ ¼ PM (8)
J¼0 0fitness ðjÞ

Where: M: is the population size. Eq. (8) is also adjusted so that the adjusted
fitness lies between 0 and 1:

1
fitness ðiÞ ¼ 00 (9)
fitness ðiÞ

Three types of points are defined in each individual: transition, function, and
function arguments. When structure-preserving crossover is performed, any point
type anywhere in the first selected individuals may be chosen as the crossover point
of the first parent. The crossover point of the second parent must be chosen only
from among points of this type. The restriction in the choice of the second crossover
points ensures the syntactic validity of the offspring.

4 Result and Discussion

4.1 Input-Output Specification

Unfortunately, when we deal with complex systems and real live problem, strong
feedback (positive as well as negative) and many interactions exist: i.e. chaotic
behavior. Thus, we need to find a way to control chaos, understand, and predict
what may happen long term. In these cases input and output specifications are self
organized, which mean that trajectory data are collected and enhanced over time,
when genetic generation process runs again and again. With high trajectory infor-
mation converge to the solution in best than these populations with little trajectory
information. Although trajectory data are changed over time, but by experiment, it
still sensitive to initial configuration (sensitivity to the initial conditions). There is a
fundamental difference between a crossovers occurring in a sub-SFSA versus one
occurring in the main-SFSA. Since the later usually contains multiple references to
the sub-SFSA(s), a crossover occurring in the sub-SFSA is usually leveraged in the
sense that it simultaneously affects the main-SFSA in several places. In contrast, a
crossover occurring in the main-SFSA provides no such leverage. In addition,
because the population is architecturally diverse, parents selected to participate in
the crossover operation will usually possess different numbers of sub-SFSA(s). The
proposed architecture-altering operations are: Creating sub-SFSA, Deleting sub-
SFSA, Adding Variables, and Deleting Variables.
604 N.M.A.A. Salami

4.2 Performance

If every run was successes to obtain a solution, the computational effort required to
get the solution depends primarily on four factors: population size M, number of
generation g (g must be less than or equal to the maximum number of generation G),
the amount of processing required for fitness measure over all fitness cases (we
assume that the processing time to measure the fitness of an individual is its run
time), and the amount of processing required for test phase e. If success occurs on
the same generation of every run, then the computational effort E would be
computed as follows:

E¼Mge (10)

Since the value of e is too small with respect to other factors, we shall not
consider it. However, in most cases, success occurs on different generations in
different runs, then the computational effort E would be computed as follows:

E ¼ M  gavr  b (11)

where gavr is the average number of executed generations. Since we use a probabi-
listic algorithm, thus the computational effort is computed in this way: first deter-
mining the number of independent runs R needed to yield a success with a certain
probability, second, multiply R by the amount of processing required for each run.
The number of independent runs R required to satisfy the success predicate by
generation i with a probability z which depends on both z and P(M, i), where z is the
probability of satisfying the success predicate by generation i at least once in R runs
defined by:

Z ¼ 1  ½1  PðM; iÞR (12)

P(M, i) is the cumulative probability of success for all the generations between
generation 0 and generation i. P(M, i) is computed after experimentally obtaining an
estimate for the instantaneous probability Y(M, i) that a particular run with a popula-
tion size M yields, for the first time, on a specified generation i, an individual is
satisfying the success predicate for the problem. This experimental measurement of
Y(M, i) usually requires a substantial number of runs. After taking logarithms, we find:
 
 logð1  zÞ 

R¼  (13)
logð1  rðM; iÞÞ

The computational effort E, is the minimal value of the total number of indivi-
duals that must be processed to yield a solution for the problem with z probability
(ex: z ¼ 99%):

E ¼ M  ðg þ 1Þ  b  R (14)
46 Analysing Multiobjective Fitness Function with Finite State Automata 605

Where *g is the first generation. The computational effort ratio, RE, is the ratio
of the computational effort without sub-SFSA to the computational effort with
sub-SFSA:

E without sub  SFSA

RE ¼ (15)
E with sub  SFSA
The fitness ratio, Rfitness, is the ratio of the average fitness without sub-SFSA to
the average fitness, with sub-SFSA, for a problem.

fitness average without sub  SFSA

Rfitness ¼ (16)
fitness average with sub  SFSA

5 Conclusion

There are many objectives to be optimized simultaneously, thus there is no longer a

single optimal solution but rather a whole set of possible solutions of equivalent
quality. Multi objective EAs can yield a whole set of potential solutions which are
all optimal in some sense, and give the engineers the option to assess the trade-offs
between different designs. The discussed evolutionary algorithm can be changed to
produce programs in different implementation languages without significantly
affecting existing problem specification, leading to an increase in the system
productivity. Multiobjective fitness measure is adopted to incorporate a combina-
tion of three objectives: correctness, parsimony, and efficiency. Trajectory data of
the system is self-organization to reflect the chaotic behavior in real live applica-
tions. Convergences time is highly sensitive to the initial input-output specification
of the problem. Programs with fewer sub-programs tend to disappear because they
accrue fitness from generation to generation, more slowly than those programs with
more sub-programs. Sub-systems can be reused to solve multiple problems. They
provide rational way to reduce software cost and increase software quality.
Although complexity parameter provide a way to bound the search space, their
effect in the induction process is relatively less than the effect of learning and
generalization parameters. System analyst must select a value for a2, so that
missing or incorrect specifications of some input-output relationships don’t lead
to un-convergence situation.

References

1. K.C. Tan, E.F. Khor, T.H. Lee, Muliobjective Evolutionary Algorithms and Applications
(Springer-Verlag, London limited, 2005
2. E. Zitzler, K. Deb, L. Thiele, Comparison of multiobjective evolutionary algorithm: empirical
result. Evol. Comput. 8(2), 173–195 (2000), (Massachustts Institute of Technology)
606 N.M.A.A. Salami

3. E. Zitzler, Evolutionary algorithm for multiobjective optimization, Evolutionary Methods for

Design, Optimization and Control (CIMNE, Barcelona, Spain, 200)
4. J.D. Schaffner, Multiple Objective Optimization with Vector Evaluated Genetic Algorithms.
Unpublished Ph.D. Thesis, Vanderbilt University, Nashville, TN, 1984
5. J.D. Schaffner, Multiple objective optimization with vector evaluated genetic algorithm, in
Proceeding of an International Conference on Genetic Algorithms and their Applications,
sponsored by Texas Instruments and the U.S. Navy Center for Applied Research in Artiffic1
intelligence (NCARAI) pp. 93–100, 1985
6. C.M. Fonseca, P.J. Fleming, Am overview of evolutionary algorithms in multiobjective
optimization. Evol. Comput. 3(1), 1–16 (1995)
7. C.A.C. Coello, A comprehensive survey of evolutionary-based multiobjective optimization.
Knowl. Inf. Syst. 1(3), 269–308, 1999
8. L. Smith, Chaos: Avery Short Introduction (Oxford University Press, UK, 2007
9. J.R. Koza, Genetic Programming: on the Programming of Computer by Means of Natural
Selection (MIT Press, Cambridge, MA, 2004)
10. D.E. Golberg, Genetic Algorithm in Search, Optimization, and Machine Learning (Addison-
Wesley, Boston, MA, 1989)
11. M. Mitchell, An Introduction to Genetic Algorithm (MIT Press, Cambridge, MA, 1996)
12. N.A. Salami, Evolutionary algorithm definition. AJEAS 2(4), 789–795 (2009)
13. J.P. Koza, Two ways of discovering the size and shape of a computer program to solve a
problem, pp. 287–294,1998
14. A. Kent, J.G. Williams, C.M. Hall, Genetic programming. Encyclopedia of Computer Science
and Technology (Marcel Dekker, New York, 1998), pp. 29–43
15. R. Poli, W.B. Langdon, N.F. McPhee, J.R. Koza, Genetic Programming: An Introductory
Tutorial and a Survey of Techniques and pplications, Technical Report CES-475 ISSN: 1744-
8050. essex.ac.uk/dces/research/publications/. . ./2007/ces475.pdf. Oct 2007
16. N.A. Salami, Genetic system generation, in Proceeding of the WCECS 2009, vol. I.
(San Francisco USA, 20–22 Oct 2009), pp. 23–27 ISBN:978-988-17012-6-8
Index

A B
Abdallah, I.B., 271 Bai, Y., 353
Abdelhalim, A., 521 Balch, T., 20
Abler, R., 85 Ball Mill, 312, 314, 316, 317
Abou, S.C., 283, 311 Bansel, R., 145
Accelerometer, 569–578 Bashah, N., 85
Accuracy, 83, 84, 144–148, 150–152, 154, 184, Basic reproductive number, 561, 564, 565, 567
187, 300, 362, 416–419, 492, 529 Battery electric vehicle, 227–241
Actuator dynamics, 214, 215, 217 Bednar, J., 131
Adaptive refinement, 158 Begg, R., 483
Adaptive selection, 112, 114–116, 502 Ben-Abdallah, I., 271
Addition chain, 507–518 Benchmarks, 112, 412, 416, 419, 465–480
Ad hoc network, 495–506 Beyls, K., 413
Adi, W., 513 Bezier, 31–45
Adly, A.S., 533 Billinghurst, M., 361
Adly, M.S., 533 Biological transport, 297
Ahmed, A.M., 85 Biomedical systems, 547–556
Akiba, Y., 523 Biped robot, 213–224
Almuallim, H., 523 Bit rate, 385, 387, 388, 391–394, 469
Al Salami, N.M.A., 595 Bluetooth, 495–497, 503–506
Analog to digital converter, 62, 63 Botta, A., 466
Anfinsen, C.B., 582 Brache, M., 197
ANFIS-GRID, 145, 147, 152–154 Breast cancer, 341–342, 348, 350
Ant colony optimization, 111–119 Budiarto, R., 83
Anti-windup, 257, 258, 260, 261, 263, 264
Arnold, W., 542 C
Aronson, J.E., 484 Calder, B., 415
Artificial neural network, 72, 73, Cano, D.C., 557
482–484, 491 Cao Jine, 122
Assimacopoulos, A., 361 Cascón, J.M., 157
Association rules mining, 312–314, 319–322, Chai, C.Y., 481
486, 487 Chandra, M.G., 373
Audio visual features, 374, 382 Chang, K., 298
Automatic annotation, 373, 375, Chan, J., 545
379–380, 382 Chellappan, C., 439
Autonomous vehicle, 31–32, 45 Chen, D., 271
Averaged model, 172–176 Chen, J.-S., 274

607
608 Index

Cheok, K.C., 213 Dual-resource constraints, 269–280

Chiu, S., 451 Dynamic neural network, 197–200
Choi, J.-W., 31
Chuah, C., 466 E
City, 97–108, 113–117 Eberhart, R.C., 74
Claypool, M., 466 ECC, 508, 509, 513, 516–518
Coaxial flow, 244 Economy model, 439–448
Collaboration, 17–22, 24–25, 27, 28 Edahiro, K., 257
Commercial break detection, 373–383 Edward, R.D., 482, 483
Complex stochastic Boolean systems, 397–409 Einhorn, M., 169
Computational economy, 53, 55 Electrode wear, 184
Computational fluid dynamics (CFD), 243, 244 Elkaim, G., 31
Computational systems, 74, 533–545 ElMaraghy, H.A., 271
Computer-aided-diagnosis, 348, 350 Endisch, C., 197
Computer algebra and mechanized reasoning, Engelbrecht, A.P., 19
557–568 Epidemic thresholds, 557
Confined jets, 243–255 Error rate, 389, 500, 502, 505, 542–544,
Connors, D.A., 413 590–592
Content based video retrieval, 380–382 Escobar, J.M., 157
Control room, 287, 338–339 Estimation, 258, 264–265, 268, 316, 353–358,
Correlation classifier, 381 389, 466, 573
Crispien, K., 360, 361 Euler-Lagrange, 244
Cruz-Cortes, N., 514 Evolutionary computing, 74
Cryptographically generated addresses (CGA),
455, 456, 459 F
Cryptography, 423, 435, 436, 507 Fannon, M., 297
Curry, R., 31 Fault detection, 4, 258, 262, 284, 285, 287
Feedback control, 260
D Fermat’s theorem, 509–513
Dao, T.-M., 311 FGF-2, 297–309
Data mining, 461–492, 521 Finite state machine, 595–605
Data quality, 143–155 Finite volume, 251
Dauzère-Pérès, S., 271 Fixed points, 126, 128, 129
Davis, E.W., 313 Floater, M.S., 158, 159
DC-DC power conversion, 169 Forsten-Williams, K., 297
Deane, R.H., 271 Free flyer, 1–14
Decision support system model, 481–492 FreeTTS, 360, 362–366
de la Fuente, J., 545 Fundamental magic square, 427–428, 431, 432
Deneubourg, J.-L., 19 Fuzzy logic, 284–289, 293, 294, 313, 314,
Deng, M., 257, 259 316–319, 323, 324
Detrended fluctuation analysis (DFA), Fuzzy modular arithmetic, 508, 509, 513–514,
547–556 516–518
D’Hollander, E.H., 413
Diagnostic, 298, 342 G
Dicke, C., 361 Gaci, O., 581
Differential susceptibility, 557 Gamess, E., 451, 465
Digital mammogram, 343, 349 Ganapathy, G., 423, 427, 507
Dimos, J.T., 545 Gargeya, V.B., 271
Direction-of-arrival (DOA), 353–358 Genetic algorithm, 271, 284, 482, 514, 582,
Discrete-event simulation, 121–129 586–588, 592, 596
Djuzenova, C., 542 Genetic programming, 597
Docking, 2, 4, 6, 7, 9–11, 13, 14 Gill, K., 71
Dorigo, M., 19 Glaser, W., 298
Double dumbbell, 386, 389–395 Golzio, M., 542, 543
Index 609

González, E., 121 IPv6, 451–463, 467, 468, 470, 472, 474, 476,
González, L., 397 478, 479
Goose, S., 361, 365 Ivanova, N.B., 545
Government, 97–108
Gragger, J.V., 169 J
Grammatical evolution, 131–142 Janiszewski, A., 413
Graphical user interface, 63–64, 471 Jassar, S., 143
Greedy mechanism, 111–119 Java, 359–370, 457, 458, 460, 462, 474
Grid scheduling, 440, 441, 445, 447 Jee, S.H., 569
Group delay, 353–358 Jianhua, Y., 86
Gulbinowicz, Z., 183 Jones, C.P., 483
Gupta, D., 466
K
H Kaashoek, F., 466
Hammer, D., 298 Kaheil, Y., 71
Hamming weight, 399, 404–406 Kamich, B.M., 482
Haptics, 2, 6, 7, 11, 13, 14 Kamruzzaman, J., 483
Hasan, M., 227 Kandil, O.A.D., 533
Haumer, A., 169 Kaneda, S., 523
Haupt, R., 270 Katabi, D., 466
Head-related impulse responses (HRIR), Katsuyama, T., 547
364–366 Kauffman, R., 145
Heartbeat, 548–552, 554, 555 Kennedy, J., 74
Heger, J., 269 Kennel, R., 197
Herberlein, L.T., 84, 86 Khalil, A., 71
Heuristic method, 48, 53–55 Kihm, J.L., 413
He, X., 298 Kim, J., 569
Hildebrandt, T., 269 Kinicki, R., 466
Hjelsvold, R., 361 Knapsack problem, 52, 53
Holland, J.H., 586 Kodlahalli, S., 361
HRIR. See Head-related impulse responses Kozak, J., 183
(HRIR) Král, A., 466
Huang, S.S., 84 Kulkarni, M., 283
Huckaby, E.D., 243 Kumar, A., 59
Huffmire, T., 412
Human spacecraft interaction, 1–14 L
Hybrid intelligence, 85–86, 92–93 Labella, T.H., 19
Hydraulic system, 283–288, 293 Laplace distribution, 134–136, 139–141
Lasserre, J.B., 271
I Lau, J., 415
Ibrahim, M.S., 533 Launder, B.E., 246
Imam, I.F., 526 L2 cache performance prediction, 412
In-car multimedia, 496–500, 505 Lee, M., 569
Incomparable binary n-tuples, 398, 407, 409 Levenberg–Marquardt, 198, 201–202, 205, 207
Inferential model, 143–145, 147, 154 Levinthal, C., 583
Inoue, A., 257 Liang, T.P., 484
Input saturation, 257–268 Liao, Z., 143
Intelligent service robot, 17–28 Liquid crystal display (LCD), 60, 62, 67, 68
Intensification and diversification, 114, Loh, R.N.K., 213
116, 119 Long-term variability, 71
Interaction networks, 582, 585–586, 588, 592 Losses, 40, 73, 123, 170, 172–180, 237, 264,
Intrinsic order, 397–409 328, 387, 389, 391–393, 395, 469, 472,
Inverse reference model, 216–217 479, 483, 503, 512, 570, 572
610 Index

Luh, P.B.c., 271 N

Lyapunov, 122, 214, 217, 218, 221, 224 Naumann, A., 249
Navier–Stokes, 244, 302, 303, 307–309
M Nested meshes, 158
Macpherson, G.B., 250 Neural networks, 53, 121–129, 143–145,
Magee, J., 482, 483 197–201, 203–204, 214, 215, 221–224,
Magic square based on properties, 428–434 284, 483
Magnitude of error, 218 Nonlinear systems, 197–210, 285, 308
Mani, K., 423, 427, 507 Non-Shannon entropy, 341–350
Marichal, R., 121 Nordin, N., 250
Marzouk, O.A., 243 NP-hard problem, 582
Mataric, M., 19
Mati, Y., 271 O
Matoušek, R., 47 O’Donoghue, K., 545
Matousek, R., 131 Omar, M.N., 83
McDermott, M., 227 On-orbit servicing (OOS), 1, 8, 11
Memory phase classification, 412–415 Ošmera, P., 47
Mesh untangling and smoothing, 158, 160, Open source software, 97–108
161, 163 Optimization, 32, 38, 42, 48, 71–81, 111–119,
Michalski, R.S., 526 131, 133, 158, 161–166, 194, 198,
Micro EDM, 184 200–202, 204, 205, 207, 208, 210, 241,
Micro SD-memory, 574 271, 321–323, 411, 423, 435, 507–518,
MILP. See Mixed integer linear program 582, 595–597, 601–603
(MILP) Ostergaard, E., 19
MIMO processes, 257–268 Osunleke, A.S., 257
MIMO systems, 257, 258, 260, 262–268
Mingzhe, L., 466 P
MiniFab, 271, 272, 274 Pan, J.C.-H., 274
Mixed integer linear program (MILP), Parker, S., 466
269, 274 Passeman, F., 122
Modelica, 177, 180 Patel, V., 271
Model reference adaptive control (MRAC), Path planning, 2, 31–45
214–222, 224 Pavageau, C., 271
Mohapatra, P., 466 Pechter, W., 361
Molecular dynamics, 256 Perceived quality, 387–389, 391–393
Montenegro, R., 157 Performance analysis, 375, 495–506
Montero, G., 157 Performance evaluation, 112, 416, 479,
Montesino, F., 466 503–505
Moon, D., 71 Pescapé, A., 466
Motor, 6, 203, 215, 220, 228, 231, Petrucci, L.S., 361
233–236, 238–241, 312–314, Pham-Trong, V., 158
316, 317, 322 Pheromone update rule, 112, 114–116
Multi-domain simulation, 169–180 Physical activity, 569–578
Multigrid, 297–309 Piñeiro, J.D., 121
Multimedia streaming, 388, 389, 395 Pitchfork bifurcation, 121–129
Multimodal haptics, 2, 6–7 Power plant, 327–339
Multi-modality, 1–14, 131 Predictive control, 257, 258
Multiobjective evolutionary algorithm, 597 Premature ventricular contractions (PVC),
Multiobjective optimization, 71–81, 601, 603 549–551
Municipal, 98–102, 105 Priority rule scheduling, 269–280
Muntean, G.-M., 385 Process control, 257–268, 335
Murugesan, G., 439 Process modeling, 322, 328, 585
Myszka, D., 298 Project management, 47
Index 611

Protein folding problem, 581–592 Schmechel, C., 466

Proxy-client-server (PCS), 388, 391 Schoenmackers, S., 415
Public-key cryptosystem, 423–436, 507, 508, Scholz-Reiter, B., 269
514, 518 Secure neighbor discovery protocol (SEND),
Pun, T., 361 452, 454–457, 460, 462, 463
PVC. See Premature ventricular Security, 83, 85, 423–436, 452, 455, 462, 483,
contractions (PVC) 507, 508
Šeda, M., 47
Q Self-organizing map (SOM), 85, 87–91, 93
Qiu, H.-H., 244 Semiconductor gas sensors, 60
Quah, T.S., 482, 483 Settle, A., 413, 420
Quality oriented adaptive scheme (QOAS), Shanmugam, I.B., 85
387–389, 392, 393, 395 Shannon entropy and texture-feature-
extraction, 341–350
R Sharda, R., 484
Radovnikovich, M., 213 Sharma, B.I., 246
Ramani, A.K., 495 Sharma, H., 495
Ranking interval, 397–409 Sharma, K., 495
Real time clock (RTC), 60, 62–63, 68 Shen, W., 297
Re-entrant processes, 272, 274 Sherwood, T., 412
Relay satellite, 8, 10–12 Shimoda, Y., 547
Remote environment, 1–14 Shi, Y., 484
Resource allocation, 439–448 Shi, Y.Y., 484
Resource-constrained scheduling, 47–56, Shi, Z., 484
269–280 Shi, Z.K., 484
Resource management, 440–442 Shop-floor control, 280
Review, 19, 48, 53, 98, 257, 270, 321, 476, Shou-Hsuan, S.H., 86
569–578 Signal processing, 353, 362, 365
Robot grouping, 18–20, 25–26, 28 Simulation, 7, 24, 25, 32, 43–45, 112, 123, 128,
Robustness, 2, 338, 382 137, 169–180, 183–194, 199, 200, 202,
Robust tracking, 257–268 215, 222–224, 227–241, 244, 250–251,
Rockefeller, B., 482, 483, 486 253, 258, 265–270, 273, 277–280, 289,
Rodic, D., 19 293–294, 303, 304, 322–323, 327–331,
Rodrı́guez, E., 157 333–335, 355–357, 387, 389–392, 411,
Roldán-Villasana, E.J., 327 413, 416, 420, 436, 502, 503, 534, 536,
Rols, M.-P., 542, 543 541–545, 581–592
Romero-Jiménez, G., 327 Simulator training, 327–339
Rotating electrode, 184, 185 Singh, A.P., 341
Roth, P., 361 Singh, B., 341
Roux, W., 271 Singh, I.P., 59
RSA, 424, 425, 435, 436, 455, 456, 459, 460, SIR model, 557, 559–560
508, 514 Sodnik, J., 359–361
Soft sensor, 143–154
S Sommerfeld, M., 244
Saad, A., 85 Spatial sound, 361, 364, 366
Saenz-Otero, A., 1 Stachowicz, M., 283
Šandera, C., 47 Staniford-Chen, S., 84, 86
Sarker, R., 483 STE. See Sum epsilon tube error (STE)
Scaling exponent, 548–555 Stepping stone detection (SSD), 83–94
Schaniel, C., 545 Stepping stone intrusion, 83, 85
Schatner, J.D., 597 Stock price prediction, 482–484, 491, 492
Schiller, L., 249 Stoll, E., 1
612 Index

Strauss, J., 466 Ventre, G., 466

Sud, S.K., 59 Virtual reality (VR), 5, 7, 10, 13
Suh, J., 17
Sukhatme, S., 19
Sukhorukov, V., 542 W
Sum epsilon tube error (STE), 131–142 Ward, D.B., 353
Sum of squares, 135, 376, 508–516 Ward, D.J., 97
Support vector machines (SVM), 72–74, 76–81 Weber, G., 361
Surface and volume parametrization, 157–166 Wei, M., 146
Swarm intelligence, 74 Weisser, R., 47
System identification, 197–210 Weitz, R., 145
Weller, H.G., 250
T Wideband, 353–358
Taib, S.M., 481 Wind energy, 72, 80, 81
Tanaka, K., 548 Wojtusiak, J., 527
Tao, E.Y., 97 Woo, C.-W., 17
Task planning, 21, 24–25, 27, 28 Wu, D., 466
Teissie, J., 542, 543 Wu, H., 84
Telepresence, 2, 5–8, 12 Wurz, W., 361
Tetrahedral mesh generation, 159–160, 163
Thames, J.L., 85 X
Time series analysis, 548–550, 552, 554 Xie, X., 271
Tomažič, S., 359, 361
Torres, J., 121 Y
Traffic generation, 468, 469, 471, 473, 479 Yazawa, T., 547
Traore, I., 521 Yoo, S.K., 569
Turban, E., 484 Yoshikawa, M., 111
Tweddle, B., 1
Two mass system, 197–210 Z
Zabre, E., 327
U Zadeh, L.A., 285
Ubik, S., 466 Zhang, C., 297
Unknown disturbances, 257–268 Zhang, J., 297
Zhang, W., 353
V Zhang, X., 353
Vehicle dynamics, 31, 235, 238 Zhan, W., 227
Velásquez, K., 465 Zhao, L., 143
Vempaty, P.K., 213 Zimmermann, U., 542
Venkataraman, H., 385 Zoghi, B., 227
Venkatesh, N., 373 Zwick, M., 411