Lecture Notes in Computer Science 6018

Commenced Publication in 1973

Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen

Editorial Board
David Hutchison
Lancaster University, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Alfred Kobsa
University of California, Irvine, CA, USA
Friedemann Mattern
ETH Zurich, Switzerland
John C. Mitchell
Stanford University, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
Oscar Nierstrasz
University of Bern, Switzerland
C. Pandu Rangan
Indian Institute of Technology, Madras, India
Bernhard Steffen
TU Dortmund University, Germany
Madhu Sudan
Microsoft Research, Cambridge, MA, USA
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max-Planck Institute of Computer Science, Saarbruecken, Germany
David Taniar Osvaldo Gervasi
Beniamino Murgante Eric Pardede
Bernady O. Apduhan (Eds.)

Computational Science
and Its Applications –
ICCSA 2010

International Conference
Fukuoka, Japan, March 23-26, 2010
Proceedings, Part III

13
Volume Editors

David Taniar
Monash University, Clayton, VIC 3800, Australia
E-mail: [email protected]

Osvaldo Gervasi
University of Perugia, 06123 Perugia, Italy
E-mail: [email protected]

Beniamino Murgante
University of Basilicata, L.I.S.U.T. - D.A.P.I.T., 85100 Potenza, Italy
E-mail: [email protected]

Eric Pardede
La Trobe University, Bundoora, VIC 3083, Australia
E-mail: [email protected]

Bernady O. Apduhan
Kyushu Sangyo University, Fukuoka 813-8503, Japan
E-mail: [email protected]

Library of Congress Control Number: 2010922807

CR Subject Classification (1998): C.2, H.4, F.2, H.3, C.2.4, F.1

LNCS Sublibrary: SL 1 – Theoretical Computer Science and General Issues

ISSN 0302-9743
ISBN-10 3-642-12178-0 Springer Berlin Heidelberg New York
ISBN-13 978-3-642-12178-4 Springer Berlin Heidelberg New York

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,
in its current version, and permission for use must always be obtained from Springer. Violations are liable
to prosecution under the German Copyright Law.
springer.com
© Springer-Verlag Berlin Heidelberg 2010
Printed in Germany
Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India
Printed on acid-free paper 06/3180
 Preface

These multiple volumes (LNCS volumes 6016, 6017, 6018 and 6019) consist of
the peer-reviewed papers from the 2010 International Conference on Computa-
tional Science and Its Applications (ICCSA2010) held in Fukuoka, Japan during
March 23–26, 2010. ICCSA 2010 was a successful event in the International Con-
ferences on Computational Science and Its Applications (ICCSA) conference se-
ries, previously held in Suwon, South Korea (2009), Perugia, Italy (2008), Kuala
Lumpur, Malaysia (2007), Glasgow, UK (2006), Singapore (2005), Assisi, Italy
(2004), Montreal, Canada (2003), and (as ICCS) Amsterdam, The Netherlands
(2002) and San Francisco, USA (2001).
Computational science is a main pillar of most of the present research, in-
dustrial and commercial activities and plays a unique role in exploiting ICT in-
novative technologies. The ICCSA conference series has been providing a venue
to researchers and industry practitioners to discuss new ideas, to share complex
problems and their solutions, and to shape new trends in computational science.
ICCSA 2010 was celebrated at the host university, Kyushu Sangyo Univer-
sity, Fukuoka, Japan, as part of the university’s 50th anniversary. We would like
to thank Kyushu Sangyo University for hosting ICCSA this year, and for in-
cluding this international event in their celebrations. Also for the first time this
year, ICCSA organized poster sessions that present on-going projects on various
aspects of computational sciences.
Apart from the general track, ICCSA 2010 also included 30 special sessions
and workshops in various areas of computational sciences, ranging from compu-
tational science technologies, to specific areas of computational sciences, such as
computer graphics and virtual reality. We would like to show our appreciation
to the workshops and special sessions Chairs and Co-chairs.
The success of the ICCSA conference series, in general, and ICCSA 2010, in
particular, was due to the support of many people: authors, presenters, partic-
ipants, keynote speakers, session Chairs, Organizing Committee members, stu-
dent volunteers, Program Committee members, Steering Committee members,
and people in other various roles. We would like to thank them all. We would
also like to thank Springer for their continuous support in publishing ICCSA
conference proceedings.

March 2010 Osvaldo Gervasi

David Taniar
 Organization

ICCSA 2010 was organized by the University of Perugia (Italy), Monash Univer-
sity (Australia), La Trobe University (Australia), University of Basilicata (Italy),
and Kyushu Sangyo University (Japan)

Honorary General Chairs

Takashi Sago Kyushu Sangyo University, Japan
Norio Shiratori Tohoku University, Japan
Kenneth C.J. Tan Qontix, UK

General Chairs
Bernady O. Apduhan Kyushu Sangyo University, Japan
Osvaldo Gervasi University of Perugia, Italy

Advisory Committee
Marina L. Gavrilova University of Calgary, Canada
Andrès Iglesias University of Cantabria, Spain
Tai-Hoon Kim Hannam University, Korea
Antonio Laganà University of Perugia, Italy
Katsuya Matsunaga Kyushu Sangyo University, Japan
Beniamino Murgante University of Basilicata, Italy
Kazuo Ushijima Kyushu Sangyo University, Japan (ret.)

Program Committee Chairs

Osvaldo Gervasi University of Perugia, Italy
David Taniar Monash University, Australia
Eric Pardede (Vice-Chair) LaTrobe University, Australia

Workshop and Session Organizing Chairs

Beniamino Murgante University of Basilicata, Italy
Eric Pardede LaTrobe University, Australia

Publicity Chairs
Jemal Abawajy Deakin University, Australia
Koji Okamura Kyushu Sangyo University, Japan
Yao Feng-Hui Tennessee State University, USA
Andrew Flahive DSTO, Australia
VIII Organization

International Liaison Chairs

Hiroaki Kikuchi Tokay University, Japan
Agustinus Borgy Waluyo Institute for InfoComm Research, Singapore
Takashi Naka Kyushu Sangyo University, Japan

Tutorial Chair
Andrès Iglesias University of Cantabria, Spain

Awards Chairs
Akiyo Miyazaki Kyushu Sangyo University, Japan
Wenny Rahayu LaTrobe University, Australia

Workshop Organizers
Application of ICT in Healthcare (AICTH 2010)
Salim Zabir France Telecom /Orange Labs Japan
Jemal Abawajy Deakin University, Australia

Approaches or Methods of Security Engineering (AMSE 2010)

Tai-hoon Kim Hannam University, Korea

Advances in Web-Based Learning (AWBL 2010)

Mustafa Murat Inceoglu Ege University (Turkey)

Brain Informatics and Its Applications (BIA 2010)

Heui Seok Lim Korea University, Korea
Kichun Nam Korea University, Korea

Computer Algebra Systems and Applications (CASA 2010)

Andrès Iglesias University of Cantabria, Spain
Akemi Galvez University of Cantabria, Spain

Computational Geometry and Applications (CGA 2010)

Marina L. Gavrilova University of Calgary, Canada

Computer Graphics and Virtual Reality (CGVR 2010)

Osvaldo Gervasi University of Perugia, Italy
Andrès Iglesias University of Cantabria, Spain
 Organization IX

Chemistry and Materials Sciences and Technologies (CMST 2010)

Antonio Laganà University of Perugia, Italy

Future Information System Technologies and Applications (FISTA

2010)
Bernady O. Apduhan Kyushu Sangyo University, Japan
Jianhua Ma Hosei University, Japan
Qun Jin Waseda University, Japan

Geographical Analysis, Urban Modeling, Spatial Statistics

(GEOG-AN-MOD 2010)
Stefania Bertazzon University of Calgary, Canada
Giuseppe Borruso University of Trieste, Italy
Beniamino Murgante University of Basilicata, Italy

Graph Mining and Its Applications (GMIA 2010)

Honghua Dai Deakin University, Australia
James Liu Hong Kong Polytechnic University, Hong Kong
Min Yao Zhejiang University, China
Zhihai Wang Beijing JiaoTong University, China

High-Performance Computing and Information Visualization

(HPCIV 2010)
Frank Dévai London South Bank University, UK
David Protheroe London South Bank University, UK

International Workshop on Biomathematics, Bioinformatics and

Biostatistics (IBBB 2010)
Unal Ufuktepe Izmir University of Economics, Turkey
Andres Iglesias University of Cantabria, Spain

International Workshop on Collective Evolutionary Systems (IWCES

2010)
Alfredo Milani University of Perugia, Italy
Clement Leung Hong Kong Baptist University, Hong Kong

International Workshop on Human and Information Space Symbiosis

(WHISS 2010)
Takuo Suganuma Tohoku University, Japan
Gen Kitagata Tohoku University, Japan
X Organization

Mobile Communications (MC 2010)

Hyunseung Choo Sungkyunkwan University, Korea

Mobile Sensor and Its Applications (MSIA 2010)

Moonseong Kim Michigan State University, USA

Numerical Methods and Modeling/Simulations in Computational

Science and Engineering (NMMS 2010)
Elise de Doncker Western Michigan University, USA
Karlis Kaugars Western Michigan University, USA

Logical, Scientific and Computational Aspects of Pulse Phenomena

in Transitions (PULSES 2010)
Carlo Cattani University of Salerno, Italy
Cristian Toma Corner Soft Technologies, Romania
Ming Li East China Normal University, China

Resource Management and Scheduling for Future-Generation

Computing Systems (RMS 2010)
Jemal H. Abawajy Deakin University, Australia

Information Retrieval, Security and Innovative Applications (RSIA

2010)
Mohammad Mesbah Usddin Kyushy University, Japan

Rough and Soft Sets Theories and Applications (RSSA 2010)

Mustafa Mat Deris Universiti Tun Hussein Onn, Malaysia
Jemal H. Abawajy Deakin University, Australia

Software Engineering Processes and Applications (SEPA 2010)

Sanjay Misra Atilim University, Turkey

Tools and Techniques in Software Development Processes (TTSDP

2010)
Sanjay Misra Atilim University, Turkey

Ubiquitous Web Systems and Intelligence (UWSI 2010)

David Taniar Monash University, Australia
Eric Pardede La Trobe University, Australia
Wenny Rahayu La Trobe University, Australia
 Organization XI

Wireless and Ad-Hoc Networking (WADNet 2010)

Jongchan Lee Kunsan National University, Korea
Sangjoon Park Kunsan National University, Korea

WEB 2.0 and Social Networks (Web2.0 2010)

Vidyasagar Potdar Curtin University of Technology, Australia

Workshop on Internet Communication Security (WICS 2010)

José Maria Sierra Camara University of Madrid, Spain

Wireless Multimedia Sensor Networks (WMSN 2010)

Vidyasagar Potdar Curtin University of Technology, Australia
Yan Yang Seikei University, Japan

Program Committee
Kenneth Adamson Ulster University, UK
Margarita Albertı́ Wirsing Universitat de Barcelona, Spain
Richard Barrett Oak Ridge National Laboratory, USA
Stefania Bertazzon University of Calgary, Canada
Michela Bertolotto University College Dublin, Ireland
Sandro Bimonte CEMAGREF, TSCF, France
Rod Blais University of Calgary, Canada
Ivan Blecic University of Sassari, Italy
Giuseppe Borruso Università degli Studi di Trieste, Italy
Martin Buecker Aachen University, Germany
Alfredo Buttari CNRS-IRIT, France
Carlo Cattani University of Salerno, Italy
Alexander Chemeris National Technical University of Ukraine
“KPI”, Ukraine
Chen-Mou Cheng National Taiwan University, Taiwan
Min Young Chung Sungkyunkwan University, Korea
Rosa Coluzzi National Research Council, Italy
Stefano Cozzini National Research Council, Italy
José A. Cardoso e Cunha Univ. Nova de Lisboa, Portugal
Gianluca Cuomo University of Basilicata, Italy
Alfredo Cuzzocrea University of Calabria, Italy
Ovidiu Daescu University of Texas at Dallas, USA
Maria Danese University of Basilicata, Italy
Pravesh Debba CSIR, South Africa
Oscar Delgado-Mohatar University Carlos III of Madrid, Spain
Roberto De Lotto University of Pavia, Italy
XII Organization

Jean-Cristophe Desplat Irish Centre for High-End Computing, Ireland

Frank Dévai London South Bank University, UK
Rodolphe Devillers Memorial University of Newfoundland, Canada
Pasquale Di Donato Sapienza University of Rome, Italy
Carla Dal Sasso Freitas UFRGS, Brazil
Francesco Gabellone National Research Council, Italy
Akemi Galvez University of Cantabria, Spain
Marina Gavrilova University of Calgary, Canada
Nicoletta Gazzea ICRAM, Italy
Jerome Gensel LSR-IMAG, France
Andrzej M. Goscinski Deakin University, Australia
Alex Hagen-Zanker Cambridge University, UK
Muki Haklay University College London, UK
Hisamoto Hiyoshi Gunma University, Japan
Choong Seon Hong Kyung Hee University, Korea
Fermin Huarte University of Barcelona, Spain
Andrès Iglesias University of Cantabria, Spain
Antonio Laganà University of Perugia, Italy
Mustafa Murat Inceoglu Ege University, Turkey
Ken-ichi Ishida Kyushu Sangyo University, Japan
Antonio Izquierdo Universidad Carlos III de Madrid, Spain
Daesik Jang Kunsan University, Korea
Peter Jimack University of Leeds, UK
Korhan Karabulut Yasar University, Turkey
Farid Karimipour Vienna University of Technology, Austria
Baris Kazar Oracle Corp., USA
Dong Seong Kim Duke University, USA
Pan Koo Kim Chosun University, Korea
Ivana Kolingerova University of West Bohemia, Czech Republic
Dieter Kranzlmueller Ludwig Maximilians University and Leibniz
Supercomputing Centre Munich, Germany
Domenico Labbate University of Basilicata, Italy
Rosa Lasaponara National Research Council, Italy
Maurizio Lazzari National Research Council, Italy
Xuan Hung Le University of South Florida, USA
Sangyoun Lee Yonsei University, Korea
Bogdan Lesyng Warsaw University, Poland
Clement Leung Hong Kong Baptist University, Hong Kong
Chendong Li University of Connecticut, USA
Laurence Liew Platform Computing, Singapore
Xin Liu University of Calgary, Canada
Cherry Liu Fang U.S. DOE Ames Laboratory, USA
Savino Longo University of Bari, Italy
Tinghuai Ma NanJing University of Information Science
and Technology, China
Antonino Marvuglia University College Cork, Ireland
 Organization XIII

Michael Mascagni Florida State University, USA

Nikolai Medvedev Institute of Chemical Kinetics and Combustion
SB RAS, Russia
Nirvana Meratnia University of Twente, The Netherlands
Alfredo Milani University of Perugia, Italy
Sanjay Misra Atilim University, Turkey
Asish Mukhopadhyay University of Windsor, Canada
Beniamino Murgante University of Basilicata, Italy
Takashi Naka Kyushu Sangyo University, Japan
Jiri Nedoma Academy of Sciences of the Czech Republic,
Czech Republic
Laszlo Neumann University of GironAlexey Rodionova, Spain
Belen Palop Universidad de Valladolid, Spain
Dimos N. Pantazis Technological Educational Institution of
Athens, Greece
Luca Paolino Università di Salerno, Italy
Marcin Paprzycki Polish Academy of Sciences, Poland
Gyung-Leen Park Cheju National University, Korea
Kwangjin Park Wonkwang University, Korea
Paola Perchinunno University of Bari, Italy
Carlo Petrongolo University of Siena, Italy
Antonino Polimeno University of Padova, Italy
Jacynthe Pouliot Université Laval, France
David C. Prosperi Florida Atlantic University, USA
Dave Protheroe London South Bank University, UK
Richard Ramaroso Harvard University, USA
Jerzy Respondek Silesian University of Technology, Poland
Alexey Rodionov Institute of Computational
Mathematics and Mathematical Geophysics,
Russia
Jon Rokne University of Calgary, Canada
Octavio Roncero CSIC, Spain
Maytham Safar Kuwait University, Kuwait
Haiduke Sarafian The Pennsylvania State University, USA
Bianca Schön University College Dublin, Ireland
Qi Shi Liverpool John Moores University, UK
Dale Shires U.S. Army Research Laboratory, USA
Olga Sourina Nanyang Technological University, Singapore
Henning Sten Copenhagen Institute of Technology, Denmark
Kokichi Sugihara Meiji University, Japan
Francesco Tarantelli University of Perugia, Italy
Jesús Téllez Universidad Carlos III de Madrid, Spain
Parimala Thulasiraman University of Manitoba, Canada
Giuseppe A. Trunfio University of Sassari, Italy
Mario Valle Swiss National Supercomputing Centre,
Switzerland
XIV Organization

Pablo Vanegas Katholieke Universiteit Leuven, Belgium

Piero Giorgio Verdini INFN Pisa and CERN, Italy
Andrea Vittadini University of Padova, Italy
Koichi Wada University of Tsukuba, Japan
Krzysztof Walkowiak Wroclaw University of Technology, Poland
Jerzy Wasniewski Technical University of Denmark, Denmark
Robert Weibel University of Zurich, Switzerland
Roland Wismüller Universität Siegen, Germany
Markus Wolff University of Potsdam, Germany
Kwai Wong University of Tennessee, USA
Mudasser Wyne National University, USA
Chung-Huang Yang National Kaohsiung Normal University, Taiwan
Albert Y. Zomaya University of Sydney, Australia

Sponsoring Organizations
ICCSA 2010 would not have been possible without the tremendous support of
many organizations and institutions, for which all organizers and participants of
ICCSA 2010 express their sincere gratitude:
University of Perugia, Italy
Kyushu Sangyo University, Japan
Monash University, Australia
La Trobe University, Australia
University of Basilicata, Italia
Information Processing Society of Japan (IPSJ) - Kyushu Chapter
and with IPSJ SIG-DPS
 Table of Contents – Part III

Workshop on Mobile Communications (MC 2010)

A Control Loop Reduction Scheme for Wireless Process Control on
Traffic Light Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Junghoon Lee, Gyung-Leen Park, In-Hye Shin, Choel Min Kim, and
Sang-Wook Kim

Performance Measurement of a Dual-Channel Intersection Switch on

the Vehicular Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Junghoon Lee, Gyung-Leen Park, In-Hye Shin, Ji-Ae Kang,
Min-Jae Kang, and Ho-Young Kwak

A Rapid Code Acquisition Scheme for Optical CDMA Systems . . . . . . . . 21

Youngyoon Lee, Dahae Chong, Chonghan Song, Youngpo Lee,
Seung Goo Kang, and Seokho Yoon

Optimal and Suboptimal Synchronization Schemes for Ultra-Wideband

Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Dahae Chong, Chonghan Song, Youngpo Lee, Myungsoo Lee,
Junhwan Kim, and Seokho Yoon

Partial Information Relaying with Multi-Layered Superposition

Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
Jingyu Kim and Dong In Kim

Performance Measurement of the Hybrid Prefetch Scheme on Vehicular

Telematics Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Junghoon Lee, Gyung-Leen Park, Youngshin Hong,
In-Hye Shin, and Sang Joon Lee

Power Control for Soft Fractional Frequency Reuse in OFDMA

System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Young Min Kwon, Ok Kyung Lee, Ju Yong Lee, and
Min Young Chung

Authentication – Based Medium Access Control to Prevent Protocol

Jamming: A-MAC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Jaemin Jeung, Seungmyeong Jeong, and Jaesung Lim

Femtocell Deployment to Minimize Performance Degradation in Mobile

WiMAX Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Chang Seup Kim, Bum-Gon Choi, Ju Yong Lee, Tae-Jin Lee,
Hyunseung Choo, and Min Young Chung
XVI Table of Contents – Part III

A Novel Frequency Planning for Femtocells in OFDMA-Based Cellular

Networks Using Fractional Frequency Reuse . . . . . . . . . . . . . . . . . . . . . . . . . 96
Chang-Yeong Oh, Min Young Chung, Hyunseung Choo, and
Tae-Jin Lee

Association Analysis of Location Tracking Data for Various Telematics

Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
In-Hye Shin and Gyung-Leen Park

An Efficient ICI Cancellation Method for Cooperative STBC-OFDM

Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
Kyunghoon Won, Jun-Hee Jang, Se-bin Im, and Hyung-Jin Choi

Low-Cost Two-Hop Anchor Node-Based Distributed Range-Free

Localization in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Taeyoung Kim, Minhan Shon, Wook Choi, MoonBae Song, and
Hyunseung Choo

SecDEACH: Secure and Resilient Dynamic Clustering Protocol

Preserving Data Privacy in WSNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Young-Ju Han, Min-Woo Park, and Tai-Myoung Chung

Avoidance of Co-channel Interference Using Switched Parasitic Array

Antenna in Femtocell Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Yeonjune Jeong, Hyunduk Kim, Byung-Sung Kim, and
Hyunseung Choo

User Policy Based Transmission Control Method in Cognitive Wireless

Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
Noriki Uchida, Yoshitaka Shibata, and Kazuo Takahata

Workshop on Rough and Soft Sets Theories and

Applications (RSSA 2010)
Development of a Hybrid Case-Based Reasoning for Bankruptcy
Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
Rong-Ho Lin and Chun-Ling Chuang

The Time Series Image Analysis of the HeLa Cell Using Viscous Fluid
Registration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
Soichiro Tokuhisa and Kunihiko Kaneko

Matrices Representation of Multi Soft-Sets and Its Application . . . . . . . . 201

Tutut Herawan, Mustafa Mat Deris, and Jemal H. Abawajy

Clustering Analysis of Water Quality for Canals in Bangkok,

Thailand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Sirilak Areerachakul and Siripun Sanguansintukul
 Table of Contents – Part III XVII

Workshop on Wireless and Ad Hoc Networking

(WADNet 2010)
A Review of Routing Protocols for UWB MANETs . . . . . . . . . . . . . . . . . . 228
Yahia Hasan Jazyah and Martin Hope

An Efficient and Reliable Routing Protocol for Wireless Mesh

Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
Jaydip Sen

A Context-Aware Service Model Based on Workflows for

u-Agriculture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Yongyun Cho, Jongbae Moon, and Hyun Yoe

A History-Based Scheduler for Dynamic Load Balancing on Distributed

VOD Server Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Jongbae Moon, Hyun-joo Moon, and Yongyun Cho

A Secure Routing Protocol for Wireless Sensor Networks . . . . . . . . . . . . . . 277

Jaydip Sen and Arijit Ukil

Efficient Pairwise Key Establishment Scheme Based on Random

Pre-distribution Keys in WSN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
Hao Wang, Jian Yang, Ping Wang, and Pu Tu

Agent Based Approach of Routing Protocol Minimizing the Number of

Hops and Maintaining Connectivity of Mobile Terminals Which Move
One Area to the Other . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Kohei Arai and Lipur Sugiyanta

Ensuring Basic Security and Preventing Replay Attack in a Query

Processing Application Domain in WSN . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Amrita Ghosal, Subir Halder, Sanjib Sur, Avishek Dan, and
Sipra DasBit

Workshop on Wireless Multimedia Sensor Networks

(WMSN 2010)
A Review of Redundancy Elimination Protocols for Wireless Sensor
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
Babak Pazand, Amitava Datta, and Rachel Cardell-Oliver

A Light-Weighted Misused Key Detection in Wireless Sensor

Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
Young-Ju Han, Min-Woo Park, Jong-Myoung Kim, and
Tai-Myoung Chung
XVIII Table of Contents – Part III

Identifying Mudslide Area and Obtaining Forewarned Time Using AMI

Associated Sensor Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
Cheng-Jen Tang and Miau Ru Dai

General Track on Information Systems and

Information Technologies
Utilization of Ontology in Health for Archetypes Constraint
Enforcement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
Anny Kartika Sari, Wenny Rahayu, and Dennis Wollersheim

Time-Decaying Bloom Filters for Efficient Middle-Tier Data

Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
Kai Cheng

Soft Decision Making for Patients Suspected Influenza . . . . . . . . . . . . . . . . 405

Tutut Herawan and Mustafa Mat Deris
Personal Identification by EEG Using ICA and Neural Network . . . . . . . . 419
Preecha Tangkraingkij, Chidchanok Lursinsap,
Siripun Sanguansintukul, and Tayard Desudchit
A Formal Concept Analysis-Based Domain-Specific Thesaurus and Its
Application in Document Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
Jihn-Chang Jehng, Shihchieh Chou, and Chin-Yi Cheng
On the Configuration of the Similarity Search Data Structure D-Index
for High Dimensional Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
Arnoldo José Müller-Molina and Takeshi Shinohara
Automatic Chinese Text Classification Using N-Gram Model . . . . . . . . . . 458
Show-Jane Yen, Yue-Shi Lee, Yu-Chieh Wu, Jia-Ching Ying, and
Vincent S. Tseng
Genetic Algorithms Evolving Quasigroups with Good Pseudorandom
Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
Václav Snášel, Jiřı́ Dvorský, Eliška Ochodková, Pavel Krömer,
Jan Platoš, and Ajith Abraham

Software Openness: Evaluating Parameters of Parametric Modeling

Tools to Support Creativity and Multidisciplinary Design Integration . . . 483
Flora Dilys Salim and Jane Burry

Dynamic and Cyclic Response Simulation of Shape Memory Alloy

Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 498
Yutaka Toi and Jie He

Adaptive Fuzzy Filter for Speech Enhancement . . . . . . . . . . . . . . . . . . . . . . 511

Chih-Chia Yao and Ming-Hsun Tsai
 Table of Contents – Part III XIX

Risk Prediction for Postoperative Morbidity of Endovascular Aneurysm

Repair Using Ensemble Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
Nan-Chen Hsieh, Chien-Hui Chan, and Hsin-Che Tsai

Further Results on Swarms Solving Graph Coloring . . . . . . . . . . . . . . . . . . 541

Manuel Graña, Blanca Cases, Carmen Hernandez, and
Alicia D’Anjou

Data Collection System for the Navigation of Wheelchair Users: A

Preliminary Report . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552
Yasuaki Sumida, Kazuaki Goshi, and Katsuya Matsunaga

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565

 A Control Loop Reduction Scheme for Wireless
Process Control on Traffic Light Networks

Junghoon Lee1 , Gyung-Leen Park1 , In-Hye Shin1 ,

Choel Min Kim2, , and Sang-Wook Kim3
1
Dept. of Computer Science and Statistics, Jeju National University
2
Dept. of Computer Education, Jeju National University
3
College of Information and Communications, Hanyang University
{jhlee,glpark,ihshin76,chkim}@jejunu.ac.kr, [email protected]

Abstract. This paper designs a loop reduction scheme and measures its
performance for the wireless process control application running on the
grid-style traffic light network, aiming at improving the response time of
the control system. Based on the WirelessHART standard, which assigns
a slot to each (sender, receiver) pair according to the routing schedule,
the allocation scheme puts all the pairs having no interference to a single
slot, reducing the loop length. For further reduction, the classic Dijkstra’s
shortest path algorithm is modified such that the number of end-to-end
paths starting from the horizontal links and the vertical link, respec-
tively, is almost same. The transmission of the controller, which initiates
all message delivery and is the main bottleneck point, will not be blocked.
The simulation result demonstrates that the proposed scheme can sig-
nificantly reduce the control loop length, and the modified path finding
algorithm further achieves about 9.8 % improvement, just sacrificing the
3.2 % of transmission success ratio.

1 Introduction

Vehicular telematics networks can keep the vehicle connected to a network even
while it is on the move. This network exploits matured wireless communication
technologies and essentially includes static elements which provide the access
point to moving vehicles. Correspondingly, each component communicates in
two-level hierarchy in vehicular networks [1]. Level 1 corresponds to connection
between static nodes and level 2 between a static node and moving vehicles
trying to access the network. Level 1 communication creates a kind of wireless
mesh networks which have been intensively researched and widely deployed into
diverse field areas, exploiting commonly available protocols such as IEEE 802.11,
Zigbee, and the like [2]. Based on this network, many vehicular applications can
be developed and serviced.

This research was supported by the MKE, Korea, under the ITRC support program
supervised by the NIPA. (NIPA-2009-C1090-0902-0040)).

Corresponding author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 1–10, 2010.

c Springer-Verlag Berlin Heidelberg 2010
2 J. Lee et al.

Traffic lights are found in every street, and they can desirably install a wire-
less communication interface as they have sufficient power provision and their
locations are highly secure. On such a traffic light network, it is possible to im-
plement a monitor-and-control application when the network includes sensors
and actuators [3]. Practically, a lot of traffic-related or environmental devices
such as speed detectors, pollution meters, and traffic signal controllers, can be
covered by the vehicular network. Moreover, vehicles can also carry a sensor
device and report the collected sensor data to a static node when connected.
The message of process control applications is necessarily time critical and how
to schedule messages is the main concern. It depends on the node distribution,
namely, topology, and the communication protocol. First, traffic lights are placed
in each intersection of Manhattan-style road networks, so the traffic light network
has grid topology. Additionally, the network protocol must provide predictable
network access.
The WirelessHART standard provides a robust wireless protocol for various
process control applications [4]. First of all, the protocol standard exploits the
slot-based access scheme to guarantee a predictable message delivery, and each
slot is assigned to the appropriate (sender, receiver) pair. In addition, for the sake
of overcoming the transient instability of wireless channels, a special emphasis is
put on reliability by mesh networking, channel hopping, and time-synchronized
messaging. This protocol has been implemented and is about to be released to
the market [5]. It can accommodate a split-merge operation to mask channel
errors [6] as well as an efficient routing scheme to find the path that is most
likely to successfully deliver the message [7].
For the timely control action, the network must deliver the sensor and control
message not only reliably but also timely, or as fast as possible. Thus, every
message transmission schedule, or slot assignment, is decided in priori. Moreover,
to speed up the response time of a control process, the length of the slot schedule
must be minimized. In the mesh network, more than one transmission can be
accomplished simultaneously as long as transmitters are far way enough not
to interfere each other. If the node is equipped with an directional antenna,
more transmissions can be overlapped. However, the slot assignment permitting
multiple transmissions is a NP hard problem, which has no polynomial time
solution. In this regard, this paper is to design and evaluate the performance of
a slot allocation scheme to the the sensor and control messages for the grid-style
traffic light control network.
The paper is organized as follows: After defining the problem in Section 1,
Section 2 introduces the background of this paper focusing on the WirelessHART
protocol and slot allocation schemes. Section 3 designs control loop reduction
schemes on the target network. The simulation result is discussed in Section 4,
and finally Section 5 summarizes and concludes this paper.

2 Background and Related Work

The WirelessHART standard is defined over the IEEE 802.15.4 GHz radioband
physical link, allowing up to 16 frequency channels spaced by 5 MHz guard
 A Control Loop Reduction Scheme for Wireless Process Control 3

band [8]. The link layer provides deterministic slot-based access on top of the
time synchronization primitives carried out continuously during the whole net-
work operation time. According to the specification, the size of a single time
slot is 10 ms, and a central controller node coordinates routing and communi-
cation schedules to meet the robustness requirement of industrial applications.
For more reliable communication, CCA (Clear Channel Assessment) [9] before
each transmission and channel blacklisting is exploited to avoid specific area of
interference and also to minimize interference to others. In some WirelessHART
implementation, the network manger traverses a communication graph or a grid
by the breadth-first search and allocates slots according to this order [8].
In each slot, a sender can try another channel if the CCA result of the channel
on the primary schedule is not clear, even though such an operation is not yet
defined in the current standard. For each destination node, a controller may
reserve two alternative paths having sufficient number of common nodes. Two
paths split at some nodes and meet again at other nodes. When two paths
split, a node can select the path according to the CCA result in a single slot
by switching to the channel associated with the secondary route. When two
paths merge, the node can receive from two possible senders by a timed switch
operation. Besides, the WirelessHART protocol can be reinforced by many useful
performance enhancement schemes such as the virtual-link routing scheme that
combines split-merge links and estimates the corresponding error rate to apply
the shortest path algorithm.
Wireless mesh networks are cost-effective solutions for ubiquitous high-speed
services, and its performance depends on the routing strategy. Routing has been
extensively studied in wireless mesh networks, and most routing schemes are
based on the shortest path algorithm [10]. The path cost is different according
to the main goal of each routing scheme, including actual distance, transmis-
sion rate, and error characteristics. In most wireless process control systems, the
central routing mechanism is generally taken, as not only the traffic character-
istics are accurately known, but also each network access is predictable. Hence,
it is possible to jointly allocate channels, route a message set, and schedule each
transmission [11]. In addition, the routing procedure is executed repeatedly ac-
cording to the link condition change and the schedule is distributed to each
node [12].

3 Routing and Scheduling Scheme

3.1 System Model
The traffic lights form a grid network in modern cities, as a traffic light node
is placed at each crossing of the Manhattan-style road network, as shown in
Figure 1(a) [13]. Each node can exchange messages directly with its vertical and
horizontal neighbors. Two nodes in the diagonal of a rectangle do not have a
direct connection, as there may be obstacles like a tall building that blocks the
wireless signal propagation. In this network, the central controller is assumed to
be located at the fringe of a rectangular area, for this architecture makes the
4 J. Lee et al.

determination of the communication schedule simple and systematic. In Figure

1(a), T0,0 is the controller node. Any grid network can be transformed into this
network by partition and rotation. In the example of Figure 1(b), four 4 × 4
grids are generated and each of them can be mapped to a grid shown in Figure
1(a), regardless of the grid dimension.

T0,0 T0,1 T0,2

(0) (1) (2)

T1,0 T1,1 T1,2

(3) (4) (5)

T2,0 T2,1 T2,2

(6) (7) (8)

(a) Traffic light network (b) Grid partition

Fig. 1. Traffic light network

A control loop consists of three phases, namely, reading state variables from
sensors, deciding the control action, and sending the value of control variables
to the actuators. The first and third steps involve the network and have their
own communication schedules.

3.2 Loop Reduction

The allocation step basically begins with the well-known Dijkstra’s shortest path
algorithm which can find the optimal path from a node to the other nodes in
the grid given as a cost graph. In this graph, each link can be specified with
a cost value such as error rate, available network bandwidth, or pass time. As
this paper focuses on the reliability of each end-to-end path, we assume that the
slot error rate is given as a link cost. Each link has its own error characteristics
due to different power level, obstacle distribution, and so on. The change of link
error characteristics can be estimated in many ways [14], but we assume that the
probing result is always correct, as the correctness of channel probing is not our
concern. Figure 1(a) can be modeled as a cost graph for the 3 × 3 grid consist of
9 nodes from T0,0 to T2,2 and 12 links connecting each adjacent nodes. Node 0
and T0,0 point the same node, and this paper will selectively use an appropriate
notation.
Now, Dijkstra’s (one-to-many) shortest path algorithm can find the best route
having the lowest error rate. First, the error rate of each link can be replaced
by the success probability by subtracting it from 1.0. Actually, they can be used
interchangeably. Then, each iteration compares the product of success probabil-
ities of a path extended by a expansion node, instead of the sum of link costs as
 A Control Loop Reduction Scheme for Wireless Process Control 5

in the original algorithm. In each node scan, the node having the highest success
probability will be selected. The path from the controller, T0,0 , to each node is
shown in Figure 2(a). The number in the parenthesis means the hop length of
the path. Each column is a time slot and each slot has a (sender, receiver) pair.
As each node has just one transmitter and one receiver module, neither the same
sender nor the same receiver can appear twice in the same column. So, the initial
allocation should be modified to avoid such a conflict.
To this end, the scheduler checks if two entries interfere. They interfere if 1)
two senders are the same, or 2) two receivers are the same, or 3) the sender of
one slot and the receiver of the other slot are the same. This criteria is valid
just for the directional antenna which can increase spatial reuse and reduce
interference by directing radio beam towards a desired direction [15]. When
the node transmits using an omnidirectional antenna, which radiates or receives
electromagnetic energy in all directions, the probability of interference increases,
as the two transmissions will collide when the two senders are less than two hops
away from each other. If two transmissions are not compatible in the same slot,
the one with a shorter path will be moved one slot back, as the longer path
is more critical to the whole control loop length. Such a move-back procedure
is repeated until there is no change. Figure 2 (b) show the final result. In this
figure, 0-3 means T0,0 → T1,0 .

Slot 1 2 3 4 Slot 1 2 3 4 5 6 7 8 9 10 11
8 (4) 0−3 3−6 6−7 7−8 8 (4) 0−3 3−6 6−7 7−8
5 (3) 0−3 3−4 4−5 5 (3) 0−3 3−4 4−5
7 (3) 0−3 3−6 6−7 7 (3) 0−3 3−6 6−7
2 (2) 0−1 1−2 2 (2) 0−1 1−2
4 (2) 0−3 3−4 4 (2) 0−3 3−4
6 (2) 0−3 3−6 6 (2) 0−3 3−6
1 (1) 0−1 1 (1) 0−1
3 (1) 0−3 3 (1) 0−3

(a) Initial setting (b) Final allocation

Fig. 2. Traffic light network

3.3 Routing Scheme

In the control loop scenario, traffic goes from and to T0,0 . Namely, each node
sends and receives a message to and from the controller node once in the con-
trol period specified by the system requirement. Even if it is desirable to take
the route which has the minimum number of hops to the destination, another
detour can be advantageous in terms of delivery ratio and transmission delay.
In process control applications, transmission reliability is most important. How-
ever, in addition to the end-to-end performance, it is also necessary to reduce
the control message exchange time. In the example of Figure 2(b), it takes 11
slots to complete the delivery of all control messages to every node in a control
round.
6 J. Lee et al.

The control loop can be reduced by overlapping transmissions. As long as the

network coordinator sticks to the classic shortest path algorithm, it can hardly
expect more loop reduction. As shown in Figure 2 (a), each end-to-end route to a
node starts from either T0,0 → T0,1 or T0,0 → T1,0 for the 3 × 3 grid. If all routes
start from the same link, say, T0,0 → T0,1 the next slot cannot be overlapped
as two transmission will inevitably collide. The first one is from the controller
to T0,1 (the receiver of the newly starting transmission) and the second one is
from T0,1 (the sender of the message arrived in the previous slot) to some other
nodes. Table 1 describes this situation. For two destinations, namely, Dest 1 and
Dest 2, suppose that Dest 1 takes the route T0,0 → T0,1 → T0,2 while Dest 2
T0,0 → T0,1 → T1,1 , as show in Table 1.

Table 1. Example assignment

(a) Invalid assignment
t t+1 t+2
Dest 1 T0,0 → T0,1 T0,1 → T0,2
Dest 2 T0,0 → T0,1 T0,1 → T1,1
(b) Valid assignment
t t+1 t+2
Dest 1 T0,0 → T0,1 T0,1 → T0,2
Dest 2 T0,0 → T1,0 T1,0 → T1,1

As T0,1 appears twice in slot t + 1, one as a sender in the first row, and
the other as a receiver in the second row. On the contrary, if the first link for
Dest 2 is T0,0 → T1,0 as in Table 1(b), and two transmissions can be done
simultaneously on slot t + 1, reducing the control loop length. In case of Figure
2(a), 6 destinations takes T0,0 → T1,0 while just 2 destinations takes T0,0 →
T0,1 for the first step. So, the length reduction in not so significant. To make
the number of two first links as close as possible, the network can sacrifice the
optimality of the shortest path algorithm.
We classify grid nodes into 3 groups. The first group, GD , has the nodes on
the diagonal from top left to bottom right. The second group, DL , has the nodes
on the lower triangle of the grid, while the third group, DU , has the node on the
upper triangle. Accordingly, each group can be specified as follows:
GD : { Ti,j | i = j }
GL : { Ti,j | i > j }
GU : { Ti,j | i < j }
For nodes in GL , we make every route start from the link T0,0 → T1,0 , and for
those in GU , from the link T0,0 → T0,1 . GD nodes can take the route according to
the normal Dijkstra’s algorithm. For GL nodes, after setting the slot error rate
of T0,0 → T0,1 to 1.0, the scheduler runs the shortest path algorithm, and vice
versa for the GU nodes. Likewise, the number of appearances for T0,0 → T0,1 and
 A Control Loop Reduction Scheme for Wireless Process Control 7

T0,0 → T1,0 becomes almost even. While this allocation cannot find the optimal
path to each node, it can maximize the simultaneous transmissions in each slot
and reduce the control loop length.

4 Performance Measurement

This section measures the performance of the proposed loop reduction scheme
via simulation using SMPL which provides abundant functions and libraries
for discrete event scheduling [16]. Only the downlink graph was considered for
simplicity, as uplink and downlink communications are symmetric and calculated
completely in the same way. For each parameter setting, 500 sets of link error
rates are generated and the success ratio and the loop length ratio are averaged.
The first experiment measures the performance of our loop reduction scheme
and compares with the upper and lower bounds. This experiment is conducted
for the classic Dijkstra’s algorithm. First, the lower bound accounts for the case
that each end-to-end transmission can start one slot after another without be-
ing blocked from the largest hop-length destination. In this case, the control loop
length is ideally smallest and thus the lower bound can be calculated as in Eq. (1).

n × n − 1 + min(hop count), (1)

where n × n − 1 is the number of noncontroller nodes in the grid, and for the
last message initiated by the controller, it takes as many slots as the hop counts
to the destination.
Next, the upper bound of the control loop length is estimated by assuming
that there is no overlapped transmission. Hence, it corresponds to the sum of
total hops to each node. For the n × n grid, the number of total transmissions
for a control round is calculated as in Eq. (2).


2(n−1)
(n − |n − i − 1|) · i, (2)
i=1

where each iteration indexed by i corresponds to the hop count from the con-
troller. Nodes having the same hop counts need the same number of slots. The
number of nodes having i hops increases by one until i reaches (n-1), and then
decreases also by one until i reaches 2(n-1).
Figure 3 shows the comparison results according to the grid dimension and
the slot error rate. First, in Figure 3(a), the average slot rate is 0.1, distributing
exponentially, while the grid dimension ranges from 3 to 15. The curves plots
the ratio of each loop length to the upper bound. Hence, the curve marked
as UpperBound always remain 1.0. The ratio of the packet length to the upper
bound is just 5.7 % larger at maximum, compared with the lower bound to upper
bound ratio. When the dimension is 3, both ratios are around 0.58, as nodes are
not sufficiently apart from each other and not so many transmissions can be
overlapped. The gap between 2 curves reaches 5.7 % at the grid dimension of 10,
8 J. Lee et al.

but decreases afterwards. It indicates that our scheme can achieve almost ideal
loop reduction without complex and time-consuming space search iterations. In
addition, Figure 3(b) shows the effect of the slot error rate, which can lead to
the variation in the length of each end-to-end path. Here, the grid dimension
is set to 5. Even though the gap between the lower bound and our packed
scheme gets smaller according to the increase of the slot error rate, its effect is
insignificant.

1 1
"UpperBound" "UpperBound"
0.9 "LowerBound" "LowerBound"
0.8 "Packed" 0.8 "Packed"
Loop length ratio

Loop length ratio

0.7
0.6 0.6
0.5
0.4 0.4
0.3
0.2 0.2
0.1
0 0
4 6 8 10 12 14 0 0.1 0.2 0.3 0.4 0.5
Square grid dimension Slot error rate

(a) Effect of grid dimension (b) Effect of slot error rate

Fig. 3. Loop length reduction

Figure 4 demonstrates the performance of the proposed routing scheme, which

sacrifices a little bit the path optimality for loop reduction. Here, the average slot
error rate is again set to 1.0 and the dimension is changed from 3 to 15. Figure
4(a) measures how much the success ratio is lost. For all square grid dimension
range, the difference is less than 3.2 %. Figure 4(b) measures how much loop
reduction we gain. The improvement reaches 9.8 % at the dimension of 3 and
decreases up to 3.2 % when the dimension is 15.

1 1
"Grid" "Grid"
"Proposed" "Proposed"
0.9 0.8
Loop length ratio
Success ratio

0.8 0.6

0.7 0.4

0.6 0.2

0.5 0
4 6 8 10 12 14 4 6 8 10 12 14
Square grid dimension Square grid dimension

(a) Success ratio decrease (b) Loop length reduction

Fig. 4. Loop reduction vs. success ratio according to grid dimension

 A Control Loop Reduction Scheme for Wireless Process Control 9

Finally, Figure 5 shows the effect of the slot error rate in the proposed routing
scheme. In this experiment, the grid dimension is also set to 5, while the slot
error rate changes from 0 to 0.5. The success ratio gap increases along with
the error rate, reaching 4.4 % at maximum. However, the proposed scheme can
achieve consistently about 8 % loop reduction in all error ranges.

1 1
"Grid" "Grid"
"Proposed" "Proposed"
0.8 0.8

Loop length ratio

Success ratio

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Slot error rate Slot error rate

(a) Success ratio decrease (b) Loop length reduction

Fig. 5. Loop reduction vs. success ratio according to slot error rate

5 Conclusion
This paper has designed a loop reduction scheme and measured its performance
for the wireless process control application running on the grid-style traffic light
network, aiming at improving the response time of the control system. Based on
the WirelessHART standard, which assigns a slot to each (sender, receiver) pair
according to the routing schedule, the allocation scheme puts all the pairs having
no interference to a single slot, reducing the loop length. For further reduction,
the classic Dijkstra’s shortest path algorithm is modified such that the number
of end-to-end paths starting from the horizontal links and the vertical link,
respectively, is almost same. The transmission of the controller, which initiates
all message delivery and is the main bottleneck point, will not be blocked. The
simulation result demonstrates that the proposed scheme can significantly reduce
the control loop length, and the modified path finding algorithm further achieves
about 9.8 % improvement, just sacrificing the 3.2 % of transmission success ratio.
As for future work, a fault-tolerant flooding scheme is expected to be very
useful in wireless process control [17]. So, we will design a slot assignment scheme
combining the split-merge operation to overcome node or link failure for the
control message broadcast.

References
1. Bucciol, P., Li, F.Y., Fragoulis, N., Vandoni, L.: ADHOCSYS: Robust and service-
oriented wireless mesh networks to bridge the digital divide. In: IEEE Globecom
Workshops, pp. 1–5 (2007)
2. Gislason, D.: ZIGBEE Wireless Networking. Newnes (2008)
10 J. Lee et al.

3. IEC/PAS 62591: Industrial communication networks - Fieldbus specifications -

WirelessHART communication network and communication profile (2008)
4. Hart Communication Foundation, Why WirelessHARTT M ? The Right Standard
at the Right Time (2007), http://www.hartcomm2.org
5. Han, S., Song, J., Zhu, X., Mok, A.K., Chen, D., Nixon, M., Pratt, W., Gond-
halekar, V.: Wi-HTest: Compliance test suite for diagnosing devices in real-time
WirelessHART network. In: The 15th IEEE Real-Time and Embedded Technology
and Applications Symposium, pp. 327–336 (2009)
6. Lee, J., Shin, I., Kim, C.: Design of a reliable traffic control system on city area
based on a wireless network. In: Gervasi, O., Taniar, D., Murgante, B., Laganà,
A., Mun, Y., Gavrilova, M.L. (eds.) Computational Science and Its Applications –
ICCSA 2009. LNCS, vol. 5592, pp. 821–830. Springer, Heidelberg (2009)
7. Lee, J., Song, H., Mok, A.K.: Design of a reliable communication system for grid-
style traffic control networks. Submitted to IEEE RTAS 2010 (2010)
8. Song, S., Han, S., Mok, A.K., Chen, D., Nixon, M., Lucas, M., Pratt, W.: Wire-
lessHART: Applying wireless technology in real-time industrial process control. In:
The 14th IEEE Real-Time and Embedded Technology and Applications Sympo-
sium, pp. 377–386 (2008)
9. Ramchandran, I., Roy, S.: Clear channel assessment in energy-constrained wide-
band wireless networks. IEEE Wireless Magazine, 70–78 (2007)
10. Kodialam, M., Nandagopal, T.: Characterizing the capacity region in multi-radio
multi-channel wireless mesh networks. In: ACM MobiCom, pp. 73–87 (2005)
11. Wang, W., Liu, X., Krishnaswamy, D.: Robust routing and scheduling in wireless
mesh networks. In: IEEE Conference on Sensor, Mesh and Ad Hoc Communications
and Networks, pp. 471–480 (2007)
12. Zaidi, Z., Landfeldt, B.: Monitoring assisted robust routing in wireless mesh net-
works. In: Mobile Network Applications, pp. 54–66 (2008)
13. Jaap, S., Bechler, M., Wolf, L.: Evaluation of routing protocols for vehicular ad
hoc networks in city traffic scenarios. In: Proceedings of the 5th International
Conference on Intelligent Transportation Systems Telecommunications (2005)
14. Chang, N., Liu, M.: Optimal channel probing and transmission scheduling for op-
portunistic spectrum access. In: Proc. ACM International Conference on Mobile
Computing and Networking, pp. 27–38 (2007)
15. Dai, H., Ng, K., Wu, M.: An overview of MAC protocols with directional antennas
in wireless ad hoc networks. In: Proc. International Conference on Computing in
the Global Information Technology, pp. 84–91 (2006)
16. MacDougall, M.: Simulating Computer Systems: Techniques and Tools. MIT Press,
Cambridge (1987)
17. Wan, P., Huang, S., Wang, L., Wan, Z., Jia, X.: Minimum latency aggregation
scheduling in multihop wireless networks. In: MobiHoc, pp. 185–193 (2009)
 Performance Measurement of a Dual-Channel
Intersection Switch on the Vehicular Network

Junghoon Lee1 , Gyung-Leen Park1, , In-Hye Shin1 ,

Ji-Ae Kang1 , Min-Jae Kang2 , and Ho-Young Kwak3
1
Dept. of Computer Science and Statistics,
Jeju National University, 690-756, Jeju Do, Republic of Korea
2
Dept. of Electronic Engineering,
Jeju National University, 690-756, Jeju Do, Republic of Korea
3
Dept. of Computer Engineering,
Jeju National University, 690-756, Jeju Do, Republic of Korea
{jhlee,glpark,ihshin76,minjk,kwak}@jejunu.ac.kr, [email protected]

Abstract. This paper designs an intensively measures the performance

of an efficient message switch scheme for the intersection area where
routing decision may be complex due to traffic concentration, aiming
at enhancing end-to-end delays and reliability for the information re-
trieval service in the vehicular network. Installed at the corner of an
intersection, each switch node opens an external interface to exchange
messages with vehicles proceeding to the intersection as well as switches
the received messages via the dual-channel internal interfaces. Based on
the slot-based MAC, at each slot time of internal interfaces, the sender
node probes the channel status of the two different destinations and then
dynamically selects the appropriate channel according to the probing re-
sult. The simulation result shows that the proposed scheme improves the
delivery ratio by up to 18.5 % for the experimental channel error rate
range as well as up to 8.1 % for the given network load distribution.

1 Introduction
According to the ongoing deployment of in-vehicle telematics devices and ve-
hicular networks, wireless vehicular communications have become an important
priority for car manufactures[1]. The vehicular network can be built in many
different ways according to the variety of available wireless communication tech-
nologies, for example, infrastructure-based cellular networks, VANET (Vehicular
Ad hoc NETwork), and roadside networks. Moreover, not just the vehicle but
also diverse communication entities are participating in this network, includ-
ing traffic lights, roadside units, sensors, gas stations, and many other facilities.
Correspondingly, abundant application scenarios are now possible. In addition
to the standard safety-related applications such as traffic accident propagation

This research was supported by the MKE, Korea, under the ITRC support program
supervised by the NIPA. (NIPA-2009-C1090-0902-0040)).

Corresponding author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 11–20, 2010.

c Springer-Verlag Berlin Heidelberg 2010
12 J. Lee et al.

and emergency warning[2], information retrieval is very useful for drivers and
passengers[3]. Specifically, a driver may want to know the current traffic con-
dition of specific road segments lying on the way to his/her destination, query
several shops to decide where to go, and check parking lot availability[4].
In the information retrieval service, queries must be delivered to the destina-
tions and the result is sent back to the query issuer. A message proceeds to its
destination according to a routing protocol each vehicle cooperatively runs. As
each vehicle can move only along the road segment and static nodes such as gas
stations and traffic lights are generally placed on the roadside[5], the message de-
livery path must trail the actual road layout[6]. Just as in the vehicle’s path, in-
tersection areas having many vehicles are important for the delay and reliability
of message transmissions, because the large number of vehicles makes the routing
decision very complex and raises intervehicle interference. Moreover, the carry and
forward strategy is preferred to cope with the disconnection problem in the sparse
vehicular network[7]. As will be described in the next section, this behavior further
increases the complexity of message routing near the intersection area.
Definitely, there exists an intersection which has much more traffic than others
and this hot area may be a bottleneck for both vehicle and data traffic. As an end-
to-end path is highly likely to involve one or more such intersection areas, how
to manage the communication in this area is very critical to the communication
performance such as the transmission delay and the delivery ratio. In this regard,
this paper is to design and measure the performance of a reliable and high-speed
message switching scheme around the hot intersection area for the vehicular
network, taking advantage of multiple channels. It is possible to create multiple
channels in a cell in many available wireless protocols such as IEEE 802.11
series[2], Zigbee, and WirelessHart[8]. Existing researches have also pointed out
that multiple network interfaces does not cost too much[9]. Moreover, the slot-
based collision-free access scheme can be employed for predictable channel access,
short delay, and reliability, while the probing-based access can efficiently reduce
the effect of the channel error inherent in the wireless frequency channel.
This paper is organized as follows: After issuing the problem in Section 1,
Section 2 describes the background of this paper and related work. Section 3
designs a wireless switch for intersection areas and performance measurement
results are demonstrated and analyzed in Section 4. Finally, Section 5 summa-
rizes and concludes this paper with a brief introduction of future work.

2 Background and Related Work

Jeju taxi telematics system keeps track of each taxi for the purpose of providing
an efficient taxi dispatch service to customer[1]. To this end, each taxi periodi-
cally reports its GPS reading via the CDMA (Code Division Multiple Access) net-
work, an instance of cellular networks serviced in the Republic of Korea. This sys-
tem is rather an expensive vehicular network in the sense that every member taxi
should pay the monthly communication fee. In the mean time, in November 2005,
the U.S. Department of Transportation began to develop and test an integrated,
vehicle-based crash warning system under the program of IVBBS (Integrated
 Performance Measurement of a Dual-Channel Intersection Switch 13

Vehicle-Based Safety Systems)[10]. Here, data transmitted from the roadside to

the vehicle could warn a driver that it is not safe to enter an intersection. Vehicles
could serve as data collectors and anonymously transmit traffic and road condi-
tion information from every major road within the transportation network. After
all, more vehicular networks are being deployed into our everyday life.
As for the wireless channel, the IEEE 802.11b standard specifies 11 channels
operating in the 2.4 GHz band with 80 MHz of reusable spectrum[11]. Even
though the number of simultaneous channels in a cell is limited to 3 due to
the channel overlap problem, it is possible to create multiple channels from
the wireless spectrum in a cell. In addition, Zigbee and Wireless HART (High-
way Addressable Remote Transducer) support channels spaced by 5 MHz guard
band, making it possible for each node to hop over channels to reduce the effect
of channel errors[8]. The link layer is based on a TDMA (Time Division Multi-
ple Access) style access scheme which runs on top of the time synchronization
mechanism carried out continuously during the whole operation time by means
of MAC PDUs. The time axis is divided into 10 ms time slots and a group of
consecutive slots are defined to be a superframe. This access scheme can provide
bounded access time for each node.
High mobility and nonuniform distribution of vehicles prevent the existing
routing schemes from applying to the vehicular network. Quite a lot of researches
have been recently conducted for vehicular networks to deal with such problems.
Basically, the carry and forward strategy is preferred to cope with disconnection
in the sparse network part[7]. Here, when a node cannot find a receiver, it stores
the message in its buffer until it enters the range of a new receiver. Even though
this scheme increases the transmission delay, it is better not to discard a message.
This leads to a large traffic load around the intersection area. As a variant
of carry and forward scheme, VADD (Vehicle-Assisted Data Delivery) exploits
predictable vehicle mobility model in which vehicle movement is limited by the
traffic pattern and road layout. Based on the traffic pattern estimation, a vehicle
decides the best next node to forward a packet[12].
For the information retrieval method, VITP (Vehicular Information Transfer
Protocol) has been proposed as an application-layer communication protocol,
which is designed to support the establishment of a distributed, ad hoc ser-
vice infrastructure over VANET[3]. The VITP infrastructure can be used to
provide location-based, traffic-oriented services to drivers, using information re-
trieved from vehicular sensors and taking advantage of on-board GPS naviga-
tion systems. In this system, vehicular-service queries must be location-sensitive,
specifying explicitly the target location of their inquiry. VITP communication
entities follow a best-effort approach in their operation, while a VITP request
is transported between protocol peers until some return condition is satisfied.
The protocol peer that detects the upholding of the return condition, creates the
VITP reply and posts it toward source-region and broadcast so that the issuer
can receive even if it has moved. In the message exchange, many protocols can
be exploited for information retrieval such as Zhao et al.’s V2VR, which select
forward and backward proxies based on the mobility pattern of the vehicle[4].
14 J. Lee et al.

3 Wireless Switch Design

Fig. 1 shows the basic idea of this paper. At the intersection, four static switch
nodes from A to D are installed at each corner as shown in Fig. 1(a). The
number of switch nodes is equal to the number of branches, whether they cross
on the same plane or in different layers. Each switch node is bound to the
vehicles heading for the intersection from the preassigned branch. For the sake
of speeding up the transit time in the intersection area, each switch node opens
an external interface to receive and send messages with other vehicles, while
internally forwarding messages to an appropriate direction via the 2 additional
channels not exposed to the vehicle. As contrast to the external interface that
must follow the standard vehicular communication protocol, we can define a
new protocol for the internal access. For the internal message exchange, this
paper suggests that each node be connected to dual frequency channels which
run slot-based MAC as shown in Fig 1. (b).

time

Intersection Ch1 A B C D A .....

switch A D
Ch2 A B C D A .....

A−D1 RTS/CTS

Ch1 Data
B C
Ch2 Data A−D2 RTS/CTS

(a) configuration of switch nodes (b) time slot operation

Fig. 1. Basic concept

When a vehicle having messages to send or relay approaches to an intersec-

tion, it checks if it can reach a switch node. If so, the message is sent directly to
the switch node without contacting any other vehicles. Otherwise, the message
is just forwarded to a vehicle in its moving direction just like in the normal
ad hoc routing protocol. A switch node, receiving a message via the external
interface, decides the next switch node to forward the message in its internal
interface according to the final destination. In the intersection area, the complex
routing among vehicles is skipped, reducing the number of relays and erroneous
transmissions between moving vehicles. For this internal operation, each switch
node transmits according to the TDMA manner to provide the predictable ac-
cess. Even though this access demands that every node have the common clock,
current GPS technology can easily achieve the global clock synchronization of
reasonable accuracy[13]. So, many ad hoc networks also exploit the slot-based
access scheme.
 Performance Measurement of a Dual-Channel Intersection Switch 15

Many vehicles passing by the intersection area may cause channel interference,
so an efficient method is necessary to cope with this problem. This problem
can be solved by dual channels. The slots of two channels are synchronized,
and two slots of the same time instance must be assigned to the same node.
At the beginning of each slot, the transmitter node gets two messages having
different destinations from its message queue. Our previous work has proposed a
channel probing mechanism from two different nodes[14]. On the contrary, this
paper makes a single node be assgined time slots on each channel and select the
appropriate one. A channel probing result its corresponding action is similar in
both schemes.
Let these two destinations be D1 and D2 , respectively. The sender probes, for
example, with RTS/CTS methods, D1 and D2 on the two channels as shown in
Fig 1 (b). The transmission on the bad channel is meaningless not solely for the
power consumption aspect. It may extend the instable energy and its duration.
Table 1 shows the probing result and corresponding actions. As shown in row
1, D1 on channel 1 and also D2 on channel 2 are both in good state, D1 and
D2 send as scheduled. In row 2, every channel status is good except the one for
D2 on channel 2. If we switch < D1 , D2 > to < D2 , D1 >, both nodes can
successfully send their messages. Otherwise, only A can succeed. In this case,
the channel switch can save one transmission loss. Row 8 describes the situation
that D1 is good only on channel 2 while D2 also only on channel 1. Switching
two channels saves 2 transmissions that might fail on the regular schedule.

Table 1.Channel status and transmission

No. Ch1−D1 Ch2−D2 Ch1−D2 Ch2−D1 Ch1 Ch2 save

1 Good Good X X D1 D2 0
2 Good Bad Good Good D2 D1 1
3 Good Bad Good Bad D1 − 0
4 Good Bad Bad X D1 − 0
5 Bad Good Good Good D2 D1 1
6 Bad Good Good Bad − D2 0
7 Bad Good Bad X − D2 0
8 Bad Bad Good Good D2 D1 2
9 Bad Bad Good Bad D2 − 1
10 Bad Bad Bad Good − D1 1
11 Bad Bad Bad Bad − − 0
X : don’t care

Each switch node maintains a message queue. The buffer space can be assumed
to have no limitation, but it is desirable to discard a message which stayed in
the buffer too much so as to obviate undue delay for the subsequent messages.
After all, we assume that a message is automatically deleted from its queue
when its waiting time exceeds the given discard interval. In addition, the rout-
ing decision outside the intersection area can be exploited by other work and
16 J. Lee et al.

there are many available schemes especially targeting at the urban area[15]. For
example, VADD can cope with the situation that a vehicle moves after issuing
a retrieval request. Finally, how to assign slots is not our concern, and we just
consider the round-robin style allocation and assume that each slot is assigned
to a switch node. There also exist some channel allocation schemes for the give
traffic requirement[16].

4 Performance Measurement
We evaluate the performance of our scheme, focusing on the effect of channel
switches, via simulation using SMPL, which provides a simple and robust library
for discrete event trace similar to ns-2 event scheduler[17]. The main performance
metrics are the delivery ratio and the access delay. The delivery ratio measures
the ratio of the number of successfully delivered messages through the intersec-
tion area to the number of total messages arrived at the 4 switch nodes. The
access delay is the time interval from the instant a message arrives at a switch
node to the instant it gets out of the intersection switch. The average of access
delays just takes the messages that have been successfully transmitted, so it is
likely to increase when the number of successful transmissions increases. After
all, this section shows the effect of the channel error probability, offered load,
and the discard interval, one by one. In the following experiments, the message is
assumed to arrive at each switch node according to the exponential distribution.
In addition, for simplicity, each message fits a single time slot, making simple to
estimate the traffic load. Moreover, every time is aligned to a slot length.
The vehicle distribution is taken from the Jeju taxi telematics system which
keeps track of each member taxi[1]. Fig. 2 shows a snapshot of vehicle distri-
bution on the real road network. This history is very useful for the vehicular

Fig. 2. Vehicle distribution at intersections

 Performance Measurement of a Dual-Channel Intersection Switch 17

6 1
"Normal" "Normal"
"Proposed" "Proposed"
Access delay (slot time)

5.8 0.8

Delivery ratio
5.6 0.6

5.4 0.4

5.2 0.2

5 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Channel error probability Channel error probability

Fig. 3. Access delay at low load Fig. 4. Delivery ratio at low load

network design and the traffic pattern analysis. In this figure, the intersection is
represented by a rectangle and the road segment is by a line, while the location
of taxis during the specific time interval is marked with a small dot. The intersec-
tion area has more dots, which appear just along the road segment, even though
the number of dots is different for each segment which meets at an intersection.
Fig. 3 and Fig. 4 show the effect of the channel error probability when the
offered load is relatively low, namely, 0.5. The curve marked as proposed is the
performance of the case of channel switching. The channel error probability is a
probability that a channel is clean when a switch node is to send its message. This
experiment assumes that a message is discarded if it is not served until 16 time
slots from its arrival. Fig. 3 reveals that a message can exit the intersection area
around in 5.6 time slots, that is, less than 1.5 round. In Fig. 4 the delivery ratio
is almost 1.0 when the channel error rate is close to 0. At the higher error rate,
we can achieve better improvement, as there are more cases channel switching
is possible. When the channel error probability is greater than 0.4, the proposed
scheme can improve the delivery ratio by up to 18.5 %.
Fig. 5 and Fig. 6 plot the effect of the channel error probability when the
offered load is 1.0. On high load, a switch node can almost always find two

12 1
"Normal" "Normal"
"Proposed" "Proposed"
Access delay (slot time)

11.8 0.8
Delivery ratio

11.6 0.6

11.4 0.4

11.2 0.2

11 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Channel error probability Channel error probability

Fig. 5. Access delay at high load Fig. 6. Delivery ratio at high load
18 J. Lee et al.

12 1

11
Access delay (slot time)

0.8
10

Delivery ratio
9 0.6

8 "Normal" 0.4 "Normal"

"Proposed" "Proposed"
7
0.2
6

5 0
0.5 0.6 0.7 0.8 0.9 1 0.5 0.6 0.7 0.8 0.9 1
Load Load

Fig. 7. Access delay vs. load Fig. 8. Delivery ratio vs. load

messages having different destinations. As can be found in Fig. 5, the access delay
for both cases stays at 11.2 time slots, namely, less than 3 rounds. This indicates
that the discard interval can control the access delay quite well regardless of
channel errors. When the channel error probability gets larger than 0.4, the
proposed scheme can improve the delivery ratio by up to 13.0 %. Fig. 6 shows
that both the delivery ratio and its performance gap are less than those in Fig.
4, indicating that the message discarded in the queue outnumbers that saved by
the channel switch.
Fig. 7 and Fig. 8 demonstrate the effect of message load to the access delay and
the delivery ratio, respectively. In this experiment, the channel error probability
is set to 0.1 and the discard interval is fixed to 16 time slots. Fig. 7 shows that
the access delays for two cases are almost same and the degree of inclination
gets smaller after the load is 0.8, as more messages are discarded for the range
of 0.8 through 1.0. In Fig. 8, the two curves show the constant gap for all load
range, while the gap is 8.1 % at maximum, indicating that the early discard
more affects the delivery ratio on higher load.
Fig. 9 and Fig. 10 show the effect of discard intervals. It can be simply expected
that the shorter the discard interval, the shorter the access delay. Fig. 9 exhibits

7 1

0.9
Access delay (slot time)

0.8
Delivery ratio

5
0.7
4 "Normal"
"Proposed" 0.6 "Normal"
"Proposed"
3
0.5

2 0.4
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
Discard interval Discard interval

Fig. 9. Effect of discard interval Fig. 10. Effect of discard interval

 Performance Measurement of a Dual-Channel Intersection Switch 19

the largely linear increase of access delays according to the discard interval. Two
curves also have no conceivable difference as in the other cases. However, Fig. 10
reveals that the effect of discard intervals is not so significant to the delivery
ratio, less than 18 % over the interval range of 5 through 35. Two curves show
a similar pattern, but the gap gets a little bit larger according to the increase of
discard intervals. For all experiments, it can be pointed out by our experiment
that the improved delivery ratio doesn’t affect the transit time.

5 Conclusion

This paper has proposed and measured the performance of a message switch
scheme for a hot intersection area in vehicular telematics network. Installed at
each corner of an intersection, switch nodes cooperatively exchange messages ac-
cording to the current channel status, speeding up the intersection transit time
and improving the delivery ratio. The simulation result shows that the proposed
scheme can improve the delivery ratio by up to 18.5 % for the given range of
channel error probability, and 8.1 % for the given range of network load. This
result comes from the fact that our scheme can avoid the complex routing and
message interference in the intersection of intensive traffic. Next, dual channels
can efficiently cope with channel errors. Even though such employment of ad-
ditional equipments brings extra cost and loses the advantage of autonomous
operations, access predictability, enhanced delivery speed, and better delivery
ratio can compensate for the cost. In short, the wireless switch can be installed
at the intersections, which have a lot of traffic especially in the urban area, to
enhance the communication quality and reliability.
As future work, we are planning to assess the performance of our scheme with
the real-life location history data obtained by the Jeju taxi telematics system[1]
and analyze the end-to-end performance characteristics to revise our algorithm
and test the diverse message scheduling policy.

References

1. Lee, J., Park, G., Kim, H., Yang, Y., Kim, P., Kim, S.: A telematics service system
based on the Linux cluster. In: Shi, Y., van Albada, G.D., Dongarra, J., Sloot,
P.M.A. (eds.) ICCS 2007. LNCS, vol. 4490, pp. 660–667. Springer, Heidelberg
(2007)
2. Society of Automotive Engineers: Dedicated short range communication message
set dictionary. Technical Report Standard J2735 (2006)
3. Dikaiakos, M., Iqbal, S., Nadeem, T., Iftode, L.: VITP: An information transfer
protocol for vehicular computing. In: ACM VANET, pp. 30–39 (2005)
4. Zhao, J., Arnold, T., Zhang, Y., Cao, G.: Extending drive-thru data access by
vehicle-to-vehicle relay. In: ACM VANET, pp. 30–39 (2008)
5. Lochert, C., Scheuermann, B., Wewetzer, C., Luebke, A., Mauve, M.: Data aggre-
gation and roadside unit placement for a VANET traffic information system. In:
ACM VANET, pp. 58–65 (2008)
20 J. Lee et al.

6. Korkmaz, G., Ekici, E., Ozguner, F., Ozguner, U.: Urban multihop broadcast pro-
tocol for inter-vehicle communication systems. In: ACM VANET, pp. 76–85 (2004)
7. Davis, J., Fagg, A., Levine, B.: Wearable computers as packet transport mech-
anism in highly-partitioned ad-hoc networks. In: Int’l Symposium on Wearable
Computing (2001)
8. Song, S., Han, S., Mok, A., Chen, D., Nixon, M., Lucas, M., Pratt, W.: Wireless
HART: Applying Wireless Technology in Real-Time Industrial Process Control.
In: The 14th IEEE Real-Time and Embedded Technology and Applications Sym-
posium, pp. 377–386 (2008)
9. Naumov, V., Gross, T.: Connectivity-aware routing (CAR) in vehicular ad hoc
networks. In: IEEE Infocom, pp. 1919–1927 (2007)
10. US Department of Transportation: Integrated Vehicle-Based Safety Systems
(IVBSS) Phase I Interim Report. DOT HS 810 952 (2008),
http://www.itsdocs.fhwa.dot.gov/JPODOCS//REPTS_TE/14434.htm
11. Kodialam, M., Nandagopal, T.: Characterizing the capacity region in multi-radio
multi-channel wireless mesh networks. In: ACM MobiCom, pp. 73–87 (2005)
12. Zhao, J., Cao, G.: VADD: Vehicle-assisted data delivery in vehicular ad hoc net-
works. In: IEEE INFOCOM (2006)
13. Lee, J., Kang, M., Park, G., Shin, S.: Design of a reliable real-time scheduling
policy for dual-channel networks. Journal of Information Science and Engineering,
1407–1419 (2007)
14. Lee, J., Park, G., Shin, S., Kang, M.: Design of intersection switches for the vehic-
ular network. In: Hong, C.S., Tonouchi, T., Ma, Y., Chao, C.-S. (eds.) APNOMS
2009. LNCS, vol. 5787, pp. 523–526. Springer, Heidelberg (2009)
15. Jaap, S., Bechler, M., Wolf, L.: Evaluation of Routing Protocols for Vehicular Ad
Hoc Networks in City Traffic Scenarios. In: Proceedings of the 5th International
Conference on Intelligent Transportation Systems Telecommunications (2005)
16. Li, H., Shenoy, P., Ramamritham, K.: Scheduling communication in real-time sen-
sor applications. In: Proc.10th IEEE Real-time Embedded Technology and Appli-
cations Symposium, pp. 2065–2080 (2004)
17. MacDougall, M.: Simulating Computer Systems: Techniques and Tools. MIT Press,
Cambridge (1987)
 A Rapid Code Acquisition Scheme for Optical CDMA
Systems

Youngyoon Lee, Dahae Chong, Chonghan Song, Youngpo Lee,

Seung Goo Kang, and Seokho Yoon

School of Information and Communication Engineering, Sungkyunkwan University,

300 Chunchun-dong, Jangan-gu, Suwon, Gyeonggi-do, 440-746, Korea
{news8876,lvjs1019,starsong83,leeyp204,
general85,syoon}@skku.edu

Abstract. In this paper, we propose a novel code acquisition scheme called mod-
ified multiple-shift (MMS) for optical code division multiple access (CDMA)
systems. By using the modified local code whose sign is positive or negative,
the proposed MMS scheme provides a shorter mean acquisition time (MAT) than
the conventional multiple-shift (MS) scheme. The simulation results demonstrate
that the MAT of the MMS scheme is shorter than that of the MS scheme in both
single-user and multi-user environments.

Keywords: Optical code division multiple access, optical orthogonal code,

multiple-shift, code acquisition, mean acquisition time.

1 Introduction

In code division multiple access (CDMA)-based systems, the data demodulation is pos-
sible only after a code synchronization is completed. Therefore, the code synchroniza-
tion is one of the most important tasks in CDMA-based systems [1]. Generally, the
code synchronization consists of two stages: code acquisition and tracking. Achieving
the code synchronization is called code acquisition and maintaining the code synchro-
nization is called tracking [2], of which the former is dealt with in this paper. In code
acquisition process, the most significant performance measure is the mean acquisition
time (MAT), which is a mean time that elapses prior to acquisition.
An optical CDMA system uses a spreading code called optical orthogonal code
(OOC) proposed by Salehi [3]. Due to its good auto-correlation and cross-correlation
properties, the OOC has been widely used for various CDMA-based systems including
optical CDMA systems [4], [5]. Keshavarzian and Salehi introduced the serial-search
(SS) [4] scheme using the OOC, which is simple; however, its MAT increases as the
code length becomes longer. Thus, the SS scheme is not suitable for rapid acquisition
of a long code used in multi-user environments. In order to overcome this drawback, in
[5], the same authors proposed the multiple-shift (MS) scheme using the OOC, which
consists of two stages and offers a shorter MAT compared with that of the SS scheme.

Corresponding author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 21–30, 2010.

c Springer-Verlag Berlin Heidelberg 2010
22 Y. Lee et al.

In this paper, we propose a novel code acquisition scheme called modified multiple-
shift (MMS). The MMS scheme comprises two stages similarly to the MS scheme,
however, by using the modified local code differently from the MS scheme, it provides
a shorter MAT than that of the MS scheme.
The remainder of this paper is organized as follows. Section 2 describes the system
model. In Section 3, we present the conventional MS and proposed MMS schemes. Sec-
tion 4 analyzes the MAT performance of the MMS scheme. In Section 5, the simulation
results show the MATs of the MS and MMS schemes in single-user and multi-user en-
vironments. Section 6 concludes this paper.

2 System Model
In an optical CDMA channel, there exist various kinds of impairments such as noise,
multipath signals, and multiple access interference (MAI). The influences of noise and
multipath signals can be almost completely mitigated by using fiber-optic medium;
however, that of MAI should be alleviated in the receiver [6], [7]. In this paper, thus,
we consider a multi-user environment without the influences of noise and multipath
signals. Then, the received signal r(t) can be written as


N
r(t) = s(n) (t − τ (n) ), (1)
n=1

where s(n) (t) is the transmitted signal of the n-th user; τ (n) ∈ [0, Tb ) denotes the
time delay of the n-th user with the bit duration Tb ; and N is the number of users. We
consider the on-off-keying (OOK) modulation and assume that the bit rate is the same
for all users. Thus, the transmitted signal s(n) (t) can be expressed as
∞
 (n)
s (n)
(t) = bi c(n) (t − iTb ), (2)
i=−∞

(n) F −1 (n)
where bi is the i-th binary data bit of the n-th user and c(n) (t) = j=0 aj p(t−jTc )
(n)
is the OOC of the n-th user with the chip duration Tc and sequence aj ∈ {0, 1} of
(n)
length F and weight K (the total number of ‘1’s in aj ), where the rectangular pulse
p(t) over [0, Tc ] is defined as

1, 0 ≤ t < Tc ,
p(t) = (3)
0, otherwise.

The OOC has four parameters F , K, λa , and λc , where λa and λc are auto-correlation
and cross-correlation constraints, respectively [3]. In order to maintain the strict orthog-
onality of the OOC, both λa and λc have to be zero; however, since an OOC consists
of 0 and 1, the ideal strict orthogonality cannot be satisfied. Thus, in this paper, both λa
and λc are set to 1.
 A Rapid Code Acquisition Scheme for Optical CDMA Systems 23

F chips

(a) The first stage local code of the MS scheme when M = 3

(b) The second stage local code of the MS scheme (original OOC)

Fig. 1. The local codes of the MS scheme when OOC of (32,4,1,1) is used

3 Code Acquisition Schemes

3.1 MS Scheme

In the MS scheme, total F cells in the search space are divided into Q groups, each of
which contains M cells. The relation of Q and M is given by
F 
Q= , (4)
M
where · denotes the ceiling operation.
The MS scheme consists of two stages. In the first stage, the received signal r(t) is
correlated with the first stage local code shown in Fig. 1. The correlation is repeated
on a group-by-group basis. If the correlation value corresponding to a certain group
exceeds a given threshold T HMS,f irst , the group is declared to be the correct group
having the time delay τ (n) and the process is transferred to the second stage. In the
second stage, the correlation-based search is performed again with the second stage
local code (original OOC) on a cell-by-cell basis over M cells in the correct group. As
in the first stage, when the correlation value corresponding to a certain cell exceeds a
given threshold T HMS,second , which is different from T HMS,f irst , the cell is declared
to be an estimate of the time delay τ (n) .
Using the definition in [5], the MAT TMS of the MS scheme is given by

Q+1 M +1
TMS = + . (5)
2 2
√
From
√ (4) and (5), we can find that the minimum value of TMS equals to F when
M = F.
24 Y. Lee et al.

F chips

(a) The first stage local code of the MMS scheme when M = 3

(b) The second stage local code of the MMS scheme (original OOC)

Fig. 2. The local codes of the MMS scheme when OOC of (32,4,1,1) is used

3.2 MMS Scheme

The MMS scheme proposed in this paper also consists of two stages similarly to the
MS scheme. However, using the modified local code, the MMS scheme can provide a
shorter MAT than that of the MS scheme.
In the first stage, the first stage local code shown in Fig. 2 is used instead of the
conventional local code used for MMS scheme. The first stage local code consists of
several chips whose sign is positive or negative. When M is an even number, the first
half of the group template is positive and its second half is negative; otherwise, the
number of positive chips is larger by one than that of negative chips.
In the first stage, the local and the received codes are correlated and the local code
is updated by M chips at every operation where the correlation value is calculated.
Unlike the MS scheme, the first stage of the MMS scheme finishes the search when the
absolute value of the correlation value is equal to or larger than the T HMMS,f irst . If the
correlation value of the correct group is positive, we only need to search the first half of
the M cells of the correct group in the second stage; otherwise, the search is performed
over the second half of the M cells. Thus, we only need to search the maximum M/2
chips regardless of the sign of the correlation value. In other words, the number of the
operations in the second stage is reduced by nearly half than that of the MS scheme by
using the modified local code. In the second stage of the MMS scheme, the correlation
is compared to the threshold T HMMS,second and the maximum number of the updates
of the local code is M/2. Therefore, the MAT of the MMS scheme TMMS is
Q+1 M +2
TMMS = + . (6)
2 4
We can notice from (5) and (6) that the MAT in the second stage is reducedby nearly
√of the MS scheme and the minimum value of TMMS equals F/2 + 1
half than that
when M = 2F .
 A Rapid Code Acquisition Scheme for Optical CDMA Systems 25

F chips

tmin

(a) Original OOC when tmin = 2 chips

(b) The first stage local code of the MS scheme when M = 4

(c) The first stage local code of the MMS scheme when M = 4

Fig. 3. The local codes of the MS and MMS schemes when M > tmin

For a correct operation of the proposed MMS and conventional MS schemes, M

has to be equal to or smaller than tmin , where tmin is the minimum chip interval of the
OOC. Otherwise, chips of local code are overlapped as shown in Fig. 3, where tmin = 2
and M = 4. In this case, the MMS scheme cannot work appropriately.
Fig. 3 shows the original OOC and the first stage local codes of the MS scheme
and the MMS schemes. In Fig. 3, the dashed ellipse means the overlapped chips of the
local code. Though the code is not synchronized, the absolute value of the correlation
value is increased (decreased) when the amplitude of the overlapped chips is increased
(decreased). Thus, the false alarm can be occurred, increasing the MAT.

4 Performance Analysis
In this section, we derive the MAT expression of the MMS scheme by modeling the
acquisition process as a discrete-time Markov process [1] with the circular flow graph
diagram shown in Figs. 4 and 5. In Fig. 4, ‘ACQ’ and ‘F A’ represent the acquisition
and false alarm states, respectively, and PD (pd ) and PF A (pf a ) denote the detection and
false alarm probabilities in the first and second stages, respectively. Fig. 5 represents the
transition process diagrams of Hdet (z) in Fig. 4.
26 Y. Lee et al.

ACQ
FA
zL
s2
pfaz h(z)
s3
Hdet(z)
s1 (1-pfa)z
H(z) Hmiss(z)
PFAz
QQ s(M/2)+1
H(z) QQ-1 FA pfaz PFAz Q1 (1-PFA)z
sM/2
L
z (1-pfa)z Q2
QQ-2 h(z)
s(M/2)+2
h(z) h(z)
H(z)
sM
s(M/2)+3
Q3

Qi+1

Qi
H(z)

Fig. 4. Markov state model for the MMS scheme

In Fig. 4, states Q1 , Q2 , · · · , QQ−1 correspond to the incorrect groups and a state

QQ corresponds to the correct group, where the acquisition can be achieved. States
s1 , s2 , · · · , sM/2 and states s(M/2)+1 , s(M/2)+2 , · · · , sM correspond to cells in the sec-
ond stage, when the correlation value of the correct group is positive and negative, re-
spectively. The gains Hdet (z) and Hmiss (z) represent the transition gains from QQ to
ACQ (detection) and from QQ to Q1 (miss), respectively. The gains H(z) and h(z)
represent the transition gains between two consecutive successful incorrect groups and
between two consecutive successful incorrect cells, respectively. L is the penalty time
(v)
factor due to the false alarm. The gains h(z), H(z), Hdet (z), and Hmiss (z) can be ex-
pressed as follows:
h(z) = (1 − pf a )z + pf a z L+1 , (7)

H(z) = (1 − PF A )z + PF A zhM/2 (z), (8)

(v)
Hdet (z) = pd PD z 2 hv−1 (z), (9)
and
Hmiss (z) = (1 − PD )z + PD (1 − pd )z 2 h(M/2)−1 (z), (10)
where v ∈ {1, 2, · · · , M/2} is distributed uniformly over [1, M/2] and represents the
correct state.
 A Rapid Code Acquisition Scheme for Optical CDMA Systems 27

(a) ACQ

pdz

h(z) (1-pd)z h(z)

h(z) s2 sv-1 sv sv+1 s(M/2)-1

s1 sM/2

PDz h(z)

QQ (1-PD)z Q1

(b) ACQ

pdz

h(z) (1-pd)z h(z)

h(z) s(M/2)+2 sv-1 sv sv+1 sM-1

s(M/2)+1 sM

PDz h(z)

QQ (1-PD)z Q1

Fig. 5. (a) Positive group from QQ to Q1 (b) Negative group from QQ to Q1

Let us assume that the search is started at Qi , then the transfer function between Qi
and ACQ nodes can be written as

(v)
(v) H Q−i (z)Hdet (z)
Ui (z) = . (11)
1 − Hmiss (z)H Q−1 (z)

In (11), i ∈ {1, 2, · · · , Q} is assumed to be distributed uniformly over [1, Q]. Thus,

(v)
after averaging Ui (z) over the probability density function (pdf) of i and v, we can
re-write (11) as

1  Q−i
Q
Hdet (z)
U (z) = H (z), (12)
1 − Hmiss (z)H Q−1 (z) Q i=1
28 Y. Lee et al.

2 M/2 (v)
where Hdet (z) = M v=1 Hdet (z). Using the moment generating function, we can
obtain the following relationship between U (Z) and TMMS .

U (z) = E(z TM M S ), (13)

where E(·) denotes the statistical expectation operation. In (13), E(TMMS ) can be
computed by its moment generating function as

dU (z) 
E(TMMS ) =  = U  (1), (14)
dz z=1
and thus, can be obtained as
 
Hdet (1) + Hmiss (1) 2 − pd PD
E(TMMS ) = + (Q − 1)H  (1) , (15)
pd PD 2pd PD
where H  (1), Hdet
 
(1), and Hmiss (1) can be expressed as

M
H  (1) = 1 + PF A (1 + Lpf a ), (16)
2

 (M/2) − 1
Hdet (1) = 2pd PD + pd PD (1 + Lpf a ), (17)
2
and

Hmiss (1) = (1 − PD ) + PD (1 − pd )[(M/2) + 1 + {(M/2) − 1}Lpf a], (18)

respectively.
When pd = PD = 1 and pf a = PF A = 0 (i.e., single-user case), (15) is re-written
as
Q+1 M +2
E(TMMS ) = + . (19)
2 4
√
Differentiating
(19), we finally obtain the optimum M and minimum TMMS as 2F
and F
2 + 1, respectively, when F  1.

5 Simulation Results
In this section, we compare the MAT of the conventional MS scheme with that of the
proposed MMS scheme in single-user and multi-user environments. Simulation pa-
rameters are as follows: F = 200; K = 5; λa = λc = 1; N = 1, 2, 3, and 4;
T HMS,f irst = T HMS,second = T HMMS,f irst = T HMMS,second = K; and L = 10.
We assume that each user transmits the data 1 or 0 with equal probability and chip
synchronization is perfect.
Fig. 6 shows the MAT performance of the MS and MMS schemes as a function
of M in a single-user environment. In Fig. 6, the dotted and solid lines represent the
simulation results of the MS and MMS schemes, respectively, and, the markers ∇ and
 A Rapid Code Acquisition Scheme for Optical CDMA Systems 29

55
MS, Simulation
50 MMS, Simulation
MS, Theory
45 MMS, Theory

40

MAT 35

30
Minimum MAT
25

20

15

10
0 5 10 15 20 25
M

Fig. 6. The MATs of the MS and MMS schemes as a function of M in a single-user environment

55
MS, 2 users
50 MMS, 2 users
MS, 3 users
45 MMS, 3 users
MS, 4 users
40 MMS, 4 users
MAT

35

30

25

20

15

10
0 5 10 15 20 25
M

Fig. 7. The MATs of the MS and MMS schemes as a function of M in a multi-user environment

O represent the theoretical results of the MS and MMS schemes, respectively. From the
figure, we can observe that the MAT of the MMS scheme is shorter than that of the MS
scheme, and√ confirm that the MS and√MMS schemes provide the minimum MAT when
M = 14 ( F
14) and M = 20 ( 2F = 20), respectively. The difference between
the MATs of the MS and MMS schemes increases as M increases. In Fig. 7, we can see
the MAT performance of the MS and MMS schemes as a function of M in a multi-user
environment. From Fig. 7, we can confirm that the MAT of the MMS scheme is shorter
than that of the MS scheme, and the difference between the MATs of the MS and MMS
schemes increases as M increases as in the single-user environment.
30 Y. Lee et al.

6 Conclusion

In this paper, we have proposed a novel code acquisition scheme called MMS for optical
CDMA systems. Exploiting the modified local code, the proposed MMS scheme can
provide a shorter MAT compared with that of the MS scheme. The performance of the
MMS scheme has been analyzed using the circular flow graph diagram. The simulation
results have confirmed that the MMS scheme offers a shorter MAT compared with that
of the MS scheme in both single-user and multi-user environments.

Acknowledgments. This research was supported by the MKE (The Ministry of Knowl-
edge Economy), Korea, under the ITRC (Information Technology Research Center)
support program supervised by the NIPA (National IT Industry Promotion Agency)
(NIPA-2009-(C1090-0902-0005)).

References
1. Polydoros, A., Weber, C.L.: A unified approach to serial-search spread spectrum code
acquisition-Part I: General theory. IEEE Trans. Commun. 32(5), 542–549 (1984)
2. Chong, D., Lee, B., Kim, S., Joung, Y.-B., Song, I., Yoon, S.: Phase-shift-network-based dif-
ferential sequential estimation for code acquisition in CDMA systems. Journal of Korea In-
form. and Commun. Society 32(3), 281–289 (2007)
3. Salehi, J.A.: Code division multiple-access techniques in optical fiber networks - Part I: Fun-
damental principles. IEEE Trans. Commun. 37(8), 824–833 (1989)
4. Keshavarzian, A., Salehi, J.A.: Optical orthogonal code acquisition in fiber-optic CDMA sys-
tems via the simple serial-search method. IEEE Trans. Commun. 50(3), 473–483 (2002)
5. Keshavarzian, A., Salehi, J.A.: Multiple-shift code acquisition of optical orthogonal codes in
optical CDMA systems. IEEE Trans. Commun. 53(3), 687–697 (2005)
6. Salehi, J.A., Brackett, C.A.: Code division multiple-access techniques in optical fiber networks
- Part II: Systems performance analysis. IEEE Trans. Commun. 37(8), 834–842 (1989)
7. Stok, A., Sargent, E.H.: Lighting the local area: Optical code-division multiple access and
quality of service provisioning. IEEE Network 14(6), 42–46 (2000)
 Optimal and Suboptimal Synchronization Schemes for
Ultra-Wideband Systems

Dahae Chong, Chonghan Song, Youngpo Lee, Myungsoo Lee,

Junhwan Kim, and Seokho Yoon

School of Information and Communication Engineering, Sungkyunkwan University,

300 Chunchun-dong, Jangan-gu, Suwon, Gyeonggi-do, 440-746, Korea
{lvjs1019,starsong83,leeyp204,maxls813,
ourlife3,syoon}@skku.edu

Abstract. The conventional ultra-wideband (UWB) synchronization schemes

could cause poor performance in receiver operations such as demodulation after
the synchronization, since unreliable timing information corresponding to low-
power multipath components might be transferred to the next stage. To solve this
problem, in this paper, we propose novel synchronization schemes for UWB sys-
tems. We first derive an optimal scheme based on the maximum likelihood (ML)
criterion and then develop a simpler suboptimal scheme. Simulation results show
that both proposed schemes provide a better synchronization performance over
the conventional scheme.

Keywords: Maximum likelihood (ML), Optimal, Suboptimal, Synchronization,

Ultra-wideband (UWB).

1 Introduction
Ultra-wideband (UWB) is a technology that has distinct features characterized by its
wide bandwidth, and has been adopted as a standard for wireless personal area networks
(WPANs) such as IEEE 802.15.3a and IEEE 802.15.4a, due to its high data rate and
low-power transmission [1]-[4].
In UWB systems, the received signal must be first synchronized with a locally gener-
ated template prior to data demodulation, and thus, a rapid synchronization is one of the
most important technical issues in UWB systems [5], [6]. For the fast synchronization
in UWB systems, several schemes have been investigated, where timing information
corresponding to any of the received multipath components is regarded to be correct
[7]-[10].
In [7], Homier and Scholts presented synchronization schemes with various search
methods, and analytically derived their mean synchronization time in a noise-free en-
vironment, and in [8], Vijayakumaran and Wong introduced a fast synchronization
scheme with the permutation search method. Ramachandran and Roy [9] analyzed the
schemes of [7] with the tapped-delay-line channel and IEEE 802.15.3a channel models,
and Arias-de-Reyna and Acha-Catalina [10] proposed a new serial search synchroniza-
tion scheme using specific search spacing based on Fibonacci sequence. However, these

Corresponding author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 31–41, 2010.

c Springer-Verlag Berlin Heidelberg 2010
32 D. Chong et al.

schemes could cause poor performance in receiver operations such as demodulation af-
ter the synchronization, if the locally generated template would be synchronized with
a low-power multipath component. Generally, the first one among multipath compo-
nents has the largest average power. Thus, the synchronization with the first multipath
component is expected to offer a good performance in receiver operations after the syn-
chronization.
In this paper, we propose new synchronization schemes for UWB systems, allowing
us to obtain the timing information of the first multipath component efficiently. First, an
optimal scheme is proposed based on a maximum-likelihood (ML) criterion. However,
the optimal scheme requires the channel information, and thus, is difficult to implement
practically. Thus, a simpler suboptimal scheme is developed. The simulation results
show that the proposed schemes provide a substantial performance improvement over
the conventional scheme.
This paper is organized as follows. Section 2 describes the system model. In Section
3, the optimal and suboptimal schemes are proposed. Section 4 presents the simula-
tion results of the conventional and proposed schemes over several multipath channel
models. Section 5 concludes this paper with a brief summary.

2 System Model

In this paper, we consider a direct sequence (DS)-UWB system [9]. During the acquisi-
tion process, we assume that an unmodulated signal (acquisition preamble) is transmit-
ted. Then, the transmitted DS-UWB signal s(t) can be expressed as

 N−1
s(t) = Ec ci p(t − iTc ), (1)
i=0

where Ec is the signal energy; ci ∈ {1, −1} is the i-th chip of a pseudo noise (PN)
sequence with a period of N chips; Tc is the chip duration; and p(t) is the second
derivative Gaussian pulse [7], [9] with a duration of Tc . At the receiver, the received
signal r(t) can be expressed as

p −1
−1 L
 N
r(t) = Ec αj ci p(t − iTc − jTc − τ Tc ) + w(t), (2)
i=0 j=0

where Lp denotes the number of multipaths; αj is the channel coefficient of the j-th
1−e−μ −(j−1)μ
multipath with the average power 1−e −μLp e , where μ is a decay factor; τ is the
time delay; and w(t) is an additive white Gaussian noise (AWGN) process with mean
zero and two-sided power spectral density N20 .
In this paper, we consider a parallel receiver for UWB synchronization shown in
Fig. 1. The receiver first yields the m-th correlator output ym by correlating r(t) with
N −1
the locally generated template signal gm (t) = n=0 cn p(t − (n + m)Tc ), where
m = 0, 1, · · · , N − 1, over a correlation time N Tc , and then, takes the absolute value
of ym to remove the influence of the signal conversion due to the channel reflections.
 Optimal and Suboptimal Synchronization Schemes for Ultra-Wideband Systems 33

r (t )

NTc y0 y0
³ dt ˜
0

g 0 (t )
NTc y1 y1
³ dt ˜
0

Synchronization
Demodulation
g1(t ) scheme

#
#
NTc yN 1 yN 1
³ dt ˜
0

g N 1(t )

Fig. 1. A parallel receiver for UWB synchronization

Subsequently, the synchronization process produces an estimate of the time delay τ em-
ploying the absolute correlator outputs, and finally, the demodulation process is started.
In the conventional scheme, an estimate τ̂c of the time delay is obtained as
τ̂c = arg max |ym |. (3)
0≤m≤N −1

However, the conventional scheme is inefficient to obtain the timing information of

the first multipath component, since it does not sufficiently exploit the signal energy
spread over multipath components. Thus, we propose novel schemes for selecting the
first multipath component efficiently.

3 Proposed Schemes
In UWB channels, the first multipath component generally has the largest power [11],
[12]. Based on which, we propose novel schemes based on the ML approach for select-
ing the first multipath component efficiently.
34 D. Chong et al.

We first calculate the mean and variance of ym to obtain the distribution of the cor-
relator output ym expressed as

N Tc
ym = r(t)gm (t)dt
0



N Tc N −1 N
 p −1
−1 L
= Ec αj ci cn
0 i=0 n=0 j=0

· p(t − (i + j + τ )Tc )p(t − (n + m)Tc )dt + wm , (4)

where wm is the noise component of the m-th correlator output with mean zero and
variance N2N0 and can be written as


N Tc N −1
wm = w(t)cn p(t − (n + m)Tc )dt. (5)
0 n=0
The value of ym changes according to the phase difference (i.e. m − τ ) between the
received and template signals as follows:
⎧ √
⎪
⎪ N Ec α0 + w0 , m−τ =0
⎪
⎪ √
⎪
⎪ N Ec α1 + w1 , m−τ =1
⎪
⎪
⎪
⎪ . ..
⎪
⎪ .. .
⎨ √
ym = N Ec αLp −1 + wLp −1 , m − τ = Lp − 1 (6)
⎪
⎪
⎪
⎪ wLp , m − τ = Lp
⎪
⎪
⎪
⎪ . ..
⎪
⎪ ..
⎪
⎪
.
⎩
wN −1 , m − τ = N − 1.
Then, the probability density function (pdf) fj of the correlator output ym corre-
sponding to the j-th multipath component can be written as
 √ 
1 (ym − N Ec αj )2
fj (ym ) = √ exp − , (7)
πN N0 N N0
√
where N Ec αj is the mean of the j-th multipath component. The pdf fw of the cor-
relator output which does not correspond to any of the multipaths can be written as
 
1 y2
fw (ym ) = √ exp − m . (8)
πN N0 N N0

The set of the correlator outputs is denoted by

y = [y0 , y1 , · · · , yN −1 ]T with yi = y(i mod N ) . (9)
Then, the pdf f of y, given that m − τ = 0, can be expressed as
−1 Lp −1

N  fj (ym+j )
f (y|m) = fw (yb ) . (10)
j=0
fw (ym+j )
b=0
 Optimal and Suboptimal Synchronization Schemes for Ultra-Wideband Systems 35

Since τ is distributed uniformly over [0, N − 1], the optimal scheme based on the ML
approach can be expressed as
τ̂o = arg max f (y|m). (11)
0≤m≤N −1

In (11), we can remove the term which is independent from m, which results in
Lp −1
 fj (ym+j )
τ̂o = arg max . (12)
0≤m≤N −1
j=0
fw (ym+j )

Using (7) and (8), (12) becomes

Lp −1 √ 2 
(ym+j − N Ec αj )2 ym+j
τ̂o = arg max exp − + . (13)
0≤m≤N −1 N N0 N N0
j=0

In (13), we can remove the terms which are independent from m and constant. Then,
the optimal scheme can be reduced as
Lp −1

τ̂o = arg max ym+j αj , (14)
0≤m≤N −1
j=0

which is optimal to select the first multipath component in UWB channel. However,
Lp −1
the optimal scheme requires information on channel coefficients {αj }j=0 , and thus,
is difficult to implement. Hence, we develop a simpler suboptimal scheme.
Lp −1
When m = τ = 0, j=0 ym+j αj of (14) is computed as
Lp −1 Lp −1
  
yj αj = N Ec α2j
j=0 j=0
Lp −1
 
= |N Ec αj ||αj |
j=0
Lp −1

= |yj ||αj | (15)
j=0

in a noise-free environment. Using (15), the optimal scheme can be re-written as

Lp −1

τ̂o = arg max |ym+j ||αj |, (16)
0≤m≤N −1
j=0

L −1
which requires the absolute values of channel coefficients. By removing {|αj |}j=0
p

and using a parameter Q instead of Lp in (16), we obtain the suboptimal scheme as


Q−1
τ̂s = arg max |ym+j | (17)
0≤m≤N −1
j=0
36 D. Chong et al.

for UWB synchronization, where Q is the combining size of the correlator outputs.
Combining the correlator outputs corresponding to the paths with low-power may lead
to an increase in the noise variance rather than in the signal strength. Thus, Q is set to
the average number of paths with power attenuated by at most 10 dB from the power of
the first path, and can be predetermined if a channel model is given.

4 Simulation Results

In this section, we compare the performance of the optimal, suboptimal, and conven-
tional schemes in terms of the probability of false synchronization defined as the proba-
bility that the estimate of the scheme does not correspond to a timing information of the
first multipath component. We define the signal to noise ratio (SNR) as Ec /N0 and use a
PN sequence with a period of 255 chips and a chip duration of 0.5 ns. We simulated the
synchronization process both in the tapped-delay line channel model with a tap spac-
ing of 0.5 ns and in the IEEE 802.15.3a channel model [11], [12], where four different
environments (CM1, CM2, CM3, and CM4) are defined. As the channel environment
changes from CM1 to CM4, the multipaths spread more widely.
Figs. 2-5 show the probabilities of false synchronization of the schemes in the
tapped-delay line channel model. As shown in figures, the proposed schemes have a
better performance compared with that of the conventional scheme. Since the optimal
scheme uses the channel information, it has the best performance. On the other hand,

0
10
Probability of false synchronization

-1
10

-2
10

-3
10

Conventional
conventional1
Optimal
data2
Suboptimal
data3
-4
10
0 5 10 15 20 25 30 35 40

Ec / N 0 (dB)

Fig. 2. Probabilities of false synchronization of the conventional, optimal, and suboptimal

schemes in the tapped-delay line channel model with CM1 coefficients
 Optimal and Suboptimal Synchronization Schemes for Ultra-Wideband Systems 37

0
10

Probability of false synchronization

-1
10

-2
10

-3
10

Conventional
conventional1
Optimal
data2
data3
Suboptimal
-4
10
0 5 10 15 20 25 30 35 40

Ec / N 0 (dB)

Fig. 3. Probabilities of false synchronization of the conventional, optimal, and suboptimal

schemes in the tapped-delay line channel model with CM2 coefficients

0
10
Probability of false synchronization

-1
10

-2
10

-3
10

Conventional
conventional1
Optimal
data2
data3
Suboptimal
-4
10
0 5 10 15 20 25 30 35 40
Ec / N 0 (dB)

Fig. 4. Probabilities of false synchronization of the conventional, optimal, and suboptimal

schemes in the tapped-delay line channel model with CM3 coefficients
38 D. Chong et al.

0
10

Probability of false synchronization

-1
10

-2
10

-3
10

Conventional
conventional1
Optimal
data2
Suboptimal
data3
-4
10
0 5 10 15 20 25 30 35 40
Ec / N 0 (dB)

Fig. 5. Probabilities of false synchronization of the conventional, optimal, and suboptimal

schemes in the tapped-delay line channel model with CM4 coefficients

0
10
Probability of false synchronization

-1
10

-2
10

-3
10

Conventional
conventional1
Optimal
data2
Suboptimal
data3
-4
10
0 5 10 15 20 25 30 35 40

Ec / N 0 (dB)

Fig. 6. Probabilities of false synchronization of the conventional, optimal, and suboptimal

schemes in the IEEE 802.15.3a CM1 channel model
 Optimal and Suboptimal Synchronization Schemes for Ultra-Wideband Systems 39

0
10

Probability of false synchronization

-1
10

-2
10

-3
10

Conventional
conventional1
Optimal
data2
Suboptimal
data3
-4
10
0 5 10 15 20 25 30 35 40

Ec / N 0 (dB)

Fig. 7. Probabilities of false synchronization of the conventional, optimal, and suboptimal

schemes in the IEEE 802.15.3a CM2 channel model

0
10
Probability of false synchronization

-1
10

-2
10

-3
10

conventional1
Conventional
data2
Optimal
data3
Suboptimal
-4
10
0 5 10 15 20 25 30 35 40
Ec / N 0 (dB)

Fig. 8. Probabilities of false synchronization of the conventional, optimal, and suboptimal

schemes in the IEEE 802.15.3a CM3 channel model
40 D. Chong et al.

0
10

Probability of false synchronization

-1
10

-2
10

-3
10

conventional1
Conventional
data2
Optimal
data3
Suboptimal
-4
10
0 5 10 15 20 25 30 35 40
Ec / N 0 (dB)

Fig. 9. Probabilities of false synchronization of the conventional, optimal, and suboptimal

schemes in the IEEE 802.15.3a CM4 channel model

since the suboptimal scheme does not use the channel information, it exhibits some
performance degradation compared with the optimal scheme.
Figs. 6-9 show the probabilities of false synchronization of the schemes in the IEEE
802.15.3a channel model. As shown in figures, the proposed schemes have a better per-
formance compared with that of the conventional scheme. Another important observation
is that the performance of the suboptimal scheme is saturated. This is because each corre-
lator output in the IEEE 802.15.3a channel model includes the interference occurred by
the other correlator outputs differently from that in the tapped-delay line channel model.
However, the optimal scheme utilizes all channel information, and thus, provides the best
synchronization performance regardless of the influence of the interference.

5 Conclusion and Future Works

In this paper, we have proposed novel synchronization schemes for UWB systems. We
first derive an optimal scheme based on the ML criterion and then develop a simpler
suboptimal scheme. Simulation results have shown that both proposed schemes have a
better synchronization performance than the conventional scheme. The optimal scheme
has the best performance; however, it requires the channel information. On the other
hand, since the suboptimal scheme does not require the channel information, it has
some performance degradation compared with the optimal scheme.
 Optimal and Suboptimal Synchronization Schemes for Ultra-Wideband Systems 41

In the practical channel environment (i.e., IEEE 802.15.3a channel model), the sub-
optimal scheme cannot overcome the influence of the interference, and thus, its perfor-
mance is saturated. In the future, we will deal with the synchronization scheme robust
to the influence of the interference.

Acknowledgments. This research was supported by the MKE (The Ministry of Knowl-
edge Economy), Korea, under the ITRC (Information Technology Research Center)
support program supervised by the NIPA (National IT Industry Promotion Agency)
(NIPA-2009-(C1090-0902-0005)).

References
1. Win, M.Z., Scholtz, R.A.: Impulse radio: how it works. IEEE Commun. Lett. 2(2), 36–38
(1998)
2. Porcino, D., Hirt, W.: Ultra-wideband radio technology: potential and challenges ahead.
IEEE Commun. Mag. 41(7), 64–74 (2003)
3. Yang, L., Giannakis, G.B.: Ultra-wideband communications: an idea whose time has come.
IEEE Sig. Process. Mag. 21(6), 26–54 (2004)
4. Gezici, S., Kobayashi, H., Poor, H.V., Molisch, A.F.: Performance evaluation of impulse ra-
dio UWB systems with pulse-based polarity randomization. IEEE Trans. Sig. Process. 53(7),
2537–2549 (2005)
5. Aedudodla, S.R., Vijayakumaran, S., Wong, T.F.: Timing acquisition in ultra-wideband com-
munication systems. IEEE Trans. Veh. Technol. 54(5), 1570–1583 (2005)
6. Wu, L., Wu, X., Tian, Z.: Asymptotically optimal receivers with noisy template: design and
comparison with RAKE. IEEE J. Sel. Areas Commun. 24(4), 808–814 (2006)
7. Homier, E.A., Scholts, R.A.: Rapid acquisition of ultra-wideband signals in the dense multi-
path channel. In: Proc. IEEE UWBST, Baltimore, MD, pp. 105–109 (2002)
8. Vijayakumaran, S., Wong, T.F.: A search strategy for ultra-wideband signal acquisition. IEEE
Trans. Commun. 53(12), 2015–2019 (2005)
9. Ramachandran, I., Roy, S.: On acquisition of wideband direct-sequence spread spectrum
signals. IEEE Trans. Commun. 5(6), 1537–1546 (2006)
10. Arias-de-Reyna, E., Acha-Catalina, J.J.: Blind and efficient serial search strategy for ultraw-
ideband signal initial acquisition. IEEE Trans. Veh. Technol. 58(6), 3053–3057 (2009)
11. Molisch, A.F., Foerster, J.R.: Channel models for ultrawideband personal area networks.
IEEE Wireless Commun. 10(6), 14–21 (2003)
12. Molisch, A.F.: Ultrawideband propagation channels - theory, measurement, and modeling.
IEEE Trans. Veh. Technol. 54(5), 1528–1545 (2005)
Partial Information Relaying with Multi-Layered
Superposition Coding

Jingyu Kim and Dong In Kim

School of Information and Communication Engineering, Sungkyunkwan University

Suwon 440-746, Korea
{greengyu,dongin}@skku.edu

Abstract. In cooperative communications, relay usually forwards a full

information of source data as received. Unlike this approach, this paper
proposes a new relaying scheme based on multi-layered superposition
coding with multiple antennas which forwards only partial information
of superposed layer(s) on top of basic layer(s) of the source data. Fast
forwarding of the partial information results in an increase in overall ca-
pacity. Simulation results are presented to show the superiority of the
proposed scheme over full information relaying and direct transmission.

Keywords: Cooperative relaying, multi-layered superposition coding,

multiple antennas, partial information relaying, spectral efficiency.

1 Introduction

Recently there have been many proposals on cooperative relaying as a useful

means to achieve better diversity and reliable link gains. Especially, when source-
to-destination (S-D) link condition is not sufficient enough to reliably transmit
data, it is possible to detour source-to-relay (S-R) and relay-to-destination (R-
D) link. Typically it is more likely to have (S-R) and (R-D) links at line-of-sight
(LOS), and it is assumed that they provide more reliable link gains than (S-D)
link. Under this assumption, relay can play an important role in achieving a
better overall data rate [2]. In half-duplexing mode, conventional two-hop relay-
ing transmission forwards a full information of source data as received, which
requires appreciable time duration for the second-hop transmission. Such longer
second-hop time duration may cause some capacity loss and degrade spectral
efficiency. Although direct transmission (DT) does not need any second-hop
transmission, we can not offer a required overall data rate when (S-D) link is so
weak.
To address the above issue, we propose partial information relaying that uti-
lizes multi-layered superposition coding (MLSC) when multiple antennas are

This research was supported by the MKE (The Ministry of Knowledge Economy),
Korea, under the ITRC (Information Technology Research Center) support pro-
gram supervised by the NIPA(National IT Industry Promotion Agency” (NIPA-
2009-(C1090-0902-0005)).

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 42–51, 2010.

c Springer-Verlag Berlin Heidelberg 2010
 Partial Information Relaying with MLSC 43

available at source, relay and desination nodes. It is reported that partial decode-
&-forward (DF) relaying with superposition coding can increase the capacity [1],
[3], [6], but these contributions considered the partial DF under single antenna
scenario. If multiple antennas are used at source, relay and destination nodes,
it increases the degree-of-freedom to form the partial information by the MLSC
more efficiently so as to increase the overall data rate. Here, the MLSC trans-
lates source data into two types of linearly combined data, such as basic layer(s)
and superposed layer(s), where relay forwards only the superposed layer(s). The
superposed layer is decoded at destination node after relay decodes and for-
wards it. Since the basic layer is not relayed, it is decoded at destination node
after first decoding the superposed layer and successfully cancelling it out. If all
data streams (layers) are superposed ones, it is same as full information relaying
(FIR) via two-hop transmission. In the other case, if all of those streams are
basic ones, it is same as DT. Hence the proposed MLSC can be viewed as the
hybrid of FIR and DT.
If the amount of partial information to be relayed is changed in a controlled
manner, the second-hop time can be adjusted adaptively [3]. As the basic layer
is dominant than the superposed layer, the second-hop time can be made shorter
than usual and the spectral efficiency can be improved. However, if (S-D) link
condition is not sufficient enough to increase the amount of basic layer, the
amount of superposed layer should be increased accordingly. Therefore, how to
split the total power between the two layers depends on individual link gains, and
a low-rate channel state information (CSI), in terms of the signal-to-interference-
plus-noise ratio (SINR) is required at source node to perform the proposed MLSC
for partial information relaying.
The rest of the paper is organized as follows. The MIMO cooperative system
model with MLSC is described in Section 2. Section 3 explains how to set the key
design parameters of MLSC in order to maximize the overall capacity. Simulation
results and discussions are given in Section 4, and then concluding remark in
Section 5.

2 System Model

As shown in Fig.1, there are single source, relay and destination nodes. Each
node has NS , NR , ND antennas, respectively. H0 , H1 and H2 are the channel
matrices of size ND × NS , NR × NS , ND × NR associated with (S-D), (S-R) and
(R-D) links, respectively. Because of half-duplexing mode assumed here, relay
can not transmit and receive the data streams at the same time.
In the first hop, source node broadcasts the signal x to relay and destination
nodes. x is composed of xs and xb where xs are the superposed streams being
broadcast from L ≤ NS source antennas, and xb are the basic streams from the
remaining (NS − L) source antennas. Destination node stores the received signal
in the buffer. Relay node decodes only the superposed streams xs .
In the second hop, relay node forwards xr which is the same as xs if decoding
is correct. Then, destination node first decodes the received signal from relay
44 J. Kim and D.I. Kim

Fig. 1. The wireless channel model with relay

node for obtaining xs , and cancels out from the received signal in the first hop.
After the successive interference cancellation (SIC), destination node is able to
decode the received signal from source node for obtaining xb .
Mathematically, the above transmission and reception can be modeled as fol-
lows: P0 and P1 represent the diagonal power matrices at source and destination
nodes, respectively. n0 , n1 , n2 are additive white Gaussian noise (AWGN) vec-
tors with zero mean and σ02 , σ12 , σ22 variances.

Fig. 2. The illustration of partial information relaying with MLSC

 
xs
x=
xb

yr = H1 P0 x + n1
1st hop :
yd,1 = H0 P0 x + n0
2nd hop : yd,2 = H2 P1 xr + n2
 Partial Information Relaying with MLSC 45

Fig. 2 shows how to forward partial information of the superposed data stream
(L = 1) on top of the basic one when NS = NR = ND = 2.

2.1 MMSE-SIC

The minimum mean-square-error (MMSE) receiver with SIC is employed to

decorrelate inter-stream interferences between basic and superposed layers after
which the basic layer is decoded using the SIC [4], [5]. To evaluate the overall
capacity, we define Rb , Rs and R2 , namely Rb is the achievable data rate of the
basic layer on (S-D) link, Rs the achievable rate of the superposed layer on (S-R)
link, and R2 the channel capacity of (R-D) link as follows:

S −L
N 
NS 
Rb = Rb,m = log2 1 + p0,m hH
0,m
m=1 m=L+1
 −1 
× H0 (m + 1)P0 (m + 1)HH
0 (m + 1) + σ02 IN h0,m (1)


L 
L 
Rs = Rs,m = log2 1 + p1,m hH
1,m
m=1 m=1
 −1 
× H1 (m + 1)P1 (m + 1)HH 2
1 (m + 1) + σ1 IN h1,m (2)


NR
R2 = R2,n
n=1
 
= log2 det IND + H2 P1 HH
2 (3)

In the above hi,m denotes the mth column of matrix Hi (i = 0, 1) with N rows
and M columns, Pi (m) = diag{pi,m , pi,m+1 , . . . , pi,M−1 , pi,M }, and Hi (m) is the
N × (M − m + 1) matrix [hi,m , hm+1 , . . . , hi,M−1 , hi,M ]. The superscript (·)H
denotes the hermitian transpose operation, and IN is the identity matrix of size
N × N . Note that Rb is evaluated under the assumption of perfect SIC while Rs
in the presence of the inter-stream interferences from basic layers.

2.2 Power Allocation

The source data is divided into a number of superposed layers and the rest of
basic layers with a total power constraint of PT . If we define the power division
factor to allocate the total power to a set of superposed layers and the rest of
basic layers, they are allocated the powers α × PT and (1 − α) × PT , respectively.
The MLSC should share a total of source antennas NS among the superposed
and basic layers, and if L source antennas are used for the superposed layers
where 0 ≤ L ≤ NS , the remaining (NS − L) antennas are assigned to the basic
layers. Assuming that the SINR estimates are available at source and relay nodes,
46 J. Kim and D.I. Kim

an optimal power allocation based on water-filling algorithm can be performed.

With the total power constraint applied, the above power allocation can be
formulated as
⎧ +
⎨ 1sc − σ12
λ |h1,m |2 for the superposition layers (0 ≤ m ≤ L)
P0,m =  0 + (4)
⎩ 1 − σ12
for the basic layers (L + 1 ≤ m ≤ Ns )
λb0 |h1,m |2
 +
1 σ22
P1,n = − (5)
λ1 |h2,n |2

L 
NS 
NR
subject to P0,m ≤ α × PT , P0,m ≤ (1 − α) × PT , P1,n ≤ PT (6)
m=1 m=L+1 n=1

where (x)+ = max(0, x) and the Lagrange multipliers λsc b

0 , λ0 and λ1 are chosen
to meet the total power constraints.

2.3 Overall Capacity

Once Rb , Rs and R2 are evaluated, we can derive the overall capacity Rsc for
the partial information relaying with MLSC. A total number of bits that source
transmits to destination is N × (Rb + Rs ) where N denotes the duration of the
first hop. Since relay forwards only N × Rs , the duration of the second hop can
be N2 = NRR2s so that the overall capacity can be derived as [3]

N (Rb + Rs ) (Rb + Rs )R2

Rsc = = (7)
N + N2 R2 + Rs

To compare Rsc with conventional schemes, such as full information relaying

(FIR) via two-hop transmission and direct transmission (DT), the corresponding
overall rates are evaluated as
 
R0 = log2 det IND + H0 P0 HH
0 (8)
 
R1 = log2 det INR + H1 P0 HH
1 (9)
 
R02 = log2 det IND + H0 P0 HH
0 + H2 P1 H2
H
(10)
R1 × R02
RF IR = (11)
R1 + R02
RDT = R0 (12)

where R0 is the capacity of (S-D) link, R1 the capacity of (S-R) link, and R02
the capacity for maximal ratio combining (MRC) of (S-D) and (R-D) links in
the second hop.
 Partial Information Relaying with MLSC 47

3 Precoding and Power Division

3.1 Precoding

In (2) we notice that relay node decodes xs in the presence of the inter-stream
interferences from xb . Given that relay node does not need to decode xb , a
precoding matrix to decorrelate the interferences may be inserted to increase
the capacity Rsc . For this the singular value decomposition is applied to H1 as

H1 = U1 Σ1 V1H (13)

where U1 and V1 are unitary matrices and Σ1 is a diagonal matrix as defined

in [7]. To decorrelate the interferences from basic layer(s), the precoding matrix
at source node could be set to V1 . The resulting channel matrices Ĥi (i = 0, 1)
are formulated as

Ĥi = Hi V1 (14)

The overall capacity with precoding is evaluated after Hi is replaced by Ĥi in

(1) and (2). Note that precoding at relay node is not performed since the power
allocation via water-filling boosts up the capacity of (R-D) link.

3.2 Choice of the Power Division Factor

The overall capacity is adjusted by two variables α and L. α is a power division

factor between superposition and basic layers and can be selected any real value
between 0 and 1. L is the number of transmit antennas for superposed layers and
modulates transmit diversity of MLSC system and can be selected any discrete
value between 0 and Ns . Since our objective is to maximize the overall capacity,
α and L can be found by solving following optimization problem.

 L Ns 
m=1 Rs,m (α) + m=L+1 Rb,m (1 − α) R2
max Rsc (α, L) = L (15)
α,L
m=1 Rs,m (α) + R2

To lower the complexity of above equation, reduce the optimization problem

from two variables to one variable. Notice that L is the number of transmit
antennas for the superposition layers and it can be decided by the allocated
power properly. Assume that large alpha (e.g. α = 0.8) is chosen. Because it
indicates that more power are allocated to the superposition layers than the
basic layers, antennas for the superposition layers should be increased. Surely
opposition can be established. For easy explanation, when there are 4 antennas
at the source, L can be determined as follows.

When there are 4 antennas at the source node

α=0→L=0
48 J. Kim and D.I. Kim

0 < α < 0.375 → L = 1

0.375 ≤ α < 0.625 → L = 2
0.625 ≤ α < 1 → L = 3
α=1→L=4
If Ns is another value (e.g. 2,3,5,...), we can set the number L as similar way.
Since L is obtained by α, optimizing problem (15) is a function of only α. To
determine the optimal α, the above problem is simplified as follows.

arg max Rsc (α)

α
subject to 0 ≤ α ≤ 1 (16)

Since this optimization problem belongs to a convex problem, this is solved

by some numerical search algorithm (e.g. bisection method) or exhaustively [8].
We have already assumed that source knows CSI of (S-D) and (S-R). But
source usually does not know exact (R-D) channel gain. If source only knows
statistical information of (R-D) channel, R2 should be replaced with the expected
value R̃2 . Although R̃2 is not the same as real R2 value, using it as suboptimal
does not cause significant error to find optimal α.

4 Simulation Results
For the simulation we consider the multiple antennas cooperation system ( Ns
= Nr = Nd = 4 antennas ). SNR is defined by the ratio of total power PT and
noise variance σ 2 . Fig. 3 shows the simulation result when SNRs of (S-R) and
(R-D) are all 20dB. The result of MLSC is compared with results of FIR and
DT in this graph. It is seen that MLSC shows much better performance than
other schemes.
Especially when SNR of (S-D) is around 10dB, performance gap of MLSC
and the others is significant. To explain the reason, we illustrate Fig. 6 showing
a trend of optimized α value. It is seen that the value of α gets smaller as the
SNR of (S-D) approaches high value. From this result, MLSC operates similary
as DT at the low SNR of (S-D) and FIR at high. Because small alpha indicates
that MLSC operates like DT, however large alpha does like FIR. Around the
middle of SNR of (S-D) axis, the value of alpha is close to 0.5. This result means
that when channel gain of (S-D) is some normal value (in this case 6-12dB
region), signal power of the superposed and basic layers are similar. Hence the
duration of second hop time is adaptively modulated and that causes significant
performance difference between MLSC and the other schemes.
For the asymmetric link, we illustrate Fig. 4 and Fig. 5. In Fig. 4, SNRs of
(S-R) and (R-D) are 15 and 20dB respectively and the capacity gain of MLSC at
low SNR region shows more prominent compared with Fig. 3 and Fig. 5. From
this result, MLSC can be more useful when SNRs of (S-D) and (S-R) are not so
different.
 Partial Information Relaying with MLSC 49

SNR (S−R) = SNR (R−D) = 20dB

22

20

18

16

Average Throughput 14

12

10

6
MLSC with Precoding
MLSC without Precoding)
4
Full Information Relaying
Direct Transmission
2
0 2 4 6 8 10 12 14 16 18 20
SNR (S−D)

Fig. 3. Average capacity when both SNRs of (S-R) and (R-D) are 20dB

SNR (S−R) = 15dB, SNR (R−D) = 20dB

18

16

14
Average Throughput

12

10

6
MLSC with precoding
4
MLSC without precoding
Full Information Relaying
Direct Transmission
2
0 5 10 15
SNR (S−D) [dB]

Fig. 4. Average capacity when SNRs (S-R) of and (R-D) are 15 and 20dB respectively

Furthermore, it is apparent that MLSC with precoding outperforms without

it. It is seen that decorrelating interferences from basic layer increases overall
capacity Rsc . This gap is more prominent when SNR of (S-D) is around 8dB
because more interferences exist without precoding.
As you see the result, the capacity of FIR is mostly affected by relay link
channels (S-R) and (R-D). On the other hand, the capacity of DT is determined
by (S-D) channel only. So the key idea of this proposed scheme is to use all the
channels adaptively depending on instant channel gains. So as we see the result,
the proposed scheme named MLSC is very useful to achieve high transmission
rate.
50 J. Kim and D.I. Kim

SNR (S−R) = 20dB, SNR (R−D) = 15dB

16

14

12

Average Throughput
10

MLSC with Precoding

4 MLSC without Precoding
Full Information Relaying
Direct Transmission
2
0 5 10 15
SNR (S−D) [dB]

Fig. 5. Average capacity when SNRs (S-R) of and (R-D) are 20 and 15dB respectively

SNR (S−R) = SNR (R−D) = 20dB

0.9

0.8

0.7
Average alpha vaule

0.6

0.5

0.4

0.3

0.2

0.1

0
0 5 10 15 20
SNR (S−D) [dB]

Fig. 6. Average α value versus SNR of (S-D) when both SNRs of (S-R) and (R-D) are
20dB

5 Conclusion

We have proposed the partial information relaying based on MLSC where relay
forwards only a part of source data. By controlling the amount of information to
be relayed, considering the asymmetric link conditions, the second-hop time at
relay can be adjusted adaptively. Although finding an optimized power division
factor causes some overhead to the proposed relaying scheme, we can achieve
much improved overall capacity. Moreover, it was shown that the precoding
method can be effective to decorrelate the inter-stream interference and yield bet-
ter capacity. Because we have only considered a single relay node, multiple-relay
 Partial Information Relaying with MLSC 51

cooperative system based on the MLSC can be a future work. Here, which and
how many links are allocated to the basic and superposed layers would be an
interesting issue to be further addressed.

References
1. Yuksel, M., Erkip, E.: Broadcast Strategies for the Fading Relay Channel. In: IEEE
Proc. MILCOM 2004, October 2004, vol. 2, pp. 1060–1065 (2004)
2. Laneman, J.N., Tse, D.N.C.: Cooperative diversity in wireless networks: efficient
protocols and outage behavior. IEEE Trans. Inform. Theory 50, 3062–3080 (2004)
3. Popovski, P., de Carvalho, E.: Improving the rates in wireless relay systems through
superposition coding. IEEE Trans. Wireless Commun. 7, 4831–4836 (2008)
4. Wolniansky, P.W., Foscini, G.J., Golden, G.D., Valenzuelar, R.A.: V-BLAST: an
architecture for realizing very high data rates over the rich-scattering wireless chan-
nel. In: Proc. of URSI International Symposium on Signals, Systems and Electronics
(ISSSE 1998), Pisa, Italy, September 1998, pp. 295–300 (1998)
5. Chung, S.T., Lozano, A., Huang, H.C., Sutivong, A., Cioffi, J.M.: Approacing the
MIMO Capacity with a Low-Rate Feedback Channel in V-BLAST. EURASIP Jour-
nal on Applied Signal Processing 5, 762–771 (2004)
6. Goparaju, A.K., Wei, S., Liu, Y.: On superposition coding based cooperative diver-
sity schemes. In: Proc. 65th IEEE Vehicular Technology Conference (VTC Spring
2007), Dublin, Ireland, April 2007, pp. 1046–1050 (2007)
7. Andrea Goldsmith: Wireless Communication, Cambridge (2005)
8. Boyd, S., Vandenberghe, L.: Convex Optimization, Cambridge (2003)
 Performance Measurement of the Hybrid
Prefetch Scheme on Vehicular Telematics
Networks

Junghoon Lee1 , Gyung-Leen Park1, , Youngshin Hong1 ,
In-Hye Shin1 , and Sang Joon Lee2
1
Dept. of Computer Science and Statistics, 2 Dept. of Computer Engineering
Jeju National University, 690-756, Jeju Do, Republic of Korea
{jhlee,glpark,yshong,ihshin76,sjlee}@jejunu.ac.kr

Abstract. This paper measures and analyzes the performance of the

hybrid prefetch scheme for gateways on the vehicular telematics network
where every information is indexed by the link (street) ID, via simula-
tion using a discrete event simulator. Combining the classic LRU and
FAR techniques, the hybrid scheme groups links according to whether
they are referenced during the given time interval or not, orders by the
Euclidean distance in each group, and fetches the referenced set first.
The exensive experiment results demonstrate that 1) the hybrid scheme
can improve the request-level instant reply ratio by up to 20.3 %, over-
coming 20 % loss of record-level hit ratio, compared with LRU, 2) this
scheme outperforms other spatial-locality-based schemes at all satisfia-
bility ranges and LRU when the satisfiability is above 0.8, also showing
stable reply ratio for the practical time tolerance range, 3) it quickly
stablizes after the prefetch memory initialization and is less affected by
the hourly data access pattern change, maintaining steady instant reply
ratio for the number of records in each request.

1 Introduction

The vehicular telematics network makes it possible to provide an information

service to drivers or passengers by keeping vehicles connected to the global net-
work even while they are driving. The most common application is data retrieval.
Traffic information, weather update, parking lot availability, local restaurant in-
formation, and the like is provided to the user [1]. A vehicle connects to a gateway
such as the RSU (Road Side Unit), which relays information from the server to
mobile hosts, maintaining a copy of original data created at the server [2]. Each
server usually updates its information at regular intervals, making the data copy
in the local memory, or cache, of gateways get stale. Gateways must reload the
data from each information server. If the information a vehicle requested is in the

This research was supported by the MKE, Korea, under the ITRC support program
supervised by the NIPA. (NIPA-2009-C1090-0902-0040)).

Corresponding author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 52–62, 2010.

c Springer-Verlag Berlin Heidelberg 2010
 Performance Measurement of the Hybrid Prefetch Scheme 53

gateway and valid, the gateway can send it immediately. Otherwise, the gateway
must refresh its local copy, either proactively or reactively.
Prefetching is a technique in which a data consumer proactively fetches from
the server the data item that is predicted to be needed in the near future [3]. The
main concern of the prefetch scheme is necessarily how to decide the prefetch
order. Not to mention, the data item many vehicles commonly need must be
prefetched first. However, the size of such information is too large to refresh the
whole set. For example, the amount of traffic information for a city or country
is too huge for a gateway to reload all data items for each update period. In the
location-dependent information service, drivers around a specific gateway are
generally interested in such information as train schedule at nearby terminals
and up-to-date traffic conditions of the route to their destinations [4]. If multiple
vehicles access the same item, the prefetch policy for the gateway must take into
account the spatial and temporal locality between the requests from different
vehicles.
Prefetch schemes are basically limited by both prediction accuracy and the
penalty for misprediction [5]. The accuracy depends on the efficiency of a prefetch
policy and the size of cache. However, considering the cost and size of installable
memory these days, the cache size is not a restriction any more. As contrast, as
the amount of data records for a city is very large, it is impossible for each gate-
way to download from the server immediately, while all records are updated at
the server almost simultaneously. Hence, the refresh order most impacts the data
validity in the gateway. The prefetch order is based on the prediction which item
is likely to be requested most and soon. In addition, the penalty of misprediction
is definitely the waste of network bandwidth, or sometimes communication cost,
and the increase in response time.
There have been a lot of cache/prefetch strategies for single mobile host, and
their concern mainly lie in mobility prediction and personalization. As contrast,
prefetch policies for the gateway must be based on the common access pattern
from the vehicles in the vicinity [6]. So, prefetch in gateways must consider
the temporal and spatial locality of data access, possibly exploiting classical
prefetch policies such as LRU (Least Recently Used) and FAR (Furthest Away
Replacement) [7]. Data prefetching and caching have many features in common,
and each policy in cache management has its counterpart in the prefetch policy.
First, in cache management, LRU evicts the item that is least recently used.
This name can be sustained in the prefetch domain if we change first to last.
That is, LRU in the prefetch policy means that the data item least recently used
will be updated last. Similarly, FAR in cache management replaces the item
which is farthest away from the reference point. In the prefetch, the item that
is closest to a reference point will be fetched first. After all, LRU and FAR can
be interchangeably used in cache and prefetch.
Combining LRU and FAR, our previous paper has proposed a hybrid prefetch
scheme which groups the data records into two sets according to whether a record
was referenced during the given previous interval or not [8]. In each set, records
are ordered by the Euclidean distance from the gateway. The hybrid scheme
54 J. Lee et al.

can fetch the closer item first in each group. Even though the record-level hit
ratio is less than the LRU scheme, the hybrid scheme can enhance the request-
level instant reply ratio. However, user requirement has a variety, not just the
basic hit or reply ratio. In this regard, this paper is to measure and analyze the
performance of the hybrid prefetch scheme in terms of diverse user satisfiability
criteria. For example, a user satisfies if at least 90 % of information is instantly
available or a user can tolerate up to 10 second delay.
This paper is organized as follows: After defining the problem in Section 1,
Section 2 describes the background of this paper and related work. Section 3
explains the experimental setup such as the target network model and the
data access pattern. After the performance measurement results are analyzed in
Section 4, Section 5 summarizes and concludes this paper with a brief introduc-
tion of future work.

2 Background and Related Work

2.1 Classic and Hybrid Prefetch Schemes
This section introduces some previous work on data prefetching in the mobile
network, focusing on the hybrid prefetch scheme. A prefetch scheme is built on
top of prediction on the future data access from a single or multiple clients.
Prediction depends on a heuristic which has been intensively developed in the
cache management domain, focusing on how to take advantage of the locality
of data access. One of the most basic cache schemes is LRU and this scheme is
known to best take advantage of temporal locality, that is, a recently referenced
item will be referenced again soon. In the mobile network, spatial locality has
been also considered to account for the characteristic that an item near the
recently referenced one will be referenced soon.
Let’s assume a service vehicles retrieve information on all links comprising
each trip. The information of the link many vehicles commonly take is necessarily
referenced frequently. Intuitively, more vehicles will likely take the road segment
closer to a gateway, as all trips pass the road near to gateway. However, not all of
close roads are taken by drivers. Some roads, even if they are close to the gateway,
may lead to a dead end or always have heavy traffic, so the navigation software
does not select them. FAR schemes cannot filter such road segments, possibly
creating not a little misprediction. However, the prefetch module can calculate
the distance, namely, either Euclidean or network distance, in priori, so this
heuristic has very small run-time overhead. Oppositely, LRU is known to have
highest hit ratio for most cases of cache management, however, it cannot give
precedence to the road segment closer to the gateway. In the road network, a road
segment, especially far away from the gateway, which has been just referenced,
cannot be referenced for a long time.
Based on the observation for LRU and FAR, the hybrid prefetch scheme
groups the links into two sets [8]. The first is the group of links referenced
during the given previous interval, say, 1 hour, and the other, those not refer-
enced. This categorization takes advantage of LRU. Actually, in the mid-size
 Performance Measurement of the Hybrid Prefetch Scheme 55

city level scenario, for a gateway on which most traffic is concentrated, the links
taken by the vehicle starting from a specific gateway location for a whole day
are below 30 %. In each set, links are ordered by the Euclidean distance to
take into account spatial locality. The first set is able to gather the links that
many drivers commonly take and will be prefetched prior to the second set. The
link far away from the gateway will be fetched later within this group. In basic
LRU, when two links are not referenced yet, it is not possible to have an item
closer to the gateway fetched first. The hybrid scheme can first fetch the closer
item to account for the higher probability to be referenced within the second
group.
For a car to reach a place far away, it must start from a link near to it. In
the vehicular telematics network, the movement of a vehicle is restricted to the
road network. So, the distance between the two points can be represented either
by the Euclidean distance or by the network distance. The Euclidean distance is
the length of the straight line connecting the two end points, while the network
distance is the sum of each road segment distance along the path. In the road
network, even if a location is very close in the Euclidean distance, its network
length can be very large when there is no direct path from the start position.
FAR-E denotes the policy which prefetches first the item closest to the given
point in the Euclidean distance, while FAR-N uses the network distance instead
of the Euclidean distance. Spatial locality is well reflected by FAR schemes.

2.2 Other Prefetch/Cache Schemes

Besides, there are some important work on prefetch for mobile host case. To
begin with, a power-controlled prefetching scheme has been designed, focusing
on how to handle the intermittent connectivity of radio links from data servers
to mobile clients and how to manage power and bandwidth resources [9]. The
system decides the item to prefetch based on the future request estimation, and
this estimation employs a statistical profile which models the user and channel
behaviors for constant time intervals based on a Markovian model. For each
time slot, the global system state is defined by user, channel interference, and
terminal buffer states. This scheme focused on which item to evict from the local
buffer of a mobile host.
Liang and et al. considered mobile users in a two-tier network, comprised of
WLANs surrounded by a ubiquitous cellular network [4]. The authors claimed
that successfully prefetched data, when a user is about to leave the coverage
area, potentially save expensive future cellular network access. Users may roam
anywhere and are not constrained to any one network. The mobility pattern of
each vehicle is known in priori, making it possible to estimate which networks
a vehicle may roam to in the near future. This work defined the total cost for
accessing a document as the sum of access cost and the penalty for access delay.
A mathematical model was developed for the cost function, by which the WLAN
gateway decides whether to prefetch an item and the prefetch order.
56 J. Lee et al.

P. Deshpande et al. have proposed a cooperative prefetch scheme based on

the location prediction of vehicles [10]. Their scheme begins with an argument
that people drive on familiar routes frequently, so the mobility and connectivity
during their drives can be predicted quite accurately by their driving history
[11]. By means of this prediction, multiple roadside APs on the route of a vehicle
cooperatively download a portion of a large object. When the vehicle approaches
and connects to an AP, the prefetched data can be directly provided. All these
schemes are developed based on the prediction of the future location of a vehicle.
Additionally, some strategies integrate an index mechanism to cache/prefetch
policies. First, a semantic cache scheme builds a Voronoi diagram and stores
the data item along with its semantic description in the mobile client [12]. By
this semantic description, it is possible to provide partial answers to the given
query, even if it does not match exactly. This scheme works quite well for the
continuous queries which keeps receiving updated information while a vehicle is
moving.

3 Experiment Setup

3.1 Network Architecture and Information Service

Prefetching is used in widely different contexts in computer systems to speed up

data retrieval. So, the design must be preceded by the specification of the target
network architecture and service. This paper shares the underlying framework
with our previous work [8]. Figure 1 illustrates the vehicular telematics network
architecture. Static gateways are installed at the place with high vehicle density.
A vehicle can connect to a gateway by means of a pre-defined network protocol
such as IEEE 802.11, DSRC (Dedicated Short Range Communication), Zigbee,
and so on [13]. Each gateway has a stable connection to the global network, be
it wired or wireless. Servers reside in the Internet domain and can send data
to the gateway via the Internet protocol, while the service provider creates and
provisions useful contents.

Server 1 Server 2 Server 3

Global network

Fig. 1. Vehicular telematics service network

 Performance Measurement of the Hybrid Prefetch Scheme 57

Apparently, one of the most frequently used services for the telematics network
is path finding. In the advanced telematics service scenario, besides the simple
route selection, a driver may want to know detailed live information along the
path, including traffic condition update, seat availability at the roadside restau-
rant, and an instant discount coupon of shopping stores that a driver may visit
during his travel. With the data, he can make a detailed plan or even try to
another route. Those data are stored in the central information server, period-
ically updated in general, and retrieved by the vehicle by way of gateways. In
the meantime, all information stored in the information server is associated and
indexed by the link ID, so the link ID is the search key to the local cache of
gateways as well as server database. Link ID, which is analogous to the street
name, is a reasonable key index, as most POI (Point Of Interest) information is
associated with a street.
In this service scenario, a vehicle first decides a path plan by the start and the
destination using a path finding algorithm such as A* embedded in the telematics
device [14]. Then, an entire route consist of link IDs is submitted to the gateway,
which searches its local cache for each link ID and waits for missing items to
be valid. The gateway cannot reply until every requested link record is ready.
Hence, the more the data miss, the poorer the response time, as more interaction
with the remote server is needed. The vehicle can request the query only when
it can reach a gateway and receive the reply when it remains connected until the
gateway returns the reply. In depends on the driver’s decision whether he will
keep waiting when the response time gets too long. However, the driver at least
knows the unavailability of information before his start.

3.2 Access Pattern Data

The realistic data access pattern is very important for the performance assess-
ment of data prefetch schemes. This paper employs the location history data
collected by the the Jeju taxi telematics system which constantly tracks the lo-
cation of vehicles [15]. Each vehicle reports its GPS reading every minute to the
central server via the cellular network. During the 2.3 months’ test period which
has the largest number of enrollments, up to 1.3 M records have been collected.
The road network of Jeju area, being represented as a graph data structure, con-
sists of about 17,000 nodes, or intersections and about 25,000 links, or the road
that connects two nodes. Our analysis system processed the commercially used
digital map to make a graph structure and detailed road shape easily accessible.
We can estimate the route taken by a vehicle and assume that this vehicle re-
trieves the information on the links belonging to the route before it has started
its trip.
The experiment is based on the trajectory of each travel obtained from the
location history data. The path, the series of links, taken by a vehicle is submitted
to the data retrieval service. The snapshot locations of a vehicle do not exactly
represent the path actually taken, however, it is possible to find start and end
points of a trip by investigating the change of taxi status connected to the
tachometer. By applying the well-known A* algorithm [14] with the start and
58 J. Lee et al.

destination of a trip, the whole set of links involved in the path taken by the car
can be generated. Next, in deciding the location of a gateway, the start positions
of all trips are first examined to find the area which has the maximum number
of starts. Actually, the city hall area is such a spot. The experiment focuses on
this single gateway, as gateways run independently in general. The simulation
takes 2,190 trips (94,546 records) that start on one day from the location within
300 km radius from the gateway location, considering the transmission range of
wireless communication interface.

4 Performance Measurement
This section measures and assesses the performance of the hybrid prefetch scheme
via simulation using SMPL [16] which provides a lot of functions and libraries for
discrete event scheduling, easily combined with the commonly used compilers. The
experiment first measures the hit ratio and the instant response ratio for the pro-
posed scheme, comparing with LRU, FAR-N, FAR-E, and random prefetch de-
scribed in Section 2. Then, the reply ratio according to the satisfiability and the
time tolerance will be measured. Finally, the instant reply ratio is further analized
according to the time flow and the hop count, or the length of each request.
To focus on the prefetch strategy, the experiment also assumes that there
is just one server and its contents are updated every ten minutes. It can be
necessarily much shorter. After all, request arrivals, periodic updates at the
server, and record refreshes at the gateway are integrated into a discrete event
scheduler. In addition, the refresh time of each record distributes exponentially,
considering the server-side load as well as the network delay between the gateway
and an information server. To parameterize the refresh time, we define update
load. If this value is 1.0, it takes one update period to refresh all link records,
that is, it takes 10 minutes to refresh 25,000 records in the Jeju city scenario. If
it is 5.0, only 20 % of entire link information can be downloaded to a gateway.

1 1
"Random"
0.8 0.8 "FAR-N"
"FAR-E"
Instant reply ratio

"LRU"
"Hybrid"
0.6 0.6
Hit ratio

0.4 0.4
"Random"
"FAR-N"
0.2 "FAR-E" 0.2
"LRU"
"Hybrid"
0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Update load Update load

Fig. 2. Hit ratio Fig. 3. Instant reply ratio

 Performance Measurement of the Hybrid Prefetch Scheme 59

Figure 2 plots the record level hit ratio for each policy. For all schemes, the hit
ratio decreases according to the increase of update load. LRU shows the highest
hit ratio and decreases slowly according to the load increase, as it can best reflect
temporal locality and refreshes first the link referenced at least once. Even in
the road network, there exists a link the drivers are most likely to take. The hit
ratio of the random scheme gets down to zero when the update load reaches 3.0,
namely, only 1/3 of link data can be refreshed on average. The hybrid scheme is
the next to LRU, as it can partially exploit the temporal locality. The hit ratio
of the hybrid scheme is better than that of FAR-E by up to 23.1 % and worse
than that of LRU by up to 20 % when the update load is 5.0.
Figure 3 shows the instant reply ratio for the respective strategies. The re-
sponsiveness of a gateway depends on how many link data are already refreshed
when a request arrives. The gateway can reply instantly when all the requested
link data are valid. This rate can be considered as the request-level hit ratio.
Even though LRU has the highest record-level hit ratio, the hybrid scheme shows
a better instant reply ratio by up to 20.3 % when the update load is 1.5. The gap
gets smaller by up to 1.9 % along with the increase of the update load. Other
spatial-locality-based schemes such as FAR-E and FAR-N also show better per-
formance than LRU until the update load becomes 1.5 and 3.5, respectively. We
can recognize that geographical affinity can better cover the whole link set in a
trip than temporal locality. The gap between FAR-E and the hybrid scheme is
13.1 % at maximum.
Figure 4 plots the reply ratio according to the satisfiability, which denotes
the bound a user satisfies. For the satisfiability of 0.9, a user is satisfied with
the result if 90 % of link information is returned instantly. This scenario is more
practical to the real-life service model. The experiment changes the satisfiability
from 0.6 to 1.0, and the update load is set to 0.5. The hybrid scheme shows the
best performance when the satisfiability lies between 0.8 and 1.0. However, LRU,
having the highest link-level hit ratio, gets better when the satisfiability goes
below 0.8. The hybrid scheme outperforms other spatial-locality-based schemes
for all ranges. All schemes except LRU show a linear increase in the reply ratio.

1 1

0.9 0.9

0.8 0.8
Reply ratio

Reply ratio

0.7 0.7 "Random"

"FAR-N"
0.6 "Random" 0.6 "FAR-E"
"FAR-N" "LRU"
"FAR-E" "Hybrid"
0.5 "LRU" 0.5
"Hybrid"
0.4 0.4
1 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 2 4 6 8 10 12 14 16 18 20
Satistiability Time tolerance (sec)

Fig. 4. Effect of satisfiability Fig. 5. Effect fo delay tolerance

60 J. Lee et al.

We think that the satisfiability is generally set to be above 0.9 for a user to get
sufficient information for his/her route.
Figure 5 plots the reply ratio according to the time tolerance. If the tolerance
is 2 seconds, a user will be satisfied with the result if he/she receives whole data
records within 2 seconds. Here, the update load is set to 0.5 again. Interestingly,
Figure 5 shows that the time tolerance has no significant influence to the instant
reply ratio. This result indicates that refreshing the core set is most important
to the performance. Anyway, the hybrid scheme shows the best instant reply
ratio for all ranges, particularly outperforming FAR-E by up to 1.2% in this
parameter setup.

180 1
"Requests"
160
140 0.9
Number of requests

Instant reply ratio

120
0.8
100
80
0.7
60 "Random"
"FAR-N"
40 0.6 "FAR-E"
"LRU"
20 "Hybrid"
0 0.5
0 5 10 15 20 0 5 10 15 20
Hour Hour

Fig. 6. Hourly number of requests Fig. 7. Hourly instant reply ratio

Figure 6 and Figure 7 show the hourly instant reply ratio. The data access
pattern has an important effect on the hit ratio and also the instant reply ratio.
In our information service model, a vehicle requests a data record according
to the destination of a trip. When trips have a common destination or more
common route pieces, the instant reply ratio will be better. For example, during
the commute, many vehicles head to an office area in the downtown and the
residential area. Figure 6 shows the hourly change in the number of requests or
trips. We can find some peaks on two commute times, namely just before and
after the office hour around 9 and around 18, respectively, in addition to the
shopping hours around 14. Around 5 o’clock, there are just 20 trips, and the
performance in this period is a little bit meaningless.
Figure 7 traces the change of the hourly instant reply ratio. The initial reply
ratio is relatively low as just a few previous access history data is available.
So, the hit ratios of FAR-E and hybrid schemes are almost same, and so are
those of random and LRU schemes. As the time passes by, each curve shows
different behavior, but he hybrid scheme seems to be less affected by the change
of access pattern. For 10 o’clock, just after the morning commute, the instant
reply ratio decrease a little bit because the destination of each trip is quite
different. Anyway, each scheme shows a slight drop in the reply ratio when
the traffic pattern changes, for example, the morning and evening commute,
shopping time, and personal time after office hour. The hybrid scheme can reach
 Performance Measurement of the Hybrid Prefetch Scheme 61

600 1
"Requests"
500
0.9

Instant reply ratio

400
Requests

0.8
300
0.7
200 "Random"
"FAR-N"
0.6 "FAR-E"
100 "LRU"
"Hybrid"
0 0.5
0 20 40 60 80 100 0 20 40 60 80 100
Hops (Number of records) Hops (Number of records)

Fig. 8. Hourly number of requests Fig. 9. Hourly instant reply ratio

the stable status after initialization. While LRU shows the 22.6 % of difference
in the instant reply ratio, the hybrid scheme shows just 13.2 % difference.
Figure 8 and Figure 9 plot the effect of hop counts, which is analogous to the
number of records in a request. The instant reply ratio is expected to decrease
when a request has more links, as the gateway can reply instantly only if all of
link records are valid. Among 2,190 requests, more than 500 requests contains 30
- 40 records, and the number of records in a request is concentrated in the range
of 20 through 50. 50 requests has less than 10 links, while 100 requests consists of
more than 100 hops, which means a vehicle must have taken a long trip. Figure
9 shows instant reply ratio according to the hop counts. Here, the update load
is set to 0.5. We can recognize that random and LRU schemes are more affected
by the number of records. On the contrary, the instant reply ratios of hybrid,
FAR-N, FAR-E schemes do not significantly change until the hop count reaches
70. Hence, the number of records in a request is not a critical factor in the spatial
locality-based prefetch schemes.

5 Conclusion
This paper has measured and analyzed the performance of the hybrid prefetch
scheme for gateways on the vehicular telematics network via simulation using
SMPL. The target service indexes every information by the link ID, analogous to
the street ID or name, and updates at reguarl intervals. To take advantage of both
LRU and FAR techniques, the hybrid scheme groups links according to whether
they are referenced during the given time interval or not, orders by the Euclidean
distance in each group, and fetches the referenced set first. The experiment
results first have revealed that the hybrid scheme can improve the request-level
instant reply ratio by up to 20.3 %, overcoming 20 % loss of record-level hit ratio,
compared with LRU. Second, this scheme outperforms other spatial-locality-
based schemes at all satisfiability ranges and LRU when the satisfiability is
above 0.8, also showing stable behavior for the practical time tolerance range.
Third and finally, it quickly stablizes after the prefetch memory initialization
and is less affected by the data access pattern change, demonstrating steady
instant reply ratio for the number of records in each request.
62 J. Lee et al.

As future work, we are to design a cooperative prefetch scheme that can

balance the workload of each gateway, as load balancing can improve prefetch
performance such as responsiveness and reply quality in most distributed appli-
cations [17]. To this end, localizing and identifying the core access set for each
gateway is most important.

References
1. Said, E., Omar, E., Robert, L.: Data prefetching algorithm in mobile environments.
European Journal of Scientific Research 28, 478–491 (2009)
2. Yu, B., Gong, J., Xu, C.: Data aggregation and roadside unit placement for a vanet
traffic information system. In: ACM VANET, pp. 49–57 (2008)
3. Denko, M.: Cooperative data caching and prefetching in wireless ad hoc networks.
International Journal of Business Data Communications and Networking 3, 1–15
(2007)
4. Liang, B., Drew, S., Wang, D.: Performance of multiuser network-aware prefetching
in heterogeneous wireless systems. Wireless Networks 15, 99–110 (2009)
5. Sato, K., Koita, T., Fukuta, A.: Broadcasted location-aware data cache for vehic-
ular applications. EURASIP Journal on Embedded Systems (2007)
6. Lee, J., Shin, I., Park, G., Kim, I., Rhee, Y.: Design of a cache management scheme
for gateways on the vehicular telematics network. In: Gervasi, O., Taniar, D., Mur-
gante, B., Laganà, A., Mun, Y., Gavrilova, M.L. (eds.) Computational Science and
Its Applications – ICCSA 2009. LNCS, vol. 5592, pp. 831–840. Springer, Heidelberg
(2009)
7. Zheng, B., Xu, J., Lee, D.: Cache invalidation and replacement strategies for
location-dependent data in mobile environments. IEEE Transactions on Comput-
ers 51, 1141–1153 (2002)
8. Lee, J., Park, G., Kim, S., Kim, H., Shin, S.: A hybrid prefetch policy for the retrieval
of line-associated information on vehicular networks. In: ACM SAC (2010) (accepted)
9. Gitzenis, S., Bambos, N.: Efficient data prefetching for power-controlled wireless
packet networks. In: International Conference on Mobile and Ubiquitous Systems:
Networking and Services, pp. 64–73 (2004)
10. Deshpande, P., Kashyap, A., Sung, C., Das, S.: Predictive methods for improved
vehicular WiFi access. In: ACM SIGMOBILE, pp. 263–276 (2009)
11. Balasubramanian, A., Mahajan, R., Venkataramani, A., Zahorjan, J.: Interactive
WiFi connectivity for moving vehicles. In: ACM SIGCOMM, pp. 427–438 (2009)
12. Zheng, B., Lee, W., Lee, D.: On semantic caching and query scheduling for mobile
nearest-neighbor search. Wireless Networks 10, 653–664 (2004)
13. Society of Automotive Engineers: Dedicated short range communication message
set dictionary. Tech. Rep. Standard J2735, SAE (2006)
14. Goldberg, A., Kaplan, H., Werneck, R.: Reach for A*: Efficient point-to-point
shortest path algorithms. MSR-TR-2005-132. Microsoft (2005)
15. Lee, J., Park, G., Kim, H., Yang, Y., Kim, P., Kim, S.: A telematics service system
based on the Linux cluster. In: Shi, Y., van Albada, G.D., Dongarra, J., Sloot, P.M.A.
(eds.) ICCS 2007. LNCS, vol. 4490, pp. 660–667. Springer, Heidelberg (2007)
16. MacDougall, M.: Simulating Computer Systems: Techniques and Tools. MIT Press,
Cambridge (1987)
17. Ting, Y., Chang, Y.: A novel cooperative caching scheme for wireless ad hoc net-
works: Group caching. In: IEEE Int’l Conference in Networking, Architecture, and
Storage, pp. 62–68 (2007)
 Power Control for Soft Fractional Frequency
Reuse in OFDMA System

Young Min Kwon, Ok Kyung Lee, Ju Yong Lee, and Min Young Chung

School of Information and Communication Engineering

Sungkyunkwan University
300, Chunchun-dong, Jangan-gu, Suwon, Kyunggi-do, 440-746, Korea

Abstract. In this paper, soft Fractional Frequency Reuse (FFR) is con-

sidered to reduce the co-channel interference and also guarantee the cell
edge users data rate. We apply power control in soft FFR and investi-
gate the system performance. Soft FFR divides whole frequency band
into multiple subbands with different kinds of transmission power levels
and allocates subband according to SINR of cell users. By controlling
allocated powers, we can increase the frequency efficiency by using the
whole frequency band and also improve the data rate for the cell edge
users remarkably by reducing the co-channel interference. If the through-
put requirement for the outer region is small, we decrease the power allo-
cated to the outer region which can save power energy while high overall
system throughput is achieved.

Keywords: OFDMA, Frequency Reuses Scheme, FFR, SFR, Soft FFR.

1 Introduction
The next generation mobile broadband wireless communication systems, such
as IEEE 802.16e Mobile WiMAX and 3GPP LTE (Long Term Evolution), are
based on Orthogonal Frequency Division Multiple Access (OFDMA) to support
the high data rate service [1][2]. Mobile WiMAX and 3GPP LTE also adopt the
frequency reuse factor of one, in which each cell serves users with whole system
bandwidth. In OFDMA, however, users with the same channel simultaneously
suffer Co-Channel Interference (CCI) of neighbor cells, which may cause severe
degradation of system performance, especially in cell edge region [3]. The CCI
can be reduced with the frequency reuse factor of more than one. However, it
may reduce the system capacity [4].
Several frequency reuse schemes have been studied to mitigate the interference
of the cell edge region and to increase the system capacity [5]−[7]. The Fractional

This work was supported by National Research Foundation of Korea Grant funded by
the Korean Government (2009-0074466) and by the MKE (The Ministry of Knowl-
edge Economy), Korea, under the ITRC (Information Technology Research Center)
support program supervised by the NIPA(National IT Industry Promotion Agency
(NIPA-2009-(C1090-0902-0005)).

Corresponding author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 63–71, 2010.

c Springer-Verlag Berlin Heidelberg 2010
64 Y.M. Kwon et al.

Frequency Reuse (FFR) [8] is proposed for mobile WiMAX in IEEE 802.16e
working group, and Soft Frequency Reuse (SFR) [9] is proposed in the 3GPP
working group. FFR scheme and SFR scheme divide the available spectrum into
two reserved parts, subbands for the inner region and subbands for the outer
region. The subband for inner User Equipment (UE) is common in each cell,
and the subband for outer UEs is different among adjacent cells. Thus, UEs in
outer region do not suffer the CCI from neighbor cells within first-tier, so that
the spectral efficiency of outer region is increased. However, since FFR does
not use whole available frequency bandwidth, overall cell throughput in a cell
is lower than reuse one. SFR has two types of subband, the subband with high
transmission power level and with low transmission power level. It assigns the
subband with high power to outer users and the subband with low power to
inner users. Since SFR can use whole system band in a cell, overall cell capacity
in a cell is higher than FFR. However, overall system capacity of SFR may be
lower than that of reuse one environment.
Soft FFR scheme has been proposed to improve overall cell throughput of
FFR [10]. While FFR do not use the subbands allocated for outer region in the
adjacent cells, soft FFR uses these subbands for the inner UEs with low transmit
power. As a result, the soft FFR also use the subband with high transmit power
level and with low transmit power level like SFR. Unlike SFR, soft FFR uses the
common subband, which can guarantee the throughput of inner users. In this
paper, we investigate the performance of soft FFR in the environment, where
subbands have several transmit power levels. An outline of the rest of this paper
is as follows. In Section 2, we propose a power control in soft FFR scheme.
The result and analysis are then presented in Section 3. Finally, conclusions are
drawn in Section 4.

2 Power Control in Soft Fractional Frequency Reuse

(Soft FFR)

When frequency reuse factor is equal to one, UEs in the edge of cell may suffer
high outage probability due to the CCI from adjacent cells. When frequency
reuse factor is more than one, adjacent cells can operate on mutually orthogonal
frequency, which yields small CCI from adjacent cells. However, since a cell can
serve UEs with a part of whole system bandwidth, the spectral efficiency is low.
To compromise these approaches, frequency reuse schemes such as FFR or soft
FFR have been proposed. Both FFR and soft FFR statically partition the cell
into two distinct geographical regions: the inner region and the outer region, as
shown in Fig. 1.
FFR divides the system bandwidth into the subband for UEs in inner region
and the subband for UEs in outer region. In Fig. 1, the system bandwidth is
divided into four subbands, F0, F1, F2, and F3. In FFR, Base Station (BS)
allocates the subband with different reuse factor according to the user’s location
in a cell. The reuse factor of F0 is equal to one and F0 is allocated to users in
the cell’s inner region. UEs in the inner region are close to the serving BS and
 Power Control for Soft Fractional Frequency Reuse in OFDMA System 65

far from interfering BSs, and they are served with the common subband. Also
one of three subbands, F1, F2, and F3 is allocated to UEs in the outer region
and the other subbands are used for adjacent cells. Thus, there is no co-channel
interference among frequency bands in the first-tier outer region. Since each cell
cannot use the subband for the other region in adjacent cells, overall capacity of
the system may be reduced. There is a trade-off between the spectral efficiency
of users in the outer region and overall capacity.
Inner region
Outer region

Outer
Cell 1

Inner
Cell 1

F0 F0 F1 F2 F3

Cell 2

Outer
Inner
F1

F0 F1 F2 F3
Cell 2

Cell 3

F0 F2 F3 F0
Cell 3

Outer
Inner

F0 F1 F2 F3

Fig. 1. Frequancy allocation and power distribution of FFR scheme

In FFR, since a cell uses partial subbands among overall system frequency
bands, the overall capacity may be decreased. To improve the overall capacity,
soft FFR scheme has been proposed, which can serve UEs with the subband for
the outer region of adjacent cell. As in Fig. 2, it divides the system bandwidth
into four subbands, F0, F1, F2, and F3, like FFR. While FFR does not use the
subband for outer region in the adjacent cells, soft FFR uses these subbands with
transmit power level, Pin , for the inner UEs. In soft FFR, UEs in the inner region
are served with additional subband with Pin as well as subband with Pcomm .
UEs in the outer region of soft FFR are served in the subband with transmission
power level, Pout . Since soft FFR can use all subbands in each cell, the overall
capacity of soft FFR may be larger than that of FFR. Due to the interference
by the subband with Pin , soft FFR may have lower spectral efficiency of outer
UEs as compared with FFR.
In soft FFR, we control the spectral efficiency of outer UEs by the transmission
power ratio (ρ), which is the ratio of the inner transmission power level to outer
transmission power level, that is, ρ = Pin /Pout . Soft FFR has common subband
with transmission power level, Pcomm . Since the common subband maintains
specific transmission power and is only used by inner UEs, it can guarantee a
certain amount of capacity for inner region.
66 Y.M. Kwon et al.

Fig. 2 explains the control of the transmission power in soft FFR. In soft
FFR, subbands with Pin are assigned to the inner UEs. On the other hand,
subbands with Pout , F1, F2, or F3, are assigned to the UEs in outer region,
which do not overlap with the subband of the adjacent cells. The SINR of outer
UEs is controlled by Pout , which can be dynamically adjusted according to the
population of outer UEs and the required data rate in outer region. In the next
section, we investigate the effect of power control where Pcomm has a fixed value
and Pout is adjusted according to the requirements of outer UEs.

Fixed
Inner region
Outer region

P comm.
Cell 1

power

P out
P in P in
Cell 1

F0 F1 F2 F3
F0 + F2 + F3

Fixed
Cell 2

P comm.
F1
power

P out
P in P in
Cell 2

Cell 3

F0 F1 F2 F3
F0 + F1 + F3 F2 F3 F0 + F1 + F2

Fixed
Cell 3

P comm.
power

P out
P in P in

F0 F1 F2 F3

Fig. 2. Power control in soft Fractional Frequency Reuse (soft FFR)

3 Simulation Results
The effect of power control in soft FFR is evaluated through system level sim-
ulation. We consider the two-tier hexagonal cellular system with nineteen cells,
where the inter-site distance is 500 meters. Each cell is divided into two regions:
the inner region and the outer region. We assume that the ratio of the inner
region area to the outer region area is two. The 48 UEs are located with a uni-
form distribution in a cell. Two thirds of whole UEs are in the inner region, and
the rest are in the outer region. A BS allocates subband to UE by estimating
the signaling from UEs. We assume that the BS knows a serving UE’s location
exactly according to signaling. We use simulation parameters in Table 1 [2].
In the established system environment, we evaluate the average cell through-
put of soft FFR in inner region and outer region. At a TTI of simulation, the
BS in a cell gathers the location information of UEs and divides UEs into two
groups, the inner group and the outer group, and it allocates subband resources
to each UE using the proportional fair scheduling [11]. In soft FFR, the subband
with Pcomm and Pin are assigned to UEs in inner group, whereas the subband
with Pout is allocated to UEs in the outer group. The throughput of an UE is
obtained based on the Signal to Interference and Noise Ratio (SINR) of UE in
 Power Control for Soft Fractional Frequency Reuse in OFDMA System 67

Table 1. Simulation Parameters

Parameter Value
Channel bandwidth 10 MHz
Carrier frequency 2 GHz
FFT size 1024
Number of subcarriers 601
Subcarrier spacing 15kHz
Cellular layout Hexagonal grid, 19 sites
Inter-site distance 500 m
Bandwidth 10 MHz
Log-normal shadowing 8dB
Penetration loss 20 dB
Propagation loss 128.1 + 37.6 log10 (R(km))
BS antenna gain 15 dBi
UE antenna gain 0 dBi
White noise power density -174dBm/Hz
Scheduling Proportional Fair
TTI 1 ms

the assigned subband. In system level simulation, SINR is determined by the

path loss and lognormal fading measured in the subband. The throughput of an
UE m is estimated using the Shannon capacity as

Tm = Wsub log2 (1 + SIN Rm ) (1)

where Wsub is the bandwidth of a subband assigned to an UE and SIN Rm is

the SINR of an UE m. The cell throughput in each region is total throughput
of UEs in the corresponding region and expressed as


M
Tcell = Tm (2)
m=1

where M is the number of UEs in a group.

To compare the cell throughput of soft FFR with that of FFR, we assume
that each frequency reuse scheme can use the total transmission power of 40W .
FFR divides the total transmission power into two and allocate the power to
inner subband and outer subband. In soft FFR, we allocate a quarter of total
transmission power to common subband. The remaining transmission power is
distributed to the subband with Pin and Pout according to the power ratio ρ.
Fig. 3 shows the cell throughput of the inner and outer region as the power ratio
changes. Since soft FFR can use the overall system band, cell throughputs in
soft FFR are larger than those in FFR. As we reduce the power ratio (ρ) to
improve the throughput in outer region, the overall throughput and the inner
throughput are decreased. With small ρ, the first tier cells use the subband with
lower power and the interference of first tier is reduced. However, since the six
68 Y.M. Kwon et al.

cells of the second tier use the same subband with the target cell, the interference
of the second tier is increased. As a result, cell throughput is affected more by
the interference from the second tier than the first tier. Therefore, as ρ decreases
from 0.1 to 0.001, the outer cell throughput of soft FFR is not increased any
more although the inner cell throughput is decreased.

Fig. 3. Cell throughput of FFR and soft FFR according to power ratio, ρ = Pin /Pout
(Pcomm = 10W , total transmission power = 40W )

Now we consider the environment that soft FFR has the fixed power ratio,
ρ = 0.1. When the transmission power of common subband is fixed to 5W ,
10W , and 20W , we increase the power of subband for outer UEs, Pout , and
estimate the cell throughput in each region as Pout changes. Since the power
ratio ρ is constant, the transmission power level of subband for inner UEs, Pin ,
is also increased as Pout increases. Fig. 4 (a) shows the cell throughput of each
region in this environment. In the inner region, although Pcomm is increased, the
cell throughput is not increased significantly. Since the common subband has
frequency reuse factor with one, the signal power of serving BS and interfering
BSs are simultaneously increased as Pcomm increases. As a result, the SINR of
UEs using the common subband is less sensitive to the change of transmission
power. In the outer region, since the power ratio (ρ) is fixed, the cell throughput
does not vary although the outer transmission power level is increased. The
overall cell throughput is less affected by the transmission power level of common
subband, Pcomm , with constant power ratio (ρ).
Fig. 4 (b) shows that total transmission power increases as Pcomm increases,
when ρ is fixed to 0.1. However, the cell throughput remains almost the same.
Therefore, while we can achieve similar throughput performance, we can save
the transmission power by reducing Pcomm .
To investigate the power control by adjusting Pout , we assign the fixed trans-
mit power to Pcomm and Pin , which are 10W and 5W , respectively. In this
 Power Control for Soft Fractional Frequency Reuse in OFDMA System 69

(a) Cell throughput

(b) Total transmission power

Fig. 4. Cell throughput and total transmission power of soft FFR according to the
outer transmission power level, Pout ( ρ = Pin /Pout = 0.1, Pcomm = 5W, 10W, 20W )

environment, we estimate the cell throughput in each region and the total trans-
mit power according to outer transmit power level, Pout . In Fig. 5 (a), when
Pout is increased, the cell throughput in outer region is increased due to the
increased signal power. On the other hand, the cell throughput in inner region
is decreased due to the increased interference. However, the total transmission
power increases to improve the cell throughput of outer UEs in Fig. 5 (b). Thus,
Pout can be controlled to the throughput requirement of outer UEs in soft FFR.
If a few UEs exist in outer region, we reduce the transmission power of the sub-
band used in outer region, which yields high throughput for inner region due to
70 Y.M. Kwon et al.

(a) Cell throughput

(b) Total transmission power

Fig. 5. Cell throughput and total transmission power according to Pout in soft FFR
(Pcomm = 10W , Pin = 5W )

the reduced interfering signal power. As a result, the overall throughput can be
increased and the soft FFR scheme can save the total transmission power.

4 Conclusion
In this paper, we investigated the effect when the power of subband for each
region changes in soft FFR. In the simulation, we observe that the transmis-
sion power level of common subband, Pcomm , does not influence the overall cell
throughput significantly. Thus, we save the total transmission power by reducing
 Power Control for Soft Fractional Frequency Reuse in OFDMA System 71

Pcomm within the required cell throughput. In addition, we also save the total
transmission power by using the relation between Pin and Pout . If the power in
the outer region increases, the throughput in the outer region increases and the
throughput in the inner region decreases. Otherwise, the throughput in the outer
region decreases and the throughput in the inner region increases. As the power
in the outer region decreases, the overall capacity increases, since the spectral
efficiency in the inner region is higher than that in the outer region. According
to the throughput requirements in the outer region, we can reduce the power
consumption for the outer region, while maintaining high system throughput.

References
1. IEEE 802.16e-2005: IEEE Standard for Local and Metropolitan Area Networks,
Part 16: Air Interface for Fixed and Mobile BroadBand Wireless Systems (October
2005)
2. 3GPP TR 25.814 ver 7.1.0, Physical layer aspects for evolved Universal Terrestrial
Radio Access (UTRA) (Release 7) (September 2006)
3. Giuliano, R., Monti, C., Loreti, P.: Wireless technologies advances for emergency
and rural communications - WiMAX fractional frequency reuse for rural environ-
ments. IEEE Wireless Communications 15(3), 60–65 (2008)
4. Wang, Z., Stirling-Gallacher, R.A.: Frequency reuse scheme for cellular OFDM
systems. Electronics Letters 38(8), 387–388 (2002)
5. IEEE 802.20 MBWA C802.20 − 05 − 69, Air Interface Spec Final Fixed (November
2005)
6. 3GPP R1-050599, Interference Mitigation Considerations and Results on Fre-
quency Reuse, Siemens (June 2005)
7. 3GPP R1-050833, Uplink Interference Mitigation via Power Control, LG Eletrinics
(May 2006)
8. Mobile WiMAX Part I: A Technical Overview and Performance Evaluation,
WiMAX Forum (February 2006)
9. 3GPP R1-050507, Soft frequency reuse scheme for UTRAN LTE, Huawei, Tech.
Rep. (May 2005)
10. IEEE 802.16 Broadband Wireless Access Working Group, IEEE C802.16m-08/782,
Fractional Frequency Reuse in Uplink, LG Electronics (August 2008)
11. Andrews, M.: Instability of the proportional fair scheduling algorithm for HDR.
IEEE Transactions on Wireless Communications 3(5), 1422–1426 (2004)
 Authentication – Based Medium Access Control
to Prevent Protocol Jamming: A-MAC

Jaemin Jeung, Seungmyeong Jeong, and Jaesung Lim*

Graduate School of Information and Communications, Ajou University, South Korea

{mnsg,aflight,jaslim}@ajou.ac.kr

Abstract. Recently, Wireless Local Area Network (WLAN) is used by enter-

prises, government, and the military, as well as small office and home offices.
Although it is convenient, it has inherent security weaknesses due to wireless
characteristics. For this reason, security-sensitive groups are still unwilling to
use WLAN. There are several types of attacks to degrade the wireless network
throughput using security weakness, especially Protocol Jamming Attacks are
critical. These attacks consume little energy and can be easily implemented. In
this paper, we introduce Authentication-based Medium Access Control (A-
MAC) to prevent Virtual Carrier Sense (VCS) Jamming Attack and De-
authentication / Disassociation Jamming Attack that are typical Protocol
Jamming Attacks in 802.11 based wireless systems. The proposed scheme can
authenticate frames using Universal Hashing Message Authentication Codes
(UMAC-32) and Hidden Sequence Number (SN). A-MAC frequently changes
the key and SN using shift row and shift column processes to overcome the
weakness of short 32 bits hashing codes. A-MAC can achieve integrity, authen-
tication, and anti-replay attack security features. It can prevent Protocol Jam-
ming Attacks that degrade wireless network throughput. Our simulation shows
A-MAC can sustain throughput under Protocol Jamming Attacks.

Keyword: Protocol Jamming, Authentication, Medium Access Control.

1 Introduction
Wireless Local Area Network (WLAN) provides users with mobility within a broad
coverage area with connection to networks. In addition, it is cost-efficient, allowing
ease of integration with other networks and network components. For these reasons,
WLAN is used by enterprises, government, and the military, as well as small office
and home offices. However, it has inherent security weaknesses due to wireless char-
acteristics. Furthermore, it uses the Industrial Scientific Medical (ISM) band to be
highly susceptible to interference. Security-sensitive groups are still unwilling to use
WLAN due to these weaknesses. If we use open source scanning software, such as

*
This research was supported by the MKE (The Ministry of Knowledge Economy), Korea, un-
der the ITRC (Information Technology Research Center) support program supervised by the
NIPA (National IT Industry Promotion Agency) (NIPA-2009-C1090-0902-0003).

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 72–84, 2010.
© Springer-Verlag Berlin Heidelberg 2010
 Authentication – Based Medium Access Control to Prevent Protocol Jamming 73

KISMET, we can easily analyze the packets [1]. Many researchers have been study-
ing these issues, since these weaknesses are well known. Thus, WEP, WPA, TKIP,
802.11i standards were formed. However, the standards cannot resolve Protocol
Jamming Attacks. Furthermore, Protocol Jamming Attacks consume little energy, and
can be easily implemented. An adversary can decrease the wireless networks
throughput for a long time, especially, when jammers are hidden.
Virtual Carrier Sense (VCS) jamming and De-authentication / Disassociation jam-
ming are typical Protocol Jamming Attacks. VCS jamming attacks exploit a weakness
in the Carrier Sense Multiple Access / Collision Avoidance (CSMA/CA). CSMA/CA
systems use VCS to avoid collision. An adversary exploits the VCS process to de-
crease the wireless networks throughput. De-authentication / Disassociation jamming
attacks exploit the protocol concerned with connection in the mobile station and AP.
All stations have to go through the process of authentication / association before send-
ing data. If an AP sends a De-authentication / Disassociation frame to a station, the
stations have to terminate the connections. Unfortunately, based on the 802.11 stan-
dards, the station cannot reject the notification of these frames. Therefore, these at-
tacks are more effective and powerful than VCS attacks [2].
We consider that the cause of Protocol Jamming Attacks is authentication prob-
lems. That is, APs and mobile stations cannot authenticate RTS / CTS / De-
authentication / Disassociation frames. Even if there is an authentication process, it
can be attacked by replay-attacks. Replay attacks also decrease network throughput.
Thus, we need authentication, integrity and anti-replay-attacks features to prevent
Protocol Jamming Attacks. In this paper, the proposed A-MAC substitutes digested
message for Cyclic Redundancy Check (CRC). The digested message is created when
the Universal Hashing Message Authentication Codes (UMAC-32) are XORed with a
Hidden Sequence Number. UMAC-32 ensures the authenticity and the integrity of
transmitted messages. In addition, it can check the transmission error. So we can sub-
stitute the digested message for CRC. The Hidden Sequence Number for anti-replay
attacks is not seen in the transmitted frames. It is stored in a Sequence Table of each
correspondent. In this approach, we can achieve integrity, authentication and anti-
replay attacks security features, and prevent Protocol Jamming Attacks that degrade
network throughput, as well as checking for transmission errors, without frame size
overhead. We also introduce Key Shift & Sequence Number Select Notification
(KSSN) that notify key and SN change by shift row and shift column processes, so
that we compensate for the short 32bits digested message. Using the KSSN, each cor-
respondent can change the key and SN, if necessary. That is, the KSSN is verified by
the A-MAC, each correspondent changes the key and SN. This method is suitable for
control and management frames that are short frame and occur frequently, because the
KSSN can use a one-way key exchange mechanism, instead of four-way hand-
shaking. Thus, we can prevent VCS and De-authentication / Disassociation jamming
attacks using the A-MAC. Our simulation shows its performance and effectiveness.
The remainder of this paper is organized as follows. Section 2 describes the back-
ground of the proposed scheme and related work. We introduce adversary models, A-
MAC and KSSN process in Section 3. Security analysis and performance analysis are
presented in Section 4. Finally we conclude the paper and outline directions for future
work in Section 5.
74 J. Jeung, S. Jeong, and J. Lim

2 Background and Related Work

2.1 Background

802.11 wireless networks use Carrier Sense Multiple Access with Collision Avoid-
ance (CSMA/CA). The Distributed Coordination Function (DCF) is the basis of the
standard CSMA/CA access mechanism to avoid collisions. It first checks to see that
the radio link is clear before transmitting. Stations use a random back-off after each
frame, with the first sender seizing the channel, to avoid collisions. The DCF may use
the Request To Send (RTS) / Clear To Send (CTS) clearing technique to further re-
duce the possibility of collisions.
Carrier sensing is used to determine if the medium is available. Two types of car-
rier sensing functions in 802.11 manage this process: Physical Carrier Sensing and
Virtual Carrier Sensing. If either carrier sensing function indicates that the medium is
busy, the MAC reports this to higher layers [3]. Unlike wired communications, Physi-
cal Carrier Sensing cannot provide all the necessary information, due to the hidden
node problem. In Fig. 1, A is about to send frames to B. However, D cannot recognize
the situations, because A’s physical signal cannot reach D. Since A and D cannot
hear each other, they may sense the media is free at the same time and both try to
transmit simultaneously; this causes collision in the network. This is the hidden node
problem. To reduce this problem, 802.11 standards adopt RTS, CTS, and VCS
mechanisms. Fig. 1 shows these mechanisms. Activity on the medium by stations is
represented by the shaded bars, and each bar is labeled with the frame type. Inter-
frame spacing is depicted by the lack of any activity. DIFS is Distributed Inter-Frame
Space and SIFS is Short Inter-Frame Space. If A wants to send frames to B, A sends
the RTS frame to B, and then B replies with the CTS frame. Neighbor stations, C and
D can overhear RTS and CTS frames, and then defer access to the medium until the

State 3
Authenticated
& Associated
A D
C

Successful
Disassociation
association
DIFS
RTS Data
Sender A State 2
SIFS
Authenticated Deauthentication
SIFS SIFS & Unassociated
CTS ACK
Receiver B

Successful
NAV (RTS) Deauthentication
authentication
C

DIFS
NAV (CTS) State 1
D Unauthenticated
& Unassociated
Defer access

Fig. 1. Virtual Carrier Sense Fig. 2. 802.11 state diagram

 Authentication – Based Medium Access Control to Prevent Protocol Jamming 75

Network Allocation Vector (NAV) elapses. VCS is provided by the NAV. The NAV
is a timer that indicates the amount of time the medium will be reserved, in microsec-
onds. It is usually presented by 2 bytes. If a station sets the timer for which they ex-
pect to use the medium, other stations count down from the NAV to 0. When the
NAV reaches 0, the VCS indicates that the medium is idle. These RTS, CTS, and
VCS mechanisms will prevent collisions.
Stations have to go through three states for connection to Access Point (AP). Fig. 2
shows the state diagram of 802.11. Stations are either authenticated or unauthenti-
cated and can be associated or unassociated. These situations can be combined into
three states. State 1 is not authenticated and not associated, State 2 is authenticated
but not associated, and State 3 is authenticated and associated. Each state is a succes-
sively higher point in the development of an 802.11 connection. All stations start in
State 1, and data can be transmitted only in State 3. That is, any stations cannot send
data in Sate 1 and State 2. When a station transfers from one state to another state, it
would be vulnerable.

2.2 Related Work

Any current or upcoming 802.11 standards would not help mitigate the risk of Proto-
col Jamming [4]. Researchers have tried to resolve these problems recently. Zhou
proposed a packet-by-packet authentication method [5]. The author encrypted all
frame content by a secret key, and then attached the encrypted content to the end of
the original frame. For example; A → B: {RTS, E(sID, dID, TS, SN)k}, B → A:
{CTS, E(sID, dID, TS, SN+1)k}. User A decrypts the encrypted attachment, verifies
sID, dID, TS, and SN+1. If the frame is correctly decrypted, A can commence data
transmission. However, this authentication scheme increases transmission overhead.
Assume TS and SN are 4 bytes each, the total encrypted attachment is 20 bytes (6
bytes each for sID and dID). Considering RTS / CTS frame are just 20 bytes / 16
bytes, the encrypted attachment is a considerable overhead. Karlof proposed the Tiny-
Auth packet format for wireless sensor networks [6]. However, it has a weakness in a
replay attack, because it does not use any kind of sequence number or time-stamp in
the TinySec-Auth packet. So, TinySec is not suitable in WLAN. Bellardo has a dif-
ferent view point [7]. In contrast to the authentication scheme, this scheme places two
different limits on the duration values accepted by stations. The low limit has a value
equal to the amount of time required to send an ACK frame, plus media access back-
off for that frame. The high limit has a value equal to the amount of time required to
send the largest data frame, plus the media access back-off for that frame. However,
this scheme still has a weakness in RTS/CTS flooding. Since each station cannot ver-
ify the adversary that frequently send false RTS/CTS frames, the stations defer their
transmission, and consequently the wireless networks throughput is drastically low-
ered. Some research prevents Protocol Jamming Attacks using the creation of a series
of protocol extension and replacements (e.g., WEP, WPA, 802.11i, 802.11w) [2].
However, these schemes need complicated key management, powerful computational
process and frame size overhead. They are not easy to implement.
76 J. Jeung, S. Jeong, and J. Lim

3 Authentication-Based Medium Access Control (A-MAC)

3.1 Malicious Adversary Models

Malicious adversary models are VCS jamming attacks and De-authentication / Disas-
sociation jamming attacks. VCS jamming attacks exploit the weakness of the virtual
carrier sensing mechanism, especially RTS/CTS process. De-authentication / Disas-
sociation jamming attacks exploit the weakness of the state diagram in Fig. 2. When a
station transfers from one state to another state, a vulnerable point occurs, especially
in the De-authentication, Disassociation frames.
Fig. 3(a) shows RTS/CTS adversary and NAV adversary in VCS jamming attacks.
If a malicious adversary frequently sends a RTS frame to A, A sets the NAV timer, as
much as the duration field of RTS frame indicates. According to the VCS mechanism,
A has to defer access to the medium until the NAV elapses, since the VCS mecha-
nism indicates that the medium is busy. If a malicious adversary repeatedly sends a
RTS frame to A, station A cannot have the opportunity to access the medium. This
scenario will decrease the total wireless network throughput. We define this adversary
as the RTS/CTS Adversary. In the case of Station B, a malicious adversary sets the
NAV timer for a longer period. According to VCS mechanism, B simply thinks that
the other station is about to send a large data set. As a result, B has to defer access to
the medium during the time of the duration field. The duration field has 16 bits. An
adversary can set the NAV duration for a 216 time slot. Assume that one NAV time
slot is 1 μs; B has to wait approximately 0.06 second. This delay is considerable. If
there are many stations around the malicious adversary, the total wireless network
throughput is drastically lowered. We define this adversary as the NAV adversary.
These two adversary models apply to the CTS transmission process too.

Fig. 3. Malicious adversary models

 Authentication – Based Medium Access Control to Prevent Protocol Jamming 77

Fig. 3(b) shows Disassociation/De-authentication jamming attacks. These attacks

exploit the weakness of the connection process. We have already learned that each
station goes through States 1, 2, and 3. If a station gets a de-authentication frame
when transferring from state 1 to State 2, the station cannot transfer to State 2. For the
same reason, if a station gets a disassociation frame when transferring from State 2 to
State 3 or sending data to the AP, the station reverts to State 2. According to the
802.11 standards, stations cannot reject the notification of disassociation and de-
authentication frames. Unfortunately, a malicious adversary can easily fake disasso-
ciation and de-authentication frames. In Fig. 3(b), assume that station B is connected
to Access Point A. If B gets a disassociation frame, B has to terminate the data com-
munication. We define this attacker as the Disassociation Adversary. The adversary
model is simple but, powerful. The total wireless throughput is to be a zero by only a
few frames / second. After terminating the connection, B tries to reestablish a new
connection. If a malicious adversary frequently sends a de-authentication frame at that
time, B cannot transfer to State 3. We define this attacker as the De-authentication
Adversary.
We define four adversary models. The characteristics of these models are easy im-
plementation, use low energy, and decrease the total wireless network throughput. It
is not easy for wireless IDS to detect the jammer, given that the jammer is hidden,
uses low signal power, and sends fake frames not frequently but appropriately. Thus,
the jammer can sustain a low network throughput that irritates network users and
managers for a long time.

3.2 A-MAC Scheme

We propose the Authentication-based Medium Access Control (A-MAC) that

achieves authentication, integrity and anti-replay attack without overhead to prevent
Protocol Jamming Attacks that degrade throughput. A-MAC compresses the frames
into a 32 bits digested message using Universal Hashing Message Authentication
Codes (UMAC-32). Then, the digested message is masked by XOR with Hidden Se-
quence Numbers. Finally, we substitute the results for Cyclic Redundancy Check
(CRC). In this scheme, we easily detect some abnormal frames that considerably de-
crease throughput. Thus, the total wireless network throughput does not decrease by
discarding the instruction of abnormal frames. Fig. 4 details the proposed A-MAC
process.
In Fig. 4, we assume that sender A transfers a message (M) to receiver B, and M
is RTS, CTS, Disassociation, or De-authentication frame without CRC. A-MAC
compresses the M into a 32 bits digested message by the UMAC-32 function and the
secret key. Then, the compressed message is XORed with the Hidden Sequence Num-
ber. The Hidden SN is not seen in the transmitted frames. It is stored in a Sequence
Table. The Hidden SN is very important in this scheme. It prevents a replay attack
and complicates the digested message. Finally, the XORed message {UMAC (K, M)
⊕ SN} is concatenated with M. Sender A sends this frame to receiver B. B also com-
presses the received M into a 32 bits digested message by the UMAC-32 function and
the secret key. Then, the compressed message is XORed with the received {UMAC
(K, M) ⊕ SN}. Receiver B derives SN from this process [UMAC (K, M) ⊕ {UMAC
(K, M) ⊕ SN} = SN]. The derived SN is compared with SN of B’s Sequence Table. If
78 J. Jeung, S. Jeong, and J. Lim

Sender A Receiver B

M K
RTS Remove CRC RTS
UMAC
CTS CTS

De-auth SN De-auth
K
Dis-ass Dis-ass
UMAC UMAC(K,M) SN SN

Sequence Table Compare

Fig. 4. Authentication-based Medium Access Control Process

the SNs agree with each other, the sender A is proved to be a legitimate station. If the
SN does not agree with each other, there are either transmission errors or Protocol
Jamming Attacks. Transmission errors are closely related to Signal to Noise Ratio
(SNR). We can easily check the SNR in 802.11 systems by checking the Received
Signal Strength Indication (RSSI). Therefore, although the RSSI is good, the SN does
not agree with each other for several times. The probability of Protocol Jamming At-
tacks becomes high.
We can apply the A-MAC to an infrastructure WLAN. For example, an adversary
sends de-authentication and disassociation frames to a certain station, the station can
verify the frames using the A-MAC mechanism. If those frames are verified, they
follow the instruction of each frame, otherwise they discard the frames. In case of
sending lots of false RTS / CTS frames to a certain station, counter measures will be
different. That is, key distribution schemes can make the method different. For exam-
ple, an AP and stations share the same secret key, the stations can derive the SN from
the frames using A-MAC. If the SNs of successive frames are incremented by one,
the station will be legitimate, otherwise the stations will be suspected as an adversary.
In contrast to the prior method, if the AP and each station share the 1:1 matching se-
cret key, neighbor stations cannot derive the SN due to a different key. But the AP
can derive the SN, and verify the frames. In that case, suppose the AP recognizes the
adversary, it can notify the protocol jamming situation to the stations. The stations re-
ceived the notification can discard the frames and cope with the situation.
A-MAC uses UMAC-32 for the digested message, instead of CRC, since UMAC-
32 ensures the authenticity and the integrity of transmitted messages. Furthermore, it
is the fastest message authentication code reported in cryptographic literature [8]. We
can derive Tag from the UMAC-32 process. Tag’s size is 32 bits, 64 bits, 96 bits, 128
bits. In this paper, A-MAC use a 32 bits Tag so that the digested message is XORed
with a 32 bits Hidden SN. Tag = H_K1 (M) ⊕ F_K1 (nonce), M is an input message,
H is a secret hash function, F is a pseudo random function, and K1 is secret random
key shared by sender and receiver. Nonce is a value that changes with each digested
message. In this paper, M is the content of the CRC removed frame, nonce is the Hid-
den SN. The proposed scheme can achieve authentication, integrity and anti-replay
 Authentication – Based Medium Access Control to Prevent Protocol Jamming 79

attack without overhead. Therefore, A-MAC can overcome the weakness of Karlof’s
scheme [6] and Bellardo’s scheme [7]. Especially, it does not expand the original
frame size. This characteristic also overcomes the problem of Zhou’s scheme [5].

3.3 Key Shift and SN Select

As using a short 32 bits digested message, A-MAC may be attacked by brute-force at-
tack that is a technique to defeat the authentication mechanism by trying successively
all the words in an exhaustive list. We propose two methods that complicate the at-
tached 32 bits message of A-MAC to mitigate this vulnerability.
At first, A-MAC does not count the initial SN from zero. If the initial SN starts
from zero, a malicious adversary can calculate {UMAC (K, M)}. Thus, the probabil-
ity of detecting the key becomes high. A-MAC derives the initial SN from the 128
bits key to mitigate this probability. Table 1. is an example of Sequence Table. We
make a 4 × 4 byte matrix using the 128 bits secret key. Each row and each column is
32 bits. The first row becomes the initial SN of the RTS frame, the second row be-
comes the initial SN of the CTS frame, the third row becomes the initial SN of the de-
authentication frame, the fourth row becomes the initial SN of the disassociation
frame. After being set, the initial SN is incremented by one every frame exchange.
In this scheme, it is not an easy for a malicious adversary to estimate the key and the
digested message.
Second, each correspondent (station or AP) can change the secret key and SN
if necessary. We define this process as Key Shift & SN Select Notification (KSSN).
Fig. 5 shows that each correspondent changes the secret key and SN simultaneously.

Table 1. Sequence Table

Sequence Table
Node De- Dis-
Key RTS CTS
ID auth ass
ABCD
EFGH
A ABCD EFGH I JK L MNOP
I JK L
MNOP

BCDE
FGH I
B BCDE FGH I JKLM NOPA
JKLM
NOPA

* * * * * *
* * * * * *

In Fig. 5, we assume that sender A wants to change the secret key and SN. A sets
the subtype of Frame Control Header to 0111. In 802.11 standards, management sub-
type 0111 and control subtype 0111 are reserved and not currently used [3]. In this
paper, we use this subtype 0111 to send the KSSN message. After setting the subtype,
sender A transmits {UMAC (K, M) ⊕ SN ⊕ KSSN} to receiver B. The received
{UMAC (K, M) ⊕ SN ⊕ KSSN} message is XORed with {UMAC (K, M) ⊕ SN} by
receiver B. B derives the KSSN from this process. B compares the last 13 bits of
KSSN with the last 13 bits of SN.
80 J. Jeung, S. Jeong, and J. Lim

Sender A Receiver B

M
K
RTS Remove CRC RTS
UMAC
CTS CTS
KSSN
De-auth De-auth

Dis-ass SN
Dis-ass
K SN UMAC(K,M) SN
KSSN KSSN

The last 13 bits The last 13

UMAC bits of KSSN
of SN
Sequence Table Compare

OK

16 bits (Key Shift Notification) 13 bits

KSSN: 3 bits (SN Select Notification) (The last 13 bits of SN) Shift Key & Select SN

Fig. 5. Key Shift & SN Select Notification Process

KSSN (32 bits)

Shift Row Shift Column Select SN The last 13 bits of SN

1 2 3 4
1 row 2 row 3 row 4 row column column columncolumn

00: 4 left shift 00: 4 down shift 0: row 00: first

01: 1 left shift 01: 1 down shift 1: column 01: second
10: 2 left shift 10: 2 down shift 10: third
11: 3 left shift 11: 3 down shift 11: fourth

Fig. 6. KSSN format

Fig. 6 shows the 16 bits Key Shift Notification, 3 bits SN Select Notification, and
the last 13 bits SN KSSN format. We form the 4 × 4 byte matrix using the 128 bits
secret key, as in Fig. 7. The front 8 bits are related to Shift Row. The first 2 bits indi-
cate the first row, 3 ~ 4 bits indicate the second row, 5 ~ 6 bits indicate the third row,
7 ~ 8 bits indicate the fourth row. 00 indicates four bytes left shift, 01 indicates one
byte left shift, 10 indicates two bytes left shift, 11 indicates three bytes left shift. The
9 ~ 16 bits are related to the Shift Column. The allocation method is similar to Shift
Row. The 17 ~ 19 bits indicate which row or column is selected as the new SN. If the
17th bit is 0, it selects the new SN from the row. If it is 1, it selects the new SN from
the column. 00 indicates the first row or column, 01 indicates the second row or col-
umn, 10 indicates the third row or column, 11 indicates the fourth row or column. The
last 13 bits are used for authentication. If the last 13 bits of KSSN and the last 13
bits of SN are verified, the receiver can change the secret key and SN. The sender
transmits the KSSN several times to overcome transmission error. This depends on
channel states.
Fig. 7 shows the Key Shift and SN Select operations. If the receiver derives the
KSSN (KSSN: 00011011 00011011 010 + 13 bits) from the received message, the re-
ceiver verifies the last 13 bits of KSSN comparing them to the last 13 bits of SN.
 Authentication – Based Medium Access Control to Prevent Protocol Jamming 81

Shift Row
Old Key

S 0,0 S 0,1 S 0,2 S 0,3 S 0,0 S 0,1 S 0,2 S 0,3

S 1,0 S 1,1 S 1,2 S 1,3

S 1,0 S 1,1 S 1,2 S 1,3 S 1,1 S 1,2 S 1,3 S 1,0

S 2,0 S 2,1 S 2,2 S 2,3

S 2,0 S 2,1 S 2,2 S 2,3 S 2,2 S 2,3 S 2,0 S 2,1

S 3,0 S 3,1 S 3,2 S 3,3 S 3,0 S 3,1 S 3,2 S 3,3 S 3,3 S 3,0 S 3,1 S 3,2

New Key

S 0,0 S 3,0 S 2,0 S 1,0 S 0,0 S 0,1 S 0,2 S 0,3

S 1,1 S 0,1 S 3,1 S 2,1 S 1,1 S 1,2 S 1,3 S 1,0

New SN S 2,2 S 1,2 S 0,2 S 3,2 S 2,2 S 2,3 S 2,0 S 2,1

S 3,3 S 2,3 S 1,3 S 0,3 S 3,3 S 3,0 S 3,1 S 3,2

Shift Column

Fig. 7. Key Shift & SN Select Operations

If they are verified, the receiver changes the secret key and SN. The first 8 bits
(00011011) indicate the first row is shifted by 4 bytes, the second row is shifted by 1
byte, the third row is shifted by 2 bytes, and the fourth row is shifted by 3 bytes. The
next 8 bits (00011011) indicate that the first column is down-shifted by 4, the second
column is down-shifted by 1 byte, the third column is down-shifted by 2 bytes, and
the fourth column is down-shifted by 3bytes. As a result of the operations, the new
key is generated. 17~19 bits (010) indicates that the third row is the new SN of the
RTS frame. Then, other rows are sequentially allocated to other frames (e.g. the
fourth row is the new SN of CTS frame, the first row is the new SN of de-
authentication frame, and the second row is the new SN of association frame).

4 Security and Performance Analysis

4.1 Security Analysis

We designed the A-MAC to achieve authentication, integrity and anti-replay attack

without overhead. The A-MAC compresses the frames into a 32 bits digested message
using UMAC-32, and then the digested message is masked by XOR with Hidden Se-
quence Numbers. We obscure the digested message using KSSN operations to over-
come the weakness of the short digested message (e.g. brute-force attack). In this
process, we can obtain some security advantages. First, to prevent replay-attacks, we
use a Hidden SN that makes it hard to estimate the next frame. If it were not for the
82 J. Jeung, S. Jeong, and J. Lim

Hidden SN, a malicious adversary could perform Denial of Service (DoS) attacks.
Assuming that A transmits a RTS control frame to B. A → B: {FC (Frame Control
Header), D (Duration), RA (Receiver Address), TA (Transmitter Address), SN,
CRC}. An adversary can attempt to predict the SN, and then frequently sends a RTS
control frame like this. Adversary → B: {FC, D, RA, TA, SN + (i = i+1), CRC}. B
will frequently send CTS or ACK frame. As a result, the total wireless network
throughput will be lowered. However, the A-MAC can prevent these DoS attacks us-
ing the Hidden SN. Furthermore, is uses the 4 byte initial SN from the 128 bits secret
key, instead of zero. This makes it more difficult to estimate the SN. Second, A-
MAC can change the secret key and SN using the KSSN. This can prevent brute-force
attacks. A malicious adversary can try to break the digested message with the prob-
ability of a 1 over 232. Assuming that an adversary repeatedly sends a forged frame,
he could be accepted by AP after about 231. On 11Mbps 802.11b, an adversary can
send about 27,000 forgery attempts per second (where RTS = 15 μs, SIFS = 10 μs)
without considering transmission delay. Therefore, it would take over 44 hours. How-
ever, the A-MAC can change the secret key and SN frequently in a short time without
a complex key exchange mechanism. In addition, A-MAC does not have any frame
size overhead. These characteristics can overcome the weakness of Zhou’s scheme [5]
and Bellardo’s scheme [7]. Therefore, the proposed A-MAC is useful for government,
the military, and enterprises.

4.2 Performance Analysis

We analyze the performance and effectiveness of the proposed A-MAC. In the simu-
lation environment, we do not consider complex key distribution systems (e.g. Re-
mote Authentication Dial-In User Service: RADIUS), because the key distribution
process does not affect our simulation, the secret keys are presumed to have been de-
livered to the AP and stations. The jammer and AP send RTS, CTS, de-authentication,
and disassociation frames. When stations receive the frames, they process the A-MAC
operation. If those frames are verified, they follow the instruction of each frame, oth-
erwise they discard the frames.
Fig. 8 shows the simulation topology. There are one AP, four stations and one
jammer. Especially, the jammer is hidden and near the AP. All stations can hear the
RTS, CTS, de-authentication, and disassociation frame of the AP and Jammer. The
jammer starts to attack at 6 seconds. Under transmission errors, although the station is
a legitimate user, it cannot access the medium. In the real world, we can check the
RSSI level and adjust the threshold. However, we just set the threshold value to one in
this simulation. That is, we do not consider transmission error. Fig. 9 shows the
throughput under the RTS/CTS Adversary model. The jammer transmits the faked
RTS frame at the rate of 100, 200 Frames Per Second (FPS). As the faked RTS frame
increases, the throughput lowers. Under a rate of 200 FPS, the throughput falls to 5.25
Mbps. However, we can sustain the throughput under the A-MAC. Fig. 10 shows the
throughput under the NAV Adversary model. The Jammer sets the duration field to
the length of 211, 212. This attack is more serious than the previous attack. Under a
length of 212, the total throughput drastically falls to 50%, since the NAV at 212 (212 ×
1μs = 0.004 sec) has a considerable delay in 802.11 systems. However, we can also
sustain the throughput under the A-MAC. Fig. 11 shows the throughput under the
 Authentication – Based Medium Access Control to Prevent Protocol Jamming 83

Fig. 8. Simulation Topology Fig. 9. RTS/CTS Adversary

Fig. 10. NAV Adversary Fig. 11. De-auth / Dis-ass Adversary

De-authentication and Disassociation Adversary. The simulation results of the De-

authentication Adversary model are similar to the results of the Disassociation Adver-
sary model. This is due to the de-authentication frame-received station also terminat-
ing the connection during data transmission. So, we just show the results of the De-
authentication Adversary model. The jammer transmits the faked frames at rates of 3,
5 FPS. These attacks are the most powerful and efficient attacks. Just a few frames
can decrease the throughput to almost zero, since all stations undergo the state transi-
tion to State 1 or State 2, as soon as they get the frames. In the four cases, the A-MAC
can prevent the attacks.

5 Conclusions
In this paper, we introduced the A-MAC to prevent VCS Attacks and De-
authentication / Disassociation Attacks that are typical Protocol Jamming Attacks. A-
MAC can authenticate frames using UMAC-32 and Hidden SN. A-MAC frequently
changes the secret key and SN by using shift row and shift column operations to pre-
vent the weakness of short 32 bits hashing codes. Therefore, A-MAC can achieve
integrity, authentication, and anti-replay-attacks security features, and can prevent
Protocol Jamming Attacks that degrade wireless network throughput. Our simulation
84 J. Jeung, S. Jeong, and J. Lim

shows that A-MAC can sustain the network throughput under Protocol Jamming At-
tacks. In the near future, we will study a physical layer anti-jamming scheme in
802.11 based systems. If we mitigate Protocol Jamming Attacks and physical layer at-
tacks in 802.11 based systems, security-sensitive groups (e.g. government, enter-
prises, and the military) have a preference to use WLAN. We hope that our research
will help that situation come true.

References
1. http://www.kismetwireless.net
2. Liu, C., Yu, J.: Rogue Access Point Based DoS attacks against 802.11 WLANs. In: The
Fourth AICT (2008)
3. Gast, M.S.: 802.11 Wireless Networks: The Definitive Guide, pp. 33–66. O’Reilly
Publisher, Sebastopol (2002)
4. Malekzadeh, M.: Empirical Analysis of Virtual Carrier Sense Flooding Attacks over Wire-
less Local Area Network. Journal of Computer Science (2009)
5. Zhou, Y., Wu, D., Nettles, S.: Analyzing and Preventing MAC-Layer Denial of Service
Attacks for Stock 802.11 Systems. In: Workshop on BWSA, Broadnets (2004)
6. Karlof, C., Sastry, N., Wagner, D.: TinySec: A link Layer Security Architecture for Wire-
less Sensor Networks. In: SenSys (2004)
7. Bellardo, J., Savage, S.: 802.11 Denial-of-Service Attacks: Real Vulnerabilities and Practi-
cal Solutions. In: USENIX Security Symposium (2003)
8. Krovetz, T.: RFC4418-UMAC: Message Authentication Code using Universal Hashing. In:
IEEE Network Working Group (March 2006)
Femtocell Deployment to Minimize Performance
Degradation in Mobile WiMAX Systems

Chang Seup Kim, Bum-Gon Choi, Ju Yong Lee, Tae-Jin Lee,

Hyunseung Choo, and Min Young Chung

School of Information and Communication Engineering

Sungkyunkwan University
300, Chunchun-dong, Jangan-gu, Suwon, Kyunggi-do, 440-746, Korea

Abstract. Femtocell is one of the promising technologies for improving

service quality and data rate of indoor users. Femtocell is short-range, low-
cost, and low-power base stations (BS) installed by the consumer indoors.
Even though femtocell can provide improved home coverage and capacity
for indoor users, it causes interference to macrocell users when femtocell
uses the same frequency band of macrocell. To reduce the interference be-
tween macrocell and femtocell, it is needed to analyze the characteristic of
macrocell and femtocell considering various interference scenarios. In this
paper, we investigate the uplink and downlink capacity of macrocell and
femtocell according to the strength of femtocell transmit power and the
number of femtocells through simulations. Simulation results show that
femtocell transmit power and the number of femtocells affect the perfor-
mance of macrocell. From the results, we find the adequate femtocell trans-
mit power which minimize the performance degradation of macrocell and
femtocell. We also investigate capacities of macrocell and femtocell accord-
ing to the locations of femtocell BS and macrocell UE.

Keywords: Femtocell, Macrocell, Co-Channel Interference, Power

control.

1 Introduction
Mobile WiMAX (IEEE 802.16e) based on IEEE 802.16e standard has been com-
mercialized in 2006 with rapid growth of various multimedia applications [1,2].
Recently, user demands for various multimedia services such as mobile IPTV
are continuously increasing. On the other hand, it is expected that 60 percent
of voice calls and 90 percent of data services will take place in indoors [3], and
in current mobile communication systems, the signal strength transmitted from
base station (BS) can be very low in indoor environments because the signal

This work was supported by National Research Foundation of Korea Grant funded by
the Korean Government (2009-0074466) and by the MKE (The Ministry of Knowl-
edge Economy), Korea, under the ITRC (Information Technology Research Center)
support program supervised by the NIPA (National IT Industry Promotion Agency)
(NIPA-2009-(C1090-0902-0005) and NIPA-2009-(C1090-0902-0046)).

Corresponding author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 85–95, 2010.

c Springer-Verlag Berlin Heidelberg 2010
86 C.S. Kim et al.

strength may be severely attenuated when it penetrates the obstacles such as

walls. Thus, providing high data rate services to indoor users is difficult only in
mobile WiMAX systems. Therefore, the mobile WiMAX has performed femtocell
standardization in two phases . The first phase of a WiMAX femtocell system,
which requires no air interface or UE change, is expected to be available in the
near future. More advanced and optimized femtocell features in phase 2 will be
available upon completion of IEEE 802.16m and as part of WiMAX Release 2.0
with target deployments [4].
Femtocell is connected to IP based backhaul such as digital subscriber line
(DSL) or cable modem, which provides lower cost than wireless mobile commu-
nication systems. Femtocells can achieve a high signal-to-interference-plus-noise
ratio (SINR) with low transmit power because the distance between transmitter
and receiver is short. In addition, the operators can save costs when femtocell
accommodates traffic concentrated to macrocell of current mobile communica-
tion. On the other hand, users can receive improved data speeds and service
quality with inexpensive costs. Although femtocell has many advantages in in-
door environments, some technical issues, such as spectrum assignment of fem-
tocell, access policy, network synchronization, handover, self-optimization, and
self-configuration, should be solved to be effectively deployed in existing mobile
WiMAX systems [5]. Especially, spectrum assignment for avoiding interference
between macrocell and femtocell is an important issue since interference may
causes severe degradation of throughput and service quality [6,7,8].
In mobile WiMAX system, throughput is determined by radius of macrocell,
distribution of femtocell, and density of femtocell [9]. The interference charac-
teristics under co-channel environments are presented in [10]. If femtocells are
deployed in mobile WiMAX systems, interference from femtocells can affect the
performance of mobile WiMAX systems. In this paper, we evaluate uplink and
downlink capacity of macrocell and femtocell by simulation under various inter-
ference scenarios. We investigate the adequate femtocell transmit power level,
which minimizes the performance degradation of mobile WiMAX system and
can achieve enough femtocell throughput. Also, we examine the effect of the
locations of femtocell BS and macrocell UE.
The rest of this paper is organized as follows. Section 2 provides interference
scenarios in the environment where femtocells and macrocells coexist. Section 3
describes the system model and propagation models. In Section 4, we examine
the uplink and downlink capacity of the macrocell and femtocell according to
the strength of femtocell transmit power, number of femtocells, and locations of
femtocell BS and macrocell UE. Finally, we conclude in Section 5.

2 Interference Scenarios of Femtocells in Mobile WiMAX

Systems
When femtocells operate in the same spectrum in mobile WiMAX systems, the
characteristic of interference depends on the method for channel assignment
[11][12]. Channel assignment method for macrocell and femtocell is classified
 Femtocell Deployment to Minimize Performance Degradation 87

into orthogonal and common channel assignment. Orthogonal channel assign-

ment does not cause co-channel interference between the macrocell and fem-
tocell because the femtocells use different channels with macrocell. However,
orthogonal channel assignment has a low spectrum efficiency because frequency
resources for the macrocell and femtocell are different. On the other hand, com-
mon channel assignment has a high spectrum efficiency because macrocell and
femtocell can use all of the spectrum. However, there exists a problem of co-
channel interference in common channel assignment method.
In common channel assignment, co-channel interference between macrocell
and femtocell should be reduced to guarantee service quality of users. We an-
alyze the co-channel interference between macrocell and femtocell according to
link direction and the locations of femtocell BS and macrocell UE. Co-channel
interference scenario between macrocell and femtocell is shown in Fig. 1. In case
of uplink, interferences of macrocell and femtocell are caused by femtocell UE
and macrocell UE, respectively. Similarly, in downlink, interferences of macrocell
and femtocell are caused by femtocell BS and macrocell BS. If macro/femtocell
UE is close to femto/macrocell BS, interference between macrocell and femtocell
becomes large. Thus, the channel capacity of macro/femtocell may be signifi-
cantly deteriorated.

Fig. 1. Co-channel interference scenario between macrocell and femtocell

Fig. 2. Co-channel interference scenario between femtocells

88 C.S. Kim et al.

When femtocells are deployed in a macrocell, we should also consider co-

channel interference among femtocells as shown in Fig. 2. Interferences are caused
by neighboring femtocell BS and femtocell UE. Especially, when the distance
between femtocells is short, the impact of interference from neighboring femtocell
may be high. That is, co-channel interference between femtocells can be a serious
problem when a lot of femtocells are concentrated in a small area.

3 System Model
To analyze the performance when femtocells are deployed in mobile WiMAX
systems, we consider a system model with 19 macrocells as shown in Fig. 3. In
our model, femtocells are uniformly deployed in the centered macrocell 1. The
attenuation of transmitted signal, called path loss, can differ according to channel
environment (e.g., indoor or outdoor). As path loss models, we apply Wireless
World Initiative New Radio (WINNER) II model to evaluate the throughput
[13]. Simulation parameters are in Table 1 [14].
To calculate the path loss between macrocell BS and UE, we use non-line of
sight (NLOS) outdoor propagation model. Also, we use NLOS indoor propaga-
tion model for calculating path loss between femtocell BS and UE. Path loss of
the indoor and outdoor signal is given as [13]:

P Lin (d) = 36.8 · log10 (d) + 37.78,

P Lout (d) = 35.74 · log10 (d) + 35.68,

Fig. 3. Cell structure for simulation

 Femtocell Deployment to Minimize Performance Degradation 89

Table 1. Simulation parameters

PARAMETER VALUE PARAMETER VALUE

Carrier frequency 2.5 GHz Channel bandwidth 10 MHz
Radius of macrocell 866 m Radius of femtocell 10 m
Maximum power of macrocell BS 20 W Maximum power of femtocell BS 0.1 W
Maximum power of UE 0.2 W Antenna gain of macrocell BS 15 dBi
Antenna gain of femtocell BS 2 dBi Antenna gain of UE -1 dBi
Lognormal std. dev. 8 dB Lognormal std. dev. 10 dB
(outdoor) (outdoor to indoor)
Lognormal std. dev. 4 dB Lognormal std. dev. 7 dB
(indoor) (indoor to outdoor)

where d is the distance between the transmitter and the receiver in meter. In or-
der to reflect the interference between macrocell and femtocell, indoor to outdoor
and outdoor to indoor path loss model are considered as follows [13].

P Lin to out (dout , din ) = P Lout (dout + din ) + 14 + 15(1 − cos(θ))2 + 0.5din ,

P Lout to in (dout , din ) = P Lout (dout + din ) + 13.8 + 0.5din ,

where dout is the distance between the outdoor node and the wall nearest to the
indoor node, din is the distance from the wall to the indoor node, and θ is the
angle between the outdoor path and the vertical direction of the wall. In the
simulation, θ is assumed to be 0 degree.
The received signal to interference ratio (SIR) is defined as the ratio of a
signal power to the interference power. Let Psr and Isr be strengths of received
signal and interference from sender s to receiver r, and M and F be the number
of macrocells and femtocells.
SIR of user l in the downlink at macrocell BS i is:
l
l PmacroBS i
SIRmacroBS i =  M ,
F l l
j=1 If emtoBS j + k=1,k=i ImacroBS k

where Ifl emtoBS j is the interference which occurs due to the signal transmited to
l
the macrocell user l in the femtocell BS j and ImacroBS k is the interference which

occurs due to the signal transmited to the macrocell user l in the macrocell BS
k. SIR of user l in the uplink at macrocell BS i is:
i
i PmacroUE l
SIRmacroUE l =  M ,
F i i
j=1 If emtoUE j + k=1,k=l ImacroUE k

where Ifi emtoUE j is the interference which occurs due to the signal scheduled
i
to transmit from the femtocell UE j to the macrocell BS i, and ImacroUE k is

the interference which occurs due to the signal scheduled to transmit from the
90 C.S. Kim et al.

macrocell UE k to the macrocell BS i. SIR of user l in the downlink at femtocell

BS i is:
Pfl emtoBS i
SIRfl emtoBS i = F M l ,
l
j=1,j=i If emtoBS j + k=1 ImacroBS k

SIR of user l in the uplink at femtocell BS i is:

Pfi emtoUE l
SIRfi emtoUE l = F M i .
i
j=1,j=l If emtoUE j + k=1 ImacroUE k

The received SIR is used to calculate channel capacity with the Shannon-Hartley
theorem. Channel capacity C is calculated as follows.

C = BW · log2 (1 + SIR),

where BW is channel bandwidth in Hz.

4 Performance Evaluation
Fig. 4 shows uplink and downlink channel capacity of macrocell according to
the number of femtocells. As the number of femtocell increases, the uplink chan-
nel capacity of macrocell steeply decreases, but downlink channel capacity of
macrocell decreases more slowly than uplink channel capacity. The reason is
that uplink transmit power of macrocell UE is lower than downlink transmit
power of macrocell BS and strength of uplink interference is similar to strength
of downlink interference. In order to maintain uplink and downlink channel ca-
pacity of macrocell high, the maximum transmit power of femtocell BS and UE
is controlled. We set the transmitted power of femtocell BS and UE to 0.1 W,
0.01 W, and 0.001 W in the simulation. When the transmitted power of femtocell
UE uses 0.001 W, the decreasing slope of uplink channel capacity of macrocell is
reduced because co-channel interference from femtocell UE is decreased. On the
other hand, macrocell UE in the downlink can get higher SIR than uplink of
macrocell due to a high signal strength of macrocell BS. However, the downlink
channel capacity of macrocell has smaller improvement than uplink of macrocell
even if the transmitted power of femtocell BS uses 0.001 W.
Compared with the channel capacity of macrocell as shown in Fig. 4, uplink
and downlink channel capacity of femtocell is high as shown in Fig. 5. Femtocell
has a high received SIR because the distance between femtocell BS and UE
is shorter than the distance between macrocell BS and UE. Since transmitted
power of macrocell UE is low compared with macrocell BS, co-channel inter-
ference from macrocell UE is relatively low. Although the transmitted power
of femtocell’s UE is weakened, the uplink channel capacity of femtocell main-
tains high. Accordingly, we can increase the uplink channel capacity of macro-
cell. The downlink channel capacity of femtocell is constant because interference
 Femtocell Deployment to Minimize Performance Degradation 91

Fig. 4. Uplink and downlink channel capacity of macrocell

Fig. 5. Uplink and downlink channel capacity of femtocell

from macrocell BS is more dominant than interference of neighboring femto-

cells. When the transmitted power of femtocell BS uses 0.001 W, the downlink
channel capacity of femtocell is about 90 Mbps. In this setting, femtocell can
provide subscriber with enough service quality and data speed while minimizing
the performance degradation of macrocell.
We consider the environment where transmitted power of femtocell BS and
UE is fixed to 0.001 W. Fig. 6 shows downlink and uplink channel capacity
92 C.S. Kim et al.

(a) Downlink

(b) Uplink

Fig. 6. Channel capacity of macrocell UE according to locations (The transmitted

power of femtocell BS and UE is 0.001 W)
 Femtocell Deployment to Minimize Performance Degradation 93

(a) Downlink

(b) Uplink

Fig. 7. Channel capacity of femtocell UE according to locations (The transmitted

power of femtocell BS and UE is 0.001 W)
94 C.S. Kim et al.

of macrocell according to the locations of macrocell UE. Macrocell UE has a

high channel capacity when the location of macrocell UE is close to macrocell
BS. The downlink channel capacity of macrocell is not influenced by co-channel
interference of femtocell because the transmitted power of macrocell BS is high
and the transmitted power of femtocell BS is low as shown in Fig. 6(a). On
the other hand, co-channel interference according to the number of femtocell
UEs highly affects the uplink channel capacity of macrocell UE since the signal
strength of macrocell UE is low as shown in Fig. 6(b).
Fig. 7(a) shows the downlink channel capacity of femtocell according to the
locations of femtocell. Femtocells close to macrocell BS has a low channel capac-
ity because the transmitted power of macrocell BS becomes a strong interference
to femtocell. If the location of femtocell is far away from macrocell BS, femtocell
has a high channel capacity. Fig. 7(b) shows uplink channel capacity of femtocell
according to the locations of femtocell. The uplink channel capacity of femtocell
does not depend on the distance between macrocell BS and femtocell. However,
the uplink channel capacity of femtocell decreases as the number of femtocells
increases.

5 Conclusion

In this paper, we investigated channel capacity of macrocell and femtocell ac-

cording to the number and power level of femtocells when femtocells are de-
ployed in mobile WiMAX systems. When femtocell uses high transmit power,
the channel capacity of macrocell severely decreases due to high interference
from femtocell as the number of femtocells increases. In our simulation environ-
ment, the adequate transmit power of femtocell BS and UE is 0.001 W. This
power level minimizes the performance degradation of macrocell while achiev-
ing enough femtocell throughput. With this femtocell transmit power, we also
examined the effect on the locations of femtocell BS and macrocell UE.

References

1. IEEE 802.16e-2005: Local and Metropolitan Networks-Part 16: Air Interface for
Fixed and Mobile Broadband Wireless Access System, Amendment 2: Physical
and Medium Access Control Layers for Combined Fixed and Mobile Operation in
Licensed Bands and Corrigendum 1 (2006)
2. WiMAX Forum: Mobile WiMAX-Part I: A Technical Overview and Performance
Evaluation (2006)
3. Gordon Mansfield: Femto Cells in the Us Market-Business Drivers and Femto Cells
in Us Market-Business Drivers and Consumer Propositions. FemtoCells Europe
(2008)
4. Kim, R.Y., Kwak, S.K., Etemad, K.: WiMAX Femtocell: Requirements, Chal-
lenges, and Solutions. IEEE Commun. Magazine 47(9), 87–91 (2009)
5. Chandrasekhar, V., Andrews, J.G., Gatherer, A.: Femtocell Networks: A Survey.
IEEE Commun. Magazine 46(9), 59–67 (2008)
 Femtocell Deployment to Minimize Performance Degradation 95

6. 3GPP R4-071661, Ericsson: Impact of HNB with controlled output power on macro
HSDPA capacity (2007)
7. 3GPP R4-080409, Qualcomm Europe: Simple Models for Home NodeB Interference
Analysis (2008)
8. Claussen, H.: Performance of macro- and co-channel femtocells in a hierarchical
cell structure. In: Proc. IEEE 18th International Symposium on Personal, Indoor
and Mobile Radio Communications (PIMRC 2007), pp. 3–7 (2007)
9. Yeh, S.-p., Talwa, S., Lee, S.-C., Kim, H.: WiMAX femtocells: A Perspective on
Network Architecture Capacity, and Coverage. IEEE Commun. Magazine 46(10),
58–64 (2008)
10. Sung, Y.-S., Jeon, N.-R., Woon, B.-W., Lee, J.-S., Lee, S.-C., Kim, S.-C.: Femto-
cell/Macrocell Interference Analysis for Mobile WiMAX System. In: IEEE VTS
Asia Pacific Wireless Communications Symposium (2008)
11. López-Pérz, D., Valcarce, A., Roche, G.D.L., Zhang, J.: OFDMA Femtocells:
A Roadmap on Interference Avoidance. IEEE Commun. Magazine 47(9), 41–48
(2009)
12. 3GPP TR25.820: 3G Home NodeB Study Item Technical Report. v8.2.0 (2008)
13. WINNER II WP1: WINNER II Part 1 Channel Models. IST-4-027756, D1.1.2,
V1.1 (2007)
14. WiMAX Forum: WiMAX System Evaluation Methodology. v.2.01 (2007)
 A Novel Frequency Planning for Femtocells in
OFDMA-Based Cellular Networks Using Fractional
Frequency Reuse

Chang-Yeong Oh, Min Young Chung, Hyunseung Choo, and Tae-Jin Lee

School of Information and Communication Engineering

Sungkyunkwan University
440-746, Suwon, South Korea, +82-31-290-7982
{ohchy,mychung,choo,tjlee}@ece.skku.ac.kr

Abstract. Femtocells are expected to be one of the emerging technologies for

next generation communication systems. For successful deployment of femto-
cells in the pre-existing macrocell networks, there are some challenges such as
the cell planning for interference management, handoff, and power control. In
this paper, we focus on frequency planning which can provide interference avoid-
ance for the co-existence of macrocells and femtocells. We propose a novel fre-
quency planning for femtocells in cellular networks using fractional frequency
reuse (FFR). We consider downlink performance of cellular systems based on
Orthogonal Frequency Division Multiplexing Access (OFDMA), e.g., WiMAX
and 3GPP Long Term Evaluation (LTE). Simulation results show that our scheme
indeed reduces the effect of additional co-channel interference (CCI) between a
given macrocell and deployed femtocells as well as neighboring macrocells.

1 Introduction
The concept of indoor cellular networks using home base stations (BSs), femtocells,
with low transmission power is drawing much interest. Femtocells are one of the emerg-
ing cost-effective technologies for both operators and users to enhance coverage and to
support higher data rate by bridging existing handsets and broadband wired networks.
However, deployment of femtocells in pre-existing cellular networks causes some prob-
lems to be addressed. One of them is that it requires intelligent frequency allocation for
femtocells and traditional macrocells when they operate simultaneously in the same
network [3]. So it is important to efficiently allocate frequency resources for femtocells
considering the effect of co-channel interference. Since frequency resources are usually
limited, some methods to enhance the efficiency are desirable.
Traditionally, there exist frequency allocation mechanisms which allocate frequency
resources according to frequency reuse factors (RFs) in multi-cell environment. One is

This research was supported by the MKE(The Ministry of Knowledge Economy), Korea, un-
der the ITRC(Information Technology Research Center) support program supervised by the
NIPA(National IT Industry Promotion Agency) (NIPA-2009-C1090-0902-0005) and (NIPA-
2009-C1090-0902-0046).

Corresponding author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 96–106, 2010.

c Springer-Verlag Berlin Heidelberg 2010
 A Novel Frequency Planning for Femtocells in OFDMA-Based Cellular Networks 97

the shared frequency allocation mechanism, in which frequency is shared with adjacent
cells, i.e., universal frequency reuse or frequency RF of 1. The other is the orthogo-
nal frequency allocation mechanism which allocates frequency according to the pre-
determined frequency patterns among multi cells with the frequency RF of 1 or more.
The former may be viewed as more efficient because all frequency resource can be uti-
lized in each of cells. However, it can suffer from performance degradation since the
performance of users located at the edge of a cell is degraded due to co-channel inter-
ference (CCI) from neighboring cells. So the frequency RF of 3 or above is generally
employed to mitigate CCI in 2G systems. The reduction of CCI is typically gained at
the cost of efficiency in frequency resource allocation [6].
Fractional frequency reuse (FFR), which is the mixture of the shared and orthogonal
frequency allocation, becomes one of the solutions to improve frequency efficiency
in OFDMA-based cellular systems, e.g., WiMAX and 3GPP Long Term Evaluation
(LTE). In most of FFR schemes, each cell is partitioned into two regions: inner region
and outer region. The inner region around a base station (BS) in a cell can use the
frequency RF of 1 due to low CCI from adjacent cells. The outer region far from a BS
in a cell uses different frequency bands with those in adjacent cells with the frequency
RF of 3 to reduce CCI [11].
FFR can be implemented easily in OFDMA-based cellular systems because the fre-
quency band of a cell is divided into subchannels and can be handled by the unit of
subchannel, a group of orthogonal subcarriers. Han et al. propose an FFR scheme to
enhance the flexibility of frequency assignment and cell performance, in which mobile
stations (MSs) in the inner region can use the subchannels assigned to both the inner
and the outer region of a cell according to the frequency partitioning [4]. Another FFR
scheme is introduced by noting that the inner region and the outer region are differently
served by not only frequency bands but also time slots [5].
While the performance of the users located at the edge of a cell is important in
the macrocell networks with FFR, interference management between macrocells and
femtocells is a main issue in femto / macro co-deployment environment with different
channel operation strategies: orthogonal channel allocation and shared channel alloca-
tion [1]. So the frequency assignment for femtocells is one of the key issues [2]. The
authors in [2] propose that femtocells in the inner region of a macrocell use different
subcahnnels with those for macrocell users to minimize interference, and femtocells in
the outer region of a macrocell use the same subcahnnels as those for macrocell users.
They suppose cell environment without FFR, in which the frequency RF of macrocells
is greater than 1.
In this paper, we propose a frequency planning mechanism with FFR for macrocells
and femtocells, in which co-channel operation is allowed in low CCI, and orthogonal
frequency allocation is adopted in high CCI. To reduce the effect of CCI we utilize
resource allocation in both frequency bands and time slots.
The remaining part of this paper is organized as follows. Section 2 describes the
basic FFR scheme for macrocells and presents the proposed frequency planning
98 C.-Y. Oh et al.

mechanism for femtocells and macrocells. Performance evaluation of the proposed

scheme is provided in Section 3. We conclude in Section 4.

2 Proposed Frequency Planning

2.1 FFR for Macrocells
In our proposed FFR for macrocells, a macrocell is grouped into two parts: inner region
and outer region. The decision on whether an MS belongs to inner or outer region is
made by the reported signal-to-interference-plus-noise-ratio (SINR) of the reference
signal at the initial configuration or at the periodic state update. If sufficient SINR is
observed, i.e., the reported SINR of an MS is above the pre-determined SINR threshold,
the BS considers that the MS belongs to the inner region. Otherwise, i.e., the reported
SINR of an MS is below the pre-determined SINR threshold, the BS considers that the
MS is in the outer region. The reported SINR of an MS is largely dependent on its
location, i.e., the farther an MS is from the BS in a cell, the larger the path-loss of the
signal from the BS becomes.
Fig. 1 shows an example, in which MSs in the inner region use the RF of 1 and those
in the outer region use the RF of 3. It illustrates the feasible frequency bands for macro-
cells according to the region that an MS belongs to. So the whole subchannels can be
allocated to the inner region while the outer region uses just 1/3 of all subchannels in a
cell. Furthermore, the service times of the inner and the outer region are separated by
time slots. The MSs in the inner region of macrocells can be served simultaneously be-
cause CCI from neighbor macro BSs is limited. They can utilize the whole subchannels

Inner Outer
Frequency
F1

F2 F0
F1

Cell 2
time

Frequency

Cell 1
F0 F2

Cell 3 time

Frequency
F0

F0
F3 F3
time

Fig. 1. Proposed frequency band/time slots allocation with FFR for macrocells
 A Novel Frequency Planning for Femtocells in OFDMA-Based Cellular Networks 99

Macrocell 1 Macrocell 2 Macrocell 3

Inner Outer Inner Outer Inner Outer
Frequency Frequency Frequency
F1

F0 F0 F2 F0
α time ratio
femtocell F3
F1+F3 time time
time
(a) Frequency band allocation for macrocell
F2+F3
Macrocell 2
Frequency Frequency Frequency
F1 F1
+
Macrocell 1 + F2
F0 or F1+F3 F2
+
F3 F3
Macrocell 3 time time time

F0 or F2+F3 (b) Frequency band allocation for inner group femtocell

F1+F2
Frequency Frequency Frequency
F0 or F1+F2 F1 F1
+
F0 F2 F0 + F0 F2
+
F3 F3

α time time time

(c) Frequency band allocation for outer group femtocell

Fig. 2. Proposed frequency planning for femtocells

during their time slots. The MSs in the outer region of macrocells can also be served
simultaneously but they must use the orthogonal subchannels in order to avoid high CCI
from neighbor macro BSs. In summary, the inner region and the outer region are allo-
cated shared or separate time slots and frequency bands to mitigate intra- and inter-cell
interference among the inner and the outer regions as well as among cells.

2.2 Frequency Planning for Femtocells

In this section, we propose a novel frequency planning for femtocells under the macro-
cells with FFR. First of all, femtocells should be deployed harmoniously in order not
to make minor impacts on the legacy macrocells but to maximize their performance. In
our proposal, femtocells are divided into the inner group and the outer group depending
on whether a femtocell is located in the inner region or the outer region of a mcarocell.
A femto BS is assumed to support auto-configuration because it has to be able to setup
a user controlled hot-spot. A femto BS reports the SINR of the reference signal of a
macrocell to the macro BS by pilot sensing, which is similar to the macro MSs when
they do in the setup procedure. Then the macro BS determines whether the femtocell
belongs to the inner or the outer group and notifies the appropriate frequency bands to
the femto BS. We assume that all macro and femto BSs are synchronized.
We now consider three factors for the frequency planning of femtocells: frequency,
time, and spatial state. The femtocells in the inner group, i.e., inner femtocells, are
served during the service time of the macro MSs in the outer region, i.e., outer service
time, and the subchannels orthogonal to the macrocell frequency bands are allocated
100 C.-Y. Oh et al.

in order to avoid intra CCI from the macrocell that the femtocell belongs to. For the
macro MSs in the outer region, only 1/3 of the overall frequency band is utilized to
reduce CCI to adjacent macrocells. So these 2/3 unused subcahnnels can be allocated
to the inner femtocells efficiently (see Fig. 2). The CCI between the inner femtocells
and the neighboring macrocells is limited due to low transmission power of femtocells
and relatively large path loss factors, e.g., high attenuation path-loss by long distance
and wall-loss by the walls in buildings.
For the femtocells in the outer group, i.e., outer femtocells, we note that they use low
transmission power and the CCI between the macrocell and the femtocells is small. So
they are allowed to use the same frequency bands and time slots, i.e., inner service time
as those of the macro MSs in the inner region. That is, the outer femtocells can utilize
all the subchannels in the given time slots. As shown in Fig. 2(c), outer femtocells use
the full frequency band during the inner service time and 2/3 of the band during the
outer service time.

3 Performance Evaluation

We consider 2-tier cellular networks consisting of M (=19) macrocells. The macro BSs
are located at the center of each cell. Femto BSs which have uniform separation with
one another are densely deployed within a macrocell (see Fig. 3). We use the modi-
fied COST-Walfish-Ikegami (WI) urban micro model for the non-line-of-sight (NLOS)
outdoor path-loss model [10].
out
Ploss [dB] = 31.81 + 40.5 log10 (d[m]) + χσout , (1)

where d is the distance between a sender and a receiver and χσout represents the outdoor
shadowing (log-normal fading), which is characterized by the Gaussian distribution

Femtocell

Macro BS
Macrocell 1
8

19 9

18 2 10

7 3

17 1 11

6 4

16 5 12

15 13

14

Fig. 3. 2-tier cellular networks and 234 uniformly separated femtocells in a macrocell
 A Novel Frequency Planning for Femtocells in OFDMA-Based Cellular Networks 101

Table 1. Modulation and Coding Scheme (MCS) table

SINR Code rate Mod. SINR Code rate Mod.

-4.34 1/12 QPSK 6.35 2/3 QPSK
-2.80 1/8 QPSK 9.50 1/2 16QAM
-1.65 1/6 QPSK 12.21 2/3 16QAM
0.31 1/4 QPSK 13.32 1/2 64QAM
1.51 1/3 QPSK 16.79 2/3 64QAM
4.12 1/2 QPSK 20.68 5/6 64QAM

Table 2. Simulation Parameters

Parameter Value
Inter macro cell distance (ICD) 1000 m
Radius of a femtocell 20 m
FFT size 1024
Total number of data subcarriers in one cell 768
Down-link symbol rate 9.76 k symbols/sec
Total frequency bandwidth in one cell 10 MHz
AWGN power density (N0 ) -174 dBm/Hz
Macro BS power 20 W
Femto BS power 20 mW
out
Outdoor Log-Normal fading (σ ) 10 dB
in
Indoor Log-Normal fading (σ ) 4 dB

with zero mean and standard deviation (σ out ). And the modified COST 231-multi wall
(MW) model one floor building [7] is employed for indoor propagation,


T
in
Ploss [dB] = 37 + 3.2 · 10 log10 (d[m]) + Ltw ntw + χσin , (2)
t=1

where d, T , Ltw , ntw , and χσin are the distance between a sender and a receiver, the
number of wall types, the wall loss according to wall type t, and the number of pen-
etrated type t walls, and the shadowing, which is characterized by the Gaussian dis-
tribution with zero mean and standard deviation (σ in ), respectively. We assume each
of the buildings with femto BSs has 1 outer (heavy) wall at 10m from the femto BS,
L1w =15dB, and 1 inner (light) wall at 5m from the femto BS, L2w =3dB. We have K data
subcarriers and Fi femtocells in macrocell i.
102 C.-Y. Oh et al.

80

70
RF of 1
60 FFR
FFR with Femtocells(proposed)
50 SINR threshold
Avg. SINR(dB)

40

30

20

10

−10

−20
0 100 200 300 400 500 600
Macro MS locations from Macro BS of a given cell(m)

Fig. 4. Average (Avg.) SINR of a macro MS according to the distance from the macro BS in a
given cell

The downlink SINR of user l in macrocell i ∈ {1, .., M } on subcarrier k ∈ {1, .., K}
is
(i)
(i) Pl,k
SIN Rl,k = , (3)

M 
Fi
cηk · Ikη + cjk · Ikj + N0 Δf
η=1, η=i j=1

(i)
where Pl,k , Ikη , Ikj , and N0 Δf are the received power of user l on subcarrier k in
macrocell i, CCI from macrocell η ∈ {1, .., M }, CCI from femtocell j ∈ {1, .., Fi } in
macrocell i, and Additive White Gaussian Noise (AWGN) on a subcareier, respectively.
And ck ∈ {0, 1} is defined as the collision coefficient having the value 1 if subcarrier k
is allocated to a specific cell so that it generates CCI to a reference cell, i.e., macrocell
i, and 0 otherwise. The downlik SINR of user m on subcarrier k in femtocell j of
macrocell i can be found as
(i,j)
(i,j) Pm,k
SIN Rm,k = , (4)

M 
Fi
cηk · Ikη + cνk · Ikν + N0 Δf
η=1 ν=1, ν=j

(i,j)
where Pm,k and Ikν are the received power of user m on subcarrier k in femtocell j of
macrocell i, CCI from femtocell ν ∈ {1, .., Fi } in macrocell i, respectively. We adopt
 A Novel Frequency Planning for Femtocells in OFDMA-Based Cellular Networks 103

6
x 10
18

16

14

12 Inner Macrocell: Proposed

Throughput(bps)

Outer Macrocell: Proposed

10 Inner Macrocell: RF=3
Outer Macrocell: RF=3
8 Total Macrocell: Proposed
Total Macrocell: RF=3
6

0
0 5 10 15 20
SINRth(dB)

Fig. 5. Downlink throughput of a macrocell according to different SINR thresholds

Modulation and Coding Scheme (MCS) in Table 1 [8] and use simulation parameters in
Table 2. The simplified proportional fairness allocation algorithm with the same power
for subcarriers is considered in [9] to allocate subcarriers.
We present the average (Avg.) SINR variation of a macro MS in Fig. 4 as it moves on
the x-axis from the center to the border when the macro BS is positioned at (x, y)=(0,0).
We consider the case where all macrocells use the same frequency band without femto-
cells, i.e., frequency RF of 1. In this case, we observe the Avg. SINR of a macro MS at
the cell edge is relatively low as we expect. On the other hand, the basic FFR scheme
without femtocells, shows that the Avg. SINR of a macro MS at the cell edge is greater
than that of the RF of 1 due to CCI avoidance from tier-1 cells. We set the SINR thresh-
old 5dB as the criterion to distinguish the inner and the outer region. We can observe
the Avg. SINR of a macro MS with 234 femtocells is similar to that without femtocells.
In the inner region there does not exist any effect of the deployment of femtocells at all
because the inner femto BSs and the macro BS operate at different time slots. The outer
femto BSs with co-channel operation do not interfere with the macro MSs much due to
their low transmission power, high attenuation by long distance, and wall-losses. In the
outer region, there is no additional CCI because all femto BSs and the macro BS are
served at different frequency bands. We can conclude the performance of macro MSs
with the numerous co-existing femtocells is not sacrificed in our proposed mechanism.
In Fig. 5, we evaluate the downlink throughput of a macrocell according to different
SINR thresholds when there exist 36 active macro MSs in the cell. We compare ours
104 C.-Y. Oh et al.

9
x 10
6
Total cell throughput: Proposed
Total cell throughput: RF=3
5.5

5
Throughput(bps)

4.5

3.5

2.5
0 5 10 15 20
SINRth(dB)

Fig. 6. Total throughput in a given cell area

with the scheme in [2] assuming that the RF of 3 is adopted. Note that the total band-
width of FFR is much wider than that of the RF of 3 since the inner region adopt the
RF of 1 and can use the full bandwidth for the macro MSs. So the throughput perfor-
mance of macrocell by using FFR is highly enhanced. In our proposal, as the SINR
threshold increases, Avg. SINR of the macro MSs in the inner region becomes higher.
Thus the spectral efficiency is improved. Moreover, the throughput performance of the
macro MSs in the outer region also increases thanks to the alleviation of CCI. However,
if the SINR threshold is set to a high value, the frequency resource may be insufficient
because many macro MSs may be included in the outer region but the bandwidth for
the outer region is limited.
The scheme in [2] serves femtocells and macrocell simultaneously, especially, femto-
cells in the outer region and the macrocell share the frequency bands but the femtocells
in the inner region and the macrocell use different frequency bands. If a macro MS is
in an outer femtocell, it may experience high interference from the femto BS due to
the co-channel operation. As the SINR threshold increases, the number of outer femto-
cells which may interfere with the macro MSs in the outer region also increases. This
degradation is observed in the inner macrocell: RF=3 in Fig. 5. However, this interfer-
ence scenario can be avoided in our proposed mechanism because macro MSs and outer
femtocells are served by orthogonal frequency bands and time slots.
Fig. 6 shows the total cell throughput for downlink, i.e., the sum of throughput of
all macro MSs and femto MSs in a given cell. Total cell throughput depends on the
 A Novel Frequency Planning for Femtocells in OFDMA-Based Cellular Networks 105

throughput of femto MSs because femtocells are densely deployed and the spectral
efficiency of femto MSs are much higher than that of macro MSs. We observe a trade-
off in Fig. 5 as the SINR threshold varies. In our proposed mechanism, a femtocell in
the outer region can use wider frequency band and longer time slots than a femtocell
in the inner region while in the scheme in [2] with the RF of 3 a femtocell in the inner
region use wider frequency bands than a femtocell in the outer region. The distribution
of inner and outer femtocells is determined by the SINR threshold and the total cell
throughput is shown to vary by the degree of distribution of femtocells.

4 Conclusion
In this paper, we have proposed a novel frequency planning for femtocells in OFDMA-
based cellular networks using FFR. FFR scheme is an efficient method not only to
enhance performance of edge users in a given cell but also to increase the spectral effi-
ciency by adopting the mixture of reuse factors according to different region in a cell.
Moreover, the basic rule of FFR, i.e., co-channel operation in low CCI or orthogonal
channel operation in high CCI, is considered as a useful policy for frequency plan-
ning of femtocells when they are deployed in pre-existing macrocells. There may ex-
ist some difficulties and risks in sharing frequency among macrocells and femtocells,
which cause performance degradation of macro MSs resulting from the effect of the
deployment of femtocells. So we utilize unused frequency band in a given macrocell to
inner femtocells for specific time slots in order to harmoniously serve macrocells and
femtocells by maximizing diversity gains.

References
1. Lȯpez-Pėrez, D., Valcarce, A., Roche, G., Zhang, J.: OFDMA Femtocells: A Roadmap on
Interference Avoidance. IEEE Communications Magazine 47(9), 41–48 (2009)
2. Guvenc, I., Jeong, M.-R., Watanabe, F., Inamura, H.: A hybrid frequency assignment for
femtocells and coverage area analysis for co-channel operation. IEEE Communications Let-
ters 12(12), 880–882 (2008)
3. Chandrasekhar, V., Andrews, J.G.: Femtocell Networks: A Survey. IEEE Wireless Commu-
nications 15(3), 1284–1536 (2008)
4. Han, S.S., Park, J., Lee, T.-J., Ahn, H.G., Jang, K.: A New Frequency Partitioning and Al-
location of Subcarriers for Fractional Frequency Reuse in Mobile Communication Systems.
IEICE Tranactions on Communications E91-B(8), 2748–2751 (2008)
5. Giuliano, R., Monti, C., Loreti, P.: WiMAX fractional frequency reuse for rural environ-
ments. IEEE Communications Magazine 46(9), 59–67 (2008)
6. Elayoubi, S.-E., Ben Haddada, O., Fourestie, B.: Performance evaluation of frequency plan-
ning schemes in OFDMA-based networks. IEEE Tranactions on Wireless Communica-
tions 7(5), 1623–1633 (2008)
7. Nihtilä, T.: Increasing Femto Cell Throughput with HSDPA Using Higher Order Modulation.
In: Proc. INCC 2008, Lahore, Pakistan, pp. 49–53 (2008)
8. Samsung Electronics, Experimental WiBro MCS Table (2007)
106 C.-Y. Oh et al.

9. Nguyen, T.-D., Han, Y.: A Proportional Fair Algorithm with QoS Provision in Downlink
OFDMA Systems. IEEE Communications Letters 10(11), 760–762 (2006)
10. Baum, D.S., Hansen, J., Salo, J.: An interim channel model for beyond-3G systems: extend-
ing the 3gpp spatial channel model (scm). In: Proc. VTC 2005-Spring, Stockholm, Sweden,
pp. 3132–3136 (2005)
11. Huawei: Soft frequency reuse scheme for UTRAN LTE, 3GPP, R1-050507 (2005)
12. Damosso, E. (ed.): Digital Mobile Radio Towards Future Generation Systems, COST 231
final report (1998)
 Association Analysis of Location Tracking Data for
Various Telematics Services*

In-Hye Shin and Gyung-Leen Park**

Dept. of Computer Science and Statistics, Jeju National University

690-756, Jeju Do, Republic of Korea
{ihshin76,glpark}@jejunu.ac.kr

Abstract. This paper proposes an approach that extracts the association

information from the location data obtained from the real fields but ignored so
far. We provide and apply the approach to the real-life location tracking data
collected from the Taxi Telematics system developed in Jeju, Korea. The
analysis aims at obtaining taxies’ meaningful moving patterns for the efficient
operations of them. The proposed approach provides the flow chart which
would not only take a glance around the overall analysis process but also help
save temporal and economic costs required to employ the same or similar data
mining analysis to similar services such as public transportations, distribution
industries, and so on. Especially, we perform an association analysis on both of
refined data and interesting factors extracted from the elementary analysis. The
paper proposes the refined association rule mining process as follow: 1)
obtaining the integrated dataset through the data cleaning process, 2) extracting
the interesting factors from the integrated dataset using frequency and
clustering method, 3) performing the association analysis, 4) extracting the
meaningful and value-added information such as moving pattern, or 5)
returning the feedback to adjust inappropriate factors. The result of the analysis
shows that the association analysis makes it possible to detect the hidden
moving patterns of vehicles that will greatly improve the quality of Telematics
services considering the business requirements.

1 Introduction
The Taxi Telematics system has operated an efficient taxi dispatch service since 2006
in Jeju island, Korea [1]. Each taxi equipped with a GPS receiver reports its location
to the central server every minute. The server is responsible for keeping such reports
from each taxi and handles the call request from a customer by finding and
dispatching a taxi closest to the call point. The tracking data obtained from the Taxi
Telematics system can be utilized for various analysis on both research and business
areas [2]. We have been doing our research to develop a data processing framework
*
This research was supported by the MKE (The Ministry of Knowledge Economy), Korea,
under the ITRC (Information Technology Research Center) support program supervised by
the IITA (Institute for Information Technology Advancement). (IITA-2009-C1090-0902-
0040).
**
Corresponding author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 107–117, 2010.
© Springer-Verlag Berlin Heidelberg 2010
108 I.–H. Shin and G.-L. Park

for an efficient analysis [3-9]. It is the exact fact that the empty taxi ratio can be
reduced by means of guiding a taxi to the spot where many passengers are waiting
for. In the project, the empty taxi ratio is around 80 % according to a survey in [6].
Thus, many taxi businesses are interested in taxis’ moving patterns in order to
decrease the consumption of fuels as well as increase the income. Today the fuels
consumption is more important to protect the environments. The real pattern analysis
using data mining is an interesting topic due to importance of the value-added
information, as shown in [10]. Especially, [4] has proposed the framework and
perform the clustering analysis for the location recommendation. The given analysis
framework consists of both of the data processing framework and the analysis
processing framework for the refined clustering. However, it does not provide
association rules effectively so far. In this regard, this paper is to propose the
association-specific analysis framework, perform the detailed analysis for association
rule mining, and provide concrete pattern information. The proposed analysis
processing framework consists of both of the given data processing framework [4]
and a novel analysis processing framework for the refined association analysis aiming
at detecting the frequent or meaningful moving patterns. This framework can help
perform even a sophisticated and a quick analysis. Particularly, the paper is to develop
the refined association analysis by means of taking into account interesting factors
such as the driving type of taxi, boarding hour, pick-up/drop-off time, pick-up/drop-
off area, boarding rate, and so on.
The result of analysis shows that the approach employed in the paper can be
extended to similar areas such as developing public transportation systems,
distribution systems, and so on.
The paper is organized as follows. Section 2 proposes the analysis process
framework for the location tracking data collected from the Taxi Telematics system.
Section 3 exhibits the summary of the results obtained from the refined association
analysis to search the taxis’ moving patterns, based on whole data, short driving data,
and long driving data, respectively. Finally, Section 4 concludes the paper.

2 Proposed Framework
This section proposes both of the data processing framework and the analysis
processing framework, in order to show effectively how to analyze the moving
patterns of taxis using association analysis. The proposed frameworks provide the
flow chart which would take a quick look around both of the data flow and the overall
analysis process. Also, it helps quickly perform the same or similar analysis, while
sparing the temporal and economic costs.

2.1 Proposed Data Processing Framework

Figure 1 outlines the overall procedure for location data processing. Taxis report their
location records to the central call server every minute. Each record includes the basic
GPS data such as timestamp, latitude, longitude, direction, speed taxi ID as well as
status fields (related with pick-up and drop-off). First of all, an analyzer transforms
the type of such records into data type readable in SAS then stores them with SAS
dataset types. We can analyze the location dataset using SAS analysis engine
including data mining analysis tool, Enterprise Miner (or E-Miner).
 Association Analysis of Location Tracking Data 109

Fig. 1. Data processing framework including SAS Data Mining

To obtain more useful information, data mining analysis process consists of the data
cleansing process, the modeling process which analyzes the refined data using a proper
method, the assessment process, the utilization process, and feedback process, as shown
in Fig. 1. As shown in Fig. 1, Modeling, fourth step of SEMMA [10], fits a predictive
model; modeling a target variable using a cluster analysis, an association analysis, a
regression model, a decision tree, a neural network, or a user-defined model. The paper
exploits an association method for data modeling. Association searches interesting
association or correlation relationships among a large set of data items.

2.2 Proposed Processing Analysis Processing for the Association Analysis

Fig. 2 shows the analysis processing framework for the refined association rule
miming based on the taxi location records to inform their valuable moving patterns
and moreover recommend the best location and time to the taxi drivers. The
framework shows an overall process of the association analysis.
The processing framework could be expansively adapted to another transport
industry, similar to a taxi business, such as delivery by changing some variables or
generating new variables, and repairing the flow chart partially. As shown in Fig. 2, a
data analyzer should load source records first, restore them as the SAS data type,
modify the SAS dataset through the variable transformation and data filtering for data
cleaning. We can create new data extracting the paired data by a pick-up and drop-off
record using data split/merge. Then we select the interesting or significant variables
form the pared dataset by the result of clustering. We adjust each variable’s category
by the result of frequency method, resulting in TaxiType (driving type of Taxi),
SArea/EArea (pick-up area/drop-off area), SDay/EDay (pick-up day/drop-off day),
STimeR/ETimeR (pick-up time/drop-off time). TaxiUseRatioR (boarding rate in
terms of hour) variable, which means the boarding rate according to a driving hour, is
obtained by dividing the number of pick-up report into the number of total report. The
110 I.–H. Shin and G.-L. Park

analyzer completes the integrated dataset by merging original dataset and extracted
interesting factors. It also needs new dedicated dataset for association analysis
consisting of ID and Target variable. We divide association dataset into two by
TaxiType variable with ShortDriveTaxi (short-driving) and the others (long-driving).
The analyzer can return the feedback to the previous step (extraction of interesting
factors), if not detect strong association rules. There is a recurrence of such a refined
data mining analysis process, adjusting the feedback factors which mean modification
of the given factor category and detection of a new factor until obtaining the good
association rules.

Fig. 2. The analysis processing framework for the association rule mining of taxi records

2.3 Extracted Interesting Factors

Table 1 shows the extracted interesting factors, or categorical variables, obtained from
clustering and frequency method using the refined and integrated SAS dataset. Note
that each frequency should be alike within each variable’s category, or not biased. As
shown in Table 1, we omit EArea, EDay, ETimeR variables (related with drop-off),
which correspond to SArea, SDay, and STimeR (related with pick-up), due to space
limitation of the paper as well as similar frequency distribution as pick-ups. Note that
STimeR variable provides additionally the average frequency per hour in parenthesis,
due to non-equaled time range.
 Association Analysis of Location Tracking Data 111

Table 1. Summary of each interesting factor

Percent
Variable Category Description Frequency
(%)
ShortDriveTaxi ~ 30minutes 77447 95.51
TaxiType LongDriveTaxi 30minutes ~ 2hours 2309 2.85
TourTaxi 2hours ~ 1331 1.64
5MinDriving 1~5minutes 44507 54.89
5-10MinDriving 5~10minutes 22254 27.44
DrivingMinR
10-15MinDriving 10~15minutes 8787 10.84
15~30MinDriving 15~30minutes 5539 6.83
LowUse 10~15% 33082 40.80
TaxiUseRatioR MiddleUse 15~20% 23389 28.84
HighUse 20~25% 24616 30.36
SAirport Ariport 3727 4.60
SOldTown Old town 43644 53.82
SArea
SNewTown New town 22929 28.28
SEtc. The others 10787 13.30
SMonday Monday 9932 12.25
STuesday Tuesday 11405 14.07
SWednesday Wednesday 12335 15.21
SDay SThursday Thursday 11582 14.28
SFriday Friday 12996 16.03
SSaturday Saturday 12033 14.84
SSunday Sunday 10804 13.32
SOfficeStart 07~10o'clock 10469 (3490) 12.91
SOfficeHour 10~18o'clock 30673 (3834) 37.83
SOfficeEnd 18~21o'clock 17434 (5811) 21.50
STimeR
SEvening 21~23o'clock 8626 (4316) 10.64
SNight 23~01o'clock 6906 (3453) 8.52
SDawn 01~07o'clock 6979 (1163) 8.61

3 Association Analysis Results

This section shows the results obtained from the refined association analysis. In the
previous section, we have split the taxi association datasets into three categories,
whole (81,087 records), short driving (77,447 records), and long driving (3,640
records). 2. Thus we present three association results according to each dataset.
Association has a support level and a confidence level, two main measures of rule
interestingness. In addition, lift is a measure of strength of the association rule.
Typically, association rules are considered interesting if they satisfy both a minimum
support threshold and a minimum confidence threshold. Such thresholds can be set by
users or domain experts. Association rule mining is a two-step process: 1) finding all
frequent itemsets; 2) generating strong association rules from the frequent itemsets.
Note that we set minimum confidence to 10% by using the default association setting.
112 I.–H. Shin and G.-L. Park

3.1 Analysis Results for Whole Taxis

Table 2 shows the meaningful and interesting association rules extracted from the
association analysis based on the whole location data records. Stronger associations

Table 2. Association rules obtained from whole taxi records

Confidence Support Lift Rule
100.00 10.64 3.29 SEvening ==> HighUse (rule 1)
70.04 15.06 2.31 SOfficeEnd ==> HighUse (rule 2)
83.53 31.60 2.05 SOfficeHour ==> LowUse (rule 3)
59.62 7.70 2.07 SOfficeStart ==> MiddleUse (rule 4)
100.00 5.69 3.29 SOldTown & SEvening ==> HighUse (rule 5)
69.94 8.22 2.30 SOldTown & SOfficeEnd ==> HighUse (rule 6)
74.85 8.79 1.46 SOldTownTown SOfficeEnd & ==> EOldTown
83.40 17.61 2.04 SOldTown & SOfficeHour ==> LowUse (rule 7)
72.40 8.50 1.46 SOldTown & SOfficeEnd ==> ShortDriveTaxi & EOldTown
83.17 8.61 2.04 SOfficeHour & SNewTown ==> LowUse (rule 8)
33.80 5.42 1.11 SFriday ==> HighUse
45.03 5.52 1.10 SMonday ==> LowUse
39.38 5.84 0.97 SSaturday ==> LowUse
41.99 5.60 1.03 SSunday ==> LowUse
40.46 5.78 0.99 SThursday ==> LowUse
41.87 5.89 1.03 STuesday ==> LowUse
40.56 6.17 0.99 SWednesday ==> LowUse
82.41 4.80 2.02 SOfficeHour & SFriday ==> LowUse
84.48 5.03 2.07 SWednesday & SOfficeHour ==> LowUse
71.78 6.31 1.40 SFriday & SOldTown ==> EOldTown
69.82 6.14 1.41 SFriday & SOldTown ==> ShortDriveTaxi & EOldTown

Fig. 3. The main association rules of whole taxi location records

 Association Analysis of Location Tracking Data 113

are selected according the support level and the confidence level. The table shows that
passengers use mostly taxi at the closing time of the office hours with 15.06% of the
support level and 70.04% of the confidence level while the support level and the
confidence level is 10.64% and 100%, respectively, for the evening hours (21~23
o'clock). Also it shows that only Friday is associated with HighUse of the taxies with
5.42% of the support level and 33.80% of the confidence level.
We depict graphs to compare those strong rules in Fig. 3 and Fig. 4. In Fig. 3, each
polygon’s shape, its color, and its size are regarded as confidence, support, lift,
respectively. Fig. 4 shows the rule comparison in terms of two important factors, the
confidence level and the support level.

3.2 Analysis Results for Short Driving Taxies

Table 3 shows the associations based on the short-driving taxi records. Eight strong
associations are selected. Fig. 5 and Fig. 6 show the comparative graph among those

Fig. 4. Confidence and support of the main association rules

Table 3. Association rules obtained from short-driving taxi records

Confidence Support Lift Rule

100.00 10.67 3.31 SEvening ==> HighUse (rule 1)
54.66 4.69 1.81 SNight ==> HighUse (rule 2)
45.34 3.89 1.57 SNight ==> MiddleUse (rule 3)
69.86 14.87 2.31 SOfficeEnd ==> HighUse (rule 4)
83.58 31.63 2.05 SOfficeHour ==> LowUse (rule 5)
69.85 8.18 2.31 SOldTown & SOfficeEnd ==> HighUse (rule 6)
70.45 4.46 2.33 SOfficeEnd & SNewTown ==> HighUse (rule 7)
100.00 3.11 3.31 SNewTown & SEvening ==> HighUse (rule 8)
75.89 13.61 1.44 SNewTown & ENewTown ==> 5MinDrving
68.69 27.33 1.30 SOldTown & EOldTown ==> 5MinDrving
76.03 8.90 1.47 SOldTown & SOfficeEnd ==> EOldTown
46.59 5.46 1.56 SOldTown & SOfficeEnd ==> EOldTown & 5MinDrving
114 I.–H. Shin and G.-L. Park

Fig. 5. The main association rules of short-driving taxies

Fig. 6. Confidence and support of the main association rules for short-driving taxies

strong rules. The main association rules mean that short-driving taxis pick up many
passengers at the close of office hours with support level 14.87% and confidence level
69.86%, respectively. Also mainly passengers move either within the old town or the
new town within 5 minutes by a taxi. Especially, many taxies picking up a passenger
move within the old town at the close of office hours.

3.3 Analysis Results for Long Driving Taxies

Taxi drivers are interested in the long driving pattern than the short due to the
economic profit and the fuel cost, because the short driving consumes more fuel than
 Association Analysis of Location Tracking Data 115

the long driving. Table 4 shows the meaningful and interesting associations based on
the long-driving taxi records. Fig. 7 and Fig. 8 depict the main association rules.
Finally, the main association indicates that long-driving taxis mainly move from old
town to old town, especially at the close of office hours.

Table 4. Association rules obtained from long-driving taxi records

Confidence Support Lift Rule

49.22 3.49 1.25 SDawn ==> LowUse (rule 1)
50.78 3.60 1.84 SDawn ==> MiddleUse (rule 2)
100.00 10.05 3.02 SEvening ==> HighUse (rule 3)
56.23 4.09 1.70 SNight ==> HighUse (rule 4)
43.77 3.19 1.59 SNight ==> MiddleUse (rule 5)
73.12 18.98 2.21 SOfficeEnd ==> HighUse (rule 6)
82.43 30.93 2.10 SOfficeHour ==> LowUse (rule 7)
40.00 4.84 1.02 SOfficeStart ==> LowUse (rule 8)
60.00 7.25 2.17 SOfficeStart ==> MiddleUse (rule 9)
100.00 4.56 3.02 SOldTown & SEvening ==> HighUse (rule 10)
100.00 3.05 3.02 SNewTown & SEvening ==> HighUse (rule 11)
71.87 8.98 2.17 SOldTown & SOfficeEnd ==> HighUse (rule 12)
82.27 14.53 2.10 SOldTown & SOfficeHour ==> LowUse (rule 13)
81.60 7.55 2.08 SOfficeHour & SNewTown ==> LowUse (rule 14)
61.18 2.86 2.22 SOldTown & SOfficeStart ==> MiddleUse (rule 15)
49.21 22.36 1.30 SOldTown ==> EOldTown
51.21 6.40 1.36 SOldTown & SOfficeEnd ==> EOldTown
28.62 6.40 1.10 SOldTown & EOldTown ==> SOfficeEnd

Fig. 7. The main association rules of long-driving taxies

116 I.–H. Shin and G.-L. Park

Fig. 8. Confidence and support of the main association rules for long-driving taxies

4 Conclusions
The paper has developed an analysis processing framework for detecting the hidden
association rules of moving objects. In order to show the effectiveness of the
framework, the framework has been applied to the real-life location history data
accumulated from the Taxi Telematics system developed in Jeju Island. The proposed
framework provides the flow chart which would have a glance around the overall
analysis process as well as help save temporal and economic costs consumed to
quickly address the same or similar mining analysis. The paper has extracted diverse
interesting categorical factors such as taxi’s driving type, driving time, driving area,
boarding rate, and so on. The approach enables us to obtain meaningful and value-
added association rules of moving objects by performing repeatedly the refined
association analysis as follow: 1) obtaining the integrated dataset through the data
cleaning process such as filtering and split-merge, 2) extracting the various
categorical variables from the integrated dataset after performing the primary analysis
such as frequency and clustering, 3) performing the method based on another
dedicated dataset created for association, and 4) deducing the meaningful and value-
added pattern information such moving pattern, or 5) returning the feedback in order
to appropriately adjust the categories of extracted factors, if desired. The approach
can be applied not only to taxi business but also to similar areas such as public
transportation services and distribution industries.

References
[1] Lee, J., Park, G., Kim, H., Yang, Y., Kim, P., Kim, S.: A telematics service system based
on the Linux cluster. In: Shi, Y., van Albada, G.D., Dongarra, J., Sloot, P.M.A. (eds.)
ICCS 2007. LNCS, vol. 4490, pp. 660–667. Springer, Heidelberg (2007)
 Association Analysis of Location Tracking Data 117

[2] Hariharan, R., Toyama, K.: Project Lachesis: Parsing and modeling location histories. In:
Egenhofer, M.J., Freksa, C., Miller, H.J. (eds.) GIScience 2004. LNCS, vol. 3234, pp.
106–124. Springer, Heidelberg (2004)
[3] Lee, J., Park, G.: Design and implementation of a movement history analysis framework
for the taxi telematics system. In: Asia-Pacific Conference on Communications, pp. 1–4
(2008)
[4] Shin, I., Park, G., Saha, A., Kwak, Y., Kim, H.: Analysis of Moving Patterns of Moving
Objects with the Proposed Framework. In: Gervasi, O., Taniar, D., Murgante, B., Laganà,
A., Mun, Y., Gavrilova, M.L. (eds.) ICCSA 2009. LNCS, vol. 5593, pp. 443–452.
Springer, Heidelberg (2009)
[5] Lee, J., Hong, J.: Design and implementation of a spatial data processing engine for the
telematics network. Applied Computing and Computational Science (2008)
[6] Lee, J.: Traveling pattern analysis for the design of location-dependent contents based on
the Taxi telematics system. In: International Conference on Multimedia, Information
Technology and its Applications (2008)
[7] Liao, Z.: Real-time taxi dispatching using global positioning systems. Communication of
the ACM, 81–83 (2003)
[8] He, H., Jin, H., Chen, J., McAullay, D., Li, J., Fallon, T.: Analysis of Breast Feeding
Data Mining Methods. In: Proc. Australasian Data Mining Conference, pp. 47–52 (2006)
[9] Madigan, E.A., Curet, O.L., Zrinyi, M.: Workforce analysis using data mining and linear
regression to understand HIV/AIDS prevalence patterns. Human Resource for Health 6
(2008)
[10] Matignon, R.: Data Mining Using SAS Enterprise Miner. Wiley, Chichester (2007)
 An Efficient ICI Cancellation Method for
Cooperative STBC-OFDM Systems

Kyunghoon Won, Jun-Hee Jang, Se-bin Im, and Hyung-Jin Choi

School of Information and Communication Enginneering,

Sungkyunkwan University,
Suwon, Korea
{kairess,hellojjh,yuner,hjchoi}@ece.skku.ac.kr

Abstract. In this paper, an efficient inter carrier interference (ICI) cancellation

method for cooperative space time block coded orthogonal frequency division
multiplexing (STBC-OFDM) system is presented. In cooperative STBC-OFDM
system, ICI cancellation is needed because ICI due to the separated local oscilla-
tors always exists. To solve the complexity problem of matrix inversion opera-
tion in zero forcing method which shows the best performance, ICI cancellation
method using sparse matrix decomposition (SMD) has been proposed. However,
overall complexity is increased in proportion to the third order of required tap
size which also increases in proportion to FFT size or carrier frequency offsets
(CFOs). Considering implementation issue, the conventional method still has not
sufficiently overcome the performance versus complexity trade-offs. Therefore,
we propose an ICI cancellation method that focuses on solving practical com-
plexity problem of conventional method. The proposed method adaptively de-
cides the required tap size of each sparse matrix through signal to interference
and noise ratio (SINR) measurement to reduce the complexity of conventional
method, and we verified that the proposed method improves the performance
versus complexity trade-offs compared with conventional method.

Keywords: Equalization, Frequency offset, ICI cancellation, STBC, OFDM.

1 Introduction
Spatial diversity offers prominent benefits in link reliability and spectral efficiency
through the use of multiple antennas at the transmitter and/or receiver side [1], [2].
Unfortunately, the use of multiple-antenna techniques might not be practical for some
reasons; especially due to the antenna size and power constraints. For this reason,
recently, there has been significant interest in a class of techniques known as coopera-
tive communication, which allows single-antenna mobiles to reap some of the bene-
fits of multiple-antenna systems by cooperation of in-cell users [3], [4].
Since space time block codes (STBC) were developed originally for frequency flat
channels [2], an effective way to use them on frequency selective channels is to use
them along with orthogonal frequency division multiplexing (STBC-OFDM) [5]. By
using space time coded with OFDM, frequency selective channel is converted into
multiple frequency flat channels.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 118–128, 2010.
© Springer-Verlag Berlin Heidelberg 2010
 An Efficient ICI Cancellation Method for Cooperative STBC-OFDM Systems 119

However, due to the distributed nature of cooperative STBC-OFDM system, each

of the transmitters has a separate local oscillator, and synchronization becomes a
difficult problem or requires feedback process to achieve. Furthermore, due to the
physically separated local oscillators, different CFOs are received in one receiver
antenna, and they cannot be easily assumed identical. Therefore, inter carrier
interference (ICI) cancellation process is always needed in cooperative STBC-OFDM
systems, because ICI caused by residual CFOs always exists at the receiver side.
Generally, zero forcing method has the best performance by multiplying the inver-
sion of ICI matrix to the received signal. However it requires very high implementa-
tion complexity for matrix inversion. Therefore, in order to solve the complexity
problem of matrix inversion operation, many proposals have been presented [6]-[9].
Some of the proposals achieve the complexity reduction, such as simplified zero
forcing (ZF) method using sparse matrix decomposition (SMD), and iterative ICI
cancellation method using the iterative operation of ICI estimation and cancellation
[8], [9].
Especially, the ICI cancellation method which uses SMD can reduce the complex-
ity of matrix inversion, because the major portion of ICI components are distributed
in the diagonal or near diagonal position of ICI matrix [7], and the inversion operation
of ICI matrix can be approximated by using SMD proposed in [10]. However, consid-
ering time-varying multi-path fading channel, it still have not sufficiently overcome
the performance versus complexity trade-offs.
Therefore, in this paper, we propose an efficient ICI cancellation method which
can efficiently reduce the practical complexity of conventional method without per-
formance loss by using adaptive sparse matrix decomposition (ASMD).
This paper is organized as follows. In Section 2, we introduce the cooperative
STBC-OFDM system model and ICI problem. The proposed ICI cancellation method
using adaptive sparse matrix decomposition method is described in Section 3. In
Section 4, the results of performance and complexity comparisons between the con-
ventional method and the proposed method are presented. A brief conclusion is drawn
in Section 5.

2 System Model
The cooperative communication system considered in this paper includes one source
node, one destination node, and one of relay node. The basic premise in this work is
that source and relay have information of their own to send, and would like to cooper-
ate in order to send this information to the receiver according to the cooperative
STBC scheme as shown in Fig. 1. That is, we do not consider signal distortion in the
source-to-relay link. In Fig. 1, two consecutive symbols on subcarrier k, Xk,0 and Xk,1,
are encoded and transmitted from the source and the relay antenna, respectively, to
obtain the diversity order 2.
During even symbol period, Xk,0 and Xk,1 are transmitted from the source antenna
and relay antenna, and during odd symbol period, (-Xk,1)* and (Xk,0)* are transmitted
from the source antenna and the relay antenna, where (·)* denotes the complex conju-
gate operation.
120 K. Won et al.

time 2 s : X k ,1
time 2 s + 1: X k*,0

H k1 = H k1,0 = H k1,1

H k0 = H k0,0 = H k0,1
time 2 s : X k ,0
time 2 s + 1: − X k*,1

Fig. 1. Schematic representation of cooperative STBC-OFDM system

With the assumption of H kj = H kj,0 = H kj,1 , which is a case of quasi-static channel

for two symbol period, the received signal Yk,i for symbol duration 0 and 1 (i=0,1) can
be expressed as

Yk ,0 = H k0 ⋅ X k ,0 + H k1 ⋅ X k ,1 + Wk0 ,
(1)
Yk ,1 = H k0 ⋅ ( − X k ,1 ) + H k1 ⋅ ( X k ,0 ) + Wk1 , 0 ≤ k ≤ N − 1
* *

where N is FFT size, H kj,i is channel frequency response (CFR) of received signal on
the j-th link, and Wk j is zero-mean white Gaussian noise with variance σ 2 .
From (1), we can easily obtain diversity gain by using simple linear combination as
described in [2]

Xˆ k ,0 = ( H k0 ) ⋅ Yk ,0 + H k1 ⋅ ( Yk ,1 )
* *

( 2
= H k0 + H k1
2
)⋅ X k ,0 + ( H k0 ) ⋅ Wk0 + H k1 ⋅ (Wk1 )
* *

(2)
Xˆ k ,1 = ( H k1 ) ⋅ Yk ,0 − H k0 ⋅ (Yk ,1 )
* *

( 2
= H k0 + H k1
2
)⋅ X k ,1 + ( H k1 ) ⋅ Wk0 − H k0 ⋅ (Wk1 ) .
* *

With the residual CFOs, the received signal for symbol duration 0 and 1 can be ex-
pressed as [6]
 An Efficient ICI Cancellation Method for Cooperative STBC-OFDM Systems 121

N −1
Yk ,0 = Γ 00 ⋅ H k0 ⋅ X k ,0 + Γ10 ⋅ H k1 ⋅ X k ,1 + ∑ ( Γ 0k − i ⋅ H i0 ⋅ X i ,0 + Γ1k − i ⋅ H i1 ⋅ X i ,1 ) + Wk0
i =0
i≠k

N −1
Yk ,1 = −Γ 00 ⋅ H k0 ⋅ X k*,1 + Γ10 ⋅ H k1 ⋅ X k*,1 + ∑ ( −Γ 0k − i ⋅ H i0 ⋅ X i*,0 + Γ1k − i ⋅ H i1 ⋅ X i*,1 ) + Wk1 (3)
i=0
i≠k

sin (π ⋅ ε j )
Γ kj = (
⋅ exp jπε j ( N − 1 N ) − jπ ( ε j − k ) N )
(
N ⋅ sin π ( ε j − k ) N )
where εj is the residual CFO from the j-th link which is normalized by the subcarrier
spacing, and the second parts in (3) means the ICI components represented by a sum
for i ≠ k on subcarrier k caused by CFOs.
Fig. 2 shows the performance degradations caused by ICI in cooperative STBC-
OFDM systems. In spite of perfect channel estimation, from Fig. 2, we can see that
there are severe performance degradations due to the loss of orthogonality between
subcarriers.

1.E+00

1.E-01

1.E-02

1Tx-1Rx (without CFO)

2Tx-1Rx (without CFOs)
BER

1.E-03 CFO=±0.1
CFO=±0.2
CFO=±0.3
1.E-04 CFO=±0.4

FFT size : 128

CP length: FFT size / 4
1.E-05 Data Modulation : QPSK
Channel : COST207-TU
Mobile speed : 3 km/h
Channel estimation : Perfect
1.E-06
0 5 10 15 20 25 30

Eb/No [dB]

Fig. 2. Performance degradations caused by ICI in cooperative STBC-OFDM systems

122 K. Won et al.

3 An Efficient ICI Cancellation Method

Equation (3) can be expressed as matrix form in (4).
⎡ Y0 ⎤ ⎡ A B ⎤ ⎡ X0 ⎤ ⎡ W0 ⎤
⎢ *⎥ = ⎢ ⎥⎢ ⎥ + ⎢ *⎥
⎣ Y1 ⎦ ⎣C D ⎦ ⎣ X1 ⎦ ⎣ W1 ⎦
⎡Y0,i ⎤ ⎡ X 0,i ⎤ ⎡W0,i ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ (4)
⎢ Y1,i ⎥ ⎢ X 1,i ⎥ ⎢ W1,i ⎥
Yi = ,X = ,W =
⎢M ⎥ i ⎢M ⎥ i ⎢M ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢⎣YN −1,i ⎥⎦ ⎢⎣ X N −1,i ⎥⎦ ⎢⎣WN −1,i ⎥⎦
where A, B, C, and D are N × N size sub-matrices.
⎡Γ 00 H 00 Γ 0−1 H10 L Γ10− N H N0 −1 ⎤ ⎡Γ10 H 01 Γ1−1H11 L Γ11− N H 1N −1 ⎤
⎢ 0 0 ⎥ ⎢ 1 1 ⎥
⎢Γ H Γ 00 H10 L Γ 02 − N H N0 −1 ⎥ ⎢Γ1 H 0 Γ10 H11 L Γ12 − N H N1 −1 ⎥
A=⎢ 1 0 ⎥ B=⎢ M ⎥ (5)
M M O M M O M
⎢ ⎥ ⎢ ⎥
⎢⎣Γ N −1H 0 Γ N − 2 H1 L Γ 00 H N0 −1
0 0 0 0
⎥⎦ ⎢⎣Γ1N −1H 01 Γ1N − 2 H11 L Γ10 H 1N −1 ⎥⎦
⎡ Γ1*
0 H0
1*
Γ1*−1 H11* L Γ11*− N H 1*N −1
⎤ ⎡ −Γ 0*
0 H0
0*
− Γ 0*
−1 H 1
0*
L − Γ10*− N H N0*−1 ⎤
⎢ 1* 1* ⎥ ⎢ ⎥
⎢ Γ1 H 0 Γ1*0 H11* L Γ1*2 − N H 1N −1 ⎥ ⎢ −Γ1 H 0 − Γ 0* L − Γ0*
0* 0* 0* 0*
0 H1 2 − N H N −1 ⎥
C=⎢ ⎥ D=⎢ M ⎥.
M M O M M O M
⎢ ⎥ ⎢ ⎥
⎢⎣ Γ1*N −1H 01* Γ1*N − 2 H11* L Γ1*0 H 1N −1 ⎥⎦ ⎢⎣ −Γ 0* 0*
N −1 H 0 − Γ 0*
N − 2 H1
0*
L − Γ 00*H N0*−1 ⎥⎦

Ignoring system complexity, for the perfect cancellation of ICI, we can simply multi-
ply the inverse of ICI matrix to the received signal matrix (zero forcing method).
However, as you can see in (4), the implementation complexity of the inversion op-
eration of 2 N × 2 N size ICI matrix is too high.
To reduce the complexity of matrix inversion operation in zero forcing method, [7]
proposed the ICI cancellation method using SMD introduced in [10]. The major por-
tion of ICI components in (5) are distributed in the diagonal or near diagonal position
of ICI matrix. Therefore, by using SMD, inversion of ICI matrix can be transmitted
into a series of inversions of its diagonal sub-blocks as illustrated in Fig. 3.
By using the SMD method, (4) can be approximated as
⎡ Yn`.0 ⎤ ⎡ A n B n ⎤ ⎡ X`n ,0 ⎤ ⎡ Wn`,0 ⎤
⎢ `* ⎥ = ⎢ ⎥⎢ ` ⎥+⎢ ` ⎥
⎢⎣ Yn,1 ⎥⎦ ⎣Cn Dn ⎦ ⎢⎣ X n ,1 ⎥⎦ ⎢⎣ Wn,1 ⎥⎦
T
X`n,i = ⎡⎣ X n ,0 X n +1,0 L X n + 2 q ,0 ⎤⎦ (6)
T
where Yn`,i = ⎡⎣ Yn ,0 Yn +1,0 L Yn + 2 q ,0 ⎤⎦ , 0 < n ≤ N − (2q + 1)
T
Wn`,i = ⎡⎣Wn,0 Wn +1,0 L Wn + 2 q ,0 ⎤⎦
where n is sparse matrix index, 2q+1 is the tap size of diagonal sub-blocks, and
An is shown in (7). an, n is the elements of A, Bn, Cn, and Dn can be obtained
similarly [7].
 An Efficient ICI Cancellation Method for Cooperative STBC-OFDM Systems 123

0 0
1
2
1

N − 2q − 2
N − 2q − 1 N − 2q − 1

Fig. 3. Sparse matrix decomposition proposed in [10]

⎡ an , n a n , n +1 L an , n + q L 0 ⎤
⎢a an +1, n +1 M ⎥
⎢ n +1, n ⎥
⎢ M 0 ⎥
⎢ ⎥
A n = ⎢ an + q , n an + q , n + 2 q ⎥ (7)
⎢ 0 M ⎥
⎢ ⎥
⎢ M an + 2 q −1, n + 2 q −1 an + 2 q −1, n + 2 q ⎥
⎢ 0 L an + 2 q , n + q L an + 2 q , n + 2 q −1 an + 2 q , n + 2 q ⎥⎦
⎣

Therefore, mitigation of ICI is accomplished by multiplying the inversion of sparse

matrix as follows

⎡Xˆ ⎤ ⎡ A B ⎤ −1 ⎡ Yn ,0 ⎤ ⎡ X ⎤ ⎡ A B ⎤ −1 ⎡ W ⎤
⎢ ⎥ = ⎢ n n⎥ ⎢ * ⎥ = ⎢
n ,0 n ,0 n ,0
⎥+⎢
n n
⎥ ⎢ ⎥. (8)
⎢⎣ X n ,1 ⎥⎦ ⎣ n n ⎦ ⎢⎣ Yn ,1 ⎦⎥ ⎣ n ,1 ⎦ ⎣ n n ⎦ ⎣ n ,1 ⎦
ˆ C D X C D W

In the case of the above SMD method, although the total number of complex multipli-
cations for the matrix inversion is reduced from (2N)3 to ( N − 2q ) ⋅ ( 2 ( 2q + 1) ) , re-
3

quired complexity remains high because the tap size q increases as increasing CFOs
and FFT size N.
However, considering time and frequency selectivity due to mobile movement and
multi-path delay, respectively, each subcarriers are unequally interfered each other.
Therefore, there is no necessity for the size of sparse matrix to be fixed. If each sparse
matrix has adequately small tap size q* within maximum tap size q depending on the
amount of ICI from each neighboring subcarriers, the total amount of complex multi-
plications for the matrix inversion is reduced to less than ( N − 2q ) ⋅ ( 2 ( 2q + 1) ) .
3

From this point of view, we propose an adaptive sparse matrix decomposition

(ASMD) method which is based on a kind of SINR measurement to adaptively decide
124 K. Won et al.

the efficient tap size q*. Without loss of generality, we assume that all transmitted
signals have average power 1, i.e., E[XHX]=I, where ( )H denotes the Hermitian

transpose.
In the proposed ASMD method, the decision procedure of efficient tap size q*n of
the n-th sparse matrix can be expressed as

⎧ ⎧ 2
⎫
⎪min ⎪u Γ ± ( u +1) ⋅ H k ± (u +1) < Ψ j ⎪ , u = 0,1,L q − 1,
j

⎪ ⎨ 2 q⎬
⎪ Γ ⋅ ⎪⎭
j
⎪ H
Ψ nj = ⎨ ⎩ 0 k

⎪
2
Γj ⋅H (9)
⎪ q, if ± ( u +1) k ± (2u +1) ≥ Ψ qj for all u = {0,1,L , q − 1}
⎪ Γ 0j ⋅ H k
⎩

(
Ψ qj = Γ qj + σ 2
2
) 2
Γ 0j , k = n + q

2 2
where Γ ±j ( u +1) ⋅ H k ± (u +1) Γ 0j ⋅ H k means signal to interference ratio (SIR) of the k-th
subcarrier to the ±(u+1)-th subcarrier in frequency selective fading channel with CFO,
Ψ qj is based on SINR of the k-th subcarrier to the (k+q)-th subcarrier in frequency flat
channel, and noise power σ 2 is inserted to the threshold Ψ qj to reduce unnecessary
operation in low signal to noise ratio (SNR) environments.
From (8) and (9), with the assumption of quasi-static channel for two symbol pe-
riod, efficient tap size q*n of n-th sparse matrix can be determined as follows

q*n = max {Ψ nj } (10)

Therefore, by using the proposed ASMD method, the total amount of complex multi-
plications for ICI cancellation in cooperative STBC-OFDM systems can be approxi-
( (
mately reduced from ( N − 2q ) ⋅ ( 2 ( 2q +1) ) to ( N − 2q) ⋅ 1+ 2⋅ 4⋅ (q* +1) + 2( 2q* +1)
3
) ),
3
q*≤q.
For each (N-2q) times of sparse matrix, 1+ 2 ⋅ 4 ⋅ (q`+1) times of complex multiplica-
2
tion is needed for the calculation of k-th subcarrier power Γ 0j ⋅ H k and SINR
comparisons.

4 Simulation Results
In this section, we investigate the performance of the proposed method on multipath
channel for cooperative STBC-OFDM system. The simulation parameters used in this
paper are shown in Table 1. The multi-path fading channel is generated by COST 207
TU model and in order to avoid inter symbol interference (ISI) we set the length of
cyclic prefix (CP) long enough.
 An Efficient ICI Cancellation Method for Cooperative STBC-OFDM Systems 125

Table 1. Simulation parameters

Parameters Values
FFT size 128
Cyclic prefix (CP) length FFT size / 4
System bandwidth FFT size * 15 kHz
Subcarrier spacing 15 kHz
Center frequency 2 GHz
Channel model COST 207 TU
Channel estimation Perfect
Data modulation QPSK
Normalized frequency offsets 0.1

1.E+00
FFT size: 128
CP length: FFT size / 4
Channel model: COST 207 TU
1.E-01
CFOs: ±0.1

1.E-02
BER

1.E-03
Without CFOs
Zero forcing

1.E-04 Conventional SMD (q=1)

Conventional SMD (q=3)
Conventional SMD (q=9)
1.E-05 Proposed ASMD (q=1)
Proposed ASMD (q=3)
Proposed ASMD (q=9)
1.E-06
0 5 10 15 20 25 30 35

SNR [dB]

Fig. 4. BER performance of ICI cancellation methods versus SNR

Fig. 4 shows the bit error rate (BER) performances comparisons between conven-
tional SMD and proposed ASMD for ICI cancellation. From Fig. 4, we can see that
zero forcing method shows almost the same performance with and without CFOs.
Also, we can see that each performance of the proposed ASMD method is very close
to that of the conventional SMD method for the same number of taps.
Fig. 5 shows the average number of complex multiplications (ACM) and relative
ACM for matrix inversion per each subcarrier. In Fig 5, in low SNR environment,
overall ACM is dramatically reduced by using proposed ASMD, because threshold
Ψ qj depends on the noise variance σ 2 (σ 2 ≈ 1) . And in high SNR environment, Ψ qj
126 K. Won et al.

1.E+05

1.E+04

1.E+03
ACM

1.E+02

Conventional SMD (q=1)

Conventional SDM (q=3)
1.E+01 Conventional SDM (q=9)
FFT size: 128 Proposed ASMD (q=1)
CP length: FFT size / 4
Proposed ASMD (q=3)
Channel model: COST 207 TU
1.E+00 CFOs: ±0.1 Proposed ASMD (q=9)

0 5 10 15 20 25 30 35

SNR [dB]
(a) ACM versus SNR

70
FFT size: 128
CP length: FFT size / 4
60 Channel model: COST 207 TU
CFOs: ±0.1

tap size q=1

50
tap size q=3
Relative ACM

tap size q=9

40

30

20

10

0
0 5 10 15 20 25 30 35

SNR [dB]
(b) Relative ACM versus SNR

Fig. 5. Complexity comparisons between conventional SMD and proposed ASMD

 An Efficient ICI Cancellation Method for Cooperative STBC-OFDM Systems 127

6.E-05
Without CFOs FFT size: 128
CP length: FFT size / 4
Conventional SMD
Channel model: COST 207 TU
5.E-05
Proposed ASMD CFOs: ±0.1
SNR: 30 dB
Performance [BER]

4.E-05

3.E-05

2.E-05

1.E-05

0.E+00
0.E+00 1.E+04 2.E+04 3.E+04 4.E+04 5.E+04 6.E+04

Complexity [ACM]
Fig. 6. BER performances versus AMC

depends on the SIR of at each subcarriers (σ 2 ≈ 0 ) . Therefore, although the ACM of

the proposed ASMD method increases in proportion to SNR, it is always lower than
the conventional SMD method, and the relative ACM ratio improves as the number of
taps increases.
Fig. 6 compares performance versus complexity trade-offs of conventional SMD
method and proposed ASMD method at SNR=30 dB environment. As shown in Fig 6,
both conventional and proposed method, the more performance is improved, the more
complexity is needed for the same amount of performance enhancement, because the
amount of ICI on the k-th subcarrier from the (k+q)-th subcarrier is described in the
form of sinc function. However, we can see that the proposed ASMD method effi-
ciently improves performance versus complexity trade-offs compared with the con-
ventional SMD method by SINR measurement and adaptive tap allocation.

5 Conclusions
In this paper, we proposed an efficient ICI cancellation method for cooperative
STBC-OFDM system. Through the frequency domain SINR measurement and adap-
tive tap allocation, we showed that the overall complexity for ICI cancellation can be
efficiently reduced with approximately same performance compared with the conven-
tional method. We expect that the proposed ASMD scheme can be applicable for
various ICI suppression or cancellation methods.
128 K. Won et al.

Acknowledgments. This research was supported by the Ministry of Knowledge

Economy, Korea, under the ITRC (Information Technology Research Center) support
program supervised by the IITA (Institute of Information Technology Assessment)
(IITA-2009-C1090-0902-0005).

References
1. Tarokh, V., Seshadri, N., Calderbank, A.R.: Space-Time Codes for High Data Rate Wire-
less Communication: Performance Criterion and Code Construction. IEEE Trans. Inf. The-
ory
44(2), 744–765 (1998)
2. Alamouti, S.M.: A Simple Transmit Diversity Technique for Wireless Communications.
IEEE J. Sel. Areas Commun. 16(8), 1451–1458 (1998)
3. Sendonaris, A., Erkip, E., Aazhang, B.: User Cooperation Diversity-Part 1. IEEE Trans.
Commun. 51(11), 1927–1937 (2003)
4. Laneman, J.N., Wornell, G.W.: Distributed Space Time Block Coded Protocols for Ex-
ploiting Cooperative Diversity in Wireless Networks. IEEE Trans. Inf. Theory 49, 2415–
2425 (2003)
5. Lee, K.F., Williams, D.B.: A Space-Time Coded Transmitter Diversity Technique for Fre-
quency Selective Fading Channels. In: IEEE Sensor Array and Multichannel Signal Proc-
essing Workshop, March 2000, pp. 149–152 (2000)
6. Mingqian, T.Z., Premkumar, A.B., Madhukmar, A.S.: Performance Investigation of
STBC-OFDM Systems with Frequency Offset and a Semi-Blind Approach for the Correc-
tion. In: IEEE Vehicular Technology Conference, May 2004, pp. 1836–1839 (2004)
7. Li, Z., Qu, D., Zhu, G.: An Equalization Technique for Distributed STBC-OFDM System
with Multiple Carrier Frequency Offsets. In: IEEE Wireless Communications and Net-
working Conference, April 2006, vol. 2, pp. 839–843 (2006)
8. Zhang, W., Qu, D., Zhu, G.: Performance Investigation of Distributed STBC-OFDM Sys-
tem with Multiple Carrier Frequency Offsets. In: IEEE Personal, Indoor and Mobile Radio
Communications Symposium, September 2006, pp. 1–5 (2006)
9. Kim, Y.Z., Lee, H., Chung, H.K., Cho, W.S.: An Iterative Decoding Technique for Coop-
erative STBC-OFDM systems with Multiple Frequency Offsets. In: IEEE Personal, Indoor
and Mobile Radio Communications Symposium, September 2007, pp. 1–5 (2007)
10. Jeon, W.G., Chang, K.H., Cho, Y.S.: An Equalization Technique for Orthogonal Fre-
quency-Division Multiplexing Systems in Time-Varying Multipath Channels. IEEE Trans.
Commun. 49, 27–32 (2001)
 Low-Cost Two-Hop Anchor Node-Based
Distributed Range-Free Localization
in Wireless Sensor Networks

Taeyoung Kim1 , Minhan Shon1 , Wook Choi2 ,

MoonBae Song3 , and Hyunseung Choo1
1
School of Information and Communication Engineering,
Sungkyunkwan Univ., Korea
[email protected],{minari95,choo}@skku.edu
2
Department of Computer Science and Engineering,
Hankuk Univ. of Foreign Studies, Korea
[email protected]
3
Digital Media and Communications Business,
Samsung Electronics, Korea
[email protected]

Abstract. Due to the fact that most of the sensor network applications
are based on the location information of sensor nodes, localization is
an essential research area. Some localization schemes in the literature
require sensor nodes to have additional devices to measure the distance
or angle between two sensor nodes, but it is not suitable for low-cost
sensor nodes. On the other hand, other localization schemes use only
the connectivity information of sensor nodes so that localization is not
much accurate enough. In this paper, we propose a range-free localization
scheme, called Low-cost Two-hop Anchor Node-based Distributed Range-
free Localization (LADL), which offers a higher accuracy with lower cost
than the previous works. LADL exploits a small portion of anchor nodes
which know their own location beforehand. In LADL, sensor nodes collect
the location information of the anchor nodes within two-hop distance and
calculate their own location using a grid-scan algorithm. The simulation
results show that LADL has a maximum of 12% lower delivery cost of
location information messages and 25% higher accuracy than DRLS [8].

Keywords: Sensor Networks; Localization; Anchor node.

1 Introduction

Wireless sensor networks (WSNs) consist of a large number of low-cost sensor

nodes with high resource constraints. A wide variety of applications and net-
working protocols in the WSNs such as environment, habitat monitoring, smart
home/spaces, intrusion detection, object tracking, and geographic routing, are
based on the position information of sensor nodes. The more accurate such po-
sition information is, the more reliable and efficient the service becomes. The

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 129–141, 2010.

c Springer-Verlag Berlin Heidelberg 2010
130 T. Kim et al.

simplest method of accurately determining the position information of sensor

nodes is to install the Global Positioning System (GPS) on each sensor node.
However, it is not suitable for the low-cost sensor nodes. Therefore, the localiza-
tion of such low-cost sensor nodes is a challenging problem to be tackled.
The localization schemes can be classified into two kinds – range-based and
range-free. The range-based localization scheme is based on a specific technique
such as Time of Arrival (ToA) [1], Time Difference of Arrival (TDoA) [2], An-
gle of Arrival (AoA) [3], and Received Signal Strength Indicator (RSSI) [4], in
order to measure the distances or angles between transmitter and receiver. In
this scheme, sensor nodes are required to have additional devices to run the
distance or angle measurement technique. In addition, the measurement accu-
racy of this technique is affected by radio interferences like noise and fading and
thus eventually having low localization accuracy. From this perspective, range-
based localization is indeed not suitable for wireless sensor networks. Unlike the
range-based schemes, in the range-free localization schemes, such as Centroid [5],
Convex Position Estimation (CPE) [6], and Distance Vector-hop (DV-hop) [7],
localization is accomplished using only the connectivity information between the
sensor nodes. However, on the downside, the range-free schemes have higher cost
of exchanging messages and calculating positions than the range-based schemes.
Moreover, their positioning accuracy is relatively high.
Recently, Distributed Range-free Localization Scheme (DRLS) [8] has been
proposed to enhance the localization accuracy. Assuming that there are a few
anchor nodes in the sensor field that knows their own location information be-
forehand, DRLS allows the normal nodes to calculate their position. Each normal
node calculates its initial position with the location information of one-hop an-
chor nodes using a grid scan algorithm, and enhances the localization accuracy
with the location information of two-hop anchor nodes by means of vector-based
refinement. To discover the anchor nodes, two-hop flooding is used but it in-
creases the cost of message transmissions between the sensor nodes and may
cause localization to be more inaccurate when the position balance between the
anchor nodes is not matched (DRLS will be discussed in detail in section 2).
This paper proposes a range-free localization scheme, called Low-cost Two-
hop Anchor-based Distributed Range-free Localization (LADL), which provides
a higher accuracy with lower transmission cost than DRLS. All the normal nodes
in LADL use the position information of a few anchor nodes to calculate their
respective positions. LADL uses two-hop flooding as in DRLS but its message
transmission cost is lower by a maximum of 12% than in DRLS as the broadcast-
ing procedure is skipped after the two-hop flooding. In addition, the localization
accuracy is higher by a maximum of 25% than in DRLS without even using the
vector-based refinement process, and LADL eliminates inappropriate computa-
tions to the sensor nodes by applying a grid scan algorithm based on the position
information of the anchor nodes within two hops.
The remainder of the paper is organized as follows. Section 2 presents DRLS
and our motivations while section 3 introduces assumptions and basic definitions.
Low-Cost Two-Hop Anchor Node-Based Distributed Range-Free Localization 131

Section 4 describes the proposed scheme, LADL in detail. Section 5 discusses

simulation results. Finally, section 6 concludes the paper.

2 Related Work
Centroid [5] is a fundamental scheme for range-free localization. In Centroid, the
normal nodes find their position by collecting location information of neighboring
anchor nodes and calculating average coordinate of them, but if the anchor
node ratio is low, the accuracy of estimated positions becomes very low. To
estimate the position of the normal nodes more accurately, CPE was proposed
which is a centralized range-free localization in [6]. Anchor nodes in this scheme
broadcast their position information, and normal nodes estimate their positions
by constructing an Estimative Rectangle (ER). However, traffic load for message
exchange is heavy, and thus it is not suitable for large scale sensor networks.
As mentioned above, DRLS [8] is a distributed range-free localization scheme.
It calculates a position of a normal node based on the fact that both one-hop
anchor nodes and two-hop anchor nodes can affect the position accuracy of
the normal node. Fig. 1 shows how a two-hop anchor node, A3 , can affect the
position estimation of N. A normal node can exist anywhere within the data
transmission radio range of two one-hop anchor nodes, A1 and A2 , but not
within that of the two-hop anchor node, A3 . The distribution of these anchor
nodes enables the normal node to calculate its position more accurately. However,
such a distribution of anchor nodes is not always possible.

Fig. 1. The influence of an anchor node within two-hop distnace

DRLS consists of three steps: beacon message exchange, grid scan, and re-
finement steps. In the beacon message exchange step, the normal node should
know the position information of the anchor nodes existing within a two-hop
distance from their neighboring anchor nodes by two-hop flooding and one ad-
ditional broadcast. Thereafter, in the grid scan step, each normal node creates
an estimative rectangle (ER) with the use of the transmission radio range of
the one-hop anchor nodes and divides the ER into a set of small grid cells.
Fig. 2 illustrates that how a normal node obtains a one-hop anchor-based ini-
tial estimated position using a grid scan algorithm. In the refinement step, the
normal node that measured the initial position decides its own final estimated
position via vector-based refinement. The direction of the vector, called virtual
132 T. Kim et al.

Fig. 2. Initial estimated position with grid scan algorithm in DRLS

Fig. 3. Location refinement based on the vector in DRLS

force (V F ), is from a two-hop anchor node to the initial estimated position, and
−−→
its size becomes | V F |= ab · N  c/ac. Fig. 3 shows how a normal node refines its
initial position based on a two-hop anchor node.
As DRLS conducts localization with the use of more information of the an-
chor nodes compared to that in the previous studies, it shows relatively high
accuracy; however, it has several problems. First, as it uses two-hop flooding to
collect the position information of two-hop anchor nodes, the message transmis-
sion cost becomes higher than that in the existing range-free methods. Second,
its square-root computation for calculating the vector in the refinement step
is not appropriate for low-cost sensor nodes [9]. Third, as it accomplishes the
vector-based refinement, its location accuracy becomes lower if the distribution
of the anchor nodes is not even. Fig.4 shows an example of the refinement which
eventually lowers the position accuracy. Suppose that the initial estimated posi-
tion of normal node N is N  . The sum of two vectors made by the anchor nodes
A3 and A4 exists in an area where the normal node cannot be placed. Thus, the
position accuracy becomes low. Followings are the motivations of this study to
enhance DRLS by tackling the above-mentioned problems:

– Normal nodes in DRLS obtain the position information of two-hop anchor

nodes by one-hop anchor nodes and ignore the position information of the
anchor nodes that can be obtained without going through the one-hop an-
chor nodes. However, if they use such ignored information, however, and
if additional broadcasting is skipped after two-hop flooding, the message
transmission cost can be lowered.
Low-Cost Two-Hop Anchor Node-Based Distributed Range-Free Localization 133

Fig. 4. The influence of farther anchors that are not distributed uniformly

– Square-root computation is not appropriate to low-cost sensor nodes [9].

Besides, vector-based refinement may cause lower position accuracy due to
the anchor nodes that are placed in similar positions. Instead of skipping
vector-based refinement, it can enhance the position accuracy and lessen the
computational cost by utilizing all the one- and two-hop anchor nodes in
grid scan step.
A summary of the differences between DRLS and LADL is shown in Table 1.

Table 1. Differences between DRLS and LADL

Head DRLS LADL

Selection of A two-hop flooding A two-hop flooding
anchor nodes and a broadcasting
Construction of Only one-hop anchor Both one- and two-hop
ER nodes anchor nodes
Grid Scan Only one-hop anchor Both one- and two-hop
nodes anchor nodes
Refinement Vector based refinement Skipping the refinement step

3 Preliminaries
We assume the followings in this study:
– Normal nodes that do not have information of their own positions, and a few
anchor nodes that have information of their positions, constitute the entire
sensor field.
– All the sensor nodes are distributed randomly in a large sensor field, and
cannot change their positions once these have been decided.
– All the sensor nodes can be identified through their unique IDs and have the
same transmission radio and sensing ranges.
– The collision that can occur during data transmission is not considered, and
there are no interferences such as fading and noise. In other words, a sensor
node receives data without fail as it transmits data to its neighbor sensor
nodes within the transmission radio range.
134 T. Kim et al.

Anchor nodes within a one-hop distance from a normal node are called one-
hop anchor nodes while anchor nodes within a two-hop distance from a normal
node are called two-hop anchor nodes. Normal nodes can enhance localization
accuracy by separating the one-hop anchor nodes from the two-hop anchor nodes
among their neighboring anchor nodes. Fig. 5 shows where normal nodes can be
located depending on the one-hop anchor nodes and the two-hop anchor nodes.
Anchor node A is within a one-hop distance from normal node N1 , thus becoming
a one-hop anchor node of N1 . On the other hand, it becomes a two-hop anchor
node of N2 since it is two hops away from normal node N2 . Accordingly, the
region where N1 can exist is a circular area with a radius of transmission radio
range r whereas the region where N2 can exist is the circular area with a radius
of 2r minus the circular area with a radius of r.

Fig. 5. The region that a normal node can be located from an anchor node

4 Proposed Scheme: LADL

LADL consists of three several steps - anchor node selection, ER construction,

and grid scan. In the anchor node selection step, each normal node collects the
position information of the anchor nodes which are located within a two-hop
distance. In the step of the ER construction, each normal node constitutes an
ER by using the collected position information of the anchor nodes. Note that
as it creates an ER using one- and two-hop anchor nodes, the two-hop anchor
node-based refinement process of DRLS is skipped. In the grid scan step, the
positions of the normal nodes are identified by dividing the ER into a set of grid
cells and then scanning each cell. Now, let us discuss the three steps of LADL
in detail below.

4.1 Anchor Node Selection

In the anchor node selection step, each anchor node transmits its own ID and
position information to its neighboring nodes. Each anchor node enables two-
hop flooding by assigning 2 to the Time to Live (TTL) value of its own position
information message. Thus, the normal node can identify either one-hop or two-
hop anchor nodes based on the value of TTL.
Fig. 6 shows an example of two-hop flooding. In Fig. 6(a), anchor nodes A1 ,
A2 , and A3 transmit their position information with TTL=2 to their neighboring
Low-Cost Two-Hop Anchor Node-Based Distributed Range-Free Localization 135

sensor nodes. In Fig. 6(b), every sensor node that has received the position
information with TTL=2 transmits the position information of neighbor anchor
nodes with TTL=1 to its neighboring sensor nodes. Via two-hop flooding, normal
node N1 selects A1 and A2 as one-hop anchor nodes, normal node N2 selects A2
as a one-hop anchor node while choosing A3 as a two-hop anchor node.

(a) (b)

Fig. 6. Two-hop flooding: (a) Flooding with TTL=2, (b) Flooding with TTL=1

4.2 ER Construction
Every normal node can construct an ER with collected position information of
both its one- and two-hop anchor nodes. To reduce the computation overhead
for the localization, the normal node uses the rectangle that circumscribes the
transmission radio range of each anchor node. Thereafter, an ER is constructed
by calculating the overlapped region of these rectangles.
Fig. 7 shows an example of the ER construction. Normal node N keeps the
position information of one-hop anchor node A3 and two-hop anchor nodes A1
and A2 , and constructs an ER (i.e., dark shaded area in the figure) applying
different rectangle drawing mechanisms to one- and two-hop anchor nodes, re-
spectively. Then, the normal node should be located inside the ER.

Fig. 7. ER made by the location information of one- and two-hop anchor nodes
136 T. Kim et al.

4.3 Distributed Grid Scan Algorithm

The ER calculated by a normal node includes a region that is larger than the
region in which the normal node can actually exist. Accordingly, the localization
accuracy can be enhanced by excluding the region where a normal node cannot
actually exist from the ER using the grid scan algorithm. After dividing the
ER into a set of grid cells, the normal node decides if which grid cells should
be excluded by scanning each cell. If the size of cell is too small, the scan com-
putation cost increases, whereas the localization accuracy decreases if it is too
large. In this paper, the length of one side of each cell is set at 0.1r (remember
that r is the data transmission radio range), in consideration of both the scan
computation cost and the localization accuracy. Additionally, the value of the
cell is set at 1, and the coordinate of the center of gravity in each cell is selected
as the coordinate value of the position of the cell. Here, 1, the value of each cell,
means that a normal node could exist in the cell.
After removing the cells where each normal node cannot exist, the normal
node applies a different criterion of scanning each cell for one- and two-hop
anchor nodes. Fig. 8 shows an example of the grid scan algorithm. Note that the
distribution of normal node N , and anchor nodes A1 , A2 , and A3 is the same
as in Fig. 7. Normal node N sets 1 to the value of each cell after dividing the
ER into a set of 0.1rx0.1r grid cells. In Fig. 8, 0 is designated as the value of
the cells whose distances from two-hop anchor nodes A1 and A2 are longer than
2r and shorter than r, and whose distances from one-hop anchor node A3 are
longer than r. That is, the cells with a value of 0 represent the excluded cells for
the final scanning. The cells with a value of 1 (i.e., dark shaded area in Fig. 8)
represents the possible cells that include the actual location of normal node N .
Finally, normal node N obtains final estimated position N  by computing the
average of the representative coordinate values in each shaded cell.
As all sensor nodes are distributed randomly in the sensor field, it may happen
that an anchor node does not exist within the two-hop distance from the nor-
mal node. Therefore, the normal node waits until other neighbor normal nodes
succeed in localization, and then accomplishes its localization by using neigh-
bors’ position information. For this purpose, the normal nodes that succeeded

Fig. 8. The result of grid scan based on the location information of anchor nodes
Low-Cost Two-Hop Anchor Node-Based Distributed Range-Free Localization 137

in localization broadcast their own position information and that of the one-hop
anchor node to its neighboring nodes.

5 Performance Evaluation
5.1 Methodology
In this section, DRLS, which has come about upon the motivation of LADL,
will be compared with LADL to analyze the performance of LADL. We will use
the following two metrics to measure the performance of LADL.
1. Localization accuracy: It is defined as the difference between the real position
and the measured position of a normal node, and the data transmission radio
range of sensor node r is the basic unit. It shows through a numerical value
how accurately the normal node calculates the position compared with the
real position.
2. Message transmission cost: It is defined by the number of position informa-
tion messages of the anchor nodes used for each normal node to obtain infor-
mation of the anchor nodes. It indicates the position information exchange
cost of the anchor node to measure the positions of the normal nodes.
The simulator in this section is realized using JAVA. The data transmission radio
range of the sensor nodes is r, and the simulation is performed in an environment
where the sensor nodes were deployed randomly in the sensor field with 10rx10r
sizes. A collision on the link, which occurs by two-hop flooding, is not considered,
and when dividing the ER into specific grid sizes, it draws a grid with the size
of each cell being 0.1rx0.1r in consideration of the position estimation accuracy,
computational cost of grid scan, and memory capacity required. Moreover, it
should be compared with the performance of DRLS by using the average value
of the results derived from the 100 simulation repeats.

5.2 Localization Accuracy

Fig. 9 shows the position estimation accuracy of Centroid, CPE, DRLS, and
LADL measured by changing the proportion of the anchor nodes from 5 to
47.5%, with the number of total sensor nodes fixed at 200. In addition, the re-
sult is about normal nodes which can get information of anchor nodes by two-hop
flooding surely. LADL measures the position by using both one- and two-hop
anchor nodes at the grid scan step; therefore, more accurate position measure-
ment is possible by eliminating the inaccuracy that occurs in the vector-based
refinement. Moreover, because the increase in the proportion of the anchor nodes
causes an increase in the amount of information of the anchor nodes, position es-
timation accuracy increases. However, the localization without the normal nodes
which does not have anchor node information is useless for applications in the
WSNs. Therefore, the localization with all the normal nodes is necessary and
its accuracy of DRLS and LADL is described in Fig. 10. The simulation result
shows that LADL has a 4-25% higher accuracy compared to DRLS.
138 T. Kim et al.

2.5

Centroid (200 nodes)

CPE (200 nodes)
DRLS (200 nodes)
2.0
LADL (200 nodes)

Mean error (r)

1.5

1.0

0.5

0.0

0 5 10 15 20 25 30 35 40 45 50

The ratio of anchors (%)

Fig. 9. Localization accuracy of normal nodes with different ratio of anchor nodes

1.6 DRLS (200 nodes)

LADL (200 nodes)
1.4

1.2
Mean error (r)

1.0

0.8

0.6

0.4

0 2
0.2

0.0

0 5 10 15 20 25 30 35 40 45 50

The ratio of anchors (%)

Fig. 10. Localization accuracy of all normal nodes with different ratio of anchor nodes

Fig. 11 shows how much the number of total sensor nodes affects the position
estimation accuracy. The result of the accuracy of the position estimation that
was done while changing the number of total anchor nodes to 200, 400, and
600 with fixing the proportion of the anchor nodes at 20% indicates that LADL
has 16-49% better accuracy compared to DRLS. The increase in the number of
total sensor nodes leads to more accurate position estimation because the higher
density of the network causes that the amount of information of the anchor nodes
that a normal node receives increases.

5.3 Message Transmission Cost

Fig. 12 shows the message transmission cost for the anchor nodes of DRLS
and LADL measured by changing the proportion of the anchor node from 5
to 47.5% in the environment where the total number of sensor nodes is 200.
DRLS transmits more messages compared to LADL because DRLS performs
additional broadcasting in addition to two-hop flooding to provide information
Low-Cost Two-Hop Anchor Node-Based Distributed Range-Free Localization 139

0.8

DRLS (anchor ratio = 20%)

LADL (anchor ratio = 20%)

0.6

Mean error (r)

0.4

0.2

0.0

200 400 600

The number of nodes

Fig. 11. Localization accuracy in terms of the total number of sensor nodes

800
DRLS (200 nodes)
700 LADL (200 nodes)

600
The number of messages

500

400

300

200

100

0
0 5 10 15 20 25 30 35 40 45 50
The ratio of anchors (%)

Fig. 12. Anchor node selection cost with different ratio of anchor nodes

of the anchor nodes to the normal nodes. Moreover, as the proportion of the
anchor nodes increases, the number of messages increases significantly due to
the fact that the number of anchor nodes which are the starting point of two-
hop flooding increases. Fig. 9 shows that LADL provides the normal nodes with
information of the anchor nodes with a 5-12% lower message transmission cost.
Fig. 13 shows how much the number of total sensor nodes affects the mes-
sage transmission cost of two-hop flooding. The number of position information
messages is measured while changing the number of total anchor nodes to 200,
400, and 600 with fixing the proportion of the anchor nodes at 20%. The result
indicates that LADL has a 7-14% lower message transmission cost compared to
DRLS. The reason for DRLS’s poorer performance compared to LADL is that
DRLS broadcasts additional position information more than LADL does. Fur-
ther, the two-hop flooding cost increases rapidly as the number of total sensor
nodes increases because the increase in the network density rapidly increases the
number of sensor nodes existing within a one-hop distance from a sensor node.
140 T. Kim et al.

2500
DRLS (anchor ratio = 20%)
LADL (anchor ratio = 20%)
2000

The number of messages

1500

1000

500

0
200 400 600
The number of sensor nodes

Fig. 13. Anchor node selection cost with changing of the total number of sensor nodes

6 Conclusions

In this paper, we proposed a range-free localization scheme, called low-cost two-

hop anchor node-based distributed range-free localization (LADL), which can
decrease the message transmission cost and can enhance the position estimation
accuracy compared to DRLS. A normal node makes an ER by utilizing the
position information of the anchor nodes within a two-hop distance from it, and
points out where it will be located in the entire sensor field by scanning the
interior of ER with a grid scan algorithm. Additionally, it decreases the cost of
exchanging the position information of the anchor nodes via two-hop flooding,
which is more effective than DRLS, to obtain the position information of the
anchor node within a two-hop distance. The merit of LADL lies in the fact that
it enables the normal node to conduct localization with a two-hop anchor node
even if it does not have any one-hop anchor node. It was proven via simulation
that LADL can conduct localization more precisely by as much as 25% with a
message transmission cost that is up to 20% smaller than that of DRLS.
In the subsequent study, it should be proven that LADL is superior to DRLS
in terms of computation cost when the square-root computation is removed from
it, by analyzing the cost required by each normal node during the execution of
the algorithm. Besides, we plan to study about identifying the most appropriate
size of the transmission radio range and the cell at the grid scan step to obtain
the higher localization accuracy.

Acknowledgments. This research was supported in part by MKE, Korea un-

der ITRC NIPA-2009-(C1090-0902-0046), in part by by MEST(Korea), under
WCU Program supervised by the KOSEF (No.R31-2008-000-10062-0) and also
supported partly by NIA (National Information Society Agency), KOREA under
the KOREN program.
Low-Cost Two-Hop Anchor Node-Based Distributed Range-Free Localization 141

References
1. Hightower, J., Borriello, G.: Location Systems for Ubiquitous Computing. Com-
puter 34(8), 57–66 (2001)
2. Pottie, G.J., Kaiser, W.J.: Wireless Integrated Network Sensors. Communications
of the ACM 43(5), 51–58 (2000)
3. Priyantha, N.B., Chakraborty, A., Balakrishnan, H.: The Cricket Location-Support
System. In: Proceedings of ACM MobiCom 2000, pp. 32–43 (2000)
4. Savvides, A., Han, C.-C., Strivastava, M.B.: Dynamic Fine-Grained Localization in
Ad-Hoc Networks of Sensors. In: Proceedings of ACM MobiCom 2001, pp. 166–179
(2001)
5. Bulusu, N., Heidemann, J., Estrin, D.: GPS-Less Low Cost Outdoor Localization
for Very Small Devices. IEEE Wireless Communications 7(5), 28–34 (2000)
6. Doherty, L., Pister, K.S.J., Ghaoui, L.E.: Convex Position Estimation in Wireless
Sensor Networks. In: Proceedings of IEEE INFOCOM 2001, vol. 3, pp. 1655–1663
(2001)
7. Niculescu, D., Nath, B.: Ad Hoc Positioning System (APS). In: Proceedings of
IEEE Globecom 2001, vol. 1, pp. 2926–2931 (2001)
8. Sheu, J.P., Chen, P.C., Hsu, C.S.: A Distributed Localization Scheme for Wireless
Sensor Networks with Improved Grid-Scan and Vector-Based Refinement. IEEE
Transactions on Mobile Computing 7(9), 1110–1123 (2008)
9. Acharya, M., Girao, J., Westhoff, D.: Secure Comparison of Encrypted Data in
Wireless Sensor Networks. In: Proceedings of IEEE Symposium on Modeling and
Optimization in Mobile, Ad Hoc, and Wireless Networks, pp. 47–53 (2005)
10. He, T., Huang, C., Blum, B.M., Stankovic, J.A., Abdelzher, T.: Range-Free Local-
ization Schemes for Large Scale Sensor Networks. In: Proceedings of ACM Mobi-
Com 2003, pp. 81–95 (2003)
11. Boukerche, A., Oliveira, H.A.B.F., Nakamura, E.F., Loureiro, A.A.F.: Localization
Systems for Wireless Sensor Networks. IEEE Wireless Communications 14(6), 6–12
(2007)
12. Akyildiz, I., Su, W., Sankarasubramaniam, Y., Cayirci, E.: A Survey on Sensor
Networks. IEEE Communications Magazine 40(8), 102–114 (2002)
13. Lederer, S., Wang, Y., Gao, J.: Connectivity-based Localization of Large Scale
Sensor Networks with Complex Shape. In: Proceedings of IEEE INFOCOM 2008,
pp. 789–797 (2008)
 SecDEACH: Secure and Resilient Dynamic Clustering
Protocol Preserving Data Privacy in WSNs

Young-Ju Han, Min-Woo Park, and Tai-Myoung Chung

Internet Management Technology Laboratory,

Department of Computer Engineering,
School of Information and Communication Engineering,
Sungkyunkwan University,
300 Cheoncheon-dong, Jangan-gu,
Suwon-si, Gyeonggi-do, 440-746, Republic of Korea
Tel.: +82-31-290-7222; Fax: +82-31-299-6673
{yjhan,mwpark}@imtl.skku.ac.kr, [email protected]

Abstract. Clustering scheme enabling the efficient utilization of the limited en-
ergy resources of the deployed sensor nodes can effectively prolong lifetime of
wireless sensor networks. Like most of the protocols for wireless sensor net-
works, a clustering scheme is vulnerable to a number of security attacks. Espe-
cially, attacks involving cluster-heads are the most damaging since cluster-based
schemes rely on their cluster-heads for routing and data aggregation. Several
studies have proposed a secure clustering scheme and a resilient cluster leader
election protocol in wireless sensor networks. However, most of resilient clus-
tering schemes do not support dynamic clustering and most of secure clustering
schemes are not robust and resilient against node(specially, cluster-heads) cap-
ture. In this paper, we present a secure and resilient clustering protocol on the
basis of dynamic clustering protocol. Our scheme provides both a secure cluster-
ing and a resilient cluster-head election preserving dynamic clustering unlike the
existing secure clustering schemes. In addition, our scheme provides data privacy
through the concealed data aggregation. The analysis shows that our scheme of-
fers the security services against both an inside attack and an outside attack while
having the little memory, computation and communication overheads. The com-
parison with related works shows that our scheme has a great improvement in
security and the reduction of the overheads compared to the existing schemes.

1 Introduction

WSNs(Wireless sensor networks) consist of a number of sensor nodes deployed over a

geographical area for the purpose of monitoring certain phenomena of interest. Wireless
sensor networks enable the reliable surveillance of a variety of environment for health,
military, home and civil applications [7,8]. Each sensor node has wireless communica-
tion capability, data aggregating ability, and self-organizing capability [8].
When embedded in critical application, WSNs are likely to be attacked [3]. Aside
from the well-known vulnerabilities due to wireless communication, WSNs lack phys-
ical protection and are usually deployed in open and unattended environments, which

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 142–157, 2010.

c Springer-Verlag Berlin Heidelberg 2010
 SecDEACH: Secure and Resilient Dynamic Clustering Protocol 143

makes them more vulnerable to attacks [3,9]. It is therefore crucial to devise security
solutions to these networks [3].
Sensor nodes are often organized into several clusters for the purpose of efficiency
and scalability. Clustering scheme enabling the efficient utilization of the limited en-
ergy resources of the deployed sensor nodes can effectively prolong lifetime of wireless
sensor networks [7,12]. Adding security to clustering scheme is challenging, as its dy-
namic and periodic rearrangement of clustering makes key distribution solutions inad-
equate [3,9]. Like most of the protocols for WSNs, a clustering scheme is vulnerable to
a number of security attacks including passive eavesdropping, jamming, spoofing and
replay attacks [3]. Especially, attacks involving CHs(cluster-heads) are the most dam-
aging since cluster scheme fundamentally relies on the CHs for data aggregation and
routing. If the adversary manages to become a CH, he can perform various attacks to
disrupt the function of the network such as routing and data aggregation. Therefore, to
guarantee for WSNs to function properly, a secure clustering protocol is required.
A number of secure routing protocols have been proposed recently in WSNs
[1,5,9,10,14]. These can be generally categorized into a secure cluster formation proto-
col [1,9] and a secure CH election protocol [5,14,10]. Secure cluster formation
protocols significantly improve the security of clustering in the presence of malicious
attacks. However, they only focus on the cluster formation and do not consider the se-
curity against disrupting CH election by the compromised node. Resilient cluster leader
election schemes in [5,14,10] are more resilient against a node capture than the above
mentioned secure cluster protocols. However, they do not support dynamic clustering
which is a heuristic method for an optimal clustering since they have assumption that a
cluster sector is pre-defined and not changed. This feature is not suitable for WSNs.
In this paper, we present a secure and resilient clustering scheme considering energy
efficiency and a concealed data aggregation in WSNs. Our scheme provides both the
resilient CH election preserving a dynamic clustering and the secure cluster formation
unlike the existing secure clustering schemes. In addition, our scheme provides data pri-
vacy through a concealed data aggregation. The comparison with related works shows
that our scheme offers the security services against both an inside attack and an outside
attack. Our scheme has a little memory and computation overhead but it is endurable in
sensor nodes.
This paper is organized as follows. Section 2 introduces the backgrounds of the pro-
posed scheme. Section 3 presents our goals and some assumptions required for describ-
ing the proposed scheme. Section 4 describes the procedures of the proposed scheme
in detail. Section 5 explains the results of analysis comparing the our proposed scheme
with the other schemes. At last, section 6 summarizes our work and discusses the future
works.

2 Backgrounds
2.1 DEACH and Its Vulnerabilities
DEACH. The basic operation of a clustering routing protocol in WSNs is to partition
network into several clusters. After clustering, the CH of each cluster aggregates the
sensed data from its member nodes for energy efficiency of data processing and then
144 Y.-J. Han, M.-W. Park, and T.-M. Chung

sends the aggregated data to the BS(base station) [11,15]. The most common technique
in famous clustering routing protocols is the probabilistic clustering algorithm based
on the randomized CH rotation for distributing the energy consumption among sensor
nodes in each cluster [11,15].
As our previous works, we have proposed DEACH(Distance based low-Energy
Adaptive Clustering Hierarchy Protocol) [16] which is an energy-efficient distance
based clustering scheme. In clustering scheme, how to control the balance of the energy
consumption among CHs can greatly affect the network lifespan which is generally de-
termined by the first dead-node [12]. Among the sources of energy consumption of a
CH, wireless data transmission is the most critical [6]. Because the energy consump-
tion of wireless data transmission increases generally in proportion with the distance
from sending node to receiving node [6], the energy consumption of CH depends on
its location from the BS. For example, in single-hop inter-cluster communication, the
CHs farther to the BS will die much faster than the other CHs, while in multi-hop
inter-cluster communication, the CH closest to the BS is burdened with a heavy relay
traffic load and die first(i.e., the hot spot problem). Most of clustering schemes such as
LEACH [15] and HEED [11] utilize mainly CH frequency or residual energy of each
node as criterion of CH election and do not consider the distance from the BS to the
CH. So, let those schemes cause the unbalanced energy consumption among CHs in
spite those schemes are able to prolong the network lifetime.
To solve this problem, DEACH considers the distance from the BS to CHs as well as
the residual energy as the criterion of CH election for the balanced energy consumption
among CHs. For a single-hop inter-cluster environment, we use Eq. 1 as probability
formula for electing a CH. That is, each sensor node has the probability, Pi , of becoming
a CH which is determined by the distance to the BS and its residual energy.
 d(Si ,BS)−dmin
Popt (i) = 32 Popt × dmax −dmin
⎧ 
⎪
⎪ Popt (i) Eres (i)
⎨  × Ci (t) = 1 (1)
Pi = 1 − Popt (i) × (r mod 
1
) Einit (i)
⎪
⎪
Popt (i)
⎩
0 Ci (t) = 0

where Popt is the optimal probability being a CH which is a system parameter, Popt (i)
is the optimal probability being a CH of sensor node i based on the distance from itself
to the BS, r is the current round of the CH election, Einit (i) and Eres (i) are the initial
energy and the residual energy of the sensor node, d(i, j) is the distance between node
i and j, dmax represents the distance of the farthest sensor node from the BS and dmin
represents the distance of the closest sensor node. Ci (t) presents whether or not sensor

node i has been a CH in the most recent r mod 1/Popt rounds. That is, Ci (t) = 0 if
node i has been a CH and one otherwise). Equation 1 shows that the closer the sensor
node is to the BS, the lower probability of becoming a CH it could have. In other words,
the CHs closer to the BS can have more members than the CHs farther to the BS since
the few CHs are elected in the area close to BS. Besides, among the nodes having the
same distance from the BS, the more the sensor node has the residual energy, the higher
probability of becoming a CH it could have. According to Eq. 1, every sensor node
 SecDEACH: Secure and Resilient Dynamic Clustering Protocol 145


elects itself as a CH only once during 1/Popt rounds on average. In this way, the CHs
are selected randomly among the sensor nodes and the energy concentration on CHs is
distributed.
DEACH provides fully distributed manner by utilizing local information and good
energy-efficiency by performing the load balanced clustering. Through simulation ex-
periments, we have showed that the proposed scheme is more effective than the other
clustering protocols in prolonging the lifespan of WSNs [16].
Its Vulnerabilities. Adding security to clustering scheme is challenging, as its dynamic
and periodic rearrangement of clustering makes key distribution solutions inadequate
[3,9]. Like most of the protocols for WSNs, a clustering scheme is vulnerable to a
number of security attacks including passive eavesdropping, jamming, spoofing and
replay attacks [3]. Especially, attacks involving CHs are the most damaging since a
cluster scheme fully relies on the CHs for data aggregation and routing. If the adversary
manages to become a CH, he can perform various attacks to disrupt the function of the
network such as routing and data aggregation.
Firstly, in existing CH election protocols, nodes rely on their local status or on re-
ported status of peer nodes for electing their CH. For example, in LEACH [15], HEED
[11] and DEACH [16], a node can independently decide to become CH based on its
energy-level or a distance. However, such approaches have substantial limitations with
the respect to security. A fundamental observation is that any election mechanism based
on a concrete election metrics such as the residual energy level or distance, can be ma-
nipulated or forged. So, the adversary can become the CH or disturb a legitimate node
not to become a CH for the upcoming epoch by forging or replaying CH announcement.
Secondly, if the CH is compromised by the adversary, the adversary can learn the
sensed data and forge the data because the adversary knows key material even though
security primitive such as encryption is used to provide confidentiality of the sensed
data. This feature makes the application in WSNs very more critical and vulnerable.

2.2 Secure Routing Protocols in WSNs

A number of secure routing protocols for sensor clustering in hostile situations have
been proposed recently in WSNs [1,5,9,10,14]. These can be generally categorized into
a secure cluster formation protocol and a secure CH election protocol.
F-LEACH [1] is proposed to protect the CH election in LEACH. A sensor itself
elects as a CH by using common keys shared with the BS and the BS authenticates
the CH declaration using the same keys. Then, the BS securely broadcasts array of
the authenticated CHs using μTESLA [2]. Sensors join only one authenticated CH.
However, this scheme cannot authenticate the sensors which join the service of CH. To
resolve this problem, Oliveiral et al. have proposed SecLEACH [9] in which the BS
authenticates the CH declaration from sensors and CHs also authenticate the joining
sensors.
However, they have some limitations. Firstly, they do not consider the security of
cluster leader election since these focus on the security issues during the formation of
initial clusters. Secondly, a malicious insider can be a CH because these do not con-
trol the frequency of CH declaration. That is, they can prevent only external attackers
not internal attackers from declaring themselves as CHs. Thirdly, the limitation related
146 Y.-J. Han, M.-W. Park, and T.-M. Chung

with a compromised CH is not to guarantee confidentiality of the sensed data from sen-
sor nodes. Even though each sensor node encrypts its sensed data for confidentiality,
the compromised CHs can decrypt the data and finally they know which information
is sensed from sensor nodes. So, a concealed data aggregation is required against a
node capture. Finally, they have a single point of failure problem because they are the
centralized mechanism which fully depends on the BS for authenticating the CH.
A number of secure cluster leader election protocols have been proposed in [5,10,14].
All of them work in a decentralized way and use the light-weighted cryptographic al-
gorithms. The main technique of their schemes is that cluster member securely elects
its CH through the agreement of a elected CH within a pre-defined cluster. [5] and [10]
use a commitment based scheme. Each sensor sends its commitment to other sensors in
the peer-to-peer manner. A commitment is an encrypted random value using a shared
key and the random value is created by each sensor. Then, each sensor sends the ful-
fillment value(that is, its random value) to other sensors within its sector. The sensors
receiving the fulfillment values verify them using the shared key and sum them to make
an agreed random value. They divide the real sum of agreed random values by the num-
ber of sensors and get the remainder which indicates the position of CH node in the
cluster. Like [5,10], Dong’s scheme [14] also has the assumption that cluster sector is
pre-defined and not changed. In Dong’s scheme, the new CH is selected based on the
remaining energy on cluster members for the purpose of the load balancing. Their idea
for a resilient clustering is to let all member within a cluster have the identical list of
the candidate CHs. To control the list of each sensor node, they use two one-way hash
chains, YES chain and NO chain. In initialization phase, every sensor node sends its
YES commitment value to the other nodes within its sector. Each sensor node receiving
all YES commitment from the other nodes creates the list of the candidate CHs through
the shuffle algorithm. The shuffle algorithm makes the list of the candidate CHs of all
sensor nodes within the same sector identical. In each round of CH election, each sen-
sor node sends YES key or NO key according to its residual energy. If sensor node
does not have energy for serving as the CH, it releases NO key, otherwise it releases
YES key. If sensor node receiving NO key, it removes identifier(ID) of the sending node
from the list of the candidate CHs. Through this mechanism, all cluster members in the

Fig. 1. The problem of pre-defined clustering

 SecDEACH: Secure and Resilient Dynamic Clustering Protocol 147

list of the candidate CHs will server as the CH in a round-robin manner unless they
release their NO keys during the CH election. When the benign cluster members are
well-connected, all benign nodes that are eligible for the CH role will be included in
the list of the candidate CHs.
Resilient cluster leader election schemes are more resilient against a node capture
than the above mentioned secure clustering protocol. However, they do not support a
dynamic clustering which is a heuristic method for an optimal clustering since they
have the assumption that cluster sector is pre-defined and not changed. Figure 1 shows
the problem of a pre-defined clustering. In Fig. 1, WSNs are divided into 4 sectors.
Nodes A,B,C and D simultaneously can become the CHs in any round. In this case,
even though they are very close to each other, they are not clustered together because
they belong to different sectors. As a result, intra-cluster communications of them cause
the heavy overhead. This feature is not suitable for WSNs.

3 System Models and The Objective

3.1 Threats Model

We assume an adversary may launch arbitrary attacks against the cluster formation
protocol and cluster based routing protocol except for completely jamming the commu-
nication channel. In this section, we classify adversaries in clustering routing protocol
as follows: external adversary and internal adversary.

– External adversary. An external adversary can randomly select some nodes and
extract key generation information from them. Then, he tries to eavesdrop the com-
munication between other ordinary nodes. Besides, he tries to inject and replay
packet to disrupt the cluster formation and cluster routing.
– Internal adversary. We assume that he already has knowledge on the algorithms or
protocols. Using this knowledge, he can selectively capture the nodes and then he
tries to eavesdrop the communication between other ordinary nodes. Then, he can
use any method to interrupt/corrupt cluster formation and to forge the sensed data or
the aggregated data, e.g., compromising some nodes, modifying the communication
message, replay attack etc.
Ensuring a good level of security for such types of networks is not a trivial task. As
WSNs use wireless communications, the threats and attacks against them are more di-
verse and often large-scale.

3.2 The Sensor Network Model and Notation

We consider a sensor network composed of N resource-constrained sensor nodes dis-

tributed uniformly on the square area of size A = a × a in a homogenous spatial
Poisson process with density λ = NA . For a cluster formation, each sensor node must
embed DEACH algorithm. For DEACH, the optimal probability of being CH, Popt , is
pre-defined, which is CHopt /N where CHopt is the optimal number of CH. The sensor
network model of the proposed scheme has the following assumptions.
148 Y.-J. Han, M.-W. Park, and T.-M. Chung

• The sensor nodes are stationary, identical, and energy-constrained.

• All sensor nodes can communicate directly with the BS and are equipped with power
control capabilities to vary their transmitted power.
• The BS has no energy constraint and is stationarily located far from sensor field.
• The communication is symmetric and based on a single-hop.
• A sensor node can compute the approximate distance based on the received signal
strength.

We denote some common notations for describing the our scheme as follows.
• Si is ith (i ∈ N ) sensor node.
• Si (id) is the identifier of the sensor node, Si .
• Si (r) is the maximum transmission range of the sensor node, Si .
• Pi is the probability of becoming a CH of the sensor node, Si .

• Popt (i) is the optimal probability of becoming a CH of the node Si referred in Eq. 1.
• CHi is ith cluster-head.
• SCH (i) is the set of the candidate CHs of the sensor node, Si .
• Eres (i) is the residual energy of the sensor node, Si .
• Eth is the minimum energy required for serving as CH.
• Sneighbor (i) is the set of neighborhood nodes of the node, Si , within Si (r).
• Qi (j) is the qualification information of the node, Sj , as a CH of the node, Si .
• Sq (i) is the set of the qualified nodes of the node, Si .
• di is the sensed data from the sensor node, Si .
• ermax is the maximum number of rounds for CH election.

3.3 The Objectives

The objective of SecDEACH is to provide for both a secure clustering protocol con-
sidering a concealed data aggregation and a resilient CH election against an internal
attacker and an external attacker. To prevent from security threats, WSNs must satisfy
the security requirements and design requirement as following:
Security Goal 1: Resilience against fabrication attacks, impersonation attacks and
replay attacks. To prevent an external attacker, a clustering protocol ensures that unau-
thorized sensor nodes should not join a cluster routing protocol. For doing this, it should
guarantee the basic security primitives such as authentication, confidentiality, freshness
and integrity.
Security Goal 2: Fault-tolerance. It is impossible to detect a compromised node act-
ing like a ordinary node. So, if a compromised node acts as a CH, the application is
damaged since a CH is responsible for data aggregation and data transmission schedule
of its member nodes. To prevent this attack, a cluster routing scheme guarantees that a
compromised node does not become a CH [14]. That is, the attacker cannot arbitrarily
increase or decrease the chance of any ordinary node being elected as a CH.
Security Goal 3: Concealed data aggregation in-networking. If a CH is compro-
mised, the compromised CH node can learn all of the sensed data from its member
nodes by decrypting them, even though its member nodes encrypt their sensed data. To
 SecDEACH: Secure and Resilient Dynamic Clustering Protocol 149

prevent this attack, a clustering routing protocol guarantees confidentiality and privacy
of the sensed data which does not depend on CH.
Design Goal: Low communication, computation overheads and low storage re-
quirement. Adding to the security primitives in a clustering protocol makes commu-
nication and computation overheads larger. The stronger robustness of security is, the
overhead is larger. So, we consider the energy-efficient mechanism preserving the ro-
bustness of security. In addition, we consider the limitation of storage resulted from the
resource constraints of the sensor node.

4 SecDEACH
In this section, we will briefly review the security primitives used in the proposed
scheme and then describe its technical detail.

4.1 The Security Primitives

Our protocol takes advantage of the following three security primitives, the one-way
key chain, the Blundo’s pairwise key pre-distribution scheme and the symmetric double
additive homomorphic encryption.
One-way key chain [14]. An one-way key chain {K0 , K1 , ...., KR } is generated by
iteratively performing the one-way hash function H(·) on the last key KR in the chain.
This key chain has the following property: given any Kj , all former keys can be derived
by computing Ki = H j−i (Kj ), 0 ≤ i < j, while none of the later keys can be
computed due to the one-wayness of function H. Therefore, with the knowledge of
K0 = H(K1 )(called the key chain commitment), anybody can verify the authenticity
of any later key by only performing hash operations. The reason is that the attacker
cannot change any hash value because the hash function in use is collision resistant. In
our protocol, the one-way key chain will be used to protect the announcement reported
by the self-elected CHs.
Blundo’s pairwise key pre-distribution scheme [4,14]. Blundo [4] proposed a
polynomial-based key pre-distribution scheme that was aimed to pre-distribute group
key. In this scheme, the BS generates a t-degree bivariate polynomial f (x, y) that has
a property f (x, y) = f (y, x) over a finite filed Fq . The q is a prime number and is
enough large to generate cryptographic keys. For each node having a unique ID i, the
BS distributes the polynomial key f (x, i). Based on this polynomial key, any two nodes
sharing the key can generate a pairwise security key. Let us have an example. When the
node i wants to communicate with a node j, it generates a key f (j, i) and the node j
can also generate a key f (i, j). Due to the property of f (x, y) = f (y, x), the two keys
are same and this key becomes a unique pairwise key between node i and node j. This
scheme can tolerate up to t compromised nodes, i.e., if an adversary wants to obtain one
polynomial key f (x, y), he must capture at least t + 1 nodes having the polynomial key
[4]. We do not consider the case where a large number of nodes are compromised since
most applications will fail anyway in such situation [14]. Therefore, we always assume
no more than t compromised nodes in the networks [14]. We will use this for resilient
CH election in our proposed scheme.
150 Y.-J. Han, M.-W. Park, and T.-M. Chung

Symmetric double additive homomorphic encryption scheme. [13] A homomorphic

encryption scheme allows arithmetic to be performed on ciphertext. One example is a
multiplicatively homomorphic scheme, whereby the multiplication of two ciphertexts
followed by a decryption operation yields the same results which is the multiplication of
the two corresponding plaintext values. Homomorphic encryption scheme is especially
useful in scenarios where someone who does not have decryption keys needs to per-
form arithmetic operations on a set of ciphertext. A more formal description of additive
homomorphic encryption scheme is as follows [13].
Let be Enc: K × P → C and Dec: K × C → P , where Enc is a symmetric encryp-
tion function, Dec is the corresponding decryption function, P is the plaintext space,
C is the ciphertext space, and K is a set of symmetric secret keys. Enc is additively
homomorphic, if and only if for given messages a1 , a2 , ..., an ∈ P and for secret keys
k1 , k2 , ..., k3 ∈ K, there exists a key k ∈ K such that the following equation holds

(a1 + a2 +, ..., +an ) = Deck (Enck1 (a1 ) + Enck2 (a2 ) + ... + Enckn (an )) (2)

A double homomorphic encryption scheme is an encryption that is homomorphic in

both the key and the plaintext. That is, it holds that

(a1 + a2 +, ..., +an ) = Deck1 +k2 +...+kn (Enck1 (a1 ) + Enck2 (a2 ) + ... + Enckn (an ))
(3)
In our proposed scheme, we will use this for concealed data aggregation.

4.2 Protocol Description

Initialization. Firstly, to use μTESLA [2] for the BS’s authenticated broadcast, the BS
creates BROADCAST key chain by using one-way hash key chain method in subsec-
tion 4.1, {K 0 , K 1 , ...., K brmax }, where brmax is the maximum number of broadcasts
from the BS. The BS creates a pairwise symmetric keys, {K1 , K2 , ..., KN }, for all
sensor nodes to protect the node-to-BS communication. The BS then maintains a sym-
metric bivariate polynomial f (·, ·) which is a key material for establishing a pairwise
key between sensor nodes after deploying and pre-distributes a polynomial share f (i, ·)
to every node Si . We will describe later how to generate sensor node id, Si (id).
Next, prior to deployment, each sensor node Si is assigned with node ID, Si (id),
two random one-way key chains, the A − key chain1 and the G − key chain2 , a
pairwise key shared with the BS, Ki , and K 0 which is shared by all members of the
network for BS’s authenticated broadcast. The G-key chain is used by this node to tell
other nodes that it has no energy left for serving as CH. Therefore, it only includes
a key chain commitment, Gi,0 , and G-key, Gi,1 . The A-key chain is used to inform
other nodes that it is a self-elected CH. Therefore, it includes a key chain commitment,
Ai,0 , and a number of A-keys ,{Ai,1 , Ai,2 , Ai,3 , ...., Ai,δ∗Popt ∗ermax }. We will discuss
in section 5.4 why we determine the number of A-keys as δ ∗ Popt ∗ ermax . Lastly, we
define that the ID of sensor node is the hash of the two key chain commitments, i.e.
Si (id) = H(Ai,0 Gi,0 ) like in [14].
1
ANNOUNCEMENT key chain.
2
GIVEUP key chain.
 SecDEACH: Secure and Resilient Dynamic Clustering Protocol 151

After deployment, the sensor node Si firstly establishes a pairwise key with neigh-
borhood sensor nodes within their communication range Si (r) by using Blundo’s pair-
wise key pre-distribution scheme and initializes their neighborhood sensor nodes list,
Sneighbor (i). A pairwise key between two sensor nodes Si and Sj is denoted as Ki,j .

Next, the every sensor node calculates its Popt (i) for the CH election. Finally, the every
sensor node needs to exchange their key chain commitments (A-key commitment, Ai,0
and G-key commitment, Gi,0 ) between each other within its transmission range. For

doing so, each sensor node sends initM sg(Si ) = {Si (id)Eki,j (Ai,0 Gi,0 Popt (i))},
within its transmission range. The sensor node receiving this message can easily verify
these commitments according to the node ID. Whenever a sensor node Si receives the
verified A-key commitment from a node Sj , Si creates Qi (j). Qi (j) means that Sj

has a qualification being a CH of Si . It consists of node ID, 1/Popt (j) rounds, the last

round that the node Sj was a CH, the number of announcements within last 1/Popt (j)
rounds and last A-key. After collecting all those authenticated A-chain commitments
from other nodes, every sensor node Si creates a qualification list, Sq (i). Sq (i) consists
of array of Qi (j). This is meaningful because a dynamic clustering formation is only
performed within its communication range.
Secure Cluster Head Election. We use DEACH in section 2.1 as a fundamental clus-
tering method. DEACH consists of a cluster-setup phase and a data transmission phase.
The secure CH election and secure cluster formation phases belong to the cluster-setup
phase. For simplicity, we assume that no collision occurred in the MAC(Media Access
Control) layers of sensors during the clustering routing.
At the beginning of this phase, if the residual energy Eres (i) of the sensor node Si
is less than the threshold Eth , the sensor node Si broadcasts the key of its G-key chain
to announce its decision. Eth is the minimum energy required for serving as CH and
pre-determined system value. If Eres (i) is greater than Eth ,this node chooses a random
number between 0 and 1. If the random number is less than its Pi in Eq. 1, this node
becomes the CH and releases its A-key from A-key chain and broadcasts its announce-
ment message, advHeadM sg(CHi ) = {Si (id)Ci,c }, β times to tolerate the channel
loss within transmission range. c is the number of the cumulative announcements sent
by Si . If a sensor node Sj receives a key in the A-key chain of Si , node Sj can verify
it by computing Ai,c−1 = H(Ai,c ) and comparing with the key Ai,c−1 received ear-
lier. If the verification succeeds, node Sj extracts Qj (i) from its Sq (j) . And then Sj
determines whether Si has a qualification being a CH by using Eq. 4.

1 , if (r − r ) ≥ r mod (1/Popt (i))
V (i) = { (4)
0 , otherwise
where r is the current round, r is the last round in Qj (i). If V (i) is one, Sj adds Si into
SCH (i) since it means that Si is a non-compromised node. Otherwise, this announce-

ment is ignored. If this announcement continues over α times within 1/Popt (i), Si is
removed permanently from Sq (j). If node Sj receives the correct G-key Gi,1 from Si ,
Sj will remove Si from Sq (j). Finally, after SCH (j) is created, Sj chooses the closest
CHi among the CHs in SCH (j) and joins to the CHi as the member. After joining, it
updates the A-key and the last round in Qj (i). Table 1 shows the CH election process
in a sensor node having 5-node neighbors within its transmission range when α = 3.
152 Y.-J. Han, M.-W. Park, and T.-M. Chung

Table 1. Illustration of the resilient CH election in a sensor node having 5-node neighbors(α = 3)

Released Information Si (Q) CH

Si (CH) Notes
1 2 3 4 5 =(list of (ida ,rb , pc ) Node
initial A1,0 A2,0 A3,0 A4,0 A5,0 Exchange Commit.
Stage G1,0 G2,0 G3,0 G4,0 G5,0 and generate Si (Q)
(1,0,3),(2,1,4),(3,0,6)
Round 1 A2,1 A4,1 2,4 2 or 4 select 2
(4,0,7),(5,0,4)
(1,0,3),(2,1,4),(3,2,6) remove 4
Round 2 A3,1 G4,1 3 3
,(5,0,4) , α for 2 =1
(1,0,3),(2,1,4),(3,2,6) select 5
Round 3 A2,2 A5,1 5 5
,(5,3,4) , α for 2 =2
select 1 and remove 2
Round 4 A1,1 A2,3 A3,2 A5,2 (1,4,3),(3,2,6),(5,3,4) 1 1
because α for 2 =3
a
Si (id).
b
Last round which it was a CH.

c
1/Popt (i).

Secure Cluster Formation. After electing the CHj , Si sends the joinM sg to the
CHj . The joinM sg(Si , CHj ) is as follows.
joinM sg(Si , CHj ) = {Si (id)Sj (id)nonceM ACki,j (Si (id)Sj (id)nonce)}
The nonce is used to prevent the replay attack. After receiving joinM sg, the CHs
send the time slot schedule to their cluster member nodes by using a pairwise key.
Secure Data Aggregation. In this step, Si encrypts the sensed data using a shared key
Ki with the BS for the concealed data aggregation of its CH and sends reportM sg.
The reportM sg(Si , CHj ) is as follows.
reportM sg(Si , CHj ) = {Si (id)Sj (id)Encki (di )nonce + l
M ACki,j (Si (id)Sj (id)Encki (di )nonce + l)}
The node-to-CH communication is protected by using the same key used to pro-
tect the joinM sg and guarantees integrity through MAC3 . A value computed from the
nonce and the reporting cycle(l) [9] is also included to prevent replay. The CHs can
now check the authenticity of sensing
 reports they receive and then perform a con-
cealed data aggregation dsum = Encki (di ) for all encrypted sensing data using
symmetric double homomorphic encryption scheme in Eq. 3. That is, not only the CHs
do not need to decrypt the encrypted sensing data but also can not the CH decrypt the
encrypted data since the encryption is performed by using the shared key with the BS.
After performing the concealed data aggregation, the CH sends reportM sg to the BS.
The reportM sg(CHi , BS) is as follows where Lmembers is the IDs list of reporting
member nodes of CHi . It is required for the BS to decrypt the sum of encrypted data
by the CHi .
reportM sg(CHi , BS) = {Si (id)Lmembers (i)dsum counter
M ACki (Si (id)Lmembers (i)dsum counter)}
3
Message authentication code as keyed hash function.
 SecDEACH: Secure and Resilient Dynamic Clustering Protocol 153

The aggregate result is protected by using the symmetric key shared between the CH
and the BS. For freshness, a counter (shared between the CH and the BS) is included
in the reporting message. The BS receiving the aggregated data can decrypt the data
using shared keys between sensor nodes using Eq. 2.

5 Analysis
5.1 Security Goal 1: Resilience against Fabrication Attacks, Impersonation
Attacks and Replay Attacks
This security goal requires authenticity, integrity, confidentiality and freshness to node-
to-node communications.
For authenticity, it has to guarantee that only legitimate cluster can participate in
cluster routing operation. In other words, we have to ensure that the unauthorized nodes
do not have to join the CH election and cannot impersonate any ordinary sensor node.
These two properties are actually achieved by using the Blundo’s key pre-distribution
protocol [4]. The ID of a sensor node is tied to its cryptographic key. It is thus not
possible for an unauthorized node to join the cluster routing procedure. That is, the
messages, initM sg, joinM sg and reportM sg, are encrypted with a pairwise key and
a successful decryption of this message allows the sensor node to conclude that the
message is originated from a legitimate node in the network. In addition, even though
advHeadM sg is not encrypted, the adversary cannot forge the announcements from a
legitimate node. The announcement contains the ID of the sensor node and the key in
A-key or G-key chains. To forge these keys, the attacker needs to either invert the one-
way hash function H or fool benign cluster members into accepting incorrect key chain
commitment. However, because of our ID assignment scheme, the ID of a sensor node,
Si (id), is the hash of its two key chain commitments, i.e., Si (id) = H(Ai,0 Gi,0 ).
Therefore, to forge the key chain commitment, the adversary has to forge the ID as
well. This requires the knowledge of keying materials related to the forged ID, which is
not available to the adversary. Therefore, the adversary has no way to make its forged
commitments accepted by any benign cluster member. Without the proper chained keys,
the adversary may try to replay the keys released earlier by the legitimate member.
However, this will not impact our scheme because our scheme uses either the fresh
A-key or the G-key in each round of CH election.
Because the messages, joinM sg and reportM sg, are encrypted by MAC using a
pairwise key and includes the nonce, the CH can also conclude that it is not changed and
it is not a stale message being replayed. The freshness of all subsequent sensor reports
from the ordinary nodes to their CH is guaranteed by nonce values that are incremented
each time. For CH-to-BS communication, the freshness is guaranteed by the counter
value shared between the CH and the BS. The counter value is also incremented each
time the CH sends a new report to the BS.

5.2 Security Goal 2: Fault-Tolerance

A compromised node can always behave normally like benign node and be elected
as the new CH at some point during the field operation. There are no effective ways
154 Y.-J. Han, M.-W. Park, and T.-M. Chung

to identify those passive malicious node since there are no evidences of them being
malicious. As a result, this goal focuses on making sure that an adversary cannot sig-
nificantly impact the chance of any benign members elected as a new CH. Thus even
if the CH is compromised, it cannot continue to control properly the cluster. The CH
election protocol should ensure the fairness of the sensor nodes being elected as CHs.

For doing so, our idea is to ensure that each sensor node becomes a CH one per 1/Popt
rounds on average.

100
F−LEACH, GS−LEACH and
The number of the compromised CHs

SecLeach
80 Dong’ Scheme and
The proposed Scheme

60

40

20

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
The probability being compromised, f

Fig. 2. The number of compromised CHs

In our scheme, the CH election is based on the neighbors list Sneighbor (i), which
is initialized at the beginning of the CH election. All nodes in the neighbors list has
the qualification serving as the CH. In our scheme, the sensor node can receive sev-
eral announcements being a CH from the neighborhood in each round of CH election.
The sensor node receiving them has to verify whether they have a qualification serving
as the CH and chooses its CH among the verified candidate CH nodes. To verify the
qualification, our scheme uses Eq. 4. If the announcement is advertised more than α

times within 1/Popt , the verification is failed. If the verification is failed for a node, the
node cannot become a CH. It ensures that each sensor node becomes a CH one per its

1/Popt rounds on average. Figure 2 shows the comparison our scheme with the exist-
ing secure cluster routing schemes [1,9] in the aspect of the compromised CH election.
Figure 2 shows the number of the compromised node being elected CH according to

the probability being compromised of the sensor node, f , where EX(1/Popt ) = 5 and
N = 100. The number of compromised CH in SLEACH, F-LEACH and SecLEACH
is very larger than that of our scheme and Dong’s scheme because the secure clustering
schemes like F-LEACH cannot know if the a self-selected CH is compromised. As we
can see in Fig. 2, our scheme provides resilient CH election like Dong’s scheme [14].

5.3 Security Goal 3: Concealed Data Aggregation In-Networking

For the concealed data aggregation, our scheme ensures that only BS can decrypt the
aggregated data from all sensor node. This property is actually achieved by using sym-
metric double additive homomorphic encryption [13]. Each sensor node encrypts the
 SecDEACH: Secure and Resilient Dynamic Clustering Protocol 155

sensed data with a shared key with only BS. After receiving all reportM sg from all
member node in its cluster, the CH just only sum all the encrypted data and then sends
the summed data to the BS. The BS can extract the sensed data from the summed en-
crypted data since the BS knows all of the shared keys with sensor nodes. It offers a
concealed data aggregation as well as data privacy.

5.4 Overhead
Before deployment, every sensor node is assigned with an ID and a t-degree polynomial
share f (i, ·) . It also needs to store two key chains along with their commitments, one
contains a single key, and the other consists of δ ∗ Popt ∗ ermax keys. These chains
are similarly used in Dong’s scheme [14]. However,in Dong’s scheme, if a node has
capability being a CH, sensor node must inform its YES key per every cluster leader
election round since a YES chain is used to announce being a candidate CH. So, the
number of keys of a YES chain in Dong’s scheme is the same as the maximum number
of rounds for CH election, ermax . On the other hands, in our scheme, the sensor node
being a self-elected CH only broadcasts its A-key to announce the fact that it becomes
a CH. Without loss of generality, the ordinary senor node is elected itself as CH one

within last 1/popt rounds on average because the sensor node elects itself as the CH by
Eq. 1 which is optimized into Popt [16]. Thus, in our scheme, each node Si only needs

Popt (i) ∗ ermax keys for A-key chain unlike Dong’s scheme. So, the number of our A-
key chain is very smaller than that of Dong’s scheme. However, there is one limitation

to consider. To calculate Popt , each node has to know its location information which

will be deployed. However, we cannot calculate Popt (i) in the initialization phase since
each sensor node will be deployed randomly. So, we determine the length of A-key

chain as δ ∗ Popt ∗ ermax based on Popt . According to Eq. 1, Popt (i) is less than or
equal to 32 Popt . As a result, the adequate value of δ can be 2.
Figure 3 shows the comparison our scheme with Dong’s scheme in the aspects of
the memory and communication overheads. Figure 3(a) and Fig. 3(b) show the mem-
ory overhead for saving A-key chain and the communication overhead for sending
announcement by the number of CH where key size = 8bytes, ermax = 50 and
N = 100. As we can see in Fig. 3, our scheme is more light-weighted and energy-
efficient than Dong’s scheme. In addition, if there are n sensor nodes within a transmis-
sion range, a sensor node will also need to buffer the neighbors list of n IDs, (n − 1)
pairwise keys, and 2(n − 1) key chain commitments. These space requirements are
usually not a problem since the transmission range is limited.
The computational overhead mainly comes from three parts. First, with the Blundo’s
key pre-distribution protocol, a sensor node needs to compute a t-degree polynomial
to establish a pairwise key and verify the ID of another sensor node. Second, a sensor
node has to perform the hash function H one time to verify the key chain commit-
ment. Finally, every sensor node will need to perform symmetric key operations such
as encryption and authentication to protect the communication between sensor nodes.
Obviously, all these operations only involve light-weighted computation.
The proposed scheme only causes small communication overhead like Dong’s scheme
[14]. The Blundo’s key pre-distribution technique does not introduce any additional
communication overhead at all for two nodes to establish a pairwise key [14]. In the
156 Y.-J. Han, M.-W. Park, and T.-M. Chung
memory overhead for saving A−key chain

450 0.03

400 Dongs Scheme

for sending announcement(J)

Dongs Scheme 0.025 The proposed Scheme

communication overhead
350 The proposed Scheme, δ=2
300 The proposed Scheme, δ=3 0.02
The proposed Scheme, δ=4
250
0.015
200
0.01
150

100 0.005
50
0
10 12 14 16 18 20 10 12 14 16 18 20
The number of CH The number of CH

(a) The memory overhead (b) The communication overhead

Fig. 3. The comparison of the overhead

initialization step, each node only needs to disclose its ID and two key chain commit-
ments. In each round of CH election, every sensor node is only required to broadcast
a single chained key, A-key. Different from centralized approaches [1,9] where sen-
sor nodes exchange messages with the BS that is far away, our scheme only involves
the communication between the sensor nodes that are physically close to each other.
Therefore, the communication cost of our protocol is not a big problem [14].

6 Conclusion and Future Works

In this paper, we have presented an secure and resilient clustering protocol preserving
data privacy. Our scheme provides both a secure clustering and a resilient CH election
preserving dynamic clustering unlike the existing secure clustering schemes. In addi-
tion, our scheme provides data privacy through the concealed data aggregation. The
analysis shows that our scheme offers the security services against both inside attack
and outside attack while having the little memory, computation and communication
overhead. The comparison with related works shows that our scheme has the improve-
ment in security and the reduction of the overheads compared to the existing scheme.
A further direction of this study will perform experiment on using our proposed
schemes and perform the network wide security analysis. In addition, we will apply
public key material into our scheme to support more strong authentication and elec-
tronic signature.

References
1. Ferreira, A.C., Vilaca, M.A., et al.: On the security of cluster-based communication protocols
for wireless sensor networks. In: Lorenz, P., Dini, P. (eds.) ICN 2005. LNCS, vol. 3420, pp.
449–458. Springer, Heidelberg (2005)
2. Perring, A., Szewczyk, R., Well, V., Culler, D., Tygar, J.D.: SPINS: Security Protocol for
Sensor Networks. Wireless Networks 8(5), 521–534 (2002)
3. Karlof, C., Wagner, D., et al.: Secure routing in wireless sensor networks: attacks and coun-
termeasures. Elsevier’s AdHoc Networks J. 1(2-3), 293–315 (2003)
 SecDEACH: Secure and Resilient Dynamic Clustering Protocol 157

4. Liu, D., Ning, P.: Establishing pairwise keys in distributed sensor networks. In: CCS 2003,
Proceeding of the 10th ACM conference on Computer and communications security, pp.
52–61. ACM Press, New York (2003), Network
5. Wang, G., Cho, G.: Secure Cluster Head Sensor Elections using Signal Strength Estima-
tion and Ordered Transmissions, Sensors. MDPI (Molecular Diversity Preservation Interna-
tional) 9(2), 4709–4727 (2009)
6. Karl, H., Willig, A.: Protocols and Architectures for Wireless sensor Networks. John Wiley
& Suns, Ltd., Chichester (2005)
7. Al-Karaki, J.N., Kamal, A.E.: Routing techniques in wireless sensor networks: a survey.
IEEE Wireless Communications 11(6), 6–28 (2004)
8. Akyildiz, L.F., et al.: A Survey on Sensor Networks. IEEE Communications Magazine 40(8),
102–114 (2002)
9. Oliveira, L.B., Ferreira, A., et al.: Secleach-on the security of clustered sensor networks.
Signal Process 87(12), 2882–2895 (2007)
10. Sirivianos, M., et al.: Non-manipulable aggregator node election for wireless sensor net-
works. In: ICST wiOpt (January 2007)
11. Younis, O., Fahmy, S.: HEED: A Hybrid, Energy-Efficient, Distributed Clustering Approach
for Ad Hoc Sensor Networks. IEEE Transactions on Mobile Computing 3(4), 660–669
(2004)
12. Younis, O., et al.: Node Clustering in Wireless Snsor Networks: Recent Developments and
Deployment Challenges. IEEE Network 20(3), 20–25 (2006)
13. Tiwari, P.: Data Aggregation in Cluster-based Wireless Sensor Networks, A master thesis,
Deemed University (July 2008)
14. Dong, Q., Liu, D.: Resilient Cluster Leader Election for Wireless Sensor Networks. In: Pro-
ceedings of The IEEE Communications Society Conference on Sensor, Mesh and Ad Hoc
Communications and Networks (SECON), June 2009, pp. 1–9 (2009)
15. heinzelman, W., et al.: An application-Specific Protocol Architecture for Wireless Microsen-
sor Networks. IEEE Transactions on Wireless Communications 1(4) (October 2002)
16. Han, Y.-J., Park, S.-H., Eom, J.-H., Chung, T.-M.: Energy-Efficient Distance based Cluster-
ing Routing Scheme for Wireless Sensor Networks. In: Gervasi, O., Gavrilova, M.L. (eds.)
ICCSA 2007, Part II. LNCS, vol. 4706, pp. 195–206. Springer, Heidelberg (2007)
 Avoidance of Co-channel Interference
Using Switched Parasitic Array Antenna
in Femtocell Networks

Yeonjune Jeong, Hyunduk Kim, Byung-Sung Kim, and Hyunseung Choo

School of Information and Communication Engineering

Sungkyunkwan University, Korea
{yjjeong83,yuby82,choo}@skku.edu,[email protected]

Abstract. Femtocells are low-cost and low-power cellular home base

stations that connect mobile users to a operator’s network by means of a
broadband backhaul. Through this approach, the network operators can
extend indoor coverage at a low-cost and reduce the operating cost. How-
ever, there exists a co-channel interference between macrocell and fem-
tocells when they use the same frequency band. This co-channel interfer-
ence causes severe performance degradation of femtocells and macrocell.
Therefore, we propose a novel scheme to reduce the co-channel interfer-
ence to macrocell without performance degradation of femtocells. The
proposed scheme avoids the co-channel interference to macrocell users
effectively using the SPA antenna which offers 8-directions to control
the antenna beam pattern and transmitting power. Performance evalu-
ation results have confirmed that the proposed scheme reduces the co-
channel interference by 80% compared to the case using omni-directional
antenna.

Keywords: We would like to encourage you to list your keywords within

the abstract section.

1 Introduction
According to recent surveys, 50 percent of phone calls and 70 percent of data
services have taken place indoors [1]. Therefore, network operators need to pro-
vide high data rate and quality of service (QoS) indoors. Femtocell has been
introduced to solve indoor coverage problem and QoS requirement. One defini-
tion of femtocell is a low-power wireless access point that operates in licensed
spectrum to connect standard mobile devices to the operators network using
a digital subscriber line (DSL) connection [2]. Femtocell provides benefits for
both subscribers and cellular poerators. Subscribers can achieve high data rate
indoors with low-cost with deployment of femtocell. Wireless operator can ex-
tend the indoor coverage with low-cost, increase cell capacity, and gain the high
frequency reuse efficiency. Moreover, cullular operators can reduce traffic for
macrocell, thus reducing infrastructure and operating cost.

Corresponding author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 158–167, 2010.

c Springer-Verlag Berlin Heidelberg 2010
 Avoidance of Co-channel Interference 159

One of the important issues of introducing femtocells is solving the interference

problem with macrocell [3-7]. If macrocell and femtocells use the same frequency
in orthogonal frequency division multiple access (OFDMA) system, there is a co-
channel interference between macrocell and femtocells. This interference causes
performance degradation of macrocell and femtocell. Cell capacity decreases as
the number of femtocells increases [3]. Since the femtocell is deployed by its users,
it is difficult for cellular operators to estimate the location of femtocells. There-
fore, cellular operator cannot solve the interference problem between macrocell
and femtocells with existing network planning and optimization methods.
To solve the above problem, a method to control femtocell coverage has been
proposed to reduce the interference between macrocell and femtocells. It defines
the area where the signal strength of femtocell is equal to that of macrocell
as interference-limited coverage area (ILCA) and control the ILCA to reduce
the amount of interference [8]. However, it reduces the femtocell coverage as
the distance between femtocell base station (FBS) and macrocell base station
(MBS) is getting closer. Another method has been introduced which controls the
transmitting power and steers the beam pattern of the femtocell to reduce the
interference to macrocell [9]. In this scheme, femtocell adjusts its transmitting
power equal to the macrocell power at the border of femtocell considering the
path loss of the macrocell and femtocell. However, it needs complicated algorithm
to reduce the interference between macrocell and femtocell.
In this paper, we propose a new physical approach adopting dynamic antenna
beam planning to avoid the interference using a switched parasitic array (SPA)
antenna. The SPA is an antenna capable of steering its main beam direction
dynamically [10-12]. Compared to the multi-element antennas used in [9], the
SPA provides the evenly distributed directional antenna patterns, which makes
it easier to estimate the receiving power and optimize the control algorithm.
FBS controls the beam pattern of antenna to support its user only with high
front-back ratio. Therefore, it can mitigate the interference to macrocell users
effectively. Additionally, it can reduce the power consumption of FBS without
any degradation of the performance of femtocell because SPA may achieve higher
antenna gain compared to omni-directional antenna.
The paper is structured as follows. In section 2, previous approaches to solve
the interference problem between macrocell and femtocells are discussed and the
structure and characteristics of SPA antenna are described. Then we propose
a method of data transmission and power control in femtocells using SPA in
section 3. In section 4, the performance of the proposed method using SPA is
evaluated compared to that of femtocell using omni-directional antenna. Finally,
conclusions are given in section 5.

2 Related Work

If the macrocell and femtocell use the same frequency band, it causes co-channel
interference between them. Figure 1 shows that the co-channel interference sce-
nario between the macrocell and femtocell in downlink. The macrocell users are
160 Y. Jeong et al.

Macrocell Femtocell Transmitted Interference

coverage coverage signal

Femto UE

Femto BS

Macro BS
Macro UE

Fig. 1. Scenario of macro and femtocell Interference in downlink

interfered from FBS, and the femtocell users are interfered from MBS as shown
in Figure 1. When a macrocell user approaches the femtocell coverage, the effect
of interference from FBS may grow up, and when the femtocell user approaches
MBS, it may suffer heavy interference from MBS. Such cases, the performance
of macrocell and femtocell are decreased. If the macrocell user gets into the
femtocell coverage, communication with MBS becomes impossible, because the
interference to macrocell user in femtocell coverage is much bigger than received
signal strength from MBS.
In the previous study, FBS uses omni-directional antenna to transmit the sig-
nal to all direction irrespective of its user’s position [7, 8]. Therefore, macrocell
users are interfered from FBS when they are located nearby the femtocell cover-
age. Additionally, co-channel interference is getting bigger as transmitting power
of FBS increases. If FBS makes beam pattern to its user’s location, the signal
level of FBS is very low in the rest of its user’s location. If the macrocell user is
located outside of femtocell coverage, the co-channel interference from FBS can
be avoided based on macrocell user’s location.
In [9], a multi-element antenna solution is proposed to decrease the interfer-
ence in macrocell users. It can control the antenna beam direction by switching
of two patch antennas and IFA antennas. The effect of interference in macrocell
users is reduced by controlling the transmission power Pf emto of FBS according
to the equation given below:

Pf emto = min (Pmacro + G (θ) − Lmacro (d) + Lf emto (r) , Pmax ) (1)

where Pmacro is the transmitting power of MBS, G(θ) is the antenna gain of
MBS, L( macro)(d) is average outdoor path loss when the distance between
macrocell and femtocell is d. Lf emto (r) is average indoor path loss when the
distance between FBS and its user is r. Pmax is the maximum transmitting
 Avoidance of Co-channel Interference 161

(a) Omni-directional antenna (a) SPA antenna

Fig. 2. Structure of antennas

power of FBS. The method of (1) means that FSB increases the transmitting
power to overcome the interference by macrocell when it is closely located to
MSB and conversely the FBS far away from MBS lowers its power.
Figure 2 shows omni-directional antenna and SPA antenna. In this paper,
FBS uses the SPA antenna in Figure 2 (b). A monopole antenna located at the
center of the ground plane, and four identical parasitic array elements are placed
surrounding the monopole symmetrically. Switches are inserted between array
elements and ground plane to control the main beam direction of the monopole
antenna by alternating the ground connection of the parasitic elements. A bias
applied to the knee in the element provides an RF shorting path to the ground
plane, and the shorted parasitic elements acts as a “director” to guide the main
beam direction to its direction. A bias applied to the end of element provides a
long RF path to the ground plane when the switch at the knee is turned-off. In
this case, the parasitic element works as a “reflector.”
If two switches in the parasitic element are turned-off, the element is “in-
visible.” Four near orthogonal beam patterns can be generated by sequentially
switching two element pair placed in a diagonal position as front director and
back reflector. In addition, eight directional patterns can be also obtained by
utilizing two adjacent reflectors and two directors. All directional beam patterns
are distinctly identified from each other according to their direction with the
same shape and easily synthesized by controlling the switches. Even radiation
pattern and reasonable front-back ratio for each direction makes it possible to
simplify the power control algorithm to reduce the interference.

3 Proposed Scheme

In this section, we explain the procedure of data transmission to its user in FBS
using SPA antenna. We assume that FBS are configured to limit access to only
a few authorized users to protect limited resources. If unauthorized users re-
quest to connect to the closest femtocell, FBS restricts connection to femtocell
for them. In this condition, our proposed scheme is as follows. The FBS mon-
itors the user location by measuring the signal strength of surrounding users.
Then, FBS performs 1-directional power control of transmitting signal with an-
tenna gain based on user’s position. Then, check the users’ ID to avoid unwanted
162 Y. Jeong et al.

data transmission to unauthorized users. Finally, FBS make the beam pattern
to transfer the signal to authorized user. These processes guarantee the same
performance of femtocell with FBS using omni-directional antenna. Addition-
ally, it reduces the co-channel interference to macrocell users nearby femtocell
coverage. The procedure of data transmission of FBS is shown in Figure 3.

8-direction Sensing
(Receive Mode)

Confirm the User’s

Position

1-direction
Power Control N

A subscriber
Certification
Y

Data Transmission

Fig. 3. Procedure of data transmission of FBS.

First, FBS measures the signal level of nearby users. Second, FBS confirms
the user’s location based on received signal strength. Third, according to the
measured distance and location, FBS controls the transmitting power as below:

Pf emto = min (Pomni − G (θ) + Lf emto (r) , Pmax ) (2)

where Pomni is transmitting power when FBS uses omni-directional antenna.

G(θ) is antenna gain of SPA in direction of the user, where θ is the angle to
FBS with respect to the user’s location. Lf emto (r) is the average path loss at
the distance r between FBS and the user. Fourth, FBS checks the user ID. In
this paper, assume that the FBS acts as closed access mode which only allows
the wireless connection to registered users. Thus, FBS can reduce the unwanted
data transmission for unregistered users. If unregistered user request the wireless
connection to femtocell, FBS goes back to receive mode. By redistricting wire-
less connection for unregistered users, femtocell can avoid the unnecessary data
transmission, and reduces co-channel interference to macrocell user effectively.
It has advantages controlling the transmitting power of FBS based on its user’s
location. First, femtocell can reduce interference to adjacent macrocell users. For
example, macrocell users get interfered from FBS without their locations when
FBS uses omni-directional antenna. However, if our proposed scheme is applied,
FBS makes the beam pattern to the certain limited area where its serving user
 Avoidance of Co-channel Interference 163

is located. If the angle of the location between macrocell user and femtocell user
is quite different, macrocell user can avoid the co-channel interference from the
FBS. Otherwise, macrocell user may get interfered from FBS as the difference
of angle with femtocell user’s location.
Second, by using SPA antenna in FBS, we can obtain higher antenna gain
than FBS using omni-directional antenna. Thus, FBS using SPA antenna has
the same performance with omni-directional antenna while transmitting power
of FBS is lower than that of omni-directional antenna. Additionally it can avoid
unnecessary data transmission due to transmit the signal only for registered
users. When FBS uses SPA antenna, power consumption can be lower than that
of omni-directional antenna.

Table 1. Simulation Parameters

Parameters Values
System bandwidth 10MHz
Inter-macrocell distance 1000m
Inter-femtocell distance 40m
Radius of femtocell 10m
Macrocell BS power 20W
Femtocell BS power 20mW
Number of subcarriers 768
Frame duration 5ms
Number of OFDM symbols 48
Thermal noise(AWGN) density -174dBm/Hz
System frequency(λ) 2GHz

4 Performance Evaluation
In this paper, we performe the system level simulation in three cases: conven-
tional OFDMA macrocell with no femtocells; macrocell with femtocells using
omni-directional antenna; and macrocell with femtocells using SPA antenna. We
consider the reuse factor 1 for macrocell as 2-tier network that has 19 cells. We
assume that MBS is located at the center (0m) of the macrocell. Then, macrocell
is divided into two regions according to the distance from MBS. 0∼409m is inner
region, and 409∼577m is outer region. There are 30 users in each macrocell. We
set the radius of femtocell to 10m; each femtocell has one user. We assume that
FBS is loacaed at the center (0m) of each femtocells. Totally, 409 Femtocells are
distributed uniformly in a macrocell and the distance between adjacent FBS is
40m. The table 1 shows parameters used in performance evaluation.
The path loss model used in performance evaluation is as follows. First, we
use the modified COST 231-Walfisch-Ikegami urban macro model to know path
loss of macrocell user [13].
Pout [dB] = 31.81 + 40.5 log10 (d) + ψout (3)
164 Y. Jeong et al.

where the d [m] is distance between MBS and the its user. ψout [dB] is outdoor
shadowing modeled as log-normal distribution with zero mean, 8dB standard
deviation.
Second, we use Modified COST 231-multiwall model to know path loss of
femtocell user [14].

Pin [dB] = 37 + 32 log10 (d) + L + ψin (4)

where the d [m] is distance between the FBS and the its user, L[dB] is path loss
by wall located in the edge of femtocell. We set to L = 8dB in this simulaiton.
ψin [dB] is indoor shadowing modeled as log-normal distribution with zero mean,
4dB standard deviation.

(a) radiation patterns with 0°, 90°, 180°, 270° (b) radiation patterns with 45°, 135°, 225°, 315°

Fig. 4. The radiation patterns of SPA antenna

Third, we use indoor-to-outdoor model to know the effect of femtocell signal to

macrocell user [15].

Pin−to−out = Pout (dout + din ) + 14 + 0.5din + ψin (5)

where Pout [dB] is outdoor path loss model in Eq. (3). dout is the distance between
macrocell user and the wall located at the edge of femtocell. din is distance
between the FBS and the wall located in the edge of femtocell.
Figure 4 shows beam patterns of SPA antenna. The mesh size is λ/10 (λ
is wavelength.), the simulation is executed by the method of moments (MoM)
which is a full wave solution of Maxwell’s equation in the frequency domain. As
we can see in Figure 4, SPA antenna can radiate in one direction which has about
10dB front-to-back ratio symmetrically [10-12]. SPA has symmetric radiation
pattern for all directions and reasonable front-back ratio for each direction. Thus,
by controlling beam pattern of FBS according to its user’s location, we can
 Avoidance of Co-channel Interference 165

1.0 1.0

0.8 no femtocell 0.8 no femtocell

with femtocell with femtocell
with femtocell with femtocell
and SPA and SPA
0.6 0.6
CDF

CDF
0.4 0.4

0.2 0.2

0.0 0.0
0 10 20 30 40 50 60 -4 -2 0 2 4 6 8
Inner cell user SINR(dB) Outer cell user SINR(dB)

Fig. 5. CDF of average macrocell user SINR

decrease interference to macrocell user with same performance that of omni-

directional antenna.
Figure 5 shows that CDF of the macrocell user’s average SINR in the following
three cases: conventional OFDMA macrocell with no femtocells; macrocell with
femtocells using omni-directional antenna; and macrocell with femtocells using
SPA antenna. First case has the highest user’s SINR with no doubt. Because,
there are only inter-cell interference among the adjacent macrocells. In conven-
tional OFDMA macrocell, outer users have low SINR because, outer users get
severe interference from neighboring cells.
Second case shows that macrocell users have the lowest SINR among three
cases. Because femtocell causes co-channel interference with macrocell, macrocell
users get interfered from FBS using omni-directional antenna. It degrades the
macrocell user’s SINR severely. In this simulation, we set the minimum required
SINR to -4.34dB for macrocell user to communicate with MBS. In outer region,
most users have SINR lower than -4dB. It denotes that most users cannot com-
municate with serving MBS because of the co-channel interference from FBS.
The third case shows the macrocell user’s SINR when our proposed scheme is
applied. Macrocell user can reduce the co-channel interference from FBS. Sim-
ulation results show that our proposed scheme improves the macrocell user’s
SINR about 60% in inner region and 80% in outer region compared to the FBS
using omni-directional antenna. It menas FBS avoids the co-channel interference
to adjacent macrocell user effectively because, FBS steers beam pattern to its
user’s location.
Figure 6 shows the cell throughput of the above conditions. First case has the
highest throughput as shown in Figure 5. Second case, conventional OFDMA
macrocell with femtocell using omni-directional antenna, has much smaller
throughput than the first case, because macrocell users are interfered from FBS
severely. Especially macrocell users in outer region get intense interference not
only inter-cell interference from adjacent macrocells, but also co-channel interfer-
ence from nearby femtocells. Since SINR is lower than the minimum requirement,
outer user’s throughput is close to zero. Third case shows the macrocell through-
put when we adopt our proposed scheme. It has higher throughput than second
166 Y. Jeong et al.

20
no femtocell
18
with femtocell
16 with femtocell and SPA

14
Throughput(Mbps)

12

10

0
inner region outer region

Fig. 6. Macrocell throughput

case, because there is less probability that macrocell users get interfered from
FBS than the second case. Third case shows that our proposed scheme improves
the throughput of macrocell users more than 4 times compared to the second
case.

5 Conclusion

Femtocells are small cellular base stations which extend indoor coverage and
provide high data rate. However, there exists co-channel interference between
macrocell and femtocells when they are operating in the same frequency band.
This co-channel interference causes severe performance degradation both macro-
cell and femtocells. Especially as FBS are using omni-directional antenna, they
cause co-channel interference to nearby macrocell users regardless of their loca-
tions. And macrocell user’s SINR gets lower as it gets closer to FBS. Therefore,
in this paper we have proposed procedure of data transmission and power control
for FBS using SPA antenna to reduce the co-channel interference with macrocell
users. In performance evaluation, we confirmed that macrocell user’s SINR is
improved by 60 ∼ 80% and macrocell throughput is increased more than four
times compared to FBS using omni-directional antenna.

Acknowledgment

This research was supported by MKE, Korea under ITRC NIPA-2009-(C1090-

0902-0046) and by MEST(Korea), under WCU Program supervised by the
KOSEF (No.R31-2008-000-10062-0). Profs. B.Kim and H.Choo are correspond-
ing authors.
 Avoidance of Co-channel Interference 167

References
1. Mansfield, G.: Femtocells in the US Market -Business Drivers and Consumer Propo-
sitions. FemtoCells Europe, ATT, London (2008)
2. Femto Forum, http://www.femtoforum.org/femto/
3. Chandrasekhar, V., Andrews, J., Gatherer, A.: Femtocell Networks: A Survey.
IEEE Comm. Mag. 46, 59–67 (2008)
4. Lopez-Perez, D., Valcarce, A., Roche, G., Zhang, J.: OFDMA Femtocells: A
Roadmap on Interference Avoidance. IEEE Comm. Mag. 47, 41–48 (2009)
5. Wang, Y., Kumar, S., Garcia, L., Pedersen, K., Lovacs, I., Frattasi, S., Marchetti,
N., Mogensen, P.: Fixed Frequency Reuse for LTE Advanced Systems in Local
Area Scenarios. In: VTC 2009-Spring, pp. 1–5 (2009)
6. Yavuz, M., Meshkati, F., Nanda, S., Pokhariyal, A., Johnson, N., Raghothaman,
B., Richardson, A.: Interference management and performance analysis of
UMTS/HSPA+ femtocells. IEEE Comm. Mag. 47, 102–109 (2009)
7. Claussen, H.: Performance of Macro- and Co-Channel Femtocells in a Hierarchical
Cell Structure. In: IEEE PIMRC 2007, pp. 1–5 (2007)
8. Guvenc, I., Jeong, M., Watanabe, F., Inamura, H.: A hybrid frequency assignment
for femtocells and coverage area analysis for co-channel operation. IEEE Commun.
Lett. 12, 880–882 (2008)
9. Claussen, H., Pivit, F.: Femtocell Coverage Optimization Using Switched Multi-
Element Antennas. In: IEEE ICC 2009, pp. 1–6 (2009)
10. Scott, N., Miles, O.: Diversity Gain from a single-Port Adaptive Antenna Using
Switched Parasitic Elements Illustrated with a Wire and Monopole Prototype.
IEEE Trans. Antennas Propagat. 47, 1–10 (1999)
11. Wennstom, M., Svantesson, T.: An antenna solution MIMO channels: The switched
parasitic antenna. Signal, Systems and Computer 2000 2, 1617–1621 (2000)
12. Mitilineos, S., Thomopoulos, S.: Development of a Compact SPA for 2.4GHz Ap-
plications using Commercially Available Elements: Design and Experimental Val-
idation. In: Loughborough Antennas and Propagation Conference 2009, pp. 16–17
(2009)
13. Baum, D., Hansen, J., Salo, J.: An interim channel model for beyond-3G systems:
extending the 3GPP spatial channel model (SCM). In: VTC 2005-Spring, pp. 3132–
3136 (2005)
14. European Cooperation in the Field of Scientific and Technical Reserch, EURO-
COST231, Digital Mobile Radio Towards Future Generation Systems, COST 231
final report, http://www.lx.it.pt/cost231/
15. IEEE 802.16m-08/004r5, IEEE 802.16m Evaluation Methodology Document, EMD
(2009)
16. Alipour, A., Hassani, H.: A Novel Omni-Directional UWB Monopole Antenna.
IEEE Trans. Antennas Propagat. 56, 3554–3857 (2008)
 User Policy Based Transmission Control Method in
Cognitive Wireless Network

Noriki Uchida1, Yoshitaka Shibata2, and Kazuo Takahata3

1
Faculty of Software and Information science, Iwate Prefectural University
152-52 Sugo, Takizawa, Iwate 020-0193 Japan
[email protected]
2
Faculty of Software and Information science, Iwate Prefectural University
152-52 Sugo, Takizawa, Iwate 020-0193 Japan
[email protected]
3
Dept. of Informational Social Studies
Saitama Institute of Technology
1690 Fuzenji, Fukaya, Saitama 369-2093 Japan
[email protected]

Abstract. Remarkable wireless networks technology developments have made

us to expect the realization of new applications like the advanced traffic system,
the disaster prevention system, and the adhoc network system. However the
resources of wireless bandwidths are not enough to use for such new
applications because it is not efficient usage. Therefore, it is necessary to
develop with new efficient wireless transmission methods like cognitive
wireless network. In this paper, the transmission control methods in cognitive
wireless network considering with cross layers including user policies are
discussed. First, at the observation stage, the physical data such as user policy,
electric field strength, bit error rate, jitter, latency, packet error rate, and
throughput are observed. Then, at the decision stage, AHP (Analytic hierarchy
process) is applied for decision making process with those parameters. Finally,
the action stage, one of the suitable link is chosen and changed links and
networks.
In the simulation, ns2 are used for the computational results to the effective-
ness of the suggested transmission methods in cognitive wireless networks.

Keywords: Cognitive Wireless Network; QoS; AHP; AODV.

1 Introduction
Recently the development of wireless networks has been growing in various fields,
such as cellar phone, digital TV, and wireless Internet services. Those technologies
make us to expect the new application like the advanced traffic system, the disaster
prevention system, and so on. However, such a remarkable spread of wireless services
cause lack and inefficient usage of radio frequency wireless spectrum, we're in front
of a spectrum explosion, that is, one may not have enough wireless resources for new
wireless applications. According to FCC reports [6], wireless resources are limited

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 168–177, 2010.
© Springer-Verlag Berlin Heidelberg 2010
 User Policy Based Transmission Control Method in Cognitive Wireless Network 169

and may not be efficient, because wireless device depends on access to the radio
frequency (RF) wireless spectrum, and spectrum has been chronically limited ever
since transmissions were first regulated in the early 20th century. Therefore, new
technologies that use spectrum more efficiently and more cooperatively, unleashed by
regulatory reforms, will soon overcome the spectrum shortage.
Cognitive wireless network (CWN) by J. Mitola III [8] from software defined
wireless was originally considered to improve spectrum utilization, and generally
considered as a technology to identify the opportunities using the “spectrum holes”
for telecommunications [1] [2]. In other words, Mitola defined that CWN is an
intelligent wireless communication system that is aware of its surrounding
environment, and uses the methodology of understanding by-building to learn from
the environment and adapt its internal states to statistical variations in the incoming
RF stimuli by making corresponding changes in certain operating parameters in real-
time, with two primary objectives in mind: highly reliable communication whenever
and wherever needed; efficient utilization of the wireless spectrum.
Also, CWN is consisted of the cogitation cycle; observing the environment,
orienting itself, creating plans, deciding, and then acting reconfigurations. The
cognition cycle is continually responding stimuli from environment, and it makes
successful link or path of network.
However, CWN still has some problems to realize like some algorisms, control
methods, and technical problems to attain efficient transmission. First of all, it is
recently known that CWN needs not only ordinal spectrum management or
Physical/Mac layer observation control but also new control method including upper
layers. Also, CWN still has technical problems like how to set radio frequency, how
to observe its environment, and how to select proper transmission protocol or radio
frequency, and so on [3][4][5]. Moreover, CWN needs the QoS algorism to select one
of wireless links and network route for each application or user environment.
In this paper, we consider the selecting method of the proper wireless link which
each node has several wireless links such as IEEE802.11a/b, and IEEE802.16
(WiMax), and the proper route in which network environment is changing through the
time. First, user policy and network parameters are observed at the observation stage.
At this stage, we set various policies for video, VoIP, text, disaster applications, and
so on. By each policy, the weight parameter is decided for the decision algorism
which based on Analytic Hierarchy Process (AHP). Network parameters like
throughput, latency, jitter, packet error rate, bit error rate, and electric field strength
are continuously measured and also used for the calculation of AHP. Secondly, at the
decision stage, the results of AHP for each communication link are compared, and the
proper link is selected when the results are changed. Also, if there is no proper
wireless links at the observing nodes, proper route is analyzed by extend Ad hoc On-
Demand Distance Vector (AODV) algorism. Finally, in the acting stage, the selected
link or route is applied for supposed network, and then simulation is carried out for
our supposed methods.
In the followings, in section 2, network model and system architecture of our proposal
communication method are defined. Section3 deals with the observation stage which is
the method of observing network parameters and user policy. Section 4 describes
wireless link selection, route selection and decision algorism which based on AHP.
Section 5 explains how to change link and route at supposed network. At section 6, the
170 N. Uchida, Y. Shibatta, and K. Takahata

simulation is held for the moodel by the calculation for the effectiveness of the suggessted
control method. Finally sectiion 6 derived our conclusion and future works.

2 Network Model
A. Network Configuration
The network configuration n of our suggested cognitive wireless network consistss of
several wireless nodes as sh
hown in Fig. 1.

Fig. 1. Wireless Network Architecture

Wireless nodes are supp posed as wireless terminal, and they have routing, adhhoc,
and multi-hop functions. Every
E node has multi wireless links like IEEE802.111a,
IEEE802.11b, IEEE802.16 (WiMax), and IMT2000. The lowest RF link is used as the
control links for the purpose of sending a network characteristic data or messagee of
switching links. Also, Antennna of each node is supposed as non-directional.
Moreover, the network condition is changed over time by node’s movementt or
radio interference like treees or buildings. A data transmission is carried out by
ordering user request from a server to a user terminal.
B. System Architecture
The system architecture is organized three layers, including the physical layer, the
network layer, and the appliication layer in Fig 2.

Fig. 2. System Architecture

 User Policy Based Transmission Control Method in Cognitive Wireless Network 171

When transmission data sent from sender to receiver, network data from each layer
is observed through link0 that is the control link as explained in the previous section.
In the application layer, a user application is detected for the decision of user
policy. A user policy is assumed to various types like video, VoIP, text, connectivity
(emergence purpose like disaster) and so on. These policies are used for the decision
formula which will be mentioned in the later section.
The system layer collects type of coding and performs coding functions including
transcoding of a video coding to another one, such as from/to Motion JPEG to/from
MPEG1.
The network layer observes network conditions such as the value of PER,
throughput, delay, jitter, BER, electric field strength, and so on.
Network data from each layer is observed, and the decision making is held with
crossing through layers. Then, the message of prefer link or route is sent through
link0, and a reconfiguration procedure is acted at both sender and receiver nodes.

3 Methodology
As the previous research [8], the cognition cycle is introduced to our proposal
method. The cognition cycle is consisted of three stages; the observation stage, the
decision stage, and the acting stage. Each stage is continuously cycled in order to
perform link or route configuration.
A. Observation Stage
Network data is continuously observed through each layer at this observation stage.
Wireless network condition varies depending on the movement of nodes or radio
interference. Therefore, CWN needs to parse these stimuli to select the available
solution for providing the performance from user requests.
In this paper, our supposed system observes application types in order to decide
user policy which is depending on the specific services or media. Also, physical
characteristics like coding, PER, throughput, delay, jitter, BER, and electric field
strength are observed. Those parameters are used for understanding stimuli from user
environment.
B. Decision Stage
Decision making is held to maintain QoS in this decision stage. When the network
condition is changed, the proposal system will seek the suitable link and route by the
calculating from the values of network characteristics like user policy, throughput,
BER, and so on. We introduce AHP for the calculation of link, and extend AODV for
the decision of suitable route.
1) Link Selection
AHP is one of multi-attribute decision making and structured techniques for dealing
with complex decisions. It was developed by Thomas L. Saaty in the 1970s [9]. By
structuring a decision problem hierarchy, AHP provides us quantifications of its
elements and evaluations of alternative solutions.
For example, when the suitable link between neighbor nodes is solved by AHP, the
hierarchy of the problem is first structured. That is, goal (To decide suitable link of
wireless node), criteria (network characteristics such as delay, PER, throughput, and
172 N. Uchida, Y. Shibatta, and K. Takahata

F 3. Example of Hierarchy on AHP

Fig.

so on), and alternatives (w

wireless links such IEEE802.11b, IMT2000, and satelllite
network, and so on).
Then, paired comparisonns of criteria and alternatives are calculated for the prioority
value on AHP. Assume thaat we are given n as the number of criteria/alternatives, and
w1, w2, …, wn as the weig ght of each criteria/alternative, and paired comparisonn of
each element is express as aij = wi/wj. These paired comparisons are expressed as the
following paired comparisoon matrix A.
⎡ w1 w1 w1 ⎤
⎢w ... ...
wj wn ⎥
⎢ 1 ⎥
⎢ ... ... ... ⎥
⎢ wi wi wi ⎥ (1)
A=⎢ ... ...
w wj wn ⎥
⎢ 1 ⎥
⎢ ... ... ... ⎥
⎢ wn wn wn ⎥
⎢w ... ...
⎣ 1 wj w n ⎥⎦
We have multiplied A on th he right by the vector of weights w=(w1, w2, …, wn)T, and
plication is λw. That is λ is the eigenvalue of A.
then the result of this multip
⎡ w1 w1 w1 ⎤
⎢w ... ...
wj wn ⎥
⎢ 1 ⎥
⎢ ... ... ... ⎥ w
⎡ ⎤ ⎡ w1 ⎤ (2)
⎢ wi wi wi ⎥ ⎢ 1 ⎥ ⎢ ... ⎥ = λw
w= ⎢ ... ... = λ
wn ⎥⎥ ⎢ ⎥
Aw ... ⎢ ⎥
⎢ w1 wj
⎢ ⎥ ⎢⎣wn ⎥⎦
⎢ ... ... ... ⎥ ⎣wn ⎦
⎢ wn wn wn ⎥
⎢ ... ... ⎥
⎣ w1 wj wn ⎦

Then, priority is defined as;;

wi
priority = (3)
∑ wi
At this time, we introduce network
n data like the following in order to acquire netw
work
environment for the alternattives.
 User Policy Based Transmission Control Method in Cognitive Wireless Network 173

⎛ n －l ⎞
Si = ⎜⎜ i i ⎟⎟ ×10 (In case of throughput) (4)
⎝ ui－li ⎠
⎛ n －l ⎞
Si = ⎜⎜1 − i i ⎟⎟ ×10 ( In case of BER, PER, latency, etc) (5)
⎝ ui－li ⎠
In the above formula, Si is the weight of each alternative, ui and li are the upper and
lower limits of the alternativ
ve, and ni is the observed value from network.
We also adopt the valu ue of consistency index (C.I.). C.I. expresses the chaara-
cteristics of polynominal of A, and we accept the estimation of w if C.I. is less tthan
certain threshold.

C.I . = (λmax − n) (n − 1) (6)

In this paper, we propose th he calculation of AHP based on each user policy like viddeo,
VoIP, or connectivity for decision
d making of the suitable link. For example, in ccase
of emergency like disaster,, connectivity is more important as user policy. Thus, the
criteria like electric field strrength and BER are weighed and calculated by the form mula
(3). Then, priorities of alteernatives is calculated by (4) and (5), and the results are
inserted for (3). Finally, eaach value of alternative is calculated by alternative prioority
multiplied with criteria prio ory, and the link with largest value will be decided as the
best suitable link.
2) Route Selection
Our proposed system would change network route if the suitable link is not foundd or
minimum requirement for user is not satisfied. When a network route needss to
change, we introduced extend AODV for the decision of suitable route.
AODV is a routing pro otocol for mobile ad hoc networks (MANETs) and otther
wireless ad-hoc networks.[10][11] It is a reactive routing protocol, which means tthat
it establishes a route to a deestination only on demand base. Therefore, the connecttion
is slower than proactive routing
r protocols like OLSR or TBRPF. But AODV V is
superior when the network condition changes so often or changes so slowly. [7].

Fig. 4. RREQ and RREP

174 N. Uchida, Y. Shibata, and K. Takahata

AODV builds routes using a route request (RREQ) and route reply (RREP). When
a source node requires a route to a destination for which it does not have a route, the
source node broadcasts RREQ packets across the network. When the other nodes
receive those packets, and then update their routing information for the source node
and set up backwards pointers to the source node in the route tables.
The RREQ contains the source node's IP address, current sequence number,
broadcast ID, and the most recent sequence number for the destination node. A node
receiving the RREQ may send RREP packets if it is either the destination or it has a
route to the destination with corresponding sequence number greater than or equal to
that contained in the RREQ. If this is the case, RREP packets are sent back to the
source node with setting the routing tables relatively.
3) Extended AOCV
We propose a extended AODV protocol by adding the link values and network
conditions on each node for RREQ and RREP packets on each node from the source
to the destination. The link values of each node are calculated by AHP as the
previous sections, and those are set to the information of next node ID. Also, network
characteristics values of each node such as delay, PER, throughput, and so on are
added to RREQ and RREP packets. Then, a destination node receives all of the
RREQ packets from the possible routes during certain unit time. Then, possible routes
are compared to select the best route,
1) that can provide the maximum End-to-End throughput among all of the routes
for video service, or
2) that can provide the minimum End-to-End delay time among all of the routes for
VoIP, and
3) that can optimize the policy based AHP for Web service.
Through those comparisons, the best suitable route is decided by the policy of
transmission data.
We propose to add the link values by AHP for RREQ and RREP packets, and those
are set in the routing tables like Fig.5.

Fig. 5. Routing Tables by Extend AODV

C. Acting Stage
After the decision making of the suitable link or the route, a link or route will be
changed in the acting stage.
In the proposed system, a control wireless link is assumed to use for a transmission
of the control information. A control link is set to the lowest frequency bandwidth,
and a node sends the information with a suitable link to the next node through the
 User Policy Based Transmission Control Method in Cognitive Wireless Network 175

control link. Then, both nodes act the link change at the same time. A route change
also acts as the same way of a link change.
The cognition cycle goes back to the observation stage after the acting stage, and it
works to a link or route change relatively.

4 Simulation
In the simulation, ns2 (Network Simulator 2) was examined to evaluate the
effectiveness of our suggested method. IEEE802.11a, IEEE802.11b, IEEE802.11g is
used as the wireless links, and IMT2000 is used as the control link. The number of
wireless nodes is set to two, and every node has the same wireless links. Then, only a
source node moved with the same speed, and video data were transmitted from the
source node to the destination node. The simulation conditions are the following
table.

Table 1. Simulated Conditions

Item Simulation Content

Node0 moves from (10,5,5) to (300,5,5)
Nodes with 2m/s
Node1 is set to (10.5.5)
Anntenna Non-directional
320x240 MJPEG (15fps, 1/15 Compressed
Transmission Data
data) is send from node0 to node1.
IEEE802.11a/b/g. Each link has one
Link Interface
channel.
Space Free Space

Observed Parameters Signal Strength, Throughput, Jitter, PER

In this paper, a link change is especially focus on the simulation. First, the
simulation of IEEE802.11a, b and g is simulated under table1 conditions.
The Fig.6 shows the results of each node simulation. At Fig.6, throughput of
IEEE802.11a is quickly down about 18sec, and the link is disconnected about 23sec.
Also, IEEE802.11g is gradually down and disconnected about 145sec. IEEE802.11b
keeps link connection although throughput is not so high.
On the contrary, the results of our proposal method are showed in Fig.7.
Fig.7 shows that our proposed method keeps high throughput because the link
selection is effectively worked under video transmission scenario. The wireless link is
switched from IEEE802.11a to IEEE802.11g at 14 sec, and also switched IEEE802.
11g to IEEE802.11b at 132 sec. This results showed the proposed methods change the
best suitable link to maximize the transmission rate with user policy.
176 N. Uchida, Y. Shibata, and K. Takahata

Fig. 6. IEEE 802.11a/b/g Throughput

Fig. 7. Proposed Cognitive Wireless Network Throughput

5 Conclusions
In this paper, the transmission control methods in cognitive radio network
considering with cross layers including user policies are introduced. Compared with
giving fixed transmission rate, the our suggested methods can determines the
transmission rate dynamically to achieve the maximum throughput by changing
wireless links. The simulation result based on the video based data show that the
proposed methods change the best suitable link to maximize the transmission rate
with user policy. Therefore, the proposed method is shown in practical, efficient and
reliable to support data transmission.
 User Policy Based Transmission Control Method in Cognitive Wireless Network 177

References
[1] Chen, K.C., Peng, Y.J., Prasad, N., Liang, Y.C., Sun, S.: Cognitive radio network
architecture: part I – general structure. In: ICUIMC 2008: Proceedings of the 2nd
international conference on Ubiquitous information management and communication
(January 2008)
[2] Chen, K.C., Peng, Y.J., Prasad, N., Liang, Y.C., Sun, S.: Cognitive radio network
architecture: part II – trusted network layer structure. In: ICUIMC 2008: Proceedings of
the 2nd international conference on Ubiquitous information management and
communication (January 2008)
[3] Cordeiro, C., Challapali, K., Ghosh, M.: Cognitive PHY and MAC layers for dynamic
spectrum access and sharing of TV bands. In: TAPAS 2006: Proceedings of the first
international workshop on Technology and policy for accessing spectrum (August
2006)
[4] Weingart, T., Sicker, D.C., Grunwald, D.: Evaluation of cross-layer interactions for
reconfigurable radio platforms. In: TAPAS 2006: Proceedings of the first international
workshop on Technology and policy for accessing spectrum (August 2006)
[5] Kliazovich, D., Granelli, F.: Packet concatenation at the IP level for performance
enhancement in wireless local area networks. Wireless Networks 14(4) (2008)
[6] Renk, T., Kloeck, C., Burgkhardt, D., Jondral, F.K., Grandblaise, D., Gault, S., Dunat,
J.-C.: Bio-inspired algorithms for dynamic resource allocation in cognitive wireless
networks. In: Mobile Networks and Applications, October 2008, vol. 13(5) (2008)
[7] Staple, G., Werbach, K.: IEEE Spectrum: The End of Spectrum Scarcity,
http://www.spectrum.ieee.org/telecom/wireless/
the-end-of-spectrum-scarcity
[8] Sugimoto, T., Yamaguhi, S., Asatani, K.: A Proposal for System Selection Scheme
Using Multiple Handover Triggers in Heterogeneous Wireless Networks, Technical
Report of IEICE
[9] He, W., Nahrstedt, K., Liu, X.: End to End Delay Control of Multumedia Applications
over Multihop Wireless Links. ACM Transaction on Multimedia Computing,
Communications and Applications 5(2), Article 16 (November 2008)
[10] Perkins, C.E., Royer, E.M.: Ad hoc On-Demand Distance Vector Routing. In:
Proceedings of the 2nd IEEE Workshop on Mobile Computing Systems and
Applications, February 1999, pp. 90–100 (1999)
[11] Chakeres, I.D., Royer, E.M.: AODV Routing Protocol Implementation Design. In:
Proceedings of the International Workshop on Wireless Ad Hoc Networking (WWAN)
(March 2004)
 Development of a Hybrid Case-Based
Reasoning for Bankruptcy Prediction

Rong-Ho Lin1,* and Chun-Ling Chuang2

1
Department of Industrial Engineering and Management,
National Taipei University of Technology, Taipei 10608, Taiwan, R.O.C.
[email protected]
2
Department of Information Management, Kainan University,
Taoyuan, 33857, Taiwan, R.O.C.
[email protected]

Abstract. This paper aims to develop an integrated model of predicting busi-

ness failure, using business financial and non-financial factors to diagnose the
status of business, thereby providing useful references for business operation.
This study applied Rough Set Theory to extract key financial and non-financial
factors and Grey Relational Analysis (GRA) as the approach of assigning
weights. In addition, Case-Based Reasoning (CBR) are adopted to propose a
new hybrid models entitled RG-CBR (combining RST and CBR with GRA) to
compare the accuracy rates in predicting failure. After exploring the TEJ (Tai-
wan Economic Journal) database and conducting various experiments with
CBR, RST-CBR and RG-CBR the study finds CBR, RST-CBR and RG-CBR
reporting an accuracy rate in predicting business failure of 49.2%, 59.8% and
83.3%respectively. The RG-CBR boasts the highest accuracy rate while also
effectively reducing Type I and Type II error rates.

Keywords: decision analysis, data mining, business failure prediction, rough

set theory, case-based reasoning.

1 Introduction
Bankruptcy prediction is one of the most important issues influencing financial and
investment decision-making [1]. While affecting the entire life span of a corporation
or an organization, a business failure can even undermine social functions and agitate
economical structures [2]. Therefore, the evaluation of business failure has emerged
as a professional area prompting academics and professionals to develop optimal pre-
diction models based on their specific interest for practical applications in the busi-
ness community.
Methods developed, such as univariate statistical model, Multiple Discriminant
analysis, Linear Probability models, Logistic Regression, and Probit analysis [3][4],
are now generally regarded as obsolete statistical methods [5]. Statistical methods

*
Corresponding author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 178–188, 2010.
© Springer-Verlag Berlin Heidelberg 2010
 Development of a Hybrid Case-Based Reasoning for Bankruptcy Prediction 179

require a few assumptions and normally being constrained by the linearity, thus they
are quietly difficult to deal with the mass and complicated data [6]. To meet the var-
ied and restrictive assumptions, such as a large number of samples, normal distributed
independent variables, and linear relationship between all variables, scholars in re-
lated fields have come up with alternatives methods to predict the risk of business
failure, notably Artificial Neural Network (ANN), Case Based-Reasoning (CBR), and
Data Mining techniques [7][8][9][2].
Although reported to perform fairly well in terms of predictive accuracy, ANN
classifies every object by its computational characteristic and is therefore often criti-
cized as a “black box” approach for its lack of transparency [10]. In contrast to the
“black box” scenario, CBR, developed by Schank and Abelson [11] in the 1970s, uses
“similar” cases stored in a case base to build a solution to a new case based on the
adaptation of solutions to past similar cases [12]. Literature indicates that CBR has
been widely used in diagnosis domain and enhanced some of the deficiencies in statis-
tical models and ANN [13][14]. However, accuracy and effectiveness may be reduced
when CBR deals with too many attributes.
To solve the problem of excessive variables (attributes), Rough Set Theory (RST),
a data mining technique that has been successfully applied to solve classification
problems, can be adopted to reduce unnecessary attributes. Attributes, once identified,
need to be further prioritized by taking their weights into due consideration and this
task can be accomplished by performing Grey relational analysis (GRA) that com-
pares quantitative analysis to the development between every variable in the grey sys-
tem dynamically. The method describes the relation between the main factor (the
business status) and other variables in the grey system.
This study accordingly proposes a hybrid model for predicting business failure,
called RG-CBR, which incorporates three stages starting with applying RST to extract
key attributes, adopting GRA to obtain the weights of the key attributes, and culmi-
nating in feeding the retrieved key attributes and weights into CBR to improve the
model’s prediction accuracy.
Following the present section as the introduction, four major sections are incorpo-
rated in the paper. Section 2 outlines the backgrounds to RST, GRA, and CBR. Sec-
tion 3 explains the research methodology. Section 4 centers on experiment results and
model evaluations while Section 5 presents our conclusions.

2 Research Background

2.1 Rough Set Theory (RST)

RST, a powerful mathematical tool introduced by Pawlak [15], can discover facts
from imperfect data [16]. Studies showed that RST has played an important role in
coping with classification problems [9][2][7]. It also has been successfully applied to
real-world classification problems with the following advantages [7]: (a). Identifying
potentially important information contained in a large number of cases and reducing
the information into a model including a generalized description of the knowledge;
(b). Requiring no interpretation as the model is a set of easily understandable
180 R.-H. Lin and C.-L. Chuang

decision rules; and (c). Demanding no additional information, such as probability in

statistics.
Many researchers [7] [17][2] have used RST to construct a prediction model or
combined it with other methods as a data preprocessor. In particular, when RST fur-
ther integrated with other methods, such as the proposed hybrid model, the results
tend to be more accurate than those obtained by using RST alone [17][2].

2.2 Grey Relational Analysis (GRA)

Deng [18] pioneered a mathematical method called Grey Relational Analysis (GRA)
which is characterized by good performance in analyzing few data and many vari-
ables to examine the relationship among factors in observable systems and construct
the prediction model. GRA has been successfully used in a wide range of fields [19],
including agriculture, traffic, industrial engineering, education. Only a few studies,
however, have applied it to tackle business prediction problems.

2.3 Case-Based Reasoning (CBR)

An analogical reasoning method developed by Schank & Abelson [11] in the 1970s,
CBR can serve as a useful research paradigm for exploring complex problems [12] by
adapting past experiences to acquire potential solutions. As a flourishing problem-
solving method, CBR has been successfully applied to a wide spectrum of different
fields, such as medical diagnosis [20], bankruptcy prediction [8], fault diagnosis [21],
risk analysis [22], and marketing [2]. The study by Bryant [8] that adopts CBR to
bankruptcy prediction modeling in particular underlines the importance of selecting
proper variables in the successful application of the method. Complete reviews of CBR
with comprehensive details and applications for further works can be found in the
study of Watson & Marir [23].

3 Research Model Development

The study first applies CBR to analyze previous experiences of failed companies
and knowledge about business failure prediction. However, it is not easy to obtain
successful results with high prediction accuracy by applying CBR alone. The meas-
ures of success of a CBR system depend on its ability to index cases and retrieve
the most relevant ones in support of the solution to a new case. RG-CBR is accord-
ingly proposed to strengthen the capacity of CBR. The model integrates RST as a
data preprocessor to identify key attributes first and proceeds to compute and assign
different weights to the identified key attributes by using GRA as an alternative
method, and then culminate in feeding the retrieved key attributes and weights
into CBR. To evaluate the prediction accuracy of RG-CBR model, the Figure 1
shows the models of CBR, RST-CBR and RG-CBR in (a), (b), and (c), respectively,
and the study further compares its performance with those of CBR, and CBR com-
bined with RST. RG-CBR model comprises three stages each of which is introduced
as follows:
 Development of a Hybrid Case-Based Reasoning for Bankruptcy Prediction 181

Data Processing

RST Attribute reduction

GRA Attribute weight

CBR Failure diagnosis

(a) CBR (b) RST-CBR (c) RG-CBR

Fig. 1. Configuration of CBR, RST-CBR, and RG-CBR

3.1 RST

At the beginning, data processing is initiated to collect samples and establish criteria
for sampling and coding. RST is adopted to reduce attributes; some redundant attrib-
utes are eliminated without any information loss to induct the minimal variables sub-
sets named “reducts.” The following sections present some basic concepts of RST.
A finite set of objects U and a finite set of attributes A can be considered as a re-
lational system S = (U , A) . Each attribute a belongs to the considered set of attributes
A ( a ∈ A ); a (x ) represents the value of attribute a of object x. For every set of at-
tributes B∈A, an indiscernibility relation Ind (B ) is defined in the following way:

Ind ( B ) = {( x , x ) ∈UXU ; b ( x ) = b ( x ) ; b ∈ B}
i j i j (1)

The equivalence class of xi in relation to Ind (B ) is represented as [xi ]ind ( B ) . Let X

denote the subset of elements of U, the upper and lower approximations can be repre-
sented as
{ }
B ( X ) = xi ∈ U | [x ]Ind ( B ) ∩ X ≠ φ (2)

B( X ) = {xi ∈U | [x ]Ind (B ) ⊆ X } (3)

The boundary of X in U is defined as Bnd ( X ) = B ( X ) − B( X ) . If Bnd ( X ) is an empty

set, then the set X is definable with respect to B; otherwise, X will be referred to as a
rough set. “Reduct” and “core” are two fundamental RST concepts. Assume that the
182 R.-H. Lin and C.-L. Chuang

set of attributes R ⊆ A depends on the set of attributes B ⊆ A in S (denotation

BÆR) iff Ind(B) ⊆ Ind(R) . The minimal subset is R⊆B⊆ A such that μ B F = μ R F 1
is called F-reduct of B, which means a reduct is the minimal subset of attributes to
represent the whole set of attributes. In addition, S might have more than one F-
reduct. Intersection of all F-reducts is called F-core of B, i.e. COREB (F ) =
∩ REDB (F) . The core is a collection of the most significant attributes in the system.
In other words, the attributes in B are absolutely necessary attributes.

3.2 GRA Attribute Weighting

Grey relational analysis (GRA) can measure the degree of interrelationships between
one major sequence and another sequence in a given system [24]. The relational grade
is regarded as high if the two sequences tend towards concordance; otherwise, the
relational grade is regarded as low. Assume that the objective sequence is defined as

X 0 (k ) = [x0 (1), x0 (2),..., x0 (k )], k = 1,2,..., n (4)

The reference sequence can be represented as follows:

X i (k ) = [xi (1), xi (2),..., xi (k )], i = 1,2,...,l k = 1,2,..., n (5)

Given the objective sequence X 0 (k ) and the reference sequence X i (k ) with the
* *

normalized form, the grey relational coefficient r0i (k ) between them at any data
point k is defined as
Δ min + ρΔ max
r0i (k ) = . (6)
Δ 0i (k ) + ρΔ max

Where Δ 0 i (k ) = x0 (k ) − xi (k ) is termed as the relative grey relational space;

* *

Δ = x *0 (k ) − x *i (k ) (7)
min min min
i k

Δ max = max max x 0* (k ) − x i* (k ) (8)

i k

Δ min and Δ max are respectively the minimum and maximum distances for all factors
in all sequences. ρ is the distinguishing coefficient (0 ≤ ρ ≤ 1) , and usually
ρ = 0.5 . After the grey relational coefficient is obtained, the grey relational grade,
i.e. the mean of the coefficient, can be measured by

1
μB F and μ R F is called the quality of approximation of partition F by set of attributes B
and R.
 Development of a Hybrid Case-Based Reasoning for Bankruptcy Prediction 183

1 n
γ (X 0 , X i ) = ∑ r0i (k )
n k =1
(9)

3.3 CBR Failure Diagnosis

A case is a contextualized piece of knowledge representing an experience. It contains

past lessons that are the content of the case and the context in which the lessons can
be used [25]. It is difficult to share experience and knowledge for future reuse, and
CBR is developed as a problem solving technique by making use of past cases and
experiences to solve a new problem. It can retrieve a set of similar cases and identify
the most similar case for reuse to help solve the problem at hand. The process of CBR
includes case representation, case retrieval, and case adaptation. In particular, the
measures of success of a CBR system depend on its ability to index cases and retrieve
the most relevant ones in support of the solution to a new case.
A case should contain both content and context, typically composed of the prob-
lem, solution, and outcome [26]. Cases can be represented in different forms. In this
study, a structured case is defined as a vector of features: x = {x1 , x2 , K , xn } , where
n is the number of features. Given a description of a problem, a retrieval algorithm,
using the indices in the case-memory, needs to be adopted to retrieve the cases most
similar to the current problem or situation. The retrieval algorithm relies on the indi-
ces and the organization of the memory to direct the search to potentially useful cases.
Shin and Han [26] categorize case indexing into three types: nearest neighbor, induc-
tive, and knowledge-guided. Nearest neighbor is the most commonly used approach
[27].
The method is used to measure the similarity between two cases and known for its
easy application to numerical data such as financial ratios. In general, use of the near-
est neighbor method leads to the retrieval time increasing linearly with the number of
cases. Therefore, this approach is more effective when the case base is relatively
small. A typical algorithm for calculating nearest neighbor matching is defined by
[12] as:
n

∑ w × sim( f
i =1
i i
I
, fiR )
(10)
Similarity Score = n

∑w
i =1
i

where sim( f i , f i ) ∈ [0,1] is the similarity function, and f i and f i

I R I R
are the val-
ues for attribute i in the input and retrieved cases, respectively.

4 Experimental Design and Results

Based on information and data acquired from Taiwan Stock Exchange Corporation
(TSEC) and Taiwan Economic Journal (TEJ) database, the firms in Taiwan’s IT
industry judged as failed ones during the period from 1999-2006 are taken as the
184 R.-H. Lin and C.-L. Chuang

samples. A qualified sample, i.e. a failed firm, refers to a company with the transac-
tion of its listed securities placed under the “altered-trading method” category, sus-
pended or terminated. More specifically, a firm is cited as failed for (a). Suffering
from credit crisis, (b). Having net operating loss, (c). Failing to pay debts, or (d). Vio-
lating related regulations.
In order to improve the comprehensiveness and accuracy in predicting, this study
will concern simultaneously non-financial (qualitative) and financial (quantitative)
attributes and then apply them to diagnose the business status with different models.
21 attributes (13 financial ones and 8 non- financial) used in previous researches are
selected for our study and coded and described in Table 1.
The sample data sets consist of an equal number of healthy and failed companies
(321 companies in each category). Sample companies are grouped by binary assign-
ment into a decision class (healthy or failed, coded by 1 and 0, respectively). The data
sampling follows the “pairwised sampling method” first proposed by Beaver [3]. Each
data set is split into two subsets: a training set of 70% and a validation set of 30% of
the data, respectively. The training data are used as a case base, and the validation data
not used in constructing the case base are adopted to test the model. Five different
validation sets, each with six entries of data, are sampled and then used to validate the
model performance; that is, a total of 97 test samples are used for validation.
As RST is concerned with discrete values, the experimental study uses the software
ROSETTA that includes a method for discretizing continuous attributes called en-
tropy heuristic approach. The approach will subject continuous data to a process of
discretization.

Table 1. List of selected attributes

Attributes Description
A1 Return on assets (%)
A2 Return on equity (%)
A3 Net income (%)-except disposed
A4 Gross margin (%)
A5 Net income (%)
A6 Current ratio (%)
A7 Acid test ratio (%)
A8 Liabilities ratio (%)
A9 TCRI credit ranking
A10 Cash flow operation to current liabilities (%)
A11 Total equity growth ratio (%)
A12 Return on total assets growth ratio (%)
A13 Accounts receivable turnover (day)
A14 Inventory turnover (%)
A15 Earning per share
A16 Added value per person
A17 Manager-director
A18 Director and supervisor shareholding
A19 Inflation rate (%)
A20 Business cycle
A21 Rediscount Rate (%)
 Development of a Hybrid Case-Based Reasoning for Bankruptcy Prediction 185

Table 2. Measuring Data of each key attribute

D A6 A7 A8 A9 A16 A18
1 57.91 27.46 26.5 8 1289.5 19.41
1 32.57 21.25 80.2 9 590.4 12.37
1 63.69 54.94 100.55 10 1188.81 7.45
1 144.5 88.71 43.55 8 2494.46 8.59
1 80.98 35.69 78.36 9 181.06 11.4
1 143.61 85.9 71.51 10 47.41 24.49
1 249.79 86.01 85.27 10 619.21 35.29
1 71.01 51.43 82.89 10 1105.08 25.36
1 2.35 1.24 94.2 10 0.78 8.89
1 512.16 438.14 11.32 8 489.64 14.74
1 199.93 151.11 44.92 5 8.68 5.7
1 117.09 63.13 52.87 9 978.6 13.46
1 225.23 143.6 44.52 6 669.41 10.47
1 30.87 64.94 97.29 10 635.85 11.61
1 209.79 158.66 50.24 10 2017.14 14.75
1 298.29 242.31 18.77 5 424.59 35.36
1 135.29 128.63 63.15 7 243.78 8.88
1 133.45 121.01 55.77 8 1377.4 9.4
1 147.06 123.46 69.68 6 747.69 16.41
1 167.64 137.2 46.55 2 1852.75 13.61

The discernibility matrix can be used to find the minimal subset(s) of attributes
called reducts who have the same quality as the complete set of attributes. As a result,
we further apply RST to reduce the redundant attributes and to combine with CBR for
lowering misclassification rate. At the beginning, there were 78 reducts computed.
These reducts were employed from the total 21 attributes. According to the results
obtained from RST, the measuring data of each key attribute is shown in Table 2. The
weights of those key attributes are developed by GRA in Excel, as described in Table
3 which shows the grey relational coefficient of each attribute adjusted its weight. The
six attributes are A6 (Current ratio), A7 (Acid test ratio), A8 (Liabilities ratio), A9
(TCRI credit ranking), A16 (Added value per person) and A18 (Director and supervi-
sor shareholding).

Table 3. Key attributes with GRA-weights

Attributes Description Weight

A6 Current ratio 0.032
A7 Acid ratio 0.048
A8 Liability ratio 0.333
A9 TCRI credit ranking 0.529
A16 Added value per person 0.005
A18 Director and supervisor shareholding 0.053
total 1.00
186 R.-H. Lin and C.-L. Chuang

The six attributes are ranked based on their weights into the following order: TCRI
credit ranking, Liabilities ratio, Director and supervisor share-holding, Acid test ratio,
Current ratio, and Added value per person. An attribute is ignored during similarity
computation if it is found to have zero weight, whereas an attribute with a higher
weight exerts a greater impact on determining similarity. As indicated above, TCRI
credit ranking emerges as the most significant attribute influencing business failure
prediction with a highest weight of 0.529.
Identifying key attributes is a sub-process of investigating the important attributes
of factors critical for identifying analogous cases and of predicting the value of the
target variable. RST for identifying key features and GRA for ranking key features by
their weights are applied to compensate the weaknesses of CBR. And the enhanced
CBR forms the core of the proposed RG-CBR model. The model then proceeds to
search in the case base the case that best matches with the target case and to adapt the
retrieved case to fit the problem and produce the solution for the problem. In the ex-
periment, the overall predictive accuracy of the validation data emerges to be 83.3%
as shown in Table 4.
To evaluate the prediction accuracy of RG-CBR model, Table 4 compares the re-
sults of the three different models of RG-CBR, CBR, and RST-CBR (combining RST
and CBR with equal weights). Among the three models, RG-CBR shows the highest
level of accuracy (83.3%) in the given validation data. Therefore, the result concludes
that RG-CBR outperforms CBR and RST-CBR and testifies to its qualification to
serve as a promising alternative for business failure prediction.

Table 4. Performance Comparison of four different Models

Model Accuracy (%) Type I error (%) Type II error (%)

CBR 49.2 20 30.8
RST-CBR 59.8 20 20.2
RG-CBR 83.3** 20.2 13.3
** signifies the best performance.

5 Conclusions
CBR, as indicated in the literature, has been widely used in diagnosis domain by re-
trieving most similar cases more accurately and effectively. The accuracy and effec-
tiveness, however, may be reduced when there are too many attributes. It is therefore
not easy to obtain successful results by adopting CBR alone to predict business fail-
ure. In order to enhance the performance of the CBR system, this study proposes a
hybrid model, RG-CBR, for predicting business failure. RG-CBR can refer the user to
the similar cases as references for making decisions, and the attributes retrieved and
assigned different weights by mathematical methods (RST with GRA) are more ob-
jective than those decided by experts’ subjective experiences. As suggested by the
results of the experiment, RG-CBR outperforms other comparative algorithms such as
CBR, and RST-CBR in terms of accuracy and serves a promising alternative for busi-
ness failure prediction.
 Development of a Hybrid Case-Based Reasoning for Bankruptcy Prediction 187

References
[1] O’Leary, D.E.: Using neural networks to predict corporate failure. International Journal
of Intelligent Systems in Accounting, Finance and Management 7, 187–197 (1998)
[2] Ahn, B.S., Cho, S.S., Kim, C.Y.: The integrated methodology of rough set theory and ar-
tificial neural network for business failure prediction. Expert Systems with Applica-
tions 18, 65–74 (2000)
[3] Beaver, W.H.: Financial ratios as predictors of failure- empirical research in accounting:
selected studies. Journal of Accounting Research 4, 71–111 (1966)
[4] Ohlson, J.A.: Financial ratios and the probabilistic prediction of bankruptcy. Journal of
Accounting Research 18, 109–131 (1980)
[5] Balcaen, S., Ooghe, H.: 35 years of studies on business failure an overview. The British
Accounting Review 38, 63–93 (2006)
[6] Akahoshi, M., Amasoki, Y., Soda, M.: Correlation between fatty liver and coronary risk
factors: a population study of elderly men and women in Nagasaki, Japan. Hypertens Re-
search 24, 337–343 (2001)
[7] Slowinski, R., Zopounidis, C.: Application of the rough set approach to evaluation of
bankruptcy risk. International Journal of Intelligent Systems in Accounting, Finance and
Management 4, 27–41 (1995)
[8] Bryant, S.M.: A case-based reasoning approach to bankruptcy prediction Modeling. In-
ternational Journal of Intelligent Systems in Accounting, Financial and Management 6,
195–214 (1997)
[9] McKee, T.E.: Developing a bankruptcy prediction model via rough set theory. Interna-
tional Journal of Intelligent Systems in Accounting, Finance and Management 9, 159–
173 (2000)
[10] Kumar, P.R., Ravi, V.: Bankruptcy prediction in banks and firms via statistical and intel-
ligent techniques – A review. European Journal of Operational Research 180, 1–28
(2007)
[11] Schank, R. (ed.): Dynamic Memory: A Theory of Learning in Computers and People.
Cambridge University Press, New York (1982)
[12] Kolodner, J.: Improving human decision making through case-based decision aiding. AI
Magazine 12, 52–68 (1991)
[13] Barletta, R.: An introduction to case-based reasoning. AI Expert 6, 42–49 (1991)
[14] Park, T.S., Han, J.Y.: Derivation and characterization of pluripotent embryonic germ
cells in chicken. Molecular Reproduction and Development 56, 475–482 (2000)
[15] Pawlak, Z.: Rough sets. International Journal of Information and Computer Sciences 11,
341–356 (1982)
[16] Walczak, B., Massart, D.L.: Rough sets theory. Chemometrics and Intelligent Laboratory
Systems 47, 1–16 (1999)
[17] McKee, T.E., Lensberg, T.: Genetic programming and rough sets: A hybrid approach to
bankruptcy classification. European Journal of Operational Research 138, 436–451 (2002)
[18] Deng, J.: Control problems of grey systems. System and Control Letters 1, 288–294
(1982)
[19] Lai, H.H., Lin, Y.C., Yeh, C.H.: Form design of product image using grey relational
analysis and neural network models. Computers & Operations Research 32, 2689–2711
(2005)
[20] Varma, A., Roddy, N.: ICARUS: Design and deployment of a case-based reasoning sys-
tem for locomotive diagnostics. Engineering Applications of Artificial Intelligence 12,
681–690 (1999)
188 R.-H. Lin and C.-L. Chuang

[21] Yang, B.S., Han, T.H., Kim, Y.S.: Integration of art-kohonen neural network and case-
based reasoning for intelligent fault diagnosis. Expert Systems with Applications 26,
387–395 (2004)
[22] Jung, C., Han, I., Suh, B.: Risk analysis for electronic commerce using case-based rea-
soning. International Journal of Intelligent Systems in Accounting, Finance and Man-
agement 8, 61–73 (1999)
[23] Watson, I., Marir, F.: Case-based reasoning: A review. The Knowledge Engineering Re-
view 9, 327–354 (1994)
[24] Hsu, Y.T., Chen, C.M.: A novel fuzzy logic system based on N-version programming.
IEEE Transactions on Fuzzy Systems 8, 155–170 (2000)
[25] Kolodner, J.: Case-based reasoning. Morgan Kaufmann, San Mateo (1993)
[26] Shin, K.S., Han, I.: A case-based approach using inductive indexing for corporate bond
rating. Decision Support Systems 32, 41–52 (2001)
[27] Burke, E.K., MacCarthy, B., Petrovic, S., Qu, R.: Structured cases in case-based reason-
ing-re-using and adapting cases for timetabling problems. Knowledge-Based Systems 13,
159–165 (2000)
 The Time Series Image Analysis of the HeLa Cell Using
Viscous Fluid Registration

Soichiro Tokuhisa and Kunihiko Kaneko

Department of Intelligent Systems, Graduate School of Information Science and Electrical

Engineering, Kyushu University
744 Motooka, Nishi, Fukuoka, Japan
{tokuhisa,kaneko}@db.is.kyushu-u.ac.jp

Abstract. Optical microscopy image analysis is important in the life science re-
search. To obtain the motion of the cell, we use the viscous fluid registration
method based on fluid dynamics. Viscous fluid registration deforms an image at
time t to the next image at time t+1. In this algorithm, there is a problem that an
object cannot be divided into two. In other words, the divided objects from one
are connected by thin line because the velocity field on the connected thin line
is zero. To solve this problem, we suggest a new viscous fluid registration algo-
rithm for the object division. This algorithm is only added similarity maximiza-
tion step to correct the displacement in the near pixels in the original viscous
fluid registration. Using this method, one object is divided into two, and divided
objects are not connected. We experiment the anaphase detection based on a
nucleus identification using laser scanning microscope HeLa cell images. Ex-
perimental result shows that 74 in 76 cells are tracking well and 6 cells in the
anaphase are detected. In three scenes in the cell division which can not be
divided into two using original viscous fluid registration, suggested algorithm
can be divided into two cells completely.

Keywords: HeLa cell, cell division, cell tracking, viscous fluid registration.

1 Introduction
Microscopy image analysis is important in the life science research. Cell phase identi-
fication and analysis of the cell features are attracted. The abnormal cell detection
method helps early detection of a cancer, diabetes, and the gene deficit disease.
Therefore early treatment for disease is enabled and helps medical care.
In the field of the cell analysis, various image processing methods are performed in
cell segmentation and the phase identification. For example, it is suggested the accu-
rate automated method that cell segmentation using watershed algorithm and the cell
phase identification using Markov model by Zhou et al. [1] and Chen et al. [2]. Dzyu-
bachyk et al. suggested the multiple level-set framework that each cell have each
level-set function [3].
We think that the cell movement can be expressed by the viscous fluid, so the cell
movement can be described as the time series image matching technique. We use the

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 189–200, 2010.
© Springer-Verlag Berlin Heidelberg 2010
190 S. Tokuhisa and K. Kaneko

image registration method applied in the medical image processing such as the CT
image and MRI image. Though this method has a large calculation cost, it is obtained
the smooth displacement because it is based on the fluid dynamics. The displacement
contains vectors xt – xt-1. It means that a pixel at position xt at time t is moved from a
pixel at position xt-1 at time t-1.
There are some kinds in a microscopy images such as an optical microscopy image
and confocal laser scanning microscope. The three-dimensional cell image can be
constructed using confocal laser scanning microscope. In this experiment, we use
these cell images in Fig.1 that are obtained by live cell imaging [4] of the HeLa cell
with mCherry-NLS[5] via the confocal laser scanning microscope created by Yuki
Tsujimura in RIKEN.
When an object is divided into two objects, viscous fluid registration cannot to be
separate correctly. So, we suggest the new algorithm that is added one calculation step
in the viscous fluid registration which step is the similarity maximization step to cor-
rect the displacement in the near pixels.
When this method is used, we can divide an object to two objects completely. Us-
ing viscous fluid registration with similarity maximization step, it will be performed a
smooth cell movement analysis. The displacement can be use for the time series im-
age analysis like Fig.2. The displacement is useful for anaphase detection because the
corresponding time series cell at time t and cell at time t+1 can be detected. The cell
in anaphase is defined the first time of the cells called daughter cell which was di-
vided one cell into two in this article. From the cell identification result, it can be
confirmed the movement of the daughter cells after the cell division. Cell identifica-
tion result shows that 74 in 76 cell in 17 images is success in tracking.

t =1 t =2 t =3 t =4 t =5 t =6

t =7 t =8 t =9 t =10 t =11 t =12

t =13 t =14 t =15 t =16 t =17

Fig. 1. Time series image data created by Yuki Tsujimura in RIKEN

 The Time Series Image Analysis of the HeLa Cell Using Viscous Fluid Registration 191

a cell is divided into two cells

Right nucleus is
moved to right
displacement
analysis

Left nucleus settled Center of the nucleus

on the place moved for a cell division

time01 time02 Overlapped image

Fig. 2. Example of the time series image analysis. The displacement calculated by viscous fluid
registration shows the white arrow in right image named “Overlapped image”. These arrows
show a pixel in time01 is moved to a pixel in time02.

2 Viscous Fluid Registration

Viscous fluid registration is the non-rigid image registration method [6] to deformed
one image to the other image based on viscous fluid dynamics. It is used for adjust-
ment the shape and appearance of two medical images. When this application imple-
mented by C++ language is given a template image, a study image and parameters as
input data, it put the output files containing a transformed image and a total displace-
ment file. The deformed image is deformed a template image into a study image.
Viscosity fluid registration is calculated three fields the number of repetition times. A
force field, a velocity field and a displacement are included in three fields (Fig.3). The
number of repetition is defined in an input parameter. In addition, a multi-resolution
method is implemented for speedup of calculation and robustness of a registration.
Using the example image set that are sample01 (contains in a black square) and sam-
ple02 (contains in a white “C”), sample01 is deformed the shape, and the application
put the deformed image that is resembles the shape of “C” (Fig.4). The parameters are
viscosity=20, rMin=1, rMax=8, calculation number=100. The parameters of the im-
age of the cell are viscosity=60, rMin=1, rMax=8, calculation number=10. There is
the explanation of the parameters to 2.1 and 2.2, and we explain the viscous fluid
registration algorithm in the section 2.1. In the section 2.2, we explain the two-
dimensional viscous image registration workflow using multi-resolution method.

force field velocity field displacement

Sample01

Sample02

Fig. 3. Example of a force field, a velocity field and a displacement calculated first time
192 S. Tokuhisa and K. Kaneko

sample01 sample02
viscosity 20
rMIN 1
rMAX 8
calculation num.100
Overlapped image Deformed sample01
parameter
input Under calculation Output

Fig. 4. Example of the viscous fluid registration

01+02 02+03 03+04 04+05 05+06 06+07

07+08 08+09 09+10 10+11 11+12 12+13

Overlapped image
Orange: current frame
Purple: next frame
White allows: displacement

13+14 14+15 15+16 16+17

Fig. 5. Overlapped current frame image (orange) on next frame image (purple)

2.1 Viscous Fluid Registration Algorithm

Viscous fluid registration is the image registration method to align two images by
deforming one image to the other [7] [8]. A viscous fluid registration is based on the
theory of a viscous fluid dynamics. This algorithm contains the calculation of three
vector fields: a force field, a velocity field, and a displacement. These vector fields are
a set of two-dimensional vectors that are x-axis and y-axis values because we perform
two-dimensional viscous fluid registration for all images. We use sum of squared
difference as an image similarity measure. Sum of squared difference is used to calcu-
late the force field. The force field is equal to the derivative of image similarity meas-
ure. The force field equation can be written as follow:
f ( x , u( x, t )) = − [T ( x - u( x, t )) − S ( x )]∇T ( x - u( x, t )) (1)
 The Time Series Image Analysis of the HeLa Cell Using Viscous Fluid Registration 193

where f(x, u(x, t)) is the force field at point x and at time t that depends on the dis-
placement u(x, t),. T(x) is the pixel value at point x on template image T, and S(x) is
the pixel value at point x on study image S. ∇T is calculated by sobel filter. The
velocity field is equal to the convolution of the force field with the Gaussian kernel.
The velocity field v(x,u(x,t)) can be written as follow:
v ( x , u( x, t ) ) = Gσ (f( x, u( x, t )) (2)

The Gaussian kernel Gσ has two parameters: window size w and standard deviation σ.
We define this standard deviation σ as the viscosity. The value of the viscosity deter-
mines the degree of freedom of the deformation to limit the movement of the pixel at
point x. The window size w is defined as follow:
6σ
w= (3)
r
where r is the resolution level that allows for reducing the image size. If the multi-
resolution method is not used, the resolution level is set to 1. Using this window size
w, the velocity at x is convolved in 99.74% force field affected to the viscosity at x
according to Gaussian distribution. Displacement u(x, t) is calculated as follow:
u( x, t i +1 ) = u( x, t i ) − (t i +1 − t i ) (I − ∇u (x, t i )) v(x, t i ) (4)
where the velocity field v(x,u(x,t)) is equal to the Lagrange derivative of displace-
ment u(x, t). Matrix I is the identity matrix. The time (ti+1−ti) is calculated by the
following equation:
MAX (| v ( x , t i ) |) (t i +1 − t i ) = du max ⋅ r (5)
where dumax is the maximal movement allowed in one iteration. We set dumax as a
fixed value of 0.7 (pixel). If |I – ∇ u(x, t)| is less than 0.5, the displacement u(x, t) is
applied to the image, and the displacement u(x, t) and time t are initially set to 0,
because the displacement is singular for large curved deformation. The total dis-
placement is calculated as follow:
utotal ( x, t i ) = u( x, t i ) + utotal ( x − u( x, t i ), t i -1 ) (6)

2.2 Two-Dimensional Viscous Fluid Registration

Viscous fluid registration algorithm requires four input parameters: the maximum
resolution level rMAX, the minimum resolution level rMin, the calculation number k,
and the viscosity σ. We use a multi-resolution method that enables fast and robust
image registration. In the multi-resolution method, image Ir in the current resolution
level r is used. Image Ir is reduced in size by resolution level r. If resolution level r is
2, the image size is reduced by half. Image Ir is defined as follows:
I r ( x r , y r ) = I 1 (r ⋅ x r , r ⋅ y r ) (7)

　　
1≤ x r ≤
width
r
　
1 ≤ yr ≤
height
r
(8)
194 S. Tokuhisa and K. Kaneko

where xr and yr are array numbers and they are constrained by (9). The current resolu-
tion level r is taken as 1, 2, 4, 8,…,2n. The values of the width and height are obtained
by the image width and image height at resolution level 1.
The viscous fluid registration method is calculated three vector field. In the first
step, it calculates the force field value using data from two images. In the second step,
it calculates the velocity field using the force field. In the third step, it calculates the
displacement using the velocity field. When these three steps are finished, it incre-
ments the iteration number and repeats the three steps. If the iteration number is equal
to the calculation number, the resolution level is divided by 2, and the iteration num-
ber is initialized to 0. After initialization, the three calculation steps are repeated. The
resolution level indicates the scale of the image. For example, if the resolution level is
1, it indicates the original image, and if the resolution parameter is 4, it indicates that
an image size is quarter. At the beginning of a calculation, the resolution level is set to
the maximum resolution level. If the resolution level is equal to the minimum resolu-
tion level, it generates output data, including the total displacement and the image
deformed using displacement.

2.3 Similarity Maximization Step

To divide one object into two, we suggest adding the similarity maximization step in
viscous fluid registration. To confirm this technique, we used two image set in Fig.6,
left image set is shown that one circle is divided into two circles, and right image set is
shown that one square is divided into two squares. Using original viscous fluid regis-
tration, two objects are connected by the thin line (see 40calculation in Fig.7). The
force field over and under the line have in the opposite direction (see Fig.8). Because
the velocity is calculated by the convolution of the force field, the velocity field on the
line is calculated to zero. And the displacement is in proportion to the velocity, dis-
placement on the connected line is zero. After all calculation, there are gaps between
the displacement on the connected line and over and under the line. Therefore I intro-
duce a similarity maximization step to approximate the displacement at the point x by
near the point x. The similarity maximization step is described as follow:
u total ( x, t i ) = MIN{(T (x - u( x, t )) − S (x) ) | x − 1 ≤ x ≤ x + 1}
2
(9)

This equation maximizes an image similarity measure by minimizing the square error
of the difference. This step is added after the calculation of the total displacement utotal
in (6). Using this step, deformed result is shown in Fig.9, and the deformed grid im-
age is shown in Fig.10. Fig.10 shows that image edge is enhanced in circle grid image
using the new step. This means that fluency of the motion vector is decreased, and the
issue of connected line was solved.

Fig. 6. two image set. Left is circle division example. Right is square division example
 The Time Series Image Analysis of the HeLa Cell Using Viscous Fluid Registration 195

䊶䊶䊶

Connected line

1calculation 4calculation 8calculation 40calculation

1calculation 4calculation 8calculation 40calculation

Fig. 7. Cconnected line is not go disappear using original viscous fluid registration. Top images
are circle image result. Bottom images are a square image result.

Fig. 8. Force field, velocity field, and displacement at 7calculation in circle image

Fig. 9. Disappear connected line adding Similarity maximization step in registration. Top line
images are circle image result. center and bottom line images are sqare image result.
196 S. Tokuhisa and K. Kaneko

Fig. 10. Deformed grid image. Left two images are deformed circle grid result. Right two im-
ages are deformed square grid result. In these set, left is not using step result, right is using step
result.

3 Experiment
We experiment the detection of the anaphase in the time series HeLa cell images.
Anaphase detection consists of the three steps: image processing, viscous fluid regis-
tration and cell identification. We use 17 bitmap images in Fig.1. These images are
350*350 size and 8bit grayscale data created by Yuki Tsujimura in RIKEN. These
images contain 76 cells. The resolution of the image is 0.286um in x and y direction
and time span between images is 1 hour.
At first we generate a label image by image processing. This step is described in
3.1. In the next step, we perform viscous fluid registration time series images. This
second step is described in 3.2. Finally, we perform time to time cell identification by
using of the displacement calculated by viscous fluid registration. If corresponding
cell is not found after the registration, these cells make pair of a nearest cell in the
previous time image. After three steps, we make cell identification in all time, so we
can detect a cell in anaphase because these cells are divided into two. This last step
and anaphase detection are described in 3.3.

3.1 Image Processing

For the cell identification, we generate a label image from grayscale image using
hybrid image processing. This image processing is implemented by the octave lan-
guage. This method contains five processes: thresholding, median filter, erosion,
labeling and dilation(Fig.11). Thresholding process make binary image from gray-
scale image. Median filter process removes white noise in the cell from binary image.
Erosion process makes the cell area shrink. Labeling process add label data on the
cell. Dilation process makes the cell area inflated. Dilation and erosion is used for
division cell because some cells are touching together. After labeling, label image is
displayed by a color using color table. So, label image have pixel values such as 0 on
the background and 1, 2, 3… on each independent cell. In this process, 17 label im-
ages are created. Regulating parameters at each time are described in Table 1.

Table 1. Parameters list in Image processing at each time

time 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
threshold 30 30 30 30 30 30 30 50 70 70 70 30 30 30 30 60 60
erosion and diration
repeat numver 0 0 0 0 0 0 0 5 6 5 5 0 0 0 0 0 0
 The Time Series Image Analysis of the HeLa Cell Using Viscous Fluid Registration 197

time11Thresholding Median filter Erosion(3x3) Labeling Dilation(3x3)

(9x9)
5 times 5 times

Fig. 11. Image processing flow example at time 10

3.2 Viscous Fluid Registration

In this step, we input two time series images into viscous fluid registration to get the
displacement. The label images are not used in this step. The result of the original
viscous fluid registration is shown in Fig.12, and the result of the suggested viscous
fluid registration added the similarity maximization step is shown in Fig.13. We can
find that a cell is divided obviously in suggested technique. As for this technique,
limitation of the transformation becomes lax, so the deformed cell shape almost same
as input study image.
By the next step, a label image in Fig.11 is deformed using the displacement, and
determined the correspondence of the pair of the cells in time t and time t+1 image. If
the cells in time t and time t+1 image exist in the same position, these cells are treated
as the correspondence of the pair. And when cell division occurs, cell at time t is
divided into two cells at time t+1, so we can distinguish the cells in Anaphase by
checking label numbers are same value or not.

deformed t=1 image deformed t=7 image deformed t=9 image

Fig. 12. Original viscous fluid registration result when cell division is occurred

deformed t=1 image deformed t=7 image deformed t=9 image

Fig. 13. Suggested viscous fluid registration result when cell division is occurred
198 S. Tokuhisa and K. Kaneko

3.3 Cell Identification and Anaphase Detection

In this step, cell identification and anaphase detection is performed. The label images
and the displacement is used for cell identification. Current time label of cell is taken
the correspondence of the next time label. The correspondence of the label is found
from the relations of a cell position between deformed current time label and next
time label. We calculate histogram of deformed current time label image and next
time label image. And next time label value is changed the value of the most over-
lapped label from deformed current time label image (Fig.14 left). Most overlapped
label is found from the histogram. If there are same labels in the next time label im-
age, it found that anaphase is detected. Once anaphase is detected in one label value,

Labeling if registration is success Labeling if registration is fault

If same label is appeared,
this frame is anaphase correct using nearest
label is neighbor label
corrected to the
most overlapped correspondence
correct label
is background correct

next time Corrected

label next time Corrected
nect time
label nect time
label label

current deformed current deformed

time label current time label time label current time label

Fig. 14. Procedure of a correcting label

Anaphase

time01 time02 tim time04 time05 time06

Anaphase
Anaphase

time07 time08 time09 time10 time11 time12

time13 time14 time15 time16 time17

Fig. 15. Result of a detection of anaphase. Dividing cell in anaphase is shown in the red circle
 The Time Series Image Analysis of the HeLa Cell Using Viscous Fluid Registration 199

create flag on this label value. This mean anaphase is already found on this label. So,
if flag is created, anaphase is not detected on this label. There is the case that registra-
tion is fault because the next time cell and deformed cell is not overlapped well. In
other words, the corrected label value candidate has the background value (Fig.14
right). When a label candidate became the background, nearest neighbor label is
given. In 17 images, two cells are detected by nearest neighbors. This nearest
neighbor process is adapted only center of the image. If the center of gravity of a label
exists on the area of 40 pixels from boundary of the image, this cell is not regard as in
anaphase because this cell moves image boundary to inside and outside. Fig.15 shows
the result of corrected all label images.

4 Time Series Image Analysis

To observe the movement of the cell division, all label images are overlapped adapted
each image in transparent percentage of 1/17, except for the cell which appeared first
is transparency 50% and the cell boundary displays in white and which appeared last
time is transparency 0% and the cell boundary displays in black. From this image, it
can be observed that three cells are divided into two, and daughter cells are moved to
different direction.

Fig. 16. Overlapped time series label images. This image shows the dividing nucleus move-
ment. The boundary line of nucleus that appears the first is white and that appear the last time is
black. A cell color from the second and before last nucleus is semitransparent. Dark blue nu-
cleus is divided and daughter nuclei are moved to left and right (yellow allow), green daughter
ones are moved to top and right-top (red allow), and blue ones are moved to top and bottom
(purple allow).

5 Conclusion
We can find that divided objects are connected by a thin line using original viscous
fluid registration. So we suggest a new viscous fluid registration algorithm for an
object division. This algorithm is to add new step in original viscous fluid registration.
200 S. Tokuhisa and K. Kaneko

We confirmed that connected line which appeared in original viscous fluid registra-
tion result is disappeared by adding new step in two sample image set and time series
HeLa cell images. We examine the anaphase detection to confirm that new algorithm
is worked well. By creating label image and deforming label image using displace-
ment, the cell identification is success 74 in 76 cells. 2 of the remainder cell can be
tracked by nearest neighbor processing. A cell in anaphase is detected be cell identifi-
cation, and daughter cell movement is observed by the overlapped time series label
image. Displacement calculated by viscous fluid registration is useful in various ways.
Using displacement, we can observe cell movement like Fig.2.

References
1. Zhou, X., Li, F., Yan, J., Wong, S.T.: A novel cell segmentation method and cell phase
identification using Markov model. IEEE transactions on information technology in bio-
medicine 13, 152–157 (2009)
2. Chen, X., Zhou, X., Wong, S.T.: Automated segmentation, classification, and tracking of
cancer cell nuclei in time-lapse microscopy. IEEE Transactions on Biomedical Engineer-
ing 53, 762–766 (2006)
3. Dzyubachyk, O., Niessen, W., Meijering, E.: Advanced level-set based multiple-cell seg-
mentation and tracking in time-lapse fluorescence microscopy images. Biomedical Imag-
ing 21, 185–188 (2008)
4. Goldman, R.D., Spector, D.L.: Live Cell Imaging: A Laboratory Manual. Cold Spring Har-
bor Laboratory Press (2005)
5. Shu, X., Shaner, N.C., Tarbrough, C.A., Tsien, R.Y., Remington, S.J.: Novel Chromophores
and Buried Charges Control Color in mFruits. Biochemistry 45, 9639–9647 (2006)
6. Zitova, B., Flusser, J.: Image registration methods: a survey. Image and Vision Comput-
ing 21, 977–1000 (2003)
7. D’Agostino, E., Maes, F., Vandermeulen, D., Suetens, P.: A Viscous Fluid Model for Mul-
timodal Non-rigid Image Registration Using Mutual Information. In: Dohi, T., Kikinis, R.
(eds.) MICCAI 2002. LNCS, vol. 2489, pp. 541–548. Springer, Heidelberg (2002)
8. Bro-Nielsen, M., Gramkow, C.: Fast fluid registration of medical images. Visualization in
Biomedical Computing, 267–276 (1996)
 Matrices Representation of Multi Soft-Sets and
Its Application

Tutut Herawan1,2, Mustafa Mat Deris1, and Jemal H. Abawajy3

1
Faculty of Information Technology and Multimedia
Universiti Tun Hussein Onn Malaysia, Johor, Malaysia
2
Department of Mathematics Education
Universitas Ahmad Dahlan, Yogyakarta, Indonesia
3
School of Engineering and Information Technology
Deakin University, Geelong, VIC, Australia
[email protected], [email protected], [email protected]

Abstract. In previous paper, we introduced a concept of multi-soft sets and

used it for finding reducts. However, the comparison of the proposed reduct has
not been presented yet, especially with rough-set based reduct. In this paper, we
present matrices representation of multi-soft sets. We define AND and OR
operations on a collection of such matrices and apply it for finding reducts and
core of attributes in a multi-valued information system. Finally, we prove that
our proposed technique for reduct is equivalent to Pawlak’s rough reduct.

Keywords: Multi-valued information system; Multi-soft sets, Matrices

representation, Reducts and core of attributes.

1 Introduction
Soft set theory [1], proposed by Molodtsov in 1999, is a new general method for
dealing with uncertain data. In recent years, research on soft set theory has been
active, and great progress has been achieved, including the works of theoretical soft
set, soft set theory in abstract algebra, parameterization reduction, decision
making and forecasting. Let S = (U , A, V , f ) be an information system as in [2].
The “standard” soft set deals with a binary-valued information system
S = (U , A, V{0,1} , f ) . For a multi valued information system, we introduced a concept of
multi soft-sets [3]. The idea is based on a decomposition of a multi-valued
information system S = (U , A,V , f ) , into A number of binary-valued information
systems S = (U , A,V{0,1} , f ) , where A denotes the cardinality of A. Consequently, the
A binary-valued information systems define multi-soft sets, denoted by
(F , A) = {(F , a i ) : 1 ≤ i ≤ A } . In [4], we used the concept of multi-soft sets and AND
operation for finding reducts in a multi-valued information system. However, the
comparison of the proposed reduct has not been presented yet, especially with rough-
set based reduct [2,5]. In this paper, we present the notion of matrices representation
of multi-soft sets. We present a definition of AND and OR operations on a collection

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 201–214, 2010.
© Springer-Verlag Berlin Heidelberg 2010
202 T. Herawan, M.M. Deris, and J.H. Abawajy

of such matrices. Further, we apply the notion of AND operation for finding reducts
and core attributes in a multi-valued information system. We prove that the proposed
technique of reduct is equivalent with Pawlak’s rough reduct.
The rest of this paper is organized as follows. Section 2 describes the notion of
information system. Section 3 describes fundamental concept of soft set theory.
Section 4 describes multi soft sets construction in a multi-valued information system.
Section 5 describes matrices representation of multi-soft sets, AND and OR operations,
as well a simple example. Section 6 describes applications of matrices representation
of multi-soft sets for finding reducts and core of attributes. Further, we prove that our
proposed technique for reduction is equivalent to Pawlak’s rough reduction. Finally,
we conclude and describe future activities of our works in section 7.

2 Information System
Data are often presented as a table, columns of which are labeled by attributes, rows
by objects of interest and entries of the table are attribute values. By an information
system, we mean a 4-tuple (quadruple) S = (U , A,V , f ) , where U = u1 , u 2 , u 3 ,L, u U { } is
{
a non-empty finite set of objects, A = a1 , a 2 , a3 ,L, a A } is a non-empty finite set of
attributes, V = U V , V is the domain (value set) of attribute a, f : U × A → V is an
a∈ A
a a

information function such that f (u , a ) ∈ V , for every (u , a ) ∈ U × A , called information

a

(knowledge) function. An information system can be intuitively expressed in terms of

an information table (refers to Table 1).

Table 1. An information system

U a1 … ak … aA
u1 f (u1 , a1 ) … f (u1 , a k ) … ( )
f u1 , a A
u2 f (u 2 , a1 ) … f (u 2 , a k ) … f (u , a )
2 A

M M O M O M
uU (
f u U , a1 ) … (
f u U , ak ) … (
f u U ,a A )
A relational database may be considered as an information system in which rows
are labeled by the objects (entities), columns are labeled by attributes and the entry in
row u and column a has the value f (u, a ) . We note that a each map
(
f (u , a ) : U × A → V is a tupple ti = f (ui , a1 ), f (ui , a2 ), f (ui , a3 ),L , f ui , a A ( )) , for
1 ≤ i ≤ U . In many applications, there is an outcome of classification that is known.
Note that the tuple t is not necessarily associated with entity uniquely (refer to objects
3 and 4 in Table 3). In an information table, two distinct entities could have the same
tuple representation (duplicated/redundant tuple), which is not permissible in
relational databases. Thus, the concepts in information systems are a generalization of
the same concepts in relational databases.
 Matrices Representation of Multi Soft-Sets and Its Application 203

This a posteriori knowledge is expressed by one (or more) distinguished attribute

called decision attribute; the process is known as supervised learning. An information
system of this kind is called a decision system. A decision system is an information
system of the form D = (U , A U {d }, V , f ) , where d ∉ A is the decision attribute. The
elements of A are called condition attributes. An example of a decision system is
given in Table 2.

3 Soft Set Theory

The theory of soft set [1] proposed, by Molodtsov 1999 is a new method for handling
uncertain data. The soft set is a mapping from parameter to the crisp subset of
universe. From such case, we may see the structure of a soft set can classify the
objects into two classes (yes/1 or no/0).
Definition 1. (See [1].) A pair (F , A) is called a soft set over U, where F is a
mapping given by
F : A → P(U ) .
In other words, a soft set over U is a parameterized family of subsets of the universe
U. For ε ∈ A , F (ε ) may be considered as the set of ε -elements of the soft set (F , A)
or as the set of ε -approximate elements of the soft set. Clearly, a soft set is not a
(crisp) set.
Based on the definition of an information system and a soft set, in this section we
show that a soft set is a special type of information systems, i.e., a binary-valued
information system.
Proposition 2. If (F , A) is a soft set over the universe U, then (F , A) is a binary-
valued information system S = (U , A, V{0,1} , f ) .

Proof. Let (F , A) be a soft set over the universe U, we define a mapping

F = { f 1 , f 2 , L, f n } ,
where
⎧1, x ∈ F (ai )
f i : U → Vi and f i (x ) = ⎨ , for 1 ≤ i ≤ A .
⎩0, x ∉ F (ai )
Hence, if V = Ua ∈AVa , where Ve = {0,1} , then a soft set (F , A) can be considered as
i i

a binary-valued information system S = (U , A, V{0,1} , f ) .

i

□
From Proposition 2, it is easily to understand that a binary-valued information system
can be represented as a soft set. Thus, we can make a one-to-one correspondence
between (F , E ) over U and S = (U , A, V{0,1} , f ) .
Definition 3. (See [6].) The class of all value sets of a soft set (F , E ) is called value-
class of the soft set and is denoted by C( F , E ) .
204 T. Herawan, M.M. Deris, and J.H. Abawajy

4 Multi-Soft Sets Construction in Information Systems

The idea of multi-soft sets is based on a decomposition of a multi-valued information
system S = (U , A,V , f ) , into A number of binary-valued information systems
S = (U , A,V{0,1} , f ) , where A denotes the cardinality of A. Consequently, the A
binary-valued information systems define multi-soft sets (F , A) = {(F , a i ) : 1 ≤ i ≤ A } .

4.1 Decomposition of Multi-valued Information Systems

The decomposition of S = (U , A, V , f ) is based on decomposition of

{
A = a1 , a 2 ,L , a A } { }
into the disjoint-singleton attribute {a1 }, {a 2 }, L , a A . Here, we
only consider for complete information systems. Let S = (U , A, V , f ) be an
information system such that for every a ∈ A , Va = f (U , A) is a finite non-empty set
and for every u ∈ U , f (u , a ) = 1 . For every a i under i th-attribute consideration,
a i ∈ A and v ∈ Va , we define the map a vi : U → {0,1} such that a vi (u ) = 1 if
f (u , a ) = v , otherwise a vi (u ) = 0 . The next result, we define a binary-valued
information system as a quadruple S i = (U , ai , V{0,1} , f ) . The information systems
S i = (U , ai , V{0,1} , f ) , 1 ≤ i ≤ A is referred to as a decomposition of a multi-valued
information system S = (U , A, V , f ) into A binary-valued information systems, as
depicted in Figure 1. Every information system S i = (U , ai ,Vai , f ) , 1 ≤ i ≤ A is a
deterministic information system since for every a ∈ A and for every u ∈ U ,
f (u , a ) = 1 such that the structure of a multi-valued information system and A
number of binary-valued information systems give the same value of attribute related
to objects.

4.2 Multi-soft Sets in Information Systems

Based on the notion of a decomposition of a multi-valued information system in the

previous sub-section, in this sub-section we present the notion of multi-soft set
representing multi-valued information systems. Let S = (U , A, V , f ) be a multi-
valued information system and S i = (U , ai ,Vai , f ) , 1 ≤ i ≤ A be the A binary-
valued information systems. From Proposition 2, we have
⎧ S 1 = (U , a1 , V{0,1} , f ) ⇔ (F , a1 )
⎪ 2
⎪S = (U , a 2 , V{0,1} , f ) ⇔ (F , a 2 )
S = (U , A, V , f ) = ⎨
⎪ M M M
( ) (
⎪ S A = U , a A , V{0,1} , f ⇔ F , a A
⎩
)
( (
= (F , a1 ), (F , a2 ),L , F , a A ))
 Matrices Representation of Multi Soft-Sets and Its Application 205

(
We define (F , A) = (F , a1 ), (F , a 2 ), L , F , a A ( )) as a multi-soft sets over universe U
representing a multi-valued information system S = (U , A, V , f ) .

U a1 a2 … ak … aA
u1 f (u1 , a1 ) f (u1 , a 2 ) … f (u1 , a k ) … ( )
f u1 , a A
u2 f (u 2 , a1 ) f (u 2 , a 2 ) … f (u 2 , a k ) … f (u , a )
2 A

u3 f (u 3 , a1 ) f (u 3 , a 2 ) … f (u 3 , a k ) … f (u , a )
3 A

M M M O M O M
uU (
f u U , a1 ) (
f u U , a2 ) … (
f u U , ak ) … (
f uU ,a A )
Binary-valued information system-1
a1
Va11 Va12 … Va1k … Va1n
U
u1 0 1 … 0 … 0
u2 0 0 … 1 … 0
u3 0 1 … 0 … 0
M M M O M O M
uU 1 0 … 0 … 0

M
Binary-valued information system- A
aA
Va A 1 Va A 2 … Va A k … Va A n
U
u1 0 0 … 1 … 0
u2 0 0 … 0 … 1
u3 0 1 … 0 … 0
M M M O M O M
uU 1 0 … 0 … 0

Fig. 1. A decomposition of information systems

Example 4. Hiring [7]: an example of a multi-valued information system is presented

in Table 2. In Table 2, six students are characterized by four conditional attributes;
Diploma, Experience, French, Reference and one decision attribute; Decision.
Let A = {Diploma, Experience, French, Reference, Decision} . Therefore, the multi-
soft set representing Table 2 is given in Figure 2. Note that the class value of every
soft set is a partition of U.
206 T. Herawan, M.M. Deris, and J.H. Abawajy

Table 2. Hiring: an information system from [7]

Student Diploma Experience French Reference Decision

1 MBA Medium Yes Excellent Accept
2 MBA Low Yes Neutral Reject
3 MCE Low Yes Good Reject
4 MSc High Yes Neutral Accept
5 MSc Medium Yes Neutral Reject
6 MSc High Yes Excellent Accept
7 MBA High No Good Accept
8 MCE Low No Excellent Reject

⎛ {MBA = {1,2,7}, MCE = {3,8}, MSc = {4,5,6}}, ⎞

⎜ ⎟
⎜ {Medium = {1,5}, Low = {2,3,8}, High = {4,6,7}}, ⎟
(F , A) = ⎜⎜ {Yes = {1,2,3,4,5,6}, No = {7,8}}, ⎟
⎟
⎜ {Excellent = {1,6,8}, Neutral = {2,4,5}, Good = {3,7}}, ⎟
⎜ ⎟
⎝ {Accept = {1,4,6,7}, Reject = {2,3,5,8}} ⎠
Fig. 2. Multi soft-sets representing Table 2

5 Matrices Representation of Multi-soft Sets

5.1 Matrix Representation

The concept of matrix representing multi-soft sets is described here. In this sub-
section, the notation (F, A) represents a multi-soft set over universe U of a multi-
valued information system S = (U , A, V , f ) .
⎧ (F , a1 ) ⇔ S 1 = (U , a1 ,V{0,1} , f ) ⇔ M a
⎪ (F , a )
⇔ S 2 = (U , a2 ,V{0 ,1} , f ) ⇔ M a
1

(F , A) ⇔ ⎪⎨ M 2 M M
2
.
⎪
( ) ( )
⎪ F , a A ⇔ S A = U , a A ,V{0,1} , f ⇔ M A
⎩
Definition 5. Matrix M a , 1 ≤ i ≤ A is called matrix representation of soft set (F , a1 )
( )
i

over universe U. The dimension of matrices is defined by dim M a = U × Vai . All i

entries of M a = [aij ] is belong to a set {0,1} , where

i

⎧⎪0, if f (u, α ) = 0
aij = ⎨ , 1 ≤ i ≤ U , 1 ≤ j ≤ Vai , u ∈ U and α ∈ Vai .
⎪⎩ 1, if f (u, α ) = 1

The collection of all matrices representing (F, A) is denoted by M A

, i.e.,

M A
{
= M a :1 ≤ i ≤ A .
i
}
 Matrices Representation of Multi Soft-Sets and Its Application 207

Definition 6. Let M a ∈ M A be a matrix representation of a soft set (F, ai ) over U.

i

The value-class of M a , i.e., class of all value sets of M a , denoted CM is defined by

i i ai

{
C M = {u : f (u , α1 ) = 1}, L, u : f u , α Va
ai
{ ( i
) = 1}},
where 1 ≤ i ≤ Vai , u ∈ U and α ∈ Vai .
Clearly CM ⊆ P(U ) .
ai

Example 7. The collection of matrices representing (F, A) , is given as

M A
{
= Ma ,Ma ,Ma ,Ma ,Ma ,
1 2 3 4 5
}
where
⎡1 0 0⎤ ⎡1 0 0⎤ ⎡1 0⎤
⎢1 0 0⎥⎥ ⎢0 1 0⎥⎥ ⎢1 0⎥
⎢ ⎢ ⎢ ⎥
⎢0 1 0⎥ ⎢0 1 0⎥ ⎢1 0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
0 0 1⎥ 0 0 1⎥ 1 0⎥
M Diploma = ⎢ , M Experience = ⎢ , M French = ⎢ ,
⎢0 0 1 ⎥ ⎢ 1 0 0⎥ ⎢1 0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢0 0 1⎥ ⎢0 0 1⎥ ⎢1 0⎥
⎢1 0 0 ⎥ ⎢0 0 1⎥ ⎢0 1 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢⎣0 1 0⎥⎦ ⎢⎣0 1 0⎥⎦ ⎢⎣0 1⎥⎦
⎡1 0 0⎤ ⎡1 0⎤
⎢0 1 0⎥⎥ ⎢0 1⎥⎥
⎢ ⎢
⎢0 0 1⎥ ⎢0 1⎥
⎢ ⎥ ⎢ ⎥
0 1 0⎥ 1 0⎥
M Reference = ⎢ and M Decision = ⎢ .
⎢0 1 0 ⎥ ⎢ 0 1⎥
⎢ ⎥ ⎢ ⎥
⎢1 0 0⎥ ⎢1 0⎥
⎢0 0 1 ⎥ ⎢1 0⎥
⎢ ⎥ ⎢ ⎥
⎢⎣1 0 0⎥⎦ ⎢⎣0 1⎥⎦
where
CM Diploma
= {{1,2,7}, {3,8}, {4,5,6}} , C M Experience
= {{1,5}, {2,3,8}, {4,6,7}} ,
CM French
= {{1,2,3,4,5,6}, {7,8}} , C M Reference
= {{1,6,8}, {2,4,5}, {3,7}} ,
and C M Decision
= {{1,4,6,7}, {2,3,5,8}} ,
respectively.

5.2 AND and OR Operations in a Collection of Matrices

The AND and OR operation of the matrices are described in this sub-section.
208 T. Herawan, M.M. Deris, and J.H. Abawajy

Definition 8. Let M a = [akl ] , 1 ≤ k ≤ U , 1 ≤ l ≤ Va and M a = [amn ] , 1 ≤ m ≤ U ,

i i j

1 ≤ n ≤ Va be two matrices in
i
M A
. The " AND" operation M a ANDM a of matrices i j

M a and M a is defined as follows

i j

M a ANDM a = M a = [a pq ] with dim M a = U × Vai × Va j ,

i j ij
( ) ij
( )
where

a p1 = min{ak 1 , am1 } , a p 2 = min{ak 1 , am 2 } ,…, a p ( Va i × Va j

{
) = min ak Va , am Va . i j
}
Example 9. Let M Diploma , M Experience ∈ M E , from Definition 8, we have

⎡1 0 0⎤ ⎡1 0 0⎤ ⎡1 0 0 0 0 0 0 0 0⎤
⎢1
⎢ 0 0⎥⎥ ⎢0
⎢ 1 0⎥⎥ ⎢⎢0 1 0 0 0 0 0 0 0⎥⎥
⎢0 1 0⎥ ⎢0 1 0⎥ ⎢0 0 0 0 1 0 0 0 0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
0 0 1⎥ 0 0 1⎥ ⎢0 0 0 0 0 0 0 0 1⎥
M Diploma ANDM Experience = ⎢ AND ⎢ = ,
⎢0 0 1 ⎥ ⎢ 1 0 0 ⎥ ⎢0 0 0 0 0 0 1 0 0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢0 0 1⎥ ⎢0 0 1 ⎥ ⎢0 0 0 0 0 0 0 0 1⎥
⎢1 0 0⎥ ⎢0 0 1 ⎥ ⎢0 0 1 0 0 0 0 0 0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢⎣0 1 0⎥⎦ ⎢⎣0 1 0⎥⎦ ⎢⎣0 0 0 0 1 0 0 0 0⎥⎦
where
CM Diploma ANDM Experience = {{1}, {2}, {3,8}, {4,6}, {5}, {7}} .

Definition 10. Let M a = [akl ] and M a = [bmn ] be two matrices in

i j
M A
. The " OR"
operation M a ORM a of matrices M a and M a is defined as follows
i j i j

M a ORM a = M a = [c pq ] with dim M a = U × Vai × Va j ,

i j ij
( )ij
( )
where
a p1 = max{a k 1 , a m1 } , a p 2 = max{a k 1 , a m 2 } ,…, a p ( Va i × Va j
{
) = max a k Va , a m Va i j
}.
Example 11. Let M Diploma , M Experience ∈ M E , from Definition 10, we have
⎡1 0 0⎤ ⎡1 0 0⎤ ⎡1 1 1 1 0 0 1 0 0⎤
⎢1
⎢ 0 0⎥⎥ ⎢0
⎢ 1 0⎥⎥ ⎢⎢1 1 1 0 1 0 0 1 0⎥⎥
⎢0 1 0⎥ ⎢0 1 0 ⎥ ⎢0 1 0 1 1 1 0 1 0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
0 0 1⎥ 0 0 1 ⎥ ⎢0 0 1 0 0 1 1 1 1⎥
M Diploma ORM Experience =⎢ OR ⎢ = ,
⎢0 0 1⎥ ⎢1 0 0 ⎥ ⎢1 0 0 1 0 0 1 1 1⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢0 0 1⎥ ⎢0 0 1 ⎥ ⎢0 0 1 0 0 1 1 1 1⎥
⎢1 0 0 ⎥ ⎢0 0 1 ⎥ ⎢1 1 1 0 0 1 0 0 1⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢⎣0 1 0⎥⎦ ⎢⎣0 1 0⎥⎦ ⎢⎣0 0 0 1 1 1 0 1 0⎥⎦
 Matrices Representation of Multi Soft-Sets and Its Application 209

where
CM Diploma ORM Experience = {{1}, {2}, {3,8}, {4,6}, {5}, {7}} .

We can see that, since dim M a = U × Vai U Va j , thus ( ) ij

M A
is not closed under
AND and OR operations, thus (M A
, AND) and (M A
, OR ) are not groupoids.

Proposition 12. Let M a , M a and M a be matrices in i j k

M A
. The following
properties are holds.
a. M a ANDM a = M a ,
i i i
idempotent
b. M a ORM a = M a , idempotent
( ) ( )
i i i

c. M a AND M a ANDM a = M a ANDM a ANDM a associative

( ) = (M ORM )ORM
i j k i j k

d. M a OR M a ORM a
i j k ai aj ak associative

AND and OR operations are not commutative.

6 Application
In this section, we present an application of matrices representation of multi-soft sets
for dimensionality reduction and finding core attributes.

6.1 Reduct and Core

In [4], we presented the application of multi-soft sets for finding reducts. The
proposed approach is based on AND operation in multi-soft sets [3]. In this sub-
section, we explore the applicability of matrices representation for finding reducts and
core attributes. We show that the reducts obtained are equivalent with that in [4]. The
notion of reduct in [4] is redefined and given as follow.
Definition 13. Let (F , A) = ((F , ai ) : 1 ≤ i ≤ A ) be multi-soft set over U representing a
multi-valued information system S = (U , A, V , f ) . A set of attributes B ⊆ A is called
a reduct for A if
CF (b ×L×b ) = CF (a ×L×a ) and CF (b ×L×b ) ≠ CF (a ×L×a ), B* ⊂ B ,
1 B 1 A 1 B* 1 A

where
(F , a × a ) = (F , a )AND(F , a ) .
i j i j

A core of A is defined as

CORE ( A) = I RED( A) ,

where RED( A) is the set of all reducts of A.

210 T. Herawan, M.M. Deris, and J.H. Abawajy

Proposition 14. Let M A

be a multi-soft set over U representing multi-soft sets
(F, A) . A set of attributes B ⊆ A is a reduct for A if only if C ANDM :b∈B = C ANDM :a∈B . b a

Proof. It is clear, since every (F, ai ) can be represented as M a , 1 ≤ i ≤ A and vice i

versa. □
Example 15. For simple example of soft reduction in multi-valued information
system, we consider a small dataset as used in [4].

Table 3. An information system

U a1 a2 a3 a4
1 low bad loss small
2 low good loss large
3 high good loss medium
4 high good loss medium
5 low good profit large

Let A = {a1 , a 2 , a 3 , a 4 } , the multi-soft set representing Table 3 is given as

{{low = 1,2,5}, {high = 3,4}}, {{bad = 1}, {good = 2,3,4,5}},

(F , A) = ⎛⎜⎜ ⎞
⎟⎟ ,
⎝ {{loss = 1,2,3,4}, {profit = 5}}, {{small = 1}, {large = 2,5}, {medium = 3,4}}⎠
and the matrices representing (F, A) is M A
{
= M a , M a , M a , M a , where
1 2 3 4
}
⎡1 0⎤ ⎡1 0 ⎤ ⎡1 0 ⎤ ⎡1 0 0⎤
⎢1 0⎥⎥ ⎢ ⎥ ⎢ ⎥ ⎢0 1 0⎥⎥
⎢ ⎢0 1 ⎥ ⎢1 0 ⎥ ⎢
M a = ⎢0 1⎥ , M a = ⎢0 1⎥ , M a = ⎢1 0⎥ and M a = ⎢0 0 1⎥ .
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
1 2 3 4

⎢0 1⎥ ⎢0 1 ⎥ ⎢1 0 ⎥ ⎢0 0 1⎥
⎢1 0⎥⎦ ⎢0 1 ⎥ ⎢0 1 ⎥ ⎢0 1 0⎥⎦
⎣ ⎣ ⎦ ⎣ ⎦ ⎣
Since A = {a1 , a2 , a3 , a4 } , then

⎡1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎤
⎢0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎥
M ANDa :1≤i≤4 = ⎢0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0⎥
i ⎢ ⎥
⎢0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0⎥
⎢⎣0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0⎥⎦

where
CM ANDai :1≤ i ≤ 4
= {{1}, {2}, {3,4}, {5}} .
 Matrices Representation of Multi Soft-Sets and Its Application 211

Let, given B = {a1 , a2 , a3 } and C = {a3 , a4 } , then we have

⎡1 0 0 0 0 0 0 0⎤
⎢0 0 1 0 0 0 0 0⎥⎥
⎢
M a ANDM a ANDM a = ⎢0 0 0 0 0 0 1 0⎥
⎢ ⎥
1 2 3

⎢0 0 0 0 0 0 1 0⎥
⎢0 0 0 1 0 0 0 0⎥⎦
⎣
and

⎡1 0 0 0 0 0⎤
⎢0 1 0 0 0 0⎥⎥
⎢
M a ANDM a = ⎢0 0 1 0 0 0⎥
⎢ ⎥
3 4

⎢0 0 1 0 0 0⎥
⎢0 0 0 0 1 0⎥⎦
⎣
where
CM a1 ANDM a 2 ANDM a3 = {{1}, {2}, {3,4}, {5}} and CM a1 ANDM a 2 ANDM a3 = {{1}, {2}, {3,4}, {5}} ,

respectively. Thus, we have {a1 , a 2 , a3 } and {a3 , a 4 } are reducts of A. Furthermore,

core is
{a , a , a }I {a , a } = {a } .
1 2 3 3 4 3

6.2 Equivalence with Pawlak’s Rough Reduct

In rough set theory [2,5], the aim of dimensionality reduction is to keep only those
attributes that preserve the indiscernibility relation and, consequently, set
approximation. The remaining attributes are redundant since their removal does not
worsen the classification. There are usually several such subsets of attributes and
those which are minimal are called reducts. While computing equivalence classes is
straightforward, finding all reducts is NP-hard. This reduction is different in soft set
theory. The existing techniques of soft parameterization are still based on a binary
information system [8,9]. Soft parameterization reduction is obtained based on the
optimal and sub-optimal choice related to each object. Thus, the idea of reduction
under rough set theory generally cannot be applied directly in reduction under soft set
theory. In this sub-section, we show that our proposed technique on dimensionality
reduction in an multi-valued information system is equivalent with that in [2,5]. Let,
firstly, we recall the notion of rough reduction as follow.
Let S = (U , A, V , f ) be an information system and let B be any subset of A. Two
elements x, y ∈ U are said to be B-indiscernible (indiscernible by the set of attribute
B) if and only if f (x, a ) = f ( y, a ) , for every a ∈ B . Obviously, every subset of A
induces unique indiscernibility relation. Notice that, an indiscernibility relation is an
equivalence relation. The partition of U induced by IND(B ) in S = (U , A, V , f ) is
denoted by U / B .
212 T. Herawan, M.M. Deris, and J.H. Abawajy

Definition 16. Let S = (U , A, V , f ) be an information system and let B be any subset

of A. A subset B ⊆ A is called a reduct of A if B satisfies the following conditions

a. U / B =U / A
b. U / (B − {b}) ≠ U / A, ∀b ∈ B

A core of A is defined as

CORE ( A) = I RED( A) ,
where RED( A) is the set of all reducts of A.
Proposition 17. Soft reduction in Definition 13 is equivalent with rough reduction in
Definition 16.
Proof. (⇒ ) It is clear that (F , A) = (U , A,V , f ) , thus CF (a ) = U / ai , 1 ≤ i ≤ A . From
i

the hypothesis, since B is a reduct of A, then

CF (b ×L×b ) = CF (a ×L×a )
(F , b )ANDL AND(F , b ) = (F , a )ANDL AND(F , a )
1 B 1 A

1 B 1 A

CF (b ) I L I C F (b ) = C F (a ) I L I C F (a )
1 B 1 A

I 1≤i ≤ B
CF (b ) = I 1≤i≤ A CF (a )
i i

I 1≤ i ≤ B
U / bi = I 1≤ i ≤ A U / ai
U / B =U / A

Say, B* = B − {b}, where b ∈ B , then

CF (b ×L×b ) ≠ CF (a ×L×a )
1 B* 1 A

U / (B − {b}) ≠ U / A

(⇐) Obvious. □

Example 18. From Table 2, let

B = {Diploma, Experience, French, Reference} .

Using rough set approach, we have

U / B = {{1}, {2}, {3}, {4}, {5}, {6}, {7}, {8}} .

Therefore, reduct of B is B* = {Experience, Reference} .

 Matrices Representation of Multi Soft-Sets and Its Application 213

For matrices approach, from Example 7, we have

M Diploma ANDM Experience ANDM French M Reference =

⎡1 0 0⎤ ⎡1 0 0⎤ ⎡1 0⎤ ⎡1 0 0⎤
⎢1 0 0⎥⎥ ⎢0 1 0⎥⎥ ⎢1 0⎥⎥ ⎢0 1 0⎥⎥
⎢ ⎢ ⎢ ⎢
⎢0 1 0⎥ ⎢0 1 0⎥ ⎢1 0⎥ ⎢0 0 1⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
=
⎢0 0 1⎥
AND ⎢
0 0 1⎥
AND ⎢
1 0⎥
AND ⎢
0 1 0⎥
⎢0 0 1 ⎥ ⎢ 1 0 0 ⎥ ⎢ 1 0 ⎥ ⎢ 0 1 0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢0 0 1⎥ ⎢0 0 1⎥ ⎢1 0⎥ ⎢ 1 0 0⎥
⎢1 0 0 ⎥ ⎢0 0 1 ⎥ ⎢0 1⎥ ⎢0 0 1⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣⎢0 1 0⎦⎥ ⎣⎢0 1 0⎦⎥ ⎣⎢0 1⎦⎥ ⎣⎢1 0 0⎦⎥
The calculation of the above AND operation is given as follows
⎛ ⎡1 0 0⎤ ⎡1 0 0⎤ ⎞ ⎛ ⎡1 0⎤ ⎡1 0 0⎤ ⎞
⎜⎢ ⎥ ⎢ ⎥⎟ ⎜⎢ ⎥ ⎢0 ⎟
⎜ ⎢1 0 0⎥ ⎢0 1 0⎥ ⎟ ⎜ ⎢1 0⎥ ⎢ 1 0⎥⎥ ⎟
⎜ ⎢⎢ ⎥ ⎢ ⎥⎟ ⎜⎢ ⎟
⎜ ⎢0 1 0⎥ ⎢0 1 0⎥ ⎟ ⎜ ⎢1 0⎥ ⎢0 0 1⎥ ⎟
⎥ ⎢ ⎥ ⎥ ⎢ ⎥
⎜ ⎢0 0 1 ⎥ ⎢0 0 1 ⎥ ⎟ ⎜ 1 0⎥ 0 1 0⎥ ⎟
= ⎜⎢ ⎥ AND ⎢ ⎥ ⎟AND⎜ ⎢ AND ⎢ ⎟
0 0⎥ ⎟ ⎢ 0⎥ ⎢0 1 0⎥ ⎟
⎜ ⎢⎢0 0 1⎥
⎥
⎢1
⎢ ⎥ ⎜ ⎢1 ⎥ ⎢ ⎥
⎜ ⎢0 0 1⎥ ⎢0 0 1⎥ ⎟ ⎜ ⎢1 0⎥ ⎢1 0 0⎥ ⎟
⎜⎢ ⎥ ⎢ ⎥⎟ ⎜⎢ ⎥ ⎢0 ⎟
⎜ ⎢⎢1 0 0⎥ ⎢0 0 1⎥ ⎟ ⎜ ⎢0 1 0 1⎥ ⎟
⎥ ⎢ ⎥ ⎥ ⎢ ⎥
⎜ ⎢0 ⎟ ⎜ ⎢0 1⎦⎥ 0 0⎦⎥ ⎟⎠
⎝⎣ 1 0⎦⎥ ⎢0
⎣
⎥
1 0⎦ ⎠ ⎝⎣ ⎣⎢1

⎛ ⎡1 0 0 0 0 0 0 0 0⎤ ⎞ ⎛ ⎡1 0 0 0 0 0⎤ ⎞
⎜⎢ ⎥⎟ ⎜⎢ ⎟
⎜ ⎢0 1 0 0 0 0 0 0 0⎥ ⎟ ⎜ ⎢0 1 0 0 0 0⎥⎥ ⎟
⎜ ⎢⎢ ⎥⎟ ⎜⎢ ⎟
⎜ ⎢0 0 0 0 1 0 0 0 0⎥ ⎟ ⎜ ⎢0 0 1 0 0 0⎥ ⎟
⎥ ⎥
⎜ ⎢0 0 0 0 0 0 0 0 1 ⎥ ⎟ ⎜ 0 1 0 0 0 0⎥ ⎟
= ⎜⎢ ⎥ ⎟ AND⎜ ⎢ ⎟
⎢ 1 0 0 0 0⎥ ⎟
⎜ ⎢⎢0 0 0 0 0 0 1 0 0⎥ ⎟
⎥ ⎜ ⎢0 ⎥
⎜ ⎢0 0 0 0 0 0 0 0 1⎥ ⎟ ⎜ ⎢1 0 0 0 0 0⎥ ⎟
⎜⎢ ⎥⎟ ⎜⎢ ⎟
⎜ ⎢⎢0 0 1 0 0 0 0 0 0⎥ ⎟ ⎜ ⎢0 0 0 0 0 1⎥ ⎟
⎥ ⎥
⎜ ⎢0 ⎥ ⎟ ⎜ ⎢0 0 0 1 0 0⎦⎥ ⎟⎠
⎝⎣ 0 0 0 1 0 0 0 0⎦ ⎠ ⎝⎣
⎡1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ⎤
⎢0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ⎥
⎢0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
⎥
= ⎢ ⎥
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0

⎢0 ⎥
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

⎢0 0⎥
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
⎢⎣0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎥⎦

Thus we have
CM Diploma ANDM Experience ANDM French M Reference = {{1}, {2}, {3}, {4}, {5}, {6}, {7}, {8}} ,
214 T. Herawan, M.M. Deris, and J.H. Abawajy

Therefore, reduct of B is B* = {Experience, Reference} . It is shown that, using rough

and matrices approaches, reduct obtained is equivalent.

7 Conclusion
The “standard” soft set deals with a binary-valued information system. For a multi-
valued information system, the notion of multi-soft sets has been proposed. The idea
of multi-soft sets is based on a decomposition of a multi-valued information system
into binary-valued information systems. In this paper, we have presented the notion of
matrices representation of such multi-sot sets. The AND and OR operations in a
collection of matrices sets also presented. The AND and OR operations satisfy
idempotent and associative properties, neither closed nor commutative. Further, we
have proven that AND operation can be applied for finding reducts. Finally, we have
proven that our proposed technique provide the same results with that rough reduct.
For future research, we further elaborate our approach for finding functional, identity
and attribute dependencies in information systems. Since, the calculation of matrices
is complex, we will use MATLAB to solve this problem. We believe that our
proposed approach can be used for feature selection and data cleansing in multi-
valued information system under soft set theory.

Acknowledgement
This work was supported by the FRGS under the Grant No. Vote 0402, Ministry of
Higher Education, Malaysia.

References
1. Molodtsov, D.: Soft Set Theory-First Results. Computers and Mathematics with
Applications 37, 19–31 (1999)
2. Pawlak, Z., Skowron, A.: Rudiments of Rough sets. Information Sciences 177(1), 3–27
(2007)
3. Herawan, T., Mustafa, M.D.: On Multi-soft Sets Construction in Information Systems. In:
Huang, D.-S., Jo, K.-H., Lee, H.-H., Kang, H.-J., Bevilacqua, V. (eds.) ICIC 2009. LNCS
(LNAI), vol. 5755, pp. 101–110. Springer, Heidelberg (2009)
4. Herawan, T., Rose, A.N.M., Mustafa, M.D.: Soft Set Theoretic Approach for
Dimensionality Reduction. In: Communication of Computer and Information Sciences,
vol. 64, pp. 180–187. Springer, Heidelberg (2009)
5. Pawlak, Z.: Rough Sets: A Theoretical Aspect of Reasoning about Data. Kluwer Academic
Publisher, Dordrecht (1991)
6. Maji, P.K., Biswas, R., Roy, A.R.: Soft Set Theory. Computers and Mathematics with
Applications 45, 555–562 (2003)
7. Komorowski, J., Ohrn, A.: Modeling prognostic power of cardiac tests using rough sets.
Artificial Intelligence in Medicine 15, 167–191 (1999)
8. Chen, D., Tsang, E.C.C., Yeung, D.S., Wang, X.: The Parameterization Reduction of Soft
Sets and its Applications. Computers and Mathematics with Applications 49, 757–763 (2005)
9. Zhi, K., Gao, L., Wang, L., Li, S.: The Normal Parameter Reduction of Soft Sets and Its
Algorithm. Computers and Mathematics with Applications 56, 3029–3037 (2008)
 Clustering Analysis of Water Quality for Canals in
Bangkok, Thailand

Sirilak Areerachakul and Siripun Sanguansintukul

Department of Mathematics, Faculty of Science, Chulalongkorn University,

Bangkok, Thailand
[email protected], [email protected]

Abstract. Two clustering techniques of water quality for canals in Bangkok

were compared: K-means and Fuzzy c-means. The result illustrated that K-
means has a better performance. As a result, K-means cluster was used to
classify 24 canals of 344 records of surface water quality within Bangkok; the
capital city of Thailand. The data was obtained from the Department of
Drainage and Sewerage Bangkok Metropolitan Administration during 2005-
2008. Water samples were collected and analyzed on 13 different parameters:
temperature, pH value (pH), hydrogen sulfide (H2S), dissolved oxygen (DO),
biochemical oxygen demand (BOD), chemical oxygen demand (COD),
substance solid (SS), total kjeldahl nitrogen (TKN), ammonia nitrogen (NH3N),
nitrite nitrogen (NO2N), nitrate nitrogen (NO3N), total phosphorous (T-P) and
total coliform. The data were analyzed and clustered. The results of cluster
analysis divided the canals into five clusters. The information from clustering
could enhance the understanding of surface water usage in the area.
Additionally, it can provide the useful information for better planning and
watershed management of canals in Bangkok.

Keywords: surface water quality, watershed management, K-means clustering.

1 Introduction
Water will become the major constraining resource for sustainable development of
large areas in the world. Water quality is one of the main characteristics of a river,
even when its purpose is other than human water supply [4]. Water quality in the
superficial waters has started to degenerate as a result of waste-water being let go to
the receiving ground and surface water without any controls [5].
Bangkok is the capital city, as well as, the economic center of Thailand. It’s
activities, which include commercial, industrial and service have caused the expansion
of the city and its population to accumulate environmental pollution to the point that
nature cannot cope with the pollution loading, especially for water quality [21].
At present, the canal water pollution is very severe as the canals still act as sewers
for direct discharge of waste-water in most areas of Bangkok even though there is
already a requirement for large buildings to treat waste-water and houses must have at
least septic tanks. However, other waste-waters are discharged without treatment to the
public sewers which drain into the canals [22].

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 215–227, 2010.
© Springer-Verlag Berlin Heidelberg 2010
216 S. Areerachakul and S. Sanguansintukul

Multivariate statistical methods have been widely applied to investigate

environmental phenomena, such as Laaksoharju et al. [11] who did hydrochemical
evaluation using a multivariate mathematical tool. Anazawa et al. [6] employed factor
analysis (FA); two major factors, to describe the chemical behaviors parameters.
Gűler and Thyne [2] applied HCA analysis to classify recharged area waters. Cluster
analysis, principle component analysis/factor analysis, time series analysis, self-
organizing maps and classification and regression trees (CART) strategy, etc., are
powerful tools for deriving useful information from complicated data about water
quality studies [1],[7],[9],[12],[14],[15]. The multivariate statistical methods have
extensive applications in characterization and evaluation of surface water quality and
are useful for evaluation of temporal and spatial variations caused by natural and
anthropogenic factors[16],[17].
The purpose of this paper is to compare two clustering techniques: K-means and
Fuzzy c-means. These two methodologies have been selected because they are the most
classical, popular and efficient methodologies for clustering surface water quality
[8],[19]. The better clustering approach will be used to summarize and analyze surface
water quality of canals in Bangkok.

2 Data and Site Description

In this study, water quality data are provided by the Department of Drainage and
Sewerage, Bangkok Metropolitan Administration, during the years 2005-2008. There
are 24 canals of 344 records of sites. Each record consists of 13 parameters, namely:
temperature, pH value (pH), hydrogen sulfide (H2S), dissolved oxygen (DO),
biochemical oxygen demand (BOD), chemical oxygen demand (COD), substance
solid (SS), total kjeldahl nitrogen (TKN), ammonia nitrogen (NH3N), nitrite nitrogen
(NO2N), nitrate nitrogen (NO3N), total phosphorous (T-P) and total coliform. Table 1
shows unit of surface water quality parameters[20].

Table 1. List of surface water quality parameters

Name of Parameters Unit of measurement

Temperature Celsius
pH value Standard Units
Hydrogen Sulfide Milligrams per liter
Dissolved Oxygen Milligrams per liter
Biochemical Oxygen Demand Milligrams per liter
Chemical Oxygen Demand Milligrams per liter
Substance Solid Milligrams per liter
Total Kjeldahl Nitrogen Milligrams per liter
Ammonia Nitrogen Milligrams per liter
Nitrite Nitrogen Milligrams per liter
Nitrate Nitrogen Milligrams per liter
Total Phosphorous Milligrams per liter
Total Coliform Most Probable Number per 100 Milliliter
 Clustering Analysis of Water Quality for Canals in Bangkok, Thailand 217

The network of canals (shown in Fig.3) is important for the daily life of the people
in Bangkok. These canals are used for consumption, transportation and recreation.
Therefore, the rapid growth of industry, condominium, high-rise and low-rise building,
and other infrastructure, have a significant effect on the canal water. The major impact
can be classified into 3 types: [21]
• Impact to tourism activities (shown in Fig.1): Bangkok is the capital city and
the center of tourism. Therefore, deterioration of canal water has direct
impact to tourism activity. As pollution in canals is in the inner area of
Bangkok, it certainly gives negative impression to the tourists who travel and
stay in Bangkok.
• Impact to aquatic life (shown in Fig.2): Basically, water pollution causes
deaths of aquatic creatures either by its toxicity or reducing dissolved oxygen
concentration. The toxicity may come from high concentration of sulfide or
ammonia. It is observed that when water is polluted there are usually fall in
the total number of species of organisms generally more sensitive than
fishes, a change in the type of species present and a change in the numbers of
individuals of each species in the water. In severe cases, where dissolved
oxygen concentration becomes zero, all aquatic creatures die and this occurs
in the highly polluted canals in Bangkok.
• Impact to public health: poor sanitary conditions prevail in many parts of
Bangkok. Records of some water related diseases generally associated with
sanitary conditions are discussed in the Annual Epidemiological Surveillance
Report of the Epidemiological Subdivision, Disease Control Division
Department of Health [23]. In case of acute diarrhea, about 45,000 and
50,000 cases were found in 1996 and 1997 respectively.

Fig. 1. The impact to tourism activities Fig. 2. The impact to aquatic life
218 S. Areerachakul and S. Sanguansintukul

Fig. 3. The network of canals in Bangkok, Thailand

3 Methodology
Clustering is an important data mining technique that has a wide range of applications
in many areas like biology, medicine, market research, image processing, and
geographical information systems, among others [24].
 Clustering Analysis of Water Quality for Canals in Bangkok, Thailand 219

In this study, the water qualities of canals in Bangkok were grouped according to
their characteristic forming clusters. The clustering process was carried out using a K-
means algorithm and Fuzzy c-means algorithm.

3.1 K-Means Algorithm

The K-means algorithm is a cluster analysis algorithm used as a partitioning method,

and was developed by MacQueen in 1967 [3]. The K-means algorithm defines a
random cluster centroid according to the initial parameter [10]. Each consecutive case
is added to the cluster according to the proximity between the mean value of the case
and the cluster centroid. The clusters are then re-analyzed to determine the new
centroid point. This procedure is repeated for each data object.
The algorithm is composed of the following steps: [24]
1. Place K point into the space represented by the objects that are being
clustered. These points represent initial group centroids.
2. Assign each object to the group that has the closest centroids.
3. When all objects have been assigned, recalculate the positions of the K
centroids.
4. Repeat step 2 and 3 until the centroids no longer move. This produces a
separation of the objects into groups from which the metric to be minimized
can be calculated.
The main idea is to define k centroids, one for each cluster. The aim of that algorithm
is to find cluster centroids for each group. The algorithm minimizes a dissimilarity
function which is given in equation (1) [3].
k n 2

J = ∑∑ x −c
i j
(1)
j =1 i =1

where x is the data point , c is the j cluster centroid , n is the number of data

x −c
i j
is a chosen distance measure between a data point x and the cluster
centroids c .

3.2 Fuzzy c-Means Algorithm

Fuzzy c-means (FCM) is a method of clustering which allows one piece of data to
belong to two or more clusters. This method is frequently used in pattern recognition
[18]. It is based on minimization of the following objective function:

where m is any real number greater than 1, is the degree of membership of in

the cluster j, is the of d-dimensional measured data, is the d-dimensional
center of the cluster and is any norm expressing the similarity between any
measured data and the center.
220 S. Areerachakul and S. Sanguansintukul

Fuzzy partitioning is carried out through an iterative optimization of the objective

function shown above with the update of membership and the cluster center by:
1
3

∑
4
∑
This iteration will stop when

where is a termination criterion between 0 and 1 and k are the iteration steps. This
procedure converges to a local minimum or a saddle point of .
The algorithm is composed of the following steps.
1. Initialize U = [ ] matrix, U(0)
2. At k-step: calculate the centers vectors C(k)=[ ] with U(k)
3. Update U(k), U(k+1)
4. If 1 < then STOP; otherwise return to step 2.

3.3 Determination of Optimize Cluster

The criteria widely accepted for partitioning a data set into a number of clusters are
the separation of the cluster and their compactness. The optimum case implies
parameters that lead to partitions that are as close as possible (in term of similarity) to
the real partitions of the data set [13]. A reliable quality assessment index should
consider both the compactness and the separation. One of the quality measures that
can be used in clustering is described as follows [8], [13].
The compactness of spatial data set , called , the value of the p-th dimension
is defined as follows: [8], [13]

1
σ x x 6
n

where is the p-th dimension of

x ∑ x , x x 7

The compactness computed from variance of cluster, Compactness of cluster i is

called and its p-th dimension defined as:
∑ x v
σ 8
n
 Clustering Analysis of Water Quality for Canals in Bangkok, Thailand 221

The total compactness of spatial data set with respect to c cluster is:

⁄
Further the term is defined as: = , where is a vector
The average compactness of c cluster, Comp is :
Comp = / (10)
The average scattering of data set compactness, Scatt_Comp is:
Scatt_Comp = Comp/ (11)
The more compact the cluster are the smaller the Scatt_Comp. Thus, for a given
spatial data set, a smaller Scatt_Comp indicates a better clustering scheme. The
distance between cluster is defined by the average distance between the centroid of
specified clusters, that is:
∑ ∑
12
1
The larger d is the more separated the clusters are according to definitions, a quality
measure for clustering was defined as follows:
_ / (13)
The definition of CD indicated that both criteria of “good” clustering (i.e.,
compactness and separation) are properly combined, enabling reliable evaluation of
clustering results [8]. Small values of CD indicate all the cluster in clustering scheme
are overall compact and separated.

4 Experimental and Results

This section discusses data preprocessing, K-means clustering technique and result of
experiment.

4.1 Data Preprocessing

At the initial stage of the experiment, data was scaled or normalized using
x − xmin
xnew = (14)
xmax − xmin
where x is the original data point, xmin and xmax are the minimum and maximum
values in the data set, respectively. This is done in order to ensure that the minimum
value in the data set is scaled to zero, and that the maximum value is scaled to one [5].

4.2 K-Means Clustering Compare with Fuzzy c-Means Clustering

The K-means clustering and Fuzzy c-means clustering methods were applied in this
study intending to divide the groups of several sites of canals. Distances between site
222 S. Areerachakul and S. Sanguansintukul

and clusters are computed as Euclidean distances in the K-dimension space, a number
of clusters N were designated arbitrary. As a result, very different means are obtained
between clusters for most K parameters used in the analysis [10].
Because the distance in the K-dimensional space is greatly affected by differences
in scale among the variables from which the distance are computed, each variable has
been standardized before implementation of cluster analysis.
The cluster analysis of each initial matrix has been carried out in several steps,
increasing the cluster number K from 2 to 6. From determination of optimized cluster
small values of CD indicate all the cluster in clustering scheme are overall compact
and separated that both criteria of “good” clustering.

Table 2. The result of optimize cluster

Number of K Value of CD
K-means Clustering Fuzzy c-means Clustering
K=2 1.599 2.256
K=3 1.445 2.433
K=4 1.193 2.463
K=5 0.682 1.963
K=6 0.965 1.971

Table 3. Descriptive mean of water quality parameters in five clusters

Cluster
Cluster Cluster Cluster Cluster Cluster
1 2 3 4 5
Parameters
Temp 29.19 29.19 27.96 28.87 29.22
pH 6.91 6.97 7.14 6.98 7.10
DO 0.04 1.08 2.17 0.37 2.77
H2S 1.76 0.11 0.08 0.76 0.01
BOD 40.92 12.40 9.33 23.81 6.05
COD 100.89 46.92 42.99 67.21 33.94
SS 26.68 33.05 39.08 23.61 28.56
TKN 13.98 7.15 4.64 9.59 4.34
NH3N 9.42 2.15 1.40 3.71 0.78
NO2N 0.04 0.12 0.11 0.05 0.11
NO3N 2.28 1.69 0.66 1.48 1.67
T-P 1.75 0.99 0.67 1.37 0.47
T.Coliform 6.42E+09 2.18E+08 1.53E+06 2.64E+08 3.03E+07

Table 2 demonstrates the comparisons between CD values of K-means and Fuzzy

c-means clustering based on number of clusters (K) from 2 to 6 clusters. It can be
seen that for each number of cluster (K), K-means has a lower CD value than the
 Clustering Analysis of Water Quality for Canals in Bangkok, Thailand 223

fuzzy c-means. Therefore, K-means provides a better performance than Fuzzy c-

means through the dataset in this experiment. In addition, the CD value for each K
using the K-means algorithm is the lowest when K is equal to 5 (in this case, the CD
value is 0.682). Thus, the optimized cluster is equal to 5 from K-means clustering.
Each cluster consists of 13 parameters. Additional information for mean
comparisons of water quality parameters in five clusters are presented in Table 3.
DO is the dissolved oxygen in the water while BOD is the Biological oxygen demand.
The low DO will relate to high BOD. In table 3, cluster 1 shows lowest DO and highest
BOD. It means that this cluster shows the worst quality of surface water. High COD
(chemical oxygen demand) also indicates bad water quality. H2S actually come from the
decay of organic matter, which also causes a stinky smell. It should be implied that a high
value of hydrogen sulfide also indicates a poor water quality. TKN (total nitrogen) comes
from fertilizer, organic rubbish or from domestic water discharge. It means that high
TKN represents a bad water quality while nitrate and nitrite can oxidize to nitrogen gas,
making TKN a more significant value of water quality index by comparison. Total
coliform shows microorganisms in the water. The high total coliform means high
microorganisms or bacteria and also bad water quality. It can be summarized, from the
table, that the best water quality is in cluster 5 and the worst is in cluster 1.
The mean of water quality indicated a relation of parameters in each cluster.
Especially, regarding the values of Biochemical Oxygen Demand (BOD) and
Dissolved Oxygen (DO), generally high value of BOD will indicate a quantity of low
DO in surface water. Fig. 4 illustrates the comparison in means of these three selected
parameters among 5 clusters. Dissolved Oxygen (DO) refers to the volume of oxygen
that is contained in water and amount of oxygen that can be held by the water depends
on the water temperature, salinity, and pressure. The amount of dissolved oxygen
often determines the number and types of organisms living in that body of water. For
example, fish like trout are sensitive to low DO levels (less than eight parts per
million) and cannot survive in warm, slow-moving streams or rivers. Decay of
organic material in water caused by either chemical processes or microbial action on

Fig. 4. A comparison of sample parameters in five clustering

224 S. Areerachakul and S. Sanguansintukul

Table 4. Water quality of records of site and number of sites in each cluster groups

Total Number of sites in each Cluster:

Canal records
Name of Canal
Number of site 1 2 3 4 5

1 Klong Kumuengderm 12 - 4 4 2 2
2 Klong Robkrung 24 1 4 7 8 4
3 Klong Mahanark 12 - 1 5 1 5
Klong
4 20 - 2 9 4 5
PadungKrungkasem
5 Klong Samsen 16 2 3 8 2 1
6 Klong Sansab 40 - 9 21 4 6
7 Klong pramprachakorn 32 1 6 10 8 7
8 Klong Bangsue 12 1 1 3 5 2
9 Klong Satorn 8 1 2 2 1 2
10 Klong Somdetjaopraya 8 - 1 4 3 -
11 Klong Huai Khwang 16 - 4 7 4 1
12 Klong Lat phrao 16 - 4 11 - 1
13 Klong Mon 8 - 2 4 2 -
14 Klong Bangkoknoi 20 - - 15 4 1
15 Klong Prapa 8 - - 5 3 -
16 Klong Bangkokyai 20 - 2 16 2 -
17 Klong Bang khen 20 - 1 15 4 -
18 Klong Bangyikon 12 - 1 11 - -
19 Klong Bang Po 8 - - 8 - -
20 Klong Tair 8 - - 8 - -
21 Klong Lat Yao 8 - - 8 - -
22 Klong Suan Lhong 8 - - 8 - -
23 Klong Bangkabue 4 - - 4 - -
24 Klong Songtonnoon 4 - - 4 - -
Total 344 6 47 197 57 37

untreated sewage or dead vegetation can severely reduce dissolved oxygen

concentration. This is the most common cause of fish death, especially in summer
months where the warmer water naturally holds less oxygen.
 Clustering Analysis of Water Quality for Canals in Bangkok, Thailand 225

In this study, water quality data were collected during 2005-2008. There are 24
canals of 344 records of sites. Bangkok’s canals are preferred for city transportation
by many locals. The network of Bangkok Canals branches out of the Chao Phraya
River, "the river of the kings", finally meeting with the ocean. The canals in Bangkok
are hundreds of miles long. Not as many canals are used nowadays, and today both
sides of the Chao Phraya River are totally different from what they used to be in the
old days. Now, alongside the older homes and buildings, there are beautiful hotels.
Table 4 shows cluster 1 consisting of 6 of 344 records of site, cluster 2 consists of 47
of 344 records of site, cluster 3 consists of 197 of 344 records of site, cluster 4
consists of 57 of 344 records of site and cluster 5 consists of 37 of 344 records of site,
respectively.
Cluster 1 represents 6 records of site with a high value of BOD and low value of
DO in surface water canals. The reasons for higher BOD and lower DO are site of
canals connecting to industrial and populated accommodations. Most of the buildings
in this area are the commercial and high rise buildings.
Cluster 2 represents 47 records of site and cluster 4 represents 57 records of site
don’t show the various site in the study area.
Cluster 3 represents 197 records of site and all sites of canal number 19 to 24
appeared in this cluster. The site of those canals are close to the sludge treatment
centers that are knowledgeable about waste-water treatment including impact of
pollution in surface water of canals. As a result, the canals of cluster 3 according to
low value of BOD and high value of DO.
Cluster 5 represents 37 records of site and each site lowest value of BOD, COD
and TKN but highest value of DO. Because, each site of canals in cluster 5 is near to
pumping stations, non-point pollution from extractive and heavy industries, and some
site collected to rainy season of high flow water, it can be summarized that the best
water quality is in cluster 5.

6 Conclusions
In this paper, a comparison between clustering methods, K-means and Fuzzy c-means
was performed. K-means algorithm is preferred since the CD values are lower than
the Fuzzy c-means for all experimental clusters (2-6 clusters). Therefore, K-means
clustering is applied to cluster surface water quality of canals in Bangkok, Thailand.
The results indicate that five relatively cluster suitable for this sample site and
conclude to parameters.
Some of the differences corresponded to different site of canals, or water surfaces
separated by plugs or gates, that were held at different elevation and differences in
source water. The BOD and DO value parameters can show water quality for each
cluster. Better comprehension of the information may be helpful in further developing
the water quality management. As a result, Cluster analysis is found to be a useful and
efficient tool for stakeholders to manage natural resources, corresponding to the
government’s policy in water management.

Acknowledgments. The authors would like to thank Suan Sunandha Rajabhat

University for scholarship support. Thanks to Department of Drainage and
Sewerage Bangkok Metropolitan Administration for the provided data.
226 S. Areerachakul and S. Sanguansintukul

References
1. Bengraine, K., Marhaba, T.F.: Using principle component analysis to monitor spatial and
temporal changes in water quality. Journal of Hazardous Materials 100(1-3), 179–195
(2003)
2. Guler, C., Thyne, G.D.: Hydrologic and geologic factors controlling surface and
groundwater chemistry in Indian wells-Owens Valley area, southeastern California.
Journal of Hydrology 285(1-4), 177–198 (2004)
3. MacQueen, J.B., Foster, I., Kesselman, C.: Some Methods for classification and Analysis
of Multivariate Observations: Procee C.: The Grid: Blueprint for a New Computing
Infrastructure. In: Proc. Fifth Berkeley Symp. on Math. Statist. and Prob., vol. 1, pp. 281–
297. Univ. of. Calif. Press, Berkeley (1967)
4. Dogan, E., Sengorur, B., Koklu, R.: Modeling biological oxygen demand of the Melen
River in Turkey using an artificial neural network technique. Journal of Environmental
Management (2008)
5. Areerachakul, S., Sanguansintukul, S.: Water Classification Using Neural Network: A
Case Study of Canals in Bangokok. In: The 4th International Conference for Internet
Technology and Secured Transactions (ICITST 2009), Thailand. IEEE Press, United
Kingdom (2009)
6. Anazawa, K., Ohmori, H.: The hydrochemistry of surface waters in Andesitic Volcanic
area, Norikura Volcano, central Japan. Chemosphere 59(5), 605–615 (2005)
7. Astel, A., Tsakovski, S., Simeonov, V., Reisenhofer, E., Piselli, S., Barbieri, P.: Multi-
variate classification and modeling in surface water pollution estimation. Analytical and
Bioanalytical Chemistry 390(5), 1283–1292 (2008)
8. Binbib, H., Fang, T., Guo, D.: Quality assessment and uncertainty handling in spatial data
mining. In: Proc. 12th Conference on Geoinformations, Sweden (2004)
9. Helena, B., Pardo, R., Vega, M., Barrado, E., Fernández, J.M., Fernández, L.: Temporal
evolution of groundwater composition in an alluvial aquifer (Pisuerga River, Spain) by
principal component analysis. Water Research 34(3), 807–816 (2000)
10. Kanungo, T., Mount, D., Netanyahu, N., Piatko, D.C., Silverman, R.: An Efficient K-
means Clustering Algorithm Analysis and Implementation. IEEE Transaction on Pattern
Analysis and Machine Intelligence 24(7) (July 2002)
11. Laaksoharju, M., Gurban, I., Skarman, C., Skarman, E.: Multivariate mixing and mass
balance (M3) calculations, a new tool for decoding hydrogeochemical information.
Applied Geochemistry 14(7), 861–871 (1999)
12. Liu, C.W., Lin, K.H., Kuo, Y.M.: Application of factor analysis in the assessment of
groundwater quality in a blackfoot disease area in Taiwan. The Science of the Total
Environment 313(1-3), 77–89 (2003)
13. Halkidi, M., Vazirgiannis, M.: Cluster validity assessment: finding the optimal partitioning
of a dataset. In: Proc. ICDM Conference, San Jose, Californaia, USA, pp. 187–194 (1996)
14. Simeonov, V., Stratis, J.A., Samara, C., Zachariadis, G., Voutsa, D., Anthemidis, A.,
Sofoniou, M., Kouimtzis, T.: Assessment of the surface water quality in Northern Greece.
Water Research 37(17), 4119–4124 (2003)
15. Simeonova, P., Simeonov, V., Andreev, G.: Water quality study of the Struma River
Basin, Bulgaria. Central European Journal of Chemistry 2, 121–136 (2003)
16. Singh, K.P., Malik, A., Sinha, S.: Water quality assessment and apportionment of pollution
sources of Gomti River (India) using multivariate statistical techniques: a case study.
Analytica Chimica Acta 538(1-2), 355–374 (2005)
 Clustering Analysis of Water Quality for Canals in Bangkok, Thailand 227

17. Vega, M., Pardo, R., Barrado, E., Deban, L.: Assessment of seasonal and polluting effects
on the quality of river water byex ploratory data analysis. Water Research 32(12), 3581–
3592 (1998)
18. Bezdek, J.C.: Fuzzy partition and relations: an axiomatic basic for clustering. Fuzzy Sets
and Systems 1, 111–127 (1978)
19. Liou, S., Lo, S., Hu, C.: Application of two-stage fuzzy set theory to river quality
evaluation in Taiwan. Water Research 37, 1406–1416 (2003)
20. Ministry of Natural Resource and Environment, http://www.mnre.go.th
21. Department of Drainage and Sewerage Bangkok Metropolitan Administration,
http://dds.bangkok.go.th/wqm/thai/home.html
22. Department of Pollution Control Bangkok, http://www.pcd.go.th
23. Ministry of Public Health, http://eng.moph.go.th
24. Luke, B.: K-Means Clustering,
http://fconyx.ncifcrf.gov/~lukeb/kmeans.html
 A Review of Routing Protocols for UWB MANETs

Yahia Hasan Jazyah and Martin Hope

Centre for Network and Telecommunications Research

The University of Salford
Salford, UK
y.h.jazyah @pgr.salford.ac.uk, [email protected]

Abstract. Mobile Ad hoc Networks (MANETs) have witnessed an increasing

amount of interest during the last decade. Hosts (or nodes) within the network
are mobile, and each node is equipped with a short range transmitter and re-
ceiver, antenna, and local power supply. Nodes then operate as a router to relay
messages from the sender to the receiver, and they can be organised into differ-
ent topologies; for example: they can be flat or hierarchical, they can move in
any direction and speed, and they can communicate with each other through
wireless routing protocols. More recently, systems based on Ultra-WideBand
(UWB) technology have become a promising candidate in application to
MANETs. This is mainly due to their powerful capabilities, such as their high
data rates and low power consumption, and although many routing protocols
have been designed for ad-hoc networks, few have considered their application
to UWB based MANETs. This paper presents a review of different types of
wireless routing protocols for MANETs with a concluding emphasis on their
application to UWB based systems.

Keywords: MANET; UWB; routing protocol.

1 Introduction
The practical application of MANETs is sometimes preferable over infrastructure base
networks due to the unavailability or high cost of wired infrastructure. They can be
used effectively in military applications such as the battle field, where the existence of
wired infrastructure is impossible. They are also very useful in industrial or commer-
cial settings involving cooperative mobile data exchange, or when establishing com-
munications in a disaster recovery situation such as a fire or rescue operation.
MANETs function in the absence of a base station where nodes communicate di-
rectly without the need of a centralised control system. Wireless routing protocols en-
able nodes within the MANET to relay data packets around the network. If nodes are
within transmission range, they can communicate directly, or if they are away from
each other, intermediate nodes are required to establish a multihop route between
sender and receiver.
Now it is the time for UWB to be given more attention due to its great capability in
high data rate, large BW, and low power consumption. Currently, hosts are equipped
with UWB system in order to achieve short range communication and exploit the

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 228–245, 2010.
© Springer-Verlag Berlin Heidelberg 2010
 A Review of Routing Protocols for UWB MANETs 229

previous mentioned advantages, besides UWB has the ability to measure distance be-
tween hosts accurately.
As a result, researchers have been encouraged to design new routing protocols for
UWB networks to make it suitable for hosts to exploit features of UWB, but still needs
more and more work to overcome drawbacks of network such as overhead, power con-
sumption, unreliable route, etc…
UWB [1], [2] is a radio technology proposed to be used in personal area networks
(PAN) and appeared in the IEEE 802.15.3a draft PAN standard. UWB systems con-
sume very low energy levels, and can be used in short-range, high BW (BW>500
MHz, 20% of the center frequency). UWB operates in range 3.1–10.6 GHz (the mini-
mum BW is 55MHz). The maximum allowable power spectral density is -41.3
dBm/MHz corresponding to average transmitted power of 0.5 mW when operating in
band range 3.1-10.6 GHz.
UWB is best used for ad-hoc and sensor networks, it is used as part of location sys-
tems and real time location systems such as hospitals and healthcare, short broadcast
time, “see-through-the-wall” precision radar imaging technology, and replacing cables
between portable multimedia Consumer Electronics (CE) devices and in next-
generation Bluetooth Technology devices [3].
It provides relatively long radio range which is around 150 meters indoor and 1 km
outdoor and high data rate in excess of 100 Mbps with bit rates 55, 110 & 200 Mbps.
UWB uses pulse coded information with sharp carrier pulses (pulse-based systems),
the signal consists of a train of pulses, one single pulse may be used to represent
one bit.
There is duration between single pulses to avoid both pulse overlapping and ambi-
guity between pulse positions [3]. Each pulse occupies the entire UWB bandwidth,
thus resist the multipath fading. UWB systems are carrierless systems based on pulses
of very short duration transmission [4].
Generically, wireless routing protocols in MANETs can be classified into topology
and position based protocols. These are presented in sections II and III, and of particu-
lar relevance to UWB systems, section IV then considers power aware routing proto-
cols and then section V highlights the issues associated with routing protocols for
UWB applications.

2 Topology Based Routing Protocols

Topology based routing protocols [5] use existing information about the network to
flood (or forward) packets. There are two main routing strategies which are classified
as topology based and these are: proactive and reactive routing protocols. In addition,
hybrid protocols also exist which mix the previous two types.
A. Proactive Routing Protocols
Proactive (or table driven) protocols [5] maintain routing information for each node in
the network and store information in routing tables. This information is then updated
whenever the topology changes and so one or more routing tables are required by each
node to store routing information. Most proactive strategies share the same features,
but they differ in the number of routing tables and frequency of topological update.
Examples of proactive routing protocols are Destination-Sequenced Distance Vector
230 Y.H. Jazyah and M. Hope

(DSDV) [6], Cluster-head Gateway Switch Routing (CGSR) [7], Wireless Routing
Protocol (WRP) [8], Optimized Link State Routing (OLSR) [9], and Hierarchical State
Routing (HSR) [10]. The first three protocols are discussed bellow:
a) Destination-Sequenced Distance Vector (DSDV)
DSDV [6] is a flat proactive routing protocol which guarantees loop freedom, single
shortest path to destination. Each node forwards its information among the network;
neighbor nodes (intermediate nodes) receive this information and update their routing
tables. Thus, all nodes have routes to all reachable nodes in the network in advance.
DSDV uses a sequence number to prevent loops; here a node will not forward in-
formation with the same sequence number twice. The higher the sequence number the
newer the route; nodes update their routing tables whenever they receive route infor-
mation that contains a higher sequence number than the stored one in routing table, or
a route with same sequence number but shorter than the stored one.
When a source node wants to send data to destination, it checks its routing table and
selects the shortest path to the destination. DSDV uses two types of packets; Full
Dump packets which carry all available routing information and can require multiple
network protocol data units (NPDUs), and a route Broadcast packet which contains the
address of destination, number of hops, and the sequence number of the information
received regarding the destination.
DSDV can not work in scalable network, especially when nodes move; this causes a
high control packet update which increases network overhead.
b) Cluster-head Gateway Switch Routing (CGSR)
CGSR [7] is a hierarchical proactive routing protocol. Nodes are arranged in clusters, a
cluster head (CH) is elected in each cluster by a distributed algorithm. Transmission is
then achieved by CH inside clusters, which reduces overhead in network, but nodes
within cluster are busy in selecting CH rather than packet relaying; which affects pro-
tocol performance. While transmission between clusters is achieved by gateway nodes
that exist between the transmission ranges of two or more CHs as shown in Fig. 1.

Gateway

Cluster Head
Member node

Fig. 1. Cluster Based in CGSR Routing Protocol

 A Review of Routing Protocols for UWB MANETs 231

Each node maintains two tables; a cluster member table for storing destination CH
of every node in the network, this table is broadcasted using the DSDV algorithm, and
another routing table for storing neighbor nodes.
c) Wireless Routing Protocol (WRP)
WRP [8] is a flat proactive routing protocol which guarantees loop freedom by consid-
ering predecessor information. It uses four routing tables; distance table, routing table,
link-cost table, and Message Retransmission List (MRL), table and thus it consumes
memory.
Each record in MRL contains a sequence number, retransmission counter, acknowl-
edgement required flag vector, and a list of update message that contains the destina-
tion, the distance to destination, and the predecessor of destination. An update message
is sent in two cases; after processing updates from neighbours or when detecting a
change in a link to neighbour.
WRP checks for nodes connectivity by sending hello messages. When a node re-
ceives hello message, it adds the node to its routing table, and then it sends a copy of
its routing table to the node. This process causes overhead and consumes power.
d) Summary
Proactive routing protocols, in general, decrease end to end delay, but provide the
shortest path. On the other hand, they increase network overhead, consume memory
and power, and so, they are only suitable to small size networks.
B. Reactive Routing Protocols
Reactive routing protocols maintain route information on demand, i.e. when source
node wants to send message to destination, it initiates route request (RREQ) to find
route to destination. When a route fails, a route maintenance process is launched to
repair the failed route.
Ad hoc On-Demand Distance Vector (AODV) [11], Dynamic Source Routing
(DSR) [12], Temporally Ordered Routing Algorithm (TORA) [8], and Associativity-
Based Routing (ABR) [8] are examples of reactive strategy.
More discussion about AODV and DSR and TORA is bellow:
a) Ad hoc On-Demand Distance Vector (AODV)
AODV [9] belongs to the class of Distance Vector (DV) routing protocols. In a DV,
every node knows about its neighbours and the costs incurred in order to reach them
using the Bellman-Ford algorithm [10].
AODV is a reactive shortest single path wireless routing protocol based on DSDV
protocol. When a source wants to send a message to a destination, it checks its routing
table, if there is a valid route to destination, it starts sending packets immediately. If
not, it broadcasts RREQ to all neighbor nodes. It should be noted that RREQ contains
the fields: hop count, source and destination sequence numbers, destination and source
addresses, RREQ ID, and other pre-determined fields. When an intermediate node re-
ceives a RREQ, it checks its routing table for a path to the destination, if it exists, it
unicasts a route reply (RREP) to the source, otherwise, it increases the hop count by
one, and adds its ID to the RREQ and then re-broadcasts it to its neighbours; the proc-
ess continues until the RREQ reaches its destination. The destination then selects the
first coming RREQ and unicasts RREP using the reverse path to source node. When
232 Y.H. Jazyah and M. Hope

the source receives several RREPs, it selects the route of the highest sequence number
and minimum hop count, and then it establishes the route and starts sending packets.
To guarantee loop freedom, the source node uses a sequence number and includes it
in RREQ. When a node receives a control message (RREQ, RREP, or RERR), it
checks its routing table for an entry to the specified destination, if there is no entry in
the routing table about the destination, it creates a new one. If there is an entry in the
routing table, the route is only updated if the new sequence number is either higher
than the destination sequence number in the routing table, the sequence numbers are
equal, but the hop count plus one is smaller than the existing hop count in the routing
table, or the sequence number is unknown. In addition, the source uses a time to live
(TTL) count to limit the flooding of RREQ packets and control the overhead associ-
ated with the network. Finally, a HELLO message is broadcasted periodically to in-
form neighbor nodes about node existence. When an active node (node on the active
route) detects a route failure (the neighbor node is unreachable; i.e. HELLO is not be-
ing received), it sends a route error (RERR) packet to the source node, which in turn,
initiates a new RREQ.
Overhead is a major drawback of AODV because of the flooding of control mes-
sages (HELLO and RREQ packets). However, the overhead is low when compared to
proactive routing protocols, and it also needs less memory, on the other hand it only
stores the routes on demand so, it is preferable in application to large scale networks.
b) Dynamic Source Routing (DSR)
DSR [13] is a reactive routing protocol; when a source node wants to send data to a
destination address; it initiates and broadcasts a RREQ that carries the full address
(every hop). When intermediate receives the RREQ, it checks its routing table (nodes
may store many routes to destination), if there is a valid path to destination, it sends
RREP to source, or it forwards the packet after adding its address if its address is not
existed in route record in order to avoid loop, and so on until RREQ reaches to destina-
tion which sends RREP to source using the reverse path.
A link failure is detected when no acknowledgement received by a node. RERR
message is sent to source, which in turn, looks for a valid route in its cache, if no valid
route is existed, source initiates RREQ again.
DSR does not use hello message, and so, it reduces the overhead and saves band-
width, which makes it suitable for scalable network. On the other hand, packet size
increases when route is long and network is large since every node adds its address to
route record of RREQ.
c) Temporally Ordered Routing Algorithm (TORA)
TORA is a distributed and reactive routing protocol. It is suitable for highly dynamic
topology network, it provides acyclic multipath but it suffers from long delay and does
not support shortest path calculation. It is based on directed acyclic graph to maintain
loop freedom.
When source want to forward packets to destination, it uses a height metric to estab-
lish direct acyclic graph and each node maintains information about one hop neighbors
in order to localize control message to a very small set of nodes near the occurrence to
topology change. When a link fails, a node generates a reference level which propa-
gates by neighbor nodes.
 A Review of Routing Protocols for UWB MANETs 233

d) Summary
Reactive routing protocols reduce network overhead, consume less bandwidth, and
need less memory when compared to proactive protocols because they establish the
route on demand. But they need a longer time than proactive protocols to establish the
route between source and destination, and the also need to route provide maintenance
because there are no ready available backup routes in general. In general, reactive rout-
ing protocols are suitable to large scale network compared to proactive ones.
Topology based routing protocols use flat addressing scheme as mentioned in the
previous examples except CGSR which uses hierarchical addressing.
C. Hybrid Routing Protocols
Most networks are partitioned into zones (or zone-based) or nodes are grouped into
trees or clusters. Hybrid routing protocols merge proactive and reactive strategies
whereas proactive strategy is applied within zone (or intra zone), while a reactive strat-
egy is applied outside zone (or inter zone). The following are examples of hybrid rout-
ing protocols:
a) Zone Routing Protocol (ZRP)
In ZRP [14], network is divided into routing zones, and each node belongs to a zone
based on its geographical location. ZRP reduces the overhead of proactive protocols and
the delay associated with reactive protocols, and so, it is suitable for large networks.

Fig. 2. Categorization of Ad hoc Routing Protocols

Zone radius is considered by number of hops. The selection of radius is a tradeoff

between the routing efficiency of proactive routing and the increasing traffic in zone.
Source node looks for destination within zone proactively, if it is not existed, it
sends RREQ reactively outside zone, if receiving node knows the path to destination, it
sends RREP to source, otherwise, it rebroadcasts RREQ after adding its address until it
reaches to destination.
In the case of route maintenance, packets can be directed around the broken link
through an active multi-hop path.
b) Zone-based hierarchical link state (ZHLS)
ZHLS [7] is a hybrid hierarchical routing protocol acting similar to ZRP. It is proactive
when Intra-zone Clustering, while it is reactive when Inter-zone Clustering. Each node
has a node ID and zone ID based on Global Positioning System (GPS).
234 Y.H. Jazyah and M. Hope

The network is divided in non overlapping zones without zone head which reduces
overhead. Zone size depends on node mobility, network density, transmission power
and propagation characteristics.
If destination is moving within the zone, no need to search in other locations since
node ID and zone ID of destination are required for routing, and so, ZHLS may scale
well to large networks; nevertheless, all nodes must have a pre-programmed static zone
map, while geographical boundary of network is dynamic.
c) Summary
Hybrid routing protocols merge both proactive and reactive approaches together by
utilizing advantages of both of them. They reduce the overhead of proactive protocols
by applying reactive routing between zones and limiting flooding inside zones, and the
delay associated with reactive protocols by applying proactive routing inside zones,
thus it is suitable for large scale networks.

3 Position-Based Routing
Position based routing focuses on two issues: a location service to determine destina-
tion's position and forwarding strategy that can be greedy forwarding, restricted direc-
tional flooding, and hierarchical routing.
D. Location Services
Location service is used by sender node to learn the position information of a certain
node in order to include it in the forwarded packet if it does not know the position of
communication target. Each node registers its current position to a location service,
when a node needs the position information of another one; it sends a request to loca-
tion service. For example, Quorum-Based Location Service [5], Home-Zone [5],
Distance Routing Effect Algorithm for Mobility (DREAM) [15], and Grid Location
Service (GLS) [5], [16]. The last two location services are discussed bellow:
a) Distance Routing Effect Algorithm for Mobility (DREAM)
Each node maintains a position database that stores position information of each other
nodes that are part of the network (all-for-all), where packets are regularly flooded to
update position information [5].
DREAM [15] combines proactive and reactive approaches. Source node has to have
positions of it, destination, one hop neighbours, and the maximum speed of destination.
Source node uses this information to determine the direction towards destination in or-
der to send packets by finding one hop neighbor within the range between source and
destination according to a certain angle determined by speed and position of destination.
b) Grid Location Service (GLS)
GLS [5], [16] divides network into a hierarchy of squares. In each hierarchy, n-order
squares contain exactly four (n – 1)-order squares, forming a "quadtree". Each mobile
node periodically updates a small set of other nodes with their current locations by
maintaining a table of all other nodes within the local first-order square. (Classified as
all-for-some)
 A Review of Routing Protocols for UWB MANETs 235

GLS uses geographical forwarding. Source must know the geographical positions of
any destination, and must label packets for that destination with its position, while in-
termediate node only needs to know its position and the positions of nearby nodes.
Geographical forwarding is similar to Cartesian routing. Each node broadcast a
HELLO message that contains its position and velocity; where neighbours within
transmission range store this information in a table. Node sends packet to the closest
neighbor of destination after checking the neighbor table. When a node does not find
closest node other than itself, it sends error indicating a hole in geographic distribution
of nodes. Greedy Perimeter Stateless Routing (GPSR) is a solution for the problem of
hole.
GLS is based on the idea that a node maintains its current location in a number of
location servers (each node acts as a location server on behalf of some other nodes)
distributed through over network, but they are no elected leaders. GLS makes balance
when location servers are selected and dense of network is considered.
c) Summary
Location service is used by source node to determine position of destination. Location
services are classified according to “how many servers maintain position information
about mobile nodes “into all for all, all for some, some for all, and some for some. For
example, when all nodes within the network have information about all nodes in the
network, then the location service is classified as “all for all”.
E. Position-Based Routing Protocols
Position based routing protocols use position information gathered by GPS system.
Each node in the wireless network knows its location, destination location, and other
geographical information in order to forward packets in geographical direction instead
of flooding as in topology based protocols.
Mainly, there are three forwarding strategies: First, greedy packet forwarding;
whereas the sender of packet includes the approximate position of recipient in the
packet, intermediate node forwards the packet to a neighbor which is lying in the gen-
eral direction of the recipient. Second, restricted directional flooding; the packet is
flooded in one-hop direction towards the destination where it is expected to be located;
this process is repeated until the packet reaches the destination. The direction is deter-
mined by calculating the expected region that covers destination. Third, hierarchical
Routing; where the complexity that each node has to handle can be reduced massively
by establishing some form of hierarchy.
Many position based routing protocols have been designed such as, Location Aided
Routing (LAR) [17], GRID [18], Greedy Perimeter Stateless Routing (GPSR) [19],
Adaptive Location Aided Roaming from Mines Protocol (ALARM) [20], Location-
Aware routing Protocol with Dynamic Adaptive of Request Zone (LARDAR)
[21],Compass [22],Angle Routing PROTOCOL (ARP) [23],Position-based Selective
Flooding (PSF) [24], Octopus [25], Improved Location-Aided Routing (ILAR) [26] ,
and Novel Routing Power Balance (NRMPB) [27] . The following is a discussion
about position based routing protocols.
a) Location Aided Routing Protocol (LAR)
LAR [17] is an on-demand routing protocol that uses the modified Dijkstra's Algorithm
to find the shortest path.
236 Y.H. Jazyah and M. Hope

Destination lies in the centre of a circular region of certain radius at a certain time
known as Expected Zone, as shown in Fig. 3, which indicates which zone of the net-
work should be reached by RREQ. GPS enables terminals to know its own position
and speed, while dissemination is performed by piggybacking location information in
all routing packets.
“Piggybacking is a bi-directional data transmission technique in the network layer.
It makes the most of the sent data frames from receiver to emitter, and adds the con-
firmation that the data frame sent by the sender was received successfully acknowl-
edge.” [28].
There are two proposed schemes of LAR:
First, LAR1 defines request zone that includes sender and receiver on opposite cor-
ner of a rectangle. Rectangle dimensions are estimated according to the receiver aver-
age speed at a certain time. Only nodes within the zone respond to the RREQ of
sender. LAR1 reduces network overhead, but causes delay when route is found.
Different shapes can be used other than rectangle such as circle, cone, and bar.
LAR2 is proposed based on the calculated distance between source position and the
estimated position of destination. When node receives RREQ, it calculates the distance
toward destination, if the distance is less than the distance from the previous sender
node to destination; it forwards the packet as shown in fig. 4. In this scheme, interme-
diate receiving node may be the closest node to destination, and so the algorithm
reaches to a dead-end.
b) GRID
GRID [18] is a reactive routing protocol which exploits location information. GRID
has strong maintenance capability; where intermediate node that attends to move per-
forms handoff operation to keep route alive.

Expected Zone

Fig. 3. Expected Zone in LAR

Destination
Source

Fig. 4. LAR Scheme 2

 A Review of Routing Protocols for UWB MANETs 237

Geographical area is divided into square grids; the nearest node to the centre of grid
is elected as a leader (gateway) to maintain routing. If leader leaves a grid, it broad-
casts routing table to the new leader. Position information of source and destination are
used to confine.
The search area, leader is responsible for searching procedure and setting up routing
table in grid by a grid manner; routing in GRID is performed grid by grid through
grids' leaders.
Source node calculates the searching range depending on its location and the loca-
tion, speed, and direction of destination in order to eliminate the effect of collection of
redundant rebroadcast, heavy contention, and collisions (problem of broadcast storm).
c) Greedy Perimeter Stateless Routing (GPSR)
In GPSR [19], the closest intermediate node to destination is selected as next hop, but
GPSR does not guarantee the existence of path, and so perimeter forwarding selects the
neighbor which is farther than itself from the destination in order to solve the dead-end
caused by greedy forwarding, GPSR switches between these two strategies. A modifi-
cation to GPSR is made by using flooding to solve the problem of dead-end.
Each node needs information about next hop only, which is exchanged by means of
periodic beacons broadcasted by each node.
d) Octopus
Octopus [25] is a fault-tolerant and efficient position-based routing protocol. Fault-
tolerance is achieved by employing redundancy (storing the location of each node at
many other nodes) and by keeping frequently refreshed soft state; nevertheless, it
achieves a low location update overhead by employing a novel aggregation technique.
Octopus is a typical routing scheme, assumes that each node gets its own location
information using like GPS. It divides the network into grid by horizontal and vertical
strips, each node maintains a table for neighbors (updated by hello message) and a ta-
ble for strips. Octopus forward RREQ in north and south directions to find destination,
if this process fails, it sends the RREQ to the horizontal direction. Octopus uses
Greedy Forwarding to forward data packets.
e) Improved Location-Aided Routing (ILAR)
ILAR [26] improves location-aided routing (LAR); when intermediate node receives
RREQ, it checks weather it is within the requested zone, if it is, it will send RREQ revise
(RREQ_R) to the transmitting node, RREQ_R includes the VDIST (vertical distance, it
is the distance from node to the line connecting source and destination; base line) as
shown in fig. 5. The node which receives RREQ_R will select the next hop which has the
smallest VDIST and has a distance to source larger than to the previous node.
This technique has less packets and better performance than LAR.
f) Novel Routing Power Balance (NRMPB)
NRMPB [27] resolves the problems of ILAR which are: the movement of nodes near
border of coverage broadcast which cause breakage of route, the other problem is se-
lecting the next hop node; NPMPB comply position information and power saving to
balance the energy of each nodes and prolong lifetime of network.
In ILAR protocol, the intermediate node which has the smallest vertical distance is
selected as next hop, but the selected node could be located near the boundary of
238 Y.H. Jazyah and M. Hope

transmission range or predecessor node, which cause unreliable (unstable) link.

NRMPB solves this problem besides considering power saving.
When source wants to send packet to destination, it initiates RREQ packet including
a field of wait time, each receiving node calculates the wait time and replaces it in the
RREQ packet before forwarding. The wait time is reduced and when it counts to zero,
node detects an error in path.
When source or intermediate node wants to find out the next hop, it sends route re-
quest transmit (RREQ_T) packet, which contains source and destination’s locations, all
nodes receive the packet will calculate their position with respect to source’s position,
if they are located within 80% of transmission range of source, they will reply by route
request revise (RREQ_R) including the power of node which is calculated by a certain
equation, the equation depends on distance between source and destination, distance
between node and source, distance between node and the line connecting source and
destination (baseline), see fig. 5, and node’s power.
Source selects the node of maximum power within the 80% of its coverage range,
and unicasts RREQ to it, the process continues until RREQ reaches to destination
which replies by RREP using the reversed path to source node.
When a node which sent RREQ can not reach the next hop, it will try to find an-
other next hop, if it fails, it will send RERR to the previous node, and so on. Another
technique is used when node does not receive RREP from destination, it will forward
RREQ_T.

80% of transmission range

Destination
VDIST

Baseline

Source

Fig. 5. VDIST in NRMPB and ILAR Routing Protocols

g) Summary
Position based routing protocols eliminates overhead of reactive protocols by directing
flood of route request towards the expected location of destination, which leads to bet-
ter performance in all aspects; power consumption, stability, reliability, maintenance,
and bandwidth utilization. These specifications make position-based routing widely
trusted to large scale networks.

4 Power-Aware Routing
Power metrics are important issues when routing protocols are designed since mobile
nodes have limited resource of power which is represented by local batteries.
 A Review of Routing Protocols for UWB MANETs 239

Generally, energy efficient issues are: 1) Sleeping mode; whereas nodes turn them-
selves off when not actively transmitting or receiving (sleep/awake). 2) Transmission
Power Control; which aims to minimize the power per bit required to transmit packet
from source to destination, this technique always selects the least-power cost routes; as
a result, nodes along those routes tend to die soon because of battery energy exhaus-
tion. 3) Power-aware route selection that find route at certain route discovery time such
that a cost function is minimized. Cost function depends on energy consumed per
packet, time to network partition, variance in node power levels, cost per packet, and
node cost metrics. And 4 ) Broadcast control.
Examples of power-aware routing protocols are: Power-Aware Multiple Access
Protocol (PAMAS) [29], Self-Healing-AODV (SHAODV) [30], Energy-Efficient Ad
hoc Routing (DEAR) [31], Energy-Aware Ad hoc On-Demand Distance Vector
(EAAODV) [32], Active Route Maintenance Protocol (ARMP) [33], and Distributed
Power Control (DPC) [34]. The following are more details about routing protocols
related to power.
a) Power-Aware Multiple Access Protocol (PAMAS)
PAMAS [29] is a shortest-cost routing algorithm, power-Aware Multiple Access pro-
tocol with Signalling, it uses a new power-aware metrics for determining routes in
wireless ad hoc networks based on battery power consumption of node.
Nodes’ radios are turned off (sleep mode) when they can not transmit or receive
packets during transmission period of other nodes; which in turn saves their power.
b) Self-Healing-AODV (SHAODV)
SHAODV [30] overcomes the problem of edge effect and route break without causing
heavy overhead. The path can be established for long lifetime by connection admission
control of route request. The unstable route can be healed without route break, flood
over the network, and overhead by link state inspection and active route local repair.
Signal to Noise Ratio (SNR) is measured and compared to a threshold in order to
overcome the overhead; it is calculated according to the received power which is based
on transmitted power, height and gain of antenna at both sides of transmitter and re-
ceiver, and distance between them.
Movable node requests the source to heal the unstable path before it moves and
causes break in the route. Process of reduction the repaired path is initiated after that by
broadcasting RREQ for reduction along the repaired path and accepting the first re-
ceived RREQ packet with link state value (SNR) greater than a threshold.
c) Energy-Efficient Ad hoc Routing (DEAR)
DEAR [31] controls the rebroadcast time of RREQ packets to prolong network lifetime
by balancing between minimum transmission energy consumption and fair node en-
ergy consumption in a distributed manner without additional control packets. It also
improves data packet delivery ratio.
The physical layer can save energy by adapting transmission power according to the
distance between nodes. While at the data link layer, energy conservation can be
achieved by sleep mode operation.
Network lifetime is defined as the time when a node runs out of its own battery
power for the first time.
240 Y.H. Jazyah and M. Hope

DEAR utilizes only the RREQ message to distribute the average residual battery
level of the entire network. Intermediate nodes control the rebroadcast time of the
RREQ packet whereas retransmission time is proportional to the ratio of average resid-
ual battery power of the entire network to its own residual battery power, and so, nodes
with larger battery energy will rebroadcast RREQ packets earlier.
Residual battery level of the entire network depends on the average residual power
of route, which in turn depends on the residual power of node and hop count. DEAR
control the rebroadcast time; if the residual battery power of node is smaller than the
average network residual power, the retransmission time will be longer and vice versa.
DEAR compromises between the min-hop path and the fair energy consumption path.
d) Energy-Aware Ad hoc On-Demand Distance Vector (EAAODV)
EAAODV [32] combines the on-demand routing capability of the AODV routing pro-
tocol with a distributed topology discovery mechanism – mobile agents, thus reaching
the energy balance condition. EAAODV eliminates the problem of low efficiency and
link breaking in routing due to the energy exhaustion of some nodes.
"A mobile agent is a software program that can suspend its execution on the host
computer, transfer itself to another agent-enabled host through the network, and
resume its execution on the new host." The key features of mobile agents which distin-
guish them from traditional distributed programming are: mobility; network aware-
ness; communication; intelligence; reactivity; autonomous; goal-oriented; temporally
continuous; learning; and flexible.
When intermediate node receives a RREQ, it decided to broadcast it according to its
residual energy, if it is under a threshold level, the node terminates the RREQ, other-
wise it broadcasts the RREQ. During the search process, the node checks its routing
table, if the table is fresh; the residual energy information of the next-hop node indi-
cated by the table is acquired by mobile agents, if the energy is above the threshold, the
routing table is valid and it will be used, otherwise, searching for the routing will be
finished, if the routing table is old, nodes will broadcast RREQ to their neighbors and
update the information of the table.
EAAODV is able to achieve longer surviving network duration and a larger
throughput compared to AODV.
e) Summary
Methods that concern power control issue can be classified into three categories: meth-
ods find the optimal transmitted power to control connectivity properties of network,
power aware routing, and algorithms based on shortest path rather than hop count. In
general, route maintenance is detected according to a threshold value which is com-
pared to a certain power metric.

5 Routing Protocols for UWB Networks

UWB is a radio technology that can be used at a very low energy levels, short-range,
high-bandwidth (>500 MHz, 20% of centre frequency) communications by using a
large portion of the radio spectrum. UWB has been proposed as technology to use in
Personal Area Networks (PANs). Many routing protocols have been designed to
exploit features of UWB system such as Piconet Formation Algorithm (PFA) [35],
 A Review of Routing Protocols for UWB MANETs 241

UWBLAR [36] , Clustered-Multi-Hop (CMulti-Hop) [37] , and as mentioned in papers

[38] [39] [40] [41] [42] . The following are more details about some of routing proto-
cols dedicated to UWB MANETs.
a) Piconet Formation Algorithm (PFA)
PFA [35] is a routing protocol for UWB networks that adopts a master-slave configu-
ration and tries to minimize the total transmission power of master nodes and interfer-
ence within piconets. Master nodes are selected according to the minimum average
distance to neighbouring nodes and forms piconets that have the minimum total emis-
sion power for a given radio range. Master-slave configuration provides good perform-
ance in terms of control overhead, power savings, and overall efficiency. On the other
hand, centralized schemes cannot be scaled to a large number of nodes or over a large
distance. Hybrid approach can overcome this weak point.
PFA can reduce average interference generated per piconet, but it suffers from a
load balancing problem; if there are too many nodes joining one cluster, the cluster
head (master) may have to limit data rates to an undesirable level, thus performance
may degrade. Moreover, a cluster with high total transmit power will generate high
interference in neighbouring clusters. The solution of this problem is to impose an up-
per limit of nodes joining the cluster as proposed in [42].
b) UWBLAR
UWBLAR [36] is a position based routing strategy that sets up a dynamic routing list
for an ad hoc network based on a UWB system.
The network is organised in a cluster manner where each node knows the exact po-
sition of every node in its cluster, when a connection request is generated; the cluster
head sets up a table of distance information by broadcasting RREQ to all network
nodes, the shortest path is then selected by the destination when it receives several
RREQ.
c) Clustered-Multi-Hop (CMulti-Hop)
CMulti-Hop [37] is a new strategy for routing in UWB based ad-hoc networks that
depend on minimizing power-dependent global cost function.
The model considers fixed nodes equipped by UWB systems, each node belongs to
a cluster and knows the position of other nodes within the cluster, where cluster’s size
depends on radio channel condition and network load.
The first node of cluster is labelled by zero, the second level by one, on so on. A
node set up a link with only a node of lower level, while no link between nodes within
the same cluster is established in order to guarantee loop freedom.
Path of the lowest cost function is selected. If two paths have the same value, the
path of minimum hop count is selected. If none of paths has a cost value greater than
predefined cost value, the connection request is rejected.
d) Other Routing Strategies for UWB MANET
Paper [38] presents a position based routing strategy which exploits the high preci-
sion ranging capabilities offered by UWB In order to reduce the emitted power and
multi-user interference.
Author considers the cost of route is the sum of the costs of all links that form the
route, which comprises signaling cost for setting-up a new link, cost for transmitting
242 Y.H. Jazyah and M. Hope

data which depends upon the requested data rate, both represent power consumption,
and distance between two nodes. Besides Signal to Noise Ratio (SNR), Symbol Error
Probability (SEP), and Packet Error Probability (PEP) are taken into account when
multi-user interference (MUI) is considered.
The above technique is applied to DSR routing protocol. Whereas each intermedi-
ate node adds the cost of previous hop to the path cost. Destination selects the path
according to the cost information included in RREQ. Also it is applied to LAR rout-
ing protocol.
Adoption of position-aware routing strategies significantly improves the efficiency
in power management for UWB networks, for example, by optimizing the path selec-
tion procedure in DSR and trade-off between performance and power efficiency can
be achieved by adopting different shapes for the Request Zone of LAR.
Paper [39] proposes power-efficient routing strategy based on the ranging capabili-
ties offered by UWB.
The cost function of link is considered as in [38]. Minimum Cost strategy leads to
more robust system in a high-interference environment; the Minimum Cost strategy is
not effective in the case of low network connectivity since it increases the average
number of hops with no advantage in system performance. When network connec-
tivity grows, the Minimum Cost effectively increases the percentage of found connec-
tions by reducing the effect of multi-user interference.
Paper [41] proposes a novel packet routing scheme for UWB ad-hoc network. This
approach utilizes the distance measurement between nodes that UWB provides,
adopting the multihop transmission, determined spreading factor and the signal to
interference and noise ratio (SINR).
Spreading Factor (SF) is the ratio of the chips to baseband information rate.
Spreading factors vary from 4 to 512 in FDD UMTS. Spreading factor in dBs indi-
cates the process gain, whereas the lower the spreading factor the higher the data rate.
Source node searches for node within its transmission range provided that the dis-
tance to destination is greater than the distance between the previous node and desti-
nation, the process is repeated by each node until destination is reached.
e) Summary
UWB is a new technology that is in MANET to transmit data packets in high speed.
Routing protocols for UWB MANET have been designed to use the high performance
and capabilities of UWB; such routing protocols are designed as master slave in general
to overcome network overhead and reduce power consumption, other routing techniques
consider a cost function that is accumulated until the RREQ reaches its destination,
whereas the destination selects the route of cost function that satisfy the criteria.
Many routing protocols have been designed for UWB MANET, nevertheless the
field needs more and more effort by researcher to design new protocols that can use
the capabilities of UWB.

6 Conclusion
This work summarises wireless routing protocols for MANETs which can be classified
into topology based and position based protocols. Topology based protocols are either
 A Review of Routing Protocols for UWB MANETs 243

proactive or reactive, while hybrid strategies merge proactive and reactive approaches
in order to enhance the drawback of proactive protocols which utilize the advantages of
reactive ones.
Position based strategies are based on positional information gathered by GPS sys-
tems. They eliminate the overhead caused by topology based strategies by directing
flooding towards the destination.
Power metrics are included in both topology and position based routing in order to
enhance the performance of routing protocols in terms of power saving. They also take
part in route maintenance as detector of route failure before it occurs, or as a metric to
select the next hop.
UWB is a radio technology proposed to use in PANs and has appeared in the IEEE
802.15.3a draft PAN standard. UWB systems consume very low energy levels, and
they are applicable to short-range high data rate systems which make it particularly
useful to MANETs and sensor networks. Many conventional routing protocols have
been designed for MANETs. However routing strategies for UWB networks need
more work but will most likely employ aspect of power aware routing and geographic
position.

References
[1] Xia, L., Lo, A., Niemegeers, I., Baugé, T.: An Ultra-Wide Band Based Ad Hoc Network-
ing Scheme for Personnel Tracking in Emergencies. In: This work was performed in the
IST Framework 6 EUROPCOM project
[2] Kuladinithi, K., Timm-Giel, A., Görg, C.: Mobile Ad Hoc Communications in AEC In-
dustry (August 2004), http://www.itcon.org/2004/22/
[3] Benedetto, M., Giancola, G., Domenicali, D.: Ultra Wide Band radio in Distributed
Wireless Networks. IEEE, Los Alamitos (2006)
[4] Project SPEARHUW, Enhanced Sensor and WLAN Networks using UWB Communica-
tions (April 2005)
[5] Mauve, M., Widmer, J., Hartenstein, H.: A Survey on Position-Based Routing in Mobile
Ad Hoc Networks. IEEE Networks 15(6), 30–39 (2001)
[6] Perkins, C.E., Bhagwat, P.: Highly Dynamic Destination-Sequenced Distance-Vector
Routing (DSDV) for Mobile Computers. ACM SIGCOMM Computer Communication
Review 24(4), 234–244 (1994)
[7] Abolhasan, M., Wysocki, T., Dutkiewicz, E.: A review of routing protocols for mobile ad
hoc networks. Ad Hoc Networks 2, 1–22 (2003)
[8] Royer, E.M., Santa, B., Toh, C.: A Review of Current Routing Protocols for Ad hoc Mo-
bile Wireless Networks. IEEE Wireless Communications 6(2), 46–55 (1999)
[9] Clausen, T., Jacquet, P., Laouiti, A., Muhlethaler, P., Qayyum, A., Viennot, L.: Opti-
mized Link State Routing Protocol. In: Proceedings of IEEE INMIC (2001)
[10] Kilinkaridis, T.: Routing Protocols for Wireless Ad Hoc Networks Hierarchical routing
protocols
[11] Perkins, C., Royer, E.B., Das, S.: Ad hoc On-Demand Distance Vector (AODV) Routing
(2003), http://www.ietf.org/rfc/rfc3561.txt
[12] Baumann, R.: AODV, Ad hoc On Demand Distance Vector Routing Protocol. AODV
Presentation at ETH, Zürich (2002)
244 Y.H. Jazyah and M. Hope

[13] Johnson, D.B., Maltz, D.A., Broch, J.: DSR: The Dynamic Source Routing Protocol for
Multi-Hop Wireless Ad Hoc Networks. In: Ad Hoc Networking, ch. 5, pp. 139–172
(2001)
[14] Beijar, N.: Zone Routing Protocol (ZRP), http://www.netlab.tkk.fi/
opetus/s38030/k02/Papers/08-Nicklas.pdf
[15] Nardis, L.D., Giancola, G., Benedetto, M.: Power-aware design of MAC and routing for
UWB networks. In: Globecom Workshops. IEEE Communications Society, Los Alami-
tos (2004)
[16] Jinyang, L., Jannotti, J., Douglas, S.J., Karger, D.R., Morris, R.: A scalable location ser-
vice for geographic ad hoc routing. In: Proc. MOBICOM, pp. 120–130 (2000)
[17] Ko, Y.B., Vaidya, N.H.: Location-Aided Routing (LAR) in mobile ad hoc networks.
Wireless Networks 6, 307–321 (2000)
[18] Liao, W.H., Tseng, Y.C., Sheu, J.P.: GRID: A fully location-aware routing protocols for
mobile ad hoc networks. Telecommunication Systems 18(1-3), 37–60 (2001)
[19] Karp, B., Kung, H.T.: GPSR: Greedy Perimeter Stateless Routing for Wireless Networks.
In: MobiCom (2000)
[20] Boleng, J., Camp, T.: Adaptive Location Aided Mobile Ad Hoc Network Routing. In:
IEEE International Conference on Performance, Computing, and Communications, pp.
423–432 (2004)
[21] Shih, T.F., Yen, H.C.: Location-aware routing protocol with dynamic adaptation of
request zone for mobile ad hoc networks, October 9. Springer, Heidelberg (2006)
[22] Kranakis, E., Singh, H., Urrutia, J.: Compass Routing on Geometric Networks. In: Proc.
11th Canadian Conf. Computational Geometry (August 1999)
[23] Banka, R.K., Xue, G.: Angle Routing Protocol: Location Aided routing for Mobile Ad-
hoc Networks Using Dynamic Angle Selection. IEEE, Los Alamitos (2002)
[24] Abolhasan, M., Wysocki, T.: GPS-based Route Discovery Algorithms for On-demand
Routing Protocols. In: MANETs (January 2004)
[25] Nakagawa, H., Ohta, T., Ishida, K., Kakuda, Y.: A Hybrid Routing with Location Infor-
mation for Mobile Ad Hoc Networks. In: Eighth International Symposium on Autono-
mous Decentralized Systems (ISADS 2007). IEEE, Los Alamitos (2007)
[26] Wang, E.N.C., Wang, S.M.: An Efficient Location-Aided Routing Protocol for Mobile
Ad Hoc Networks. In: Proceedings of the 2005 11th International Conference on Parallel
and Distributed Systems (ICPADS 2005). IEEE, Los Alamitos (2005)
[27] Li, J.P., Wu, L.Y., Lin, S.Y., Wu, G.M.: A Novel Routing Protocol Integrate Power Bal-
ance for Ad Hoc Networks. In: Third International Conference on Networking and Ser-
vices (ICNS 2007). IEEE, Los Alamitos (2007)
[28] Ucan, E., Thompson, N., Gupta, I.: A PiggyBacking Approach to Reduce Overhead in
Sensor Network Gossiping. In: MidSens 2007, Newport Beach, USA, November 26-30,
ACM, New York (2007); Copyright 2007, 978-1-59593-929-6/07/11
[29] Singh, S., Mghavendra, M.W.C.S.: Power-Aware Routing in Mobile Ad Hoc Networks.
In: Proceedings of Fourth Annual ACM IEEE International Conference on Mobile Com-
puting and Networking, Dallas, TX, October 1998, pp. 181–190 (1998)
[30] Feng, M., Cheng, S., Zhang, X., Ding, W.: A Self-Healing Routing Scheme Based on
AODV in Ad hoc Networks. In: The Fourth International Conference on Computer and
Information Technology CIT 2004, September 14-16, pp. 616–620 (2004)
[31] Gil, H., Yoo, J., Lee, J.: An On-demand Energy-efficient Routing Algorithm for Wireless
Ad hoc Networks. In: Chung, C.-W., et al. (eds.) HSI 2003. LNCS, vol. 2713, pp. 302–
311. Springer, Heidelberg (2003)
 A Review of Routing Protocols for UWB MANETs 245

[32] Chenchen, Z., Zhen, Y.: A New EAAODV Routing Protocol based on Mobile Agent.
In: International Conference on Systems and Networks Communications, ICSNC 2006
(October 2006)
[33] Chang, C.Y., Chang, C.T., Tu, S.C., Hsieh, T.T.: Active Route-Maintenance Protocol for
Signal-Based Communication Path in Ad-Hoc Networks. In: Proceedings of Ninth IEEE
International Conference on Networks, October 10-12, pp. 25–30 (2001)
[34] Bergamo, P., Giovanardi, A., Travasoni, A., Maniezzo, D., Mazzini, G., Zorzi, M.: Dis-
tributed Power Control for Energy Efficient Routing in Ad Hoc Networks. ACMKluwer
WINET 10(1) (January 2004)
[35] Gong, M.X., Midkiff, S.F., Buehrer, R.M.: A New Piconet Formation Algorithm for
UWB Ad Hoc Networks. In: IEEE Conference on Ultra Wideband Systems and Tech-
nologies, November 16-19, pp. 180–184 (2003)
[36] Huang, Y., Zeng, W.: A reliable Routing Protocol for Multi-hop Ad hoc Networks. In:
International Conference on Wireless Communications, Networking and Mobile Com-
puting, WiCom, September 21-25, pp. 1656–1659 (2007)
[37] Nardis, L., Baldi, P., Benedetto, M.G.: UWB Ad-Hoc Networks. In: IEEE Conference on
Ultra Wideband Systems and Technologies, Baltimore, Maryland, USA, May 20-23, pp.
219–223 (2002)
[38] Nardis, L.D., Giancola, G., Benedetto, M.G.D.: A position based routing strategy for
UWB networks. IEEE, Los Alamitos (2003)
[39] Nardis, L.D., Giancola, G., Benedetto, M.G.D.: A power-efficient routing metric for
UWB wireless mobile networks. In: Procs. of Vehicular Technology Conference, IEEE
VTS Fall VTC 2003, October 2003, pp. 3105–3109 (2003)
[40] Horie, W., Sanada, Y.: Novel Routing Schemes Based on Location Information for UWB
Ad-Hoc Networks. Wiley Periodicals, Inc., Chichester (2004)
[41] Home, W., Sanada, Y.: Novel Packet Routing Scheme based on Location Information for
UWB Ad-hoc Network. IEEE, Los Alamitos (2003)
[42] Gong, M.X., Midkiff, S.F., Buehrer, R.M.: A Self-Organized Clustering Algorithm for
UWB Ad Hoc Networks. In: WCNC 2004. IEEE Communications Society, Los Alamitos
(2004)
 An Efficient and Reliable Routing Protocol for Wireless
Mesh Networks

Jaydip Sen

Innovation Lab, Tata Consultancy Services Ltd., Bengal Intelligent Park,

Salt Lake Electronics Complex, Kolkata, India
[email protected]

Abstract. Wireless mesh networks (WMNs) have emerged as a key technology

for next generation wireless networks showing rapid progress and inspiring
numerous applications. The persistence driving force in the development of
WMNs comes from their envisioned advantages including extended coverage,
robustness, self-configuration, easy maintenance, and low cost. However, to
support real-time applications with stringent quality of support (QoS), WMNs
must be equipped with a robust, reliable and extremely efficient routing proto-
col so that packets can be routed through them with minimum delay. In this pa-
per, we focus on the critical factors in designing a routing protocol for WMNs,
and propose an efficient and reliable routing protocol. The protocol is based on
a reliable estimation of available bandwidth in a wireless path and end-to-end
delay measurements. Simulations carried out on the protocol demonstrate that it
is more efficient than some of the current routing protocols.

Keywords: Wireless mesh networks, quality of service, routing, end-to-end de-

lay, bandwidth estimation, selfish nodes.

1 Introduction
Wireless mesh networking has emerged as a promising concept to meet the challenges
in next-generation wireless networks such as providing flexible, adaptive, and recon-
figurable architecture while offering cost-effective solutions to service providers.
WMNs are multi-hop wireless networks formed by mesh routers (which form a wire-
less mesh backbone) and mesh clients. The mesh routers provide a rich radio mesh
connectivity which significantly reduces the up-front deployment cost of the network.
Mesh routers are typically stationary and do not have power constraints. However, the
clients are mobile and energy-constrained. Some mesh routers are designated as
gateway routers which are connected to the Internet through a wired backbone. A
gateway router provides access to conventional clients and interconnects ad hoc, sen-
sor, cellular, and other networks to the Internet. A mesh network can provide multi-
hop communication paths between wireless clients, thereby serving as a community
network, or can provide multi-hop paths between the client and the gateway router,
therby providing broadband Internet access to the clients.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 246–257, 2010.
© Springer-Verlag Berlin Heidelberg 2010
 An Efficient and Reliable Routing Protocol for Wireless Mesh Networks 247

As WMNs become an increasingly popular replacement technology for last-mile

connectivity to the home networking, community and neighborhood networking, it is
imperative to design an efficient resource management system for these networks.
Routing is one of the most challenging issues in resource management for supporting
real-time applications with stringent QoS requirements. However, most of the existing
routing protocols for WMNs are extensions of protocols originally designed for
MANETs and thus they perform sub-optimally.
The paper presents an efficient routing protocol for WMNs that is able to handle
stringent QoS requirements of real-time applications. It utilizes two metrics- (i)
minimum bandwidth (Bmin) and (ii) maximum delay (Tmin) that can be tolerated by the
applications that it supports. While issues such as end-to-end route discovery and
QoS-aware routing have been proposed for WMNs in [1], the protocol proposed in
this paper is more efficient than those schemes as shown in the simulation results. The
QoS-awareness in the protocol is achieved by a robust estimation of the available
bandwidth of the wireless channel and a proactive discovery of the routing path.
However, if the bandwidth estimation process indicates that it is not possible to guar-
antee the required QoS, the protocol does not admit the flow. In the event of any node
or link failure, the protocol adapts itself and re-discovers an alternate path.
The key contributions of the paper are as follows: (i) It provides an accurate esti-
mation of the end-to-end delay in a routing path. This estimated value is then used to
check whether the routing can guarantee the application QoS. (ii) It computes a link
quality estimator and utilizes it in route selection. (iii) It provides a framework for
reliable estimation of available bandwidth in a routing path so that flow admission
with guaranteed QoS satisfaction can be made. In addition the proposed protocol
helps in identifying and isolating selfish nodes [2].
The rest of this paper is organized as follows. Section 2 describes related work on
routing in WMNs. Section 3 describes the network model and the design objectives.
Next, Section 4 describes the details of the proposed routing protocol. Simulation
results are presented in Section 5. Section 6 concludes the paper.

2 Related Work
Although significant work has been done on routing in wireless networks, very little
work has been done for WMNs. Most of the routing protocols for MANETs such as
AODV and DSR use hop-count as the routing metric. However, this is not well-suited
for WMNs. The basic idea in minimizing the hop-count is that it reduces delay and
maximizes the throughput. But the assumption here is that the links in the path are
either perfect or do not work at all, and all links are of equal bandwidth. A routing
scheme that uses the hop-count metric does not take link quality into consideration. A
minimum hop-count path has, on the average, longer links between the nodes present
in the path compared to a higher hop-count path. This reduces the signal strength
received by the nodes in that path and thereby increases the loss ratio at each link [3].
Hence, it is always possible that a two-hop path with a good link quality provides
higher throughput than a one-hop path with a poor link quality. Moreover, wireless
links usually have asymmetric loss rate [4]. Hence, new routing metrics based on link
248 J. Sen

quality are proposed such expected transmission count (ETX), per-hop round-trip
time (RTT), and per-hop packet pair. Different approaches have been suggested by
researchers for designing routing protocols for WMNs. In [5], a QoS routing over
OLSR protocol has been proposed that takes into account metrics such as bandwidth
and delay where the source node proactively changes a flow’s next hop in response to
the change in available bandwidth on its path. In [6], the authors have proposed a link
quality source routing (LQSR) protocol. It is based on DSR and uses ETX as the
routing metric. A new routing protocol called multi-radio link quality source routing
(MR-LQSR) is proposed in [7]. The process of neighbor node discovery and propaga-
tion of link metric are same as those in DSR. However, assignment of link weight and
computation of the path weight is different. A QoS enabling routing algorithm for
mesh-based wireless LAN architecture has been proposed in [8] where the wireless
users form an ad hoc peer-to-peer network. The authors also have proposed a protocol
for MANET called ad hoc QoS on-demand routing (AQOR) [9]. However, both these
algorithms are incapable of guaranteeing QoS [1]. In [10], the authors have shown
that if a weighted cumulative expected transmission time [7] is used in a link state
routing protocol, it does not satisfy the isotonicity property of the routing protocol and
leads to formation of routing loops. To avoid routing loops, an algorithm is proposed
that uses metric of interference and channel switching (MIC) as the routing metric.
In contrast to most of the above approaches, the proposed protocol performs on-
demand route discovery using multiple metrics like bandwidth, delay, and reliability
of the links and provides a routing framework that can support application QoS.

3 Network Model and Challenges

This section presents the model of the network and discusses the design key chal-
lenges for routing protocols in WMNs.

Fig. 1. The hierarchical architecture of a WMN

 An Efficient and Reliable Routing Protocol for Wireless Mesh Networks 249

3.1 Design challenges

Two major design challenges for design of a mesh routing protocol are detecting and
avoiding unreliable routes, and accurately estimating the end-to-end delay of a flow.
1. Measuring link reliability: It has been observed that in wireless ad hoc networks
neighbor sensing with broadcast messages introduces communication gray zones [11].
In such zones, data messages cannot be exchanged although the Hello messages indi-
cate neighbor reachability. This leads to a systematic mismatch between the route
state and the real world connectivity, resulting in disruptive behavior. Since the rout-
ing protocols such as AODV and WMR [8] rely on control packets like RREQ, these
protocols are highly unreliable for estimating the quality of wireless links. Due to
communication gray zone problem, nodes that are able to send and receive bidirec-
tional RREQ packets may not be able to send/receive high bit-rate data packets. These
fragile links trigger link repairs resulting in high control overhead for routing proto-
cols. Therefore, designing a robust and reliable link quality estimator is crucial for
reducing the control overhead of a routing protocol in WMNs.
2. End-to-end delay estimation: An important issue in a routing protocol is end-to-
end delay estimation. Current protocols estimate end-to-end delay by measuring the
time taken to route RREQ and RREP packets along the given path. However, RREQ
and RREPs are quite different from normal data packets, and therefore are unlikely to
experience the same levels of traffic delay and loss as data packets. It has been ob-
served through simulation that a RREP-based estimator overestimates while a hop-
count-based estimator underestimates the actual delay experienced by the data packets
[1]. The reason for the significant deviation of a RREP-based estimator from the ac-
tual end-to-end delay is interference of signals. The RREQ packets are flooded in the
network resulting in a heavy burst of traffic. This heavy traffic causes inter-flow inter-
ference in the paths. The unicast data packets do not cause such events. Moreover, as
a stream of packets traverse along a route, due to the broadcast nature of wireless
links, different packets in the same flow interfere with each other resulting in per-
packet delays. Since the control packets do not experience per-packet delay, the esti-
mate based on control packet delay deviate widely from the actual delay experienced
by the data packets. Computing a reliable estimate of the end-to-end delay is, there-
fore, another challenge in design of a routing protocol for WMNs.

4 The Proposed Routing Protocol

The goal of the proposed routing protocol is to establish a route from the source to the
destination that can allow traffic flow to satisfy the QoS of the application, i.e., the
flow should be able to deliver packets with a guaranteed minimum bandwidth Bmin
and maximum allowable end-to-end delay Tmax.
The proposed protocol is a reactive routing protocol, in which during the route dis-
covery phase, each intermediate node uses an admission control scheme to check
whether the flow can be accepted or not. If a flow is accepted, an entry is created for
the flow in table (called the flow table) maintained locally by the node. The admission
control scheme used is the one proposed in [9]. The novel features of the proposed
protocol are discussed in the remainder of this section.
250 J. Sen

4.1 Estimating Reliability of Routing Paths

Every node estimates the reliability of each of its wireless links to its one-hop
neighbor nodes. For computing the reliability of a link, the number of control packets
that a node receives in a given time window is used as a base parameter. An exponen-
tially weighted moving average (EWMA) method is used to update the link reliability
estimate. If the percentage of control packets received by a node over a link in the last
interval of measurement of link reliability is Nt, and if Nt-1 is the historical value of the
link reliability before the last measurement interval, α = 0.5 is the weighting parame-
ter, then the updated link reliability (R) is computed as:

R = α * N t + (1 − α ) * N t −1 (1)

Every node maintains the estimates of the reliability of each of its links with its
neighbors in a link reliability table. The reliability for an end-to-end routing path is
computed by taking the average of the reliability values of all the links on the path.
Computation of the link reliability values are based on the RREQ packets on the re-
verse path and the RREP packets on the forward path. The use of routing path with
the highest reliability reduces the overhead of route repair and makes the routing
process more efficient.

4.2 Use of Network Topological Information in Route Discovery

The proposed protocol makes use of the knowledge of network topology by utilizing
selective flooding of control messages in a portion of the network. In this way,
broadcasting of control messages is avoided and thus the chances of network conges-
tion and disruption to the flows in the network are reduced. If both the source and the
destination are under the control of the same mesh router (Fig. 1), the flooding of the
control messages are confined within the portion of the network served by the mesh
router only. However, if the source and the destination are under different mesh
routers, the control traffic is limited to the two mesh groups.
To further reduce the overhead of control message and enhance the reliability in
routing, the nodes accept broadcast control messages from only those neighbors which
have the link reliability value greater than 0.5 (i.e., on average 50% of the control
packets sent from those nodes have been received by the node). This ensures that path
with less reliability values are not discovered and therefore not considered for routing.

4.3 Estimating End-to-End Delay in a Routing Path

For accurate estimation of end-to-end delay in a routing path, an approach similar to

the one proposed in [1] has been taken. For addressing the issue of differential delays
experienced by the control and the data packets, the proposed protocol makes use of
some probe packets during the route discovery phase. When a source node receives
RREP packets from the destination in response to its RREQ, it stores in a table, the
records for all the RREP packets together with the path through which the packets
have arrived at it. However, instead of randomly selecting a path to send probe pack-
ets to the destination as suggested in [1], the packets are sent along the path from
which the RREP messages have arrived at the source first. This ensures that the probe
 An Efficient and Reliable Routing Protocol for Wireless Mesh Networks 251

packets are sent along the path which is likely to induce less end-to-end delay result-
ing in a better performance of the protocol as observed from the simulation results
presented in Section 5. The probe packets are identical to data packets so far as their
size, priority and flow rates are concerned. The objective of sending probe packets is
to simulate the data flow and observe the delay characteristics in the routing path.
Number of probe packets is kept limited to 2H for a path consisting of H hops to
make a tradeoff between control overhead and measurement accuracy [1].
A destination node sets on a timer after it receives the first probe packet from the
source node. The timer duration is based on the estimated time for receiving all the
probe packets and is computed statistically. The destination computes the average
delay experienced by all the probe packets it has received, and sends the computed
value to the source node piggybacking it on a RREP message. If the computed value
is within the limit of tolerance of the application QoS, the source selects the route and
sends data packets. If the delay exceeds the required limit, the source selects the next
best path (based on the arrival of RREP packets) from its table and tries once again.
Since the routing path is set up based on the probe packets rather than the naïve RREP
packets, the proposed protocol has higher route establishment The proposed algorithm
has higher setup time due to sending of the probe packets and selection of the best
path based on the estimated end-to-end delay. However, since the selected paths have
high end-to-end reliability, the delay and the control packet overhead are reduced
because of minimal subsequent route breaks.

4.4 Estimation of Available Network Bandwidth

In addition to computation of path reliability and end-to-end delay, it is also necessary

that the effective bandwidth in a routing path is reliably estimated. This is extremely
important to support real-time applications since these applications require a guaran-
tee for a minimum available bandwidth for their flows. In the proposed protocol the
available bandwidth for a wireless link is estimated using its end-to-end delay and the
loss of packets in it due to congestion. The packet-loss due to congestion in the link is
estimated as follows.
In a wireless link packet loss may happen due to two reasons: (i) loss due to faulty
wireless links and (ii) loss due to network congestion. The radio link control (RLC)
layer segments an IP packet into several RLC frames before transmission and reas-
sembles them into an IP packet at the receiver side. An IP packet loss occurs when
any RLC frame belonging to an IP packet fails to be delivered. When this happens,
the receiver knows the RLC frames reassembly has failed and the IP packet has been
lost due to wireless error. Meanwhile, the sender detects retransmission time out
(RTO) of the frame and discards all the RLC frames belonging to the IP packet. This
enables the sender to compute packet drop rate in the wireless links. Moreover, using
the sequence numbers of the IP packets received at the receiver, it possible to differ-
entiate the packet loss due to link error and packet loss due to congestion [Yang pa-
per]. For example, while receiving two incoming packets with sequence number i and
i + 2, if the receiver finds an IP packet assembly failure in RLC layer, the packet with
sequence number i + 1 is lost due to wireless channel. Once the packet loss ratio due
to congestion (Pcongestion) is estimated, the available bandwidth in the wireless link,
estrat, is computed as follows [12]:
252 J. Sen

PacketSize
estrat = (2)
X +Y
where,

2Pcongestion (3)
X = RTT
3
3Pcongestion
Y = RTO * min(1,3 * Pcongestion(1 + 32 Pcongestion
2
) (4)
8
In (2), RTT is the average round trip time for a control packet. RTO is the retransmis-
sion time out for a packet, and is computed using (5).
− −−− − −− −−−
RTO = RTT + k * RTT Var (5)
−− −− − −− −−−
where, RTT and RTT Var are the mean and variance respectively of RTTs and k is
set to 4. This bandwidth estimator is employed to dynamically compute the available
bandwidth in the wireless links on a routing path so that the guaranteed minimum
bandwidth for the flow is always maintained throughout the application life-time.

4.5 Identifying Selfishness in Wireless Nodes

The proposed routing protocol also enforces cooperation among the nodes by identi-
fying the selfish nodes in the network and isolating them. Selfishness is an inherent
problem associated with any capacity-constrained multi-hop wireless networks like
WMNs. A mesh router can behave selfishly owing to various reasons such as: (i) to
obtain more wireless or Internet throughput, or (ii) to avoid path congestion. A selfish
mesh router increases the packet delivery latency, and also increases the packet loss
rate. A selfish node while utilizing the network resources for routing its own packet,
avoids forwarding packets for others to conserve its energy. Identification of selfish
nodes is therefore, a vital issue.
Several schemes proposed in the literature to mitigate the selfish behavior of nodes
in wireless networks such as credit-based schemes, reputation-based schemes and
game theory-based schemes [13]. However, to keep the overhead of computation and
communication at the minimum, the proposed protocol employs a simple mechanism
to discourage selfish behavior and encourage cooperation among nodes. To punish the
selfish nodes, each node forwards packets to its neighbor node for routing only if the
link reliability of that node is greater than a threshold (set at 0.5). Since the link reli-
ability of a selfish node is 0, the packets arriving from this node will not be for-
warded. Therefore, to keep its link reliability higher than the threshold, each node has
to participate and cooperate in routing. The link reliability serves a dual purpose of
enhancing reliability and enforcing node cooperation in the network.

4.6 QoS Violation and Recovery

The proposed protocol detects failure to guarantee QoS along a path with the help of
reservation timeouts in flow tables records maintained in the nodes, and by detection
of non-availability of minimum bandwidth as estimated along its outbound wireless
 An Efficient and Reliable Routing Protocol for Wireless Mesh Networks 253

link. Failure to guarantee QoS may occur in three different scenarios. In the first case, a
node receives a data packet for which it does not find a corresponding record in its
flow table. This implies that a reservation time-out has happened for that flow. The
node, therefore, sends a route error (RERR) to the source which re-initiates route dis-
covery. In the second scenario, a destination node detects from its flow table records
that the data packets received have exceeded the maximum allowable delay (Tmax). To
restore the path, the destination broadcasts a new RREP back to the source, and the
source starts rerouting the packets via the same path of the RREP. This approach is
similar to the one proposed in [9]. In the third case, an intermediate node on the routing
path may find that the estimated bandwidth (using (2)) in its forwarding link is less
than the guaranteed minimum (Bmin) value. In this case, the intermediate node sends a
route error (RERR) to the source which re-initiates the route discovery process. The
real-time estimation of the bandwidth in the next-hop wireless link at each node on the
routing path makes the proposed protocol more robust and reliable compared to other
existing routing protocols for WMNs. For example, the protocol presented in [1] does
not provide any bandwidth estimation mechanism at intermediate nodes, and therefore
cannot provide delivery of all packets for every admitted flow in the system.

5 Simulation Results
The proposed protocol has been implemented in Qualnet network simulator [14]. The
WMN topology is shown in Fig. 2. The simulated network consists of 50 nodes of
which 5 are mesh routers and remaining nodes are mesh clients. Each mesh router has
9 mesh clients associated with it. To compare the performance of the proposed proto-
col with the protocol presented in [1] by Kone et al, parameters in the simulation are
chosen to be identical as those used in [1]. The simulation parameters are presented in
Table 1. The MAC protocol used is 802.11b, which is based on CSMA/CA protocol.
No wireless fading model is used scheme to keep it simple, yet realistic.

Fig. 2. The network topology of the simulated WMN

254 J. Sen

Table 1. Simulation parameters

Parameter Values
Simulation area 1300 m * 1300 m
MAC protocol 802.11
PHY-model 802.11b
Propagation channel frequency 2.4 GHz
Raw channel bandwidth 11 Mbps
Simulation duration 15 s
Traffic type CBR UDP
Packet size 512 bytes
Max. no. of packets from a source node 10000
Hello packet broadcast interval 200 ms
Wireless fading model None
Node mobility None
Propagation limit of each node -111.dBm
IP queue scheduler Strict priority

5.1 QoS Routing Behavior

To evaluate the effectiveness of the admission control module of the protocol, the
performance of the protocol in a heavy traffic condition is observed. Due to heavy
traffic, some flows in the network will not have their bandwidth and end-to-end delay
requirement satisfied, and hence will be rejected by the protocol. AODV, on the other
hand being non-aware about the QoS parameters, will try to route packets from all the
sources. The experimental results have shown that the proposed protocol has rejected
admission of new flows after the traffic level in the network has gone above a certain
limit. However, for flows that have been admitted, all packets are routed within the
required end-to-end delay. AODV, on the other hand has admitted all the flows and
routed them with increasing delay as the congestion in the network increased to due
more number of control packets. An important point worth mentioning here is that the
protocol in [1] is not able to achieve 100% packet delivery ratio for all flows. In con-
trast, the proposed protocol has been found to achieve 100% packet delivery. This is
due to bandwidth estimation at each intermediate node as discussed in Section 4.4.

5.2 Control Overhead

For computing the overhead due to control packets, only the RREQ messages and the
probe packets in the proposed algorithm are considered since these are broadcast
messages and largely contribute to the control overhead of the protocols.
Fig. 3(a) shows the overhead of RREQ packets in AODV and the proposed proto-
col for different data rates. The control overhead of the proposed protocol is first
evaluated only with the RREQ packets and then with RREQ packets and the probe
packets together. This also gives an idea of the additional overhead introduced due to
the probe packets. It can be easily observed that the proposed protocol has a very
limited overhead even with the probe packets compared to naïve AODV protocol. It is
also worth observing that the proposed protocol has about 20% less control overhead
than the protocol proposed in [1] due to its robust bandwidth estimation of links.
 An Efficient and Reliable Routing Protocol for Wireless Mesh Networks 255

Fig. 3(a). Control overhead vs. data rate

Fig. 3 (b). Control overhead with no. of flows

Fig. 3(b) shows that AODV has a very high overhead of control packets for in-
creasing number of flows in the system. AODV always tries to establish routing paths
based on minimum hop-counts. It does not pay any attention to the link reliability.
This results in frequent selection of unreliable links and consequent link break and re-
discovery of route. This leads to large overhead of control information. In contrast,
the proposed protocol has a very limited control overhead since paths with higher link
reliability are only used for routing. The algorithm proposed in [1] has similar per-
formance as the proposed protocol in this case.

5.3 End-to-End Delay Estimation

To demonstrate the effectiveness of the end-to-end delay estimation mechanism by

probe packets, delays estimated by naïve RREP approach and the probe packet ap-
proach are compared with the actual end-to-end delay in the routing paths. The re-
cords are observed for different flow rates with each flow having a minimum band-
width requirement of Bmin = 50 kbps. Fig. 4 shows that the probe packet-based
mechanism very accurately estimates the actual delay, and has similar performance as
that of the algorithm in [1].
256 J. Sen

Fig. 4. End-to-end delay estimation for different no. of flows

5.4 Detection of Selfish Nodes

As mentioned in Section 4.4, the proposed protocol has the ability to detect selfish
nodes which try to use network resources without contributing to the cooperative
framework of routing. To evaluate its detection capability of selfish nodes, two types
of flows are distinguished: selfish and honest. A flow is considered selfish if either its
source or destination is a selfish node, otherwise the flow is considered to be honest.
Some mesh clients are selected randomly and configured as selfish nodes. In each of
the 20 run of the simulation, 10 flows of 50 kbps data rate are generated randomly.
It may be observed from Fig. 5 that AODV cannot restrict the traffic along the self-
ish flows. The selfish nodes can fully exploit the routing process to have their packets
routed in the network. However, the proposed protocol reduces the flows along the
selfish paths. In fact, the performance of AODV is not very much affected by pres-
ence of selfish nodes, since it never establishes routing path based on hello packets.
Since the proposed protocol establishes route based on hello packets received form
neighbors, its performance is affected by the presence of selfish nodes. However, its
performance is not substantially affected, since most in most cases, these nodes are
not allowed to participate in the routing, because of the low values of their link reli-
ability. It may also be mentioned that the proposed protocol is able to maintain, on
average, 35% more throughput in presence of selfish nodes when compared with the
protocol proposed in [1]. The large difference is due to its ability to detect selfish
nodes faster by its effective bandwidth estimation in the route where non-forwarding
of packets is treated as packet drops due to congestion.

Fig. 5. Avg. throughput with different protocols for varying percentage of selfish nodes
 An Efficient and Reliable Routing Protocol for Wireless Mesh Networks 257

6 Conclusion
WMNs present a promising technology for providing low-cast last mile broadband
Internet connectivity through multi-hop communications. Designing an efficient and
reliable routing protocol for WMNs is a challenging issue. This paper has presented a
QoS-aware routing protocol that is capable of supporting real-time applications. By
robust estimation of available bandwidth in the wireless route and a reliable computa-
tion of end-to-end delay along a path, the protocol is able to sustain a high level of
throughput even in presence of a large proportion of selfish nodes in a WMN. More-
over, it is able to reduce control overhead substantially by exploiting the knowledge
of network topology. Simulation results have shown that the proposed protocol is
more efficient than some of the current routing protocols for WMNs.

References
1. Kone, V., Das, S., Zhao, B.Y., Zheng, H.: Quorum: Quality of Service in Wireless Mesh
Networks. Journal of Mobile Networks and Applications 12(5), 358–369 (2007)
2. Mahajan, R., Rodrig, M., Wetherall, D., Zahorjan, J.: Sustaining Cooperation in Multi-Hop
Wireless Networks. Proc. of NSDI 2, 231–244 (2005)
3. Cuto, D.S.J.D., Aguqayo, D., Bricket, J., Morris, R.: A High-Throughput Path Metric for
Multi-Hop Wireless Routing. In: Proc. of ACM MOBICOM, pp. 134–146 (2003)
4. Aguayo, D., Bicket, J., Biswas, S., Judd, G., Morris, R.: Link-Level Measurements from
an 802.11b Mesh Network. In: Proc. of ACM SIGCOMM, pp. 121–132 (2004)
5. Badis, H., Gawedzki, I., Al Agha, K.: QoS Routing in Ad Hoc Networks using QOLSR
with no need of Explicit Reservation. In: Proc. of VTC (September 2004)
6. Draves, R., Padhya, J., Zill, B.: Comparison of Routing Metrics for Static Multi-Hop
Wireless Networks. In: Proc. of ACM SIGCOMM, pp. 133–144 (2004)
7. Draves, R., Padhye, J., Zill, B.: Routing in Multi-Radio, Multi-Hop Wireless Mesh
Networks. In: Proc. of ACM MOBICOM, pp. 114–128 (2004)
8. Xue, Q., Ganz, A.: QoS Routing for Mesh-Based Wireless LANs. International Journal of
Wireless Information Networks 9(3), 179–190 (2002)
9. Xue, Q., Ganz, A.: Ad Hoc QoS On-Demand Routing (AQOR) in Mobile Ad Hoc
Networks. Journal of Parallel and Distributed Computing 63(2), 154–165 (2003)
10. Yang, Y., Wang, J., Kravets, R.: Interference-Aware Load Balancing for Multi-Hop Wire-
less Networks. Technical Report UIUCDCS-R-2005-2526, Dept. of Computer Science,
University of Illinois at Urbana-Champaign (2005)
11. Lundgren, H., Nordstrom, E., Tschudin, C.: The Gray Zone Problem in IEEE 802.11b
Based Ad Hoc Networks. MC2R 6(3), 104–105 (2002)
12. Yang, F., Zhang, Q., Zhu, W., Zhang, Y.-Q.: End-to-End TCP-Friendly Streaming Proto-
col and Bit Allocation for Scalable Video over Wireless Internet. IEEE Journal on Selected
Areas in Communications 22(22), 777–790 (2004)
13. Santhanam, L., Xie, B., Agrawal, D.: Selfishness in Mesh Networks: Wired Multi-Hop
MANETs. IEEE Journal on Wireless Communication 15(4), 16–23 (2008)
14. Qualnet Simulator, http://www.scalable-networks.com
 A Context-Aware Service Model Based on
Workflows for u-Agriculture

Yongyun Cho1 , Jongbae Moon2, , and Hyun Yoe1

1
Information and Communication Engineering, Sunchon National University,
413 Jungangno, Suncheon, Jeonnam 540-742, Korea
{yycho,yhyun}@sunchon.ac.kr
2
Supercomputing Center, Korea Institute of Science and Technology Information,
52-11 Eoeun-dong, Yuseong-gu, Daejeon 305-806, Korea
[email protected]

Abstract. In agricultural processes, many works may need to be au-

tomated, because those are very hard labors or time-consuming jobs
for farmers. Workflow technologies, which have successfully been a good
model for a service automation in various computing environments, can
be used as a possible service automation model in agriculture. A work-
flow for u-agriculture may need various contexts sensed from real sensor
networks for service automation. Recently, many researches have ap-
plied workflow technologies into the various fields such as smart home,
u-health care, u-city, u-port, and u-agriculture. However, many current
workflow service have difficulty to control work processes and execute
services according to context information in u-agricultural environment.
This paper proposes a context-aware service model for u-agriculture,
which is based on workflows and is aimed at supporting smart workflow
services based on ubiquitous sensor networks in u-agriculture. With the
proposed context-aware service model, developers can easily integrate
various service demands into a service workflow, and can easily develop
a context-aware workflow service for u-agriculture. Therefore, the pro-
posed service model can be greatly helpful in the development of smart
applications or the work automation in the fields of u-agriculture.

1 Introduction

Service automation is commonly centered in the service execution without inter-

vention of humans according to the changes of surrounding situation information
[1]. Because agricultural environment is hard to work and labor-intensive in most
case and areas, various efforts for the work automation and smart cultivation is
invaluable. In recent years, because workflow is a very good model for service

This research was supported by the MKE(Ministry of Knowledge Economy),
Korea, under the ITRC(Information Technology Research Center) support program
supervised by the IITA(Institute of Information Technology Advancement)” (IITA-
2009-(C1090-0902-0047)).

Corresponding author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 258–268, 2010.

c Springer-Verlag Berlin Heidelberg 2010
 A Context-Aware Service Model Based on Workflows for u-Agriculture 259

automation in the business computing environments, some studies have tried to

use workflow techniques to automate services according to context information
or situation information in ubiquitous computing environment or pervasive com-
puting environment [2,3,4]. And, some researches have tried to apply workflow
and/or context-aware technologies in the field of agriculture [5,6,7].
A context-aware service is a kind of a smart service which can be automatically
processed according to various situation information. A context in a ubiquitous
environment means any information that can be used to characterize the situ-
ation of an entity [1,2]. For example, in u-agricultural environments, contexts
may be the plant’s rate of growth, the amount of sunshine, the percentage of hu-
midity and the indoor temperature of greenhouse and so on. These contexts can
be an important service execution information for work automation or smart
cultivation in u-agriculture environment. Therefore, context-aware service for
u-agriculture can consider contexts around agricultural environment with any
work automation model.
This paper introduces a new context-aware service model based on workflow
as a service model for work automation in u-agriculture environment. The sug-
gested approach uses uWDL [2], which is an academic context-aware workflow
language, as a workflow model for work automation. So, the developers can eas-
ily describe context information directly into a workflow service scenario, and
can correctly execute the best one of the services described into the workflow
according to the contexts. Therefore, with the suggested model, context-aware
service developers can easily development various context-aware applications in
the fields of u-agriculture environments. And, because a developed workflow can
be easily reused for other context-aware services, the developers can raise devel-
opment efficiency and reusability.

2 Related Work
2.1 Context-Aware Workflow Languages
Workflow languages such as XLANG [8], WSFL [9] and WS-BPEL were devel-
oped by international IT companies such as Microsoft, IBM, Bea, Oracle etc.
in order to support service automation in business processes. The BPEL4WS
[10] is a workflow language based on web-service technologies and includes the
advantages of both WSFL and XLANG technologies. XLANG is one of the ini-
tial workflow languages using a directed graph approach and WSFL is a block-
structured workflow language for describing web services. However, They do not
have any element in their language schema or grammar to describe context in-
formation as transition conditions of services. The workflow languages can use
simple return values through XPath, XLink, Xpoint to express current service’s
transition conditions. Therefore, they are not suitable for a workflow language
for context-aware services in ubiquitous environment.
uWDL [2] is a workflow language for ubiquitous computing environment. It
includes <constraints>, <contexts>, and <profiles> elements to describe var-
ious contexts as service transition conditions in its language schema. And, For
260 Y. Cho, J. Moon, and H. Yoe

the automatic execution of context-aware workflow services, it has also ¡flows¿

element, which supports various service flow patterns such as a single-flow, a
multi-flow, and a sub-flow. Even if uWDL is a kind of academic and exper-
imental context-aware workflow language, it can enough describe contexts as
transition conditions of services directly in a uWDL workflow scenario. To do
that, it offers a structural context model based on RDF [11] triplet-subject, verb,
and object.

2.2 Context-Aware Workflow Systems

A workflow system manages and controls flows of subtasks using state transi-
tion constraints specified in a workflow language. WorkSco [12] is an situation-
adaptable workflow system that can support service demands generated dynam-
ically in a business process. It is based on a micro workflow model, a dynamic
evolution and an open-point adaptation techniques to dynamically handle user’s
requests, which may be generated in various business domains. However, it does
not yet give an explicit method to do that. Even though WorkSco considers
dynamic handling for user’s requests in a workflow system, because it does not
consider situation information or contexts as user’s requests, it is basically not
adequate for ubiquitous computing environments.
CAWE [13] is a framework for the management of context-aware workflow
systems based on Web Services. CAWE supports a synthetic and extensible
specification of context-sensitive workflows, which can be executed by standard
workflow engines. The system explained an initial prototype of the CAWE frame-
work using jBPM, which is a business process management system based on Petri
Netmodel implemented in Java. However, even if it proposed a systematic and
valuable architecture for context-aware workflow services, it does not offer an
explicit method to express the contexts into a workflow scenario as service ex-
ecution condition, and doesn’t enough describe how it can use user’s situation
information or contexts, which can be dynamically occurred in real ubiquitous
environments.
uFlow [14] is an another ubiquitous workflow framework to support a context-
aware service based on a uWDL. uFlow is based a workflow scenario, and offers a
method to handle the changes of user’s demands or situation information which
may be dynamically generated during service processing. POESIA [7] is a re-
sult of the effort to use a ontology-based workflow to compose Web services
in agriculture. The approach described how to compose Web services, using
domain-specific multidimensional ontologies in the field of agricultural environ-
ment. POESIA is invaluable for introducing a systematic model or method to
use ontologies, web-services, and workflow technologies to support context-aware
services in agriculture. However, it seems to be mainly aimed to define Web ser-
vices using ontologies, but it is not a main purpose to describe contexts into a
workflow scenario, and to control context-aware service flows according the sce-
nario and real context information for work automation in agriculture. Therefore,
it may be inappropriate to intactly use POESIA or uFlow as a service model for
automatic and context-aware services in agriculture.
 A Context-Aware Service Model Based on Workflows for u-Agriculture 261

3 A Model of Context-Aware Workflow Services for

u-Agriculture

3.1 A Suggested Service Model Architecture

All things considered for a context-aware service in u-agriculture, it is desirable

to be autonomic and context-aware. So, a context-aware service model for u-
agriculture has to offer a good model to control execution of services for work
automation and a method to handle context information for a smart work with-
out human’s intervention.

Fig. 1. An architecture for the suggested context-aware service model.

In Figure 1, the suggested service model architecture has a context model, a

workflow generator and a workflow processor. The workflow model is based on
a RDF-based structural context model. The situation information received from
USNs/Sensor Networks the model is transformed as a context data through the
workflow model. The workflow generator includes an expanded-uWDL as a au-
tomation model for service execution, a OWL-based ontology for u-agricultural
environment, and a GUI-based workflow editor for generating a context-aware
workflow scenario. With the workflow generator, a developer can explicitly de-
scribe the contexts into a workflow scenario as service execution information.
The workflow processor is to recognize a workflow scenario and to process ser-
vices described into the scenario. In this case, the workflow processor determines
whether a service defines the sensed context as its execution condition and con-
trols the service flows according to the sensed contexts.
262 Y. Cho, J. Moon, and H. Yoe

3.2 A Context Model for u-Agriculture

Contexts in u-agriculture may be very various and mainly sensor-based. For

example, in u-agricultural environments contexts may vary from the plant’s rate
of growth, the amount of sunshine, the percentage of humidity and the indoor
temperature of greenhouse to a farmer’s job schedule and his individual profile
information. In Figure 1, the module recognizes the contexts sensed from USNs
or sensor networks in u-agricultural environment.

Fig. 2. A context model for u-agriculture

A service in u-agriculture has to be executed according to contexts so that

it becomes context-aware. Commonly, A data in real agricultural environment
may occur mainly in real-time and is continuously transferred from various kinds
of sensors. Therefore, its value and type need to be transferred together for any
context-aware service. For example, if the numeric value 60 is transferred from
anyone of temperature sensors in a greenhouse, it needs to show that a data type
of it is Fahrenheit and a data value is 60. It means that the value is related with
just temperature, not anything else. In a context-aware service, the additional
semantic information can be meaningful as a data unit to make more conceptual
and higher level data. Figure 2 shows a context model in this paper.
 A Context-Aware Service Model Based on Workflows for u-Agriculture 263

In Figure 2, the data set that consists of a value and type is an entity. A
few of entities form a constraint, which is based on RDF triplet. And, a context
includes several constraints using rule-based operation. For example, let suppose
that a numeric value 60 is sensed from the current temperature of a section A
in a greenhouse 1. A farmer may wants a service scenario for a heating service,
that when the current temperature of a section A in a greenhouse 1 is less
than Fahrenheit 50, start a heating system. In that case, a context may needs
constraints consisting of 3 entities, < Type:Fahr., value: 60>, <Type:greenhouse,
value: no.1>, and <Type:Position., value: (x, y)>. Therefore, the farmer only
has to describe the context as execution conditions of the heating service into
his workflow scenario. Then, when the workflow processor in Figure 1 finds the
context among numerous entities sensed from a real environment, it will start
the heating service.

3.3 Workflow Patterns for Context-Aware Services

Figure 3 describes workflow service patterns [15] added into the uWDL schema.
The suggested service model uses a workflow model as an work automation model
based on uWDL.

Fig. 3. Workflow patterns based on the extended uWDL

264 Y. Cho, J. Moon, and H. Yoe

However, uWDL didn’t support enough workflow patterns to satisfy various

flow pattern of context-aware services. And, uWDL has difficulties in composing
a new workflow using some of workflows, because it provide only a sequential
flow pattern and a simple workflow pattern. So, this paper expands the schema of
uWDL. A workflow based on the expanded uWDL can support various service
flows and is reused as a part of a new workflow. This can raise development
efficiency and workflow reusability.
In Figure 3, (a) represents a sequential flow model. It means a deterministic
flow that various service nodes are connected in a sequence. However, in many
cases of real u-agricultural environment, the sequential flow model is not enough
to meet various service requirements. Figure 3(b) is a parallel service flow is
based on the basic workflow patterns such as the parallel split, synchronization
and simple merge pattern. This means that a workflow is deterministic, but it
has the flows that split into two or more. Parallel workflow model also is used to
build more complicated workflow services. This means that it is able to be used to
execute two or more workflow services independently and concurrently. In Figure
3(c), a subflow model is based on both exclusive choice and synchronization
pattern. It means that other workflows, which may are developed already, can
be called and attached on one of the nodes on flow B. So, with the pattern, a
developer can quickly and easily compose new context-aware service flows by
reusing pre-developed valuable workflows. In Figure 3(c), the service 3 and N-1
represent the workflow reusing with the subflow model.
The flow model (a), (b), and (c) in Figure 3 can be composed for more com-
plicated and large service flows. There are various service domains in ubiquitous
computing environments. Thus, service developer must be able to provide mul-
tiple workflow services suitable to users by describing many context-aware work-
flow documents. Each workflow service document can be used independently for
a user or be composed in a shape of a new integrated workflow for many users.
And service developer is able to express a composite workflow service, which is
composed of some of workflows, by reusing a workflow document or more.

4 A Experimental Scenario for Context-Aware Services

in u-Agriculture

For an experiment using the suggested service model, we will take experimental
scenario for context-aware Services in u-agriculture using the expanded uWDL.
To do this, let’s suppose a work scenario in u-agricultural environment. The
scenario is as like this:
There is a farmer to cultivate of mushrooms in his/her greenhouse. Mushrooms
is one of crops that are very sensitive to temperature, humidity, and PH of soil.
From morning to midnight everyday, the farmer checks the cultivation factors
for his mushrooms. And, the farmer has to take proper works according to the
results of the examination. Furthermore, he/she has to incubate the mycelia
of mushrooms and shelter the cultivation soil for the continuous cultivation of
mushrooms. And, the farmer has to control the speed and the amount of the
 A Context-Aware Service Model Based on Workflows for u-Agriculture 265

works according the market demands and the production of mushrooms. Clearly,
it is very labor-intensive jobs, and the farmer will want to reduce the amount
of the works which he/she has to do. To do this, let’s suppose a next scenario.
The scenario is based on a workflow which describes the works as the services.
Figure 4 shows the parts of the required services and the workflow scenario.
In Figure 4(a), the services are based on Web services and exist independently.
The service H may requests the current sensed values of humidity, temperature,
PH of soil, a pollution degree of air in the greenhouse, sunshine, and illuminance,
to various sensors positioned around the mushroom greenhouse. In Figure 4(b)
shows the possible service flows that can be connected with the services of Figure
4(a) in order to achieve a specific work. The service H is needed in the basic
cultivation control flow and the pollution control flow. The service C is used in
all of the three flows in Figure 4(b). Of cause, There can be other possible and
reasonable flows in addition to those in Figure 4. And, the procedural orders of
the flows in Figure 4 don’t have to be absolutely like that.

Fig. 4. The parts of the demanded services and the possible service flows
266 Y. Cho, J. Moon, and H. Yoe

Now, let’s make a context-aware workflow service model by using the three
flows with contexts, which may occur in the greenhouse. Figure 5 represents the
context-aware workflow service model. Figure 5(a) shows a possible composition
of the basic cultivation condition control flow and the pollution control flow in
Figure 4. The service H is needed for the basic cultivation condition control flow
and the pollution control flow.
The service H selects a next service flow according to contexts that the service
received from the requested sensors. For example, in case of a context data that
can provoke the basic cultivation services is sensed from the sensors, the service
H will move to the service A. And, when the sensed data is related with the
pollution control services, the service H will move to the service J. Of cause, if
the contexts for the two flows occur all together, the service H will go to the
two flows in a type of parallel. However, the ventilation service or the sunshine

Fig. 5. The possible composition for a context-aware service in u-agriculture

 A Context-Aware Service Model Based on Workflows for u-Agriculture 267

service to purify the air of the greenhouse has to be executed just before the
heating service. So, it may be more reasonable that the service E connects to
the service B just before. Figure 5(b) shows that. Now, the growth control flow
has to be added to the Figure 5(b). First, the sequential flow like Figure 5(c) can
be considered simply. However, because the order check service or the product
checking service can affect the watering service or heating service, it is desired
that the services are started as soon as possible. And, the service I and K may
probably generate air pollution of the greenhouse. Therefore, those service can
reuse the existing service flow, which consists of J, E, and B, as their subflow.
The service J in Figure 5(d) illustrates the process to reuse the section X in
Figure 5(c) as a subflow. The expended uWDL can grammatically support the
method to reuse any partition of the existing service flows as a subflow.

5 Conclusion
A research for service automation and context-aware services is a valuable sub-
ject in u-agricultural environment like in the other fields. However, most existing
context-aware service models are not suitable for work automation and smart
cultivation in u-agriculture, because they have troubles to use or express contexts
to automatically execute a service. This paper introduced a context-aware service
model based on workflows for u-agriculture. The suggested service model offers a
structural context-model to describe contexts in u-agriculture through a entity,
a RDF-based constraint, and rule-based operators. And, it offers flow models of
an expanded uWDL as a service automation model. With the proposed context-
aware service model, developers can easily develope various context-aware service
flows into a service workflow for u-agriculture. And, a developer can reuse the
existing service flows in composition of a new service flow with the flow models.
Therefore, the proposed service model is expected to reduce the time and effort
to develop context-aware applications or the work automation in the fields of
u-agriculture.

References
1. Abowd, G.D., Dey, A.K., Brown, P.J., Davies, N., Smith, M., Steggles, P.: Towards
a Better Understanding of Context and Context-Awareness. In: HUC 1999, pp.
304–307 (1999)
2. Han, J., Cho, Y., Choi, J.: Context-Aware Workflow Language based on Web
Services for Ubiquitous Computing. In: Gervasi, O., Gavrilova, M.L., Kumar, V.,
Laganá, A., Lee, H.P., Mun, Y., Taniar, D., Tan, C.J.K. (eds.) ICCSA 2005. LNCS,
vol. 3481, pp. 1008–1017. Springer, Heidelberg (2005)
3. Aqeel-ur-Rehman, Shaikh, Z.A.: Intelligent Workflow: A Step towards Intelligent
Agent based Academic Workflow System. In: 2nd All Pakistan UTECH and ACM
Technical Conference (UTECH 2007), Karachi, Pakistan, November 23-24 (2007)
4. Tang, F., Guo, M., Dong, M., Li, M., Guan, H.: Towards Context-Aware Work-
flow Management for Ubiquitous Computing. In: Proceedings of ICESS 2008,
pp. 221–228 (2008)
268 Y. Cho, J. Moon, and H. Yoe

5. Lauser, B., Sini, M., Liang, A., Keizer, J., Katz, S.: From AGROVOC to the
Agricultural Ontology Service / Concept Server - An OWL model for creating
ontologies in the agricultural domain. In: Proceedings of the OWLED*06 Workshop
on OWL: Experiences and Directions, Athens, Georgia (USA), pp. 10–11 (2006)
6. Wark, T., Corke, P., Sikka, P., Klingbeil, L.: Transforming Agriculture through
Pervasive Wireless Sensor Networks. IEEE Pervasive Computing, 50–57 (2007)
7. Fileto, R., Liu, L., Pu, C., Assad, E.D., Medeiros, C.B.: POESIA: An Ontologi-
cal Workflow Approach for Composing Web Services in Agriculture. The VLDB
J. 12(4), 352–367 (2003)
8. Satish, T.: XLANG: Web Services for Business Process Design. Microsoft Corp.
(2001)
9. Leymann, F.: Web Services Flow Language (WSFL 1.0), Distinguished Engineer
Member IBM Academy of Technology, IBM Software Group (May 2001)
10. Tony, A., Francisco, C., et al.: Business Process Execution Language for Web Ser-
vices, BEA Systems. Microsoft Corp. IBM Corp., Version 1.1 (2003)
11. W3C: RDF/XML Syntax Specification, W3C Recommendation, 30-39 (2004/2006)
12. Vieira, P., Rito-Silva, A.: Adaptive Workflow Management in WorkSCo. In: 16th
International Workshop on Database and Expert Systems Applications (DEXA
2005), pp. 640–645 (2005)
13. Ardissono, L., Furnari, R., Goy, A., Petrone, G., Segnan, M.: Context-
awareworkflow management. In: Baresi, L., Fraternali, P., Houben, G.-J. (eds.)
ICWE 2007. LNCS, vol. 4607, pp. 47–52. Springer, Heidelberg (2007)
14. Han, J., Cho, Y., Kim, E., Choi, J.: A Ubiquitous Workflow Service Framework.
In: Proceedings of the 2006 International Conference on Computational Science
and its Application (2006)
15. van der Aalst, W.M.P., ter Hofstede, A.H.M., Kiepuszewski, B., Barros, A.P.:
Workflow Patterns. Distributed and Parallel Databases 14(3), 5–51 (2003)
 A History-Based Scheduler for Dynamic Load Balancing
on Distributed VOD Server Environments

Jongbae Moon1, Hyun-joo Moon2, and Yongyun Cho3,*

1
Supercomputing Center, Korea Institute of Science and Technology Information,
52-11, Eoeun-dong, Yuseong-gu, Daejeon, 305-806, Korea
[email protected]
2
Dept. of Cultural Contents, Hankuk University of Foreign Studies,
270 Imun 2-dong, Dongdaemun-gu, Seoul, 130-082, Korea
[email protected],
3
Information and Communication Engineering, Sunchon National University,
413 Jungangno, Suncheon, Jeonnam 540-742, Korea
[email protected]

Abstract. As computer and network technology advance, multimedia data can

be transferred in real time on the Internet. The increasing user demands for
various multimedia data make VOD (Video-on-Demand) services to be devel-
oped. VOD services are being used in lots of fields, such as entertainment,
distance learning, home shopping, and interactive news. In comparison to the
existing HTTP services, these VOD services have special features that the time
for their services is longer and the services request more disks and network
bandwidths. Therefore, compared to HTTP service environments, VOD ser-
vices have some different workload patterns. In these VOD service environ-
ments, the existing load balancing algorithms researched before are not proper.
In this paper, we propose a new load balancing algorithm that is based on
the history of past user access patterns to make server loads even on hierarchi-
cally distributed VOD system environments. This algorithm uses a dynamic
genetic algorithm. The proposed distributed VOD system environment consists
of a number of VOD servers, which are distributed geographically, and
control servers that manage each group of VOD servers. User requests are dis-
tributed to prevent convergence by distributing VOD servers geographically.
We use a genetic algorithm based on history data to distribute user requests in a
local service area. The information of user requests and services is stored and
referred in a database as history data. By applying these history data to the fit-
ness function of genetic algorithms, we implemented the genetic algorithm
and operations for VOD systems. The load balancing algorithm proposed in this
paper can distributed workloads by predicting workloads precisely on VOD
environments.

Keywords: VOD, distributed system, load balancing, genetic algorithm.

*
Corresponding Author.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 269–276, 2010.
© Springer-Verlag Berlin Heidelberg 2010
270 J. Moon, H.-j. Moon, and Y. Cho

1 Introduction
Recently, the Internet is growing increasingly due to the development of computer
networks. In addition, there were many attempts to transfer audio and video data
through the Internet, and it can be possible to transfer movies via the World Wide
Web and high-speed Internet. Video-on-Demand (VOD) service is activating due to
these attempts and technologies. VOD allows users to select and watch/listen to video
or audio contents on demand over high speed networks, while sitting at home, and
have almost as good a control over the viewing of the video as when using a conven-
tional VCR [1]. VOD services have some different characteristics from HTTP service
as well access patterns. VOD services have long service time more than HTTP ser-
vices, and request more network bandwidth and hardware disk. Moreover, VOD is
more QoS-sensitive service than HTTP service, so users cancel their requests when
the service is delayed. Therefore, the scheduling algorithms that used in distributed
environments are not suitable for VOD system environments.
There are many researches on scheduling algorithm for distributed system
[2][3][4]. We have to consider dynamic loads information about servers in dynamic
service environments. Moreover, analysis of users request pattern must be considered
to predict system loads and balance loads of VOD system. For examples, when we
know the access pattern of users for a day, we are able to get information about the
time of day that resources are most frequently used. CDN (Content Distribution Net-
work) service uses such access pattern to store contents that are frequently used into
cache memory. If we record histories of user requests and then analyze the history,
the user request pattern is recognized by using some data mining techniques. Such
users request pattern operates as an important part of load balancing for VOD system
environments. In this paper, we suggest a scheduling algorithm for hierarchical VOD
system environment to provide high QoS and load balancing. The proposed schedul-
ing algorithm uses users request pattern, and selects appropriate servers to assign
various requests by using genetic algorithm. We developed a fitness function based on
history of user requests for the genetic algorithm.

2 System Model
VOD systems divided into three categories according to how to manage VOD servers:
central, distributed, and hierarchical server model. The central server model consists of
many VOD servers and a central server which controls users request to the VOD serv-
ers. This model is easy to set up and provides fast response time due to very simple
scheduling algorithm. However, the model results in lower quality when users far from
the central server. Moreover, the front-end server may cause a Single Point of Failure
(SPOF) because the central server does not work when user requests increase rapidly.
Distributed server model distributes VOD servers managing user requests by them-
selves and has no central managing server. This model provides fast response time
due to assigning user requests to the nearest VOD server. When the VOD server gets
high loads, user requests are forwarded to another VOD server nearest. Therefore,
there is no SPOF and provides scalability easily. However, VOD server has to keep
track of other servers’ loads information to forward user requests to another server
 A History-Based Scheduler for Dynamic Load Balancing 271

which has low load and located near. The number of distributed servers increase, the
amount of information to keep track of as well as network overhead also increase.
Hierarchical distributed server model combines the advantages of central and dis-
tributed server model. This model distributes VOD servers geographically, and con-
trol servers that located in each site manage user requests and VOD servers. This
model can reduce control servers’ loads by assigning user requests to the nearest
server group. Therefore, there are no SPOF and scalability is provided by adding
VOD server in each site.
In this paper, we construct a hierarchical distributed VOD system model to develop
history based scheduling algorithm. Fig. 1 shows the system model which consists of a
global DNS, VOD servers, control servers, and database server. Each server group can
be composed of many VOD servers and a control server which manage the servers in
the group. And the control server performs load balancing by distributing user requests.

Fig. 1. Hierarchical Distributed VOD System Model

3 Proposed Scheduling Algorithm

In this section, we suggest a two level load balancing algorithm for hierarchical dis-
tributed VOD system. The first step is a global load distribution by using global DNS,
and the second is local load distribution. Fig. 2 shows the global load balancing algo-
rithm which keeps track of best VOD server list and is performed by the global DNS
[5]. We assumed that there are N servers which are registered in the global.
272 J. Moon, H.-j. Moon, and Y. Cho

while (1) {
for (i = 1 to n)
{
Ti , Load = α × Li , CPU + β × Li , MEM + γ × Li , NT

if (( S i is in DNS List ) and (Ti , Load > TH ))

then remove S i from DNS list

else if (( Si is not in DNS list ) and (Ti , Load < TH ))

then add S i to DNS list

}
}
Ti , Load = total loads of ith control server
α , β , γ : weighting value (α + β + γ = 100), TH : threshold value
S i = ith control server
Li ,CPU = CPU loads of ith Server , Li ,MEM = Memory loads of ith server
Li , NT = Network loads of ith Server

Fig. 2. Global load balancing algorithm

First of all, we need to calculate total load of ith control server Si by estimating
CPU, memory, and network utilization rate of each server. A monitoring module
running on the control server collects each server’s CPU, memory, and network loads.
We assign a different weighting value according to the service provided by distributed
VOD system. Some services that transfer large amounts of data such as online broad-
casting or VOD service have more increase in network overhead than CPU and mem-
ory, while some services that request large amount of computing power such as online
game have more increase in CPU and memory utilization than network. Therefore, we
put more weight on the resources that have a big increase according to the type of
service.
When a control server Si is in the global DNS list and total load is larger than
the threshold value TH, the Si is removed from the list. When the Si is not in the list
and total load is less than the threshold value, Si is inserted in the list again. Then the
DNS manage the list with best servers. The threshold value is determined by using
heuristics and is assigned by administrator manually. We provide the following four
scheduling policies based on load information that is collected by monitoring module
running on control servers, and can modify the policy dynamically.
 A History-Based Scheduler for Dynamic Load Balancing 273

• CPU based : α =100, β =0, γ =0, when a service is CPU intensive

• Memory based : α =100, β =0, γ =0, when a service is Memory intensive
• Network based : α =0, β =0, γ =100, when a service is Network intensive
• Combined based : α + β + γ =100, when a service uses all of the resources, the
weight is assigned by administrator

Local load distribution is a method to distribute load between VOD servers managed
by a control server. We adopt a genetic algorithm to local distribution because there is
no general solution to find best servers under the various types of service and lots of
requests. When the requests come into the queue, we need to sort the requests as the
service type. After sorting, we generate a population of candidate solutions. We create
candidate solutions from 10% of all m× n cases randomly. A fitness function is a
particular type of objective function that prescribes the optimality of a solution in a
genetic algorithm so that that particular chromosome may be ranked against all the
other chromosomes. Optimal chromosomes, or at least chromosomes which are more
optimal, are allowed to breed and mix their datasets by any of several techniques. And
the following is a fitness function and related expressions.
m
F(x) = ∑T
i =1
i

Ti = current loads on Pi × working time on Pi × Load History Value

( Pi : ith process, Ti : expected working time)
(Loads on Pi = CPU Utilization × Memory Utilization × Network Utilization on Pi )
(Load History Value = Average load of jobs that executed before)
1
Fitness value =
F ( x)

4 Performance Evaluation
We implemented a simulator which is based on OPNET Modeler [6] and conducted
some comparison experiments with GA (genetic algorithm), RR (round-robin), and
Random scheduling algorithm when users increase. The first experiment is about
changes of average response time when users increase from 1 to 3,000, and Fig. 3
shows the result. As you can see from Fig. 3, the average response time is pretty much
the same to some extent which is about 800. However, there is a growing gap among
scheduling algorithms. Random scheduling shows the largest increase because highly
loaded servers are getting user requests continuously. In this experiment, the average
response time of GA, the proposed scheduling algorithm, is under 40 seconds until
users are 3,000.
Fig. 4 shows the cancellation rate when users increase. There is a little probability
that users cancel popular VODs and there is strong probability that users cancel non-
popular VODs. The cancellation rate also increases when the response time is getting
274 J. Moon, H.-j. Moon, and Y. Cho

Fig. 3. Average Response Time when users increase

Fig. 4. Service Cancellation Rate when users increase

Fig. 5. Average CPU Utilization when users increase

 A History-Based Scheduler for Dynamic Load Balancing 275

too long. Compared to response time, it looks like Random algorithm has a high can-
cellation rate. This means that there are many users who cancel the VOD due to long
response time because the response time had a lot of ups and downs. There is no big
change of the average response times of RR and GA. However, the gap is growing
over 2,600 when using RR. In this experiment, Random algorithm shows 77%, RR
shows 13%, and GA shows 2.2% of cancellation rate. The result shows that there is a
little probability that users cancel the service using GA excepting when the user can-
cel the service due to lack of interest.
The next experiment is about average CPU utilization when the users increase.
Fig. 5 shows the result that the CPU utilization rate is growing when users increase.
The CPU utilization rate increase sharply when we use Random scheduling because
the requests may be assigned into highly loaded server. In response to this we can see
the response time increase in Fig. 3. The average CPU utilization of RR and GA
scheduling increases slowly and steadily. GA scheduling shows the very gently in-
crease because lots of requests are evenly distributed and the gap between servers’
loads is getting close.
We also conducted an experiment using user request pattern and history during a
day. The users’ pattern can get from the service history. In this experiment, we as-
sumed the user requests increase dramatically at 6 P.M. and the numbers of users
increase from 1 to 1,000.
Fig. 6 shows the result of the changes of response time as the user request pattern.
The x-axis refers to the time from 10 A.M. to 10 P.M. The response time is getting
longer when the rush hours of about 6 P.M. Random and RR scheduling shows the
response time is getting longer sharply. Random scheduling has a large gap between
response times when the system has not many requests, and moreover the response
time is longer when the requests increase rapidly. However, GA scheduling predicts
the loads of servers and distributes the loads evenly among servers, so the response
time increases slowly. Therefore, we can see that the proposed scheduling algorithm
provides more enhanced QoS.

Fig. 6. Response time as time passed

276 J. Moon, H.-j. Moon, and Y. Cho

5 Conclusions and Future Works

In this paper, we suggested a dynamic load balancing algorithm in hierarchical dis-
tributed VOD system environment. The hierarchical distributed VOD system model
consists of a global DNS, VOD servers, control servers, and database server. Each
server group can be composed of many VOD servers and a control server which man-
age the servers in the group. And the control server performs load balancing by dis-
tributing user requests. The dynamic load balancing algorithm is based on genetic
algorithm and history information of user request pattern. We implemented a simula-
tor based on OPNET. And we conducted some comparison experiments with GA, RR
and Random scheduling method. According to the results, we can see that the pro-
posed algorithm provides more enhanced QoS.

References
1. Ma, K., Shin, K.G.: Multicast Video-On-Demand Services. ACM SIGCOMM Computer
Communication Review 32(1), 31–43 (2002)
2. Loukopoulos, T., Ahmad, I.: Static and Adaptive Distributed Data Replication Using
Genetic Algorithms. J. of Parallel and Distributed Computing 64(11), 1270–1285 (2004)
3. Rothlauf, F.: Representations for Genetic and Evolutionary Algorithms. In: Studies on
Fuzziness and Soft Computing, vol. 104 (2002)
4. Mundur, P., Simon, R., Sood, A.K.: End-to-End Analysis of Distributed Video-on-Demand
Systems. IEEE Transactions on Multimedia 6(1), 129–141 (2004)
5. Moon, J.-B., Kim, M.-H.: Dynamic Load Balancing Method Based on DNS for Distributed
Web Systems. In: Bauknecht, K., Pröll, B., Werthner, H. (eds.) EC-Web 2005. LNCS,
vol. 3590, pp. 238–247. Springer, Heidelberg (2005)
6. OPNET Technologies Inc., http://www.opnet.com/
A Secure Routing Protocol for Wireless Sensor Networks

Jaydip Sen and Arijit Ukil

Innovation Lab, Tata Consultancy Services Ltd., Bengal Intelligent Park,

Salt Lake Electronics Complex, Kolkata, India
{Jaydip.Sen,Arijit.Ukil}@tcs.com

Abstract. Wireless sensor networking has been a subject of extensive research

efforts in the recent years, and has been well recognized as a ubiquitous and
general approach for some emerging applications such as a real-time traffic
monitoring, ecosystem and battlefield surveillance. Since these networks deal
with sensitive data, it is imperative that they are made secure against various
types of attacks such as node capture, physical tampering, eavesdropping, de-
nial of service etc. This paper presents a secure routing mechanism for wireless
sensor networks. The protocol is resilient in the presence of malicious nodes
that may launch selective packet dropping attack on the routing path. The
scheme employs single-path routing, and therefore, it is energy-efficient. While
the packets are forwarded towards the base station, if any node fails to forward
a packet, it is isolated immediately. Packets are then routed around the node.
Simulation conducted on the scheme clearly demonstrates that it is more effi-
cient than some of the existing similar schemes.

Keywords: Wireless sensor networks, routing, malicious packet dropping,

neighborhood watch, energy-efficiency, reliability.

1 Introduction
Recent convergence of technological and application trends has resulted in an excep-
tional level of interest in wireless ad hoc networks, and in particular, wireless sensor
networks (WSNs). The push was provided by rapid progress in computation and
communication technology as well as the emerging field of low cost, reliable,
MEMS-based sensors, while the pull was provided by numerous applications that can
be summarized under the umbrella of computational worlds, where the physical world
can be observed and influenced through the Internet and WSN infrastructures. Typi-
cally, WSNs contain hundreds or thousands of sensor nodes that have the ability to
communicate among them and also to a base station (BS) [1]. While the sensor nodes
have limited sensing region, processing power and energy, networking a large number
of such nodes makes a robust, reliable and accurate sensor network covering a wider
region. The WSNs are envisioned to play an important role in a wide variety of appli-
cations ranging from critical military surveillance to forest fire monitoring and the
building security monitoring. Since these networks are mostly deployed in harsh and
often hostile operating environments, security and reliability are two important as-
pects for their operations. Design of a robust and secure routing protocol for WSNs is

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 277–290, 2010.
© Springer-Verlag Berlin Heidelberg 2010
278 J. Sen and A. Ukil

a particularly challenging problem due to limited computational resources of the

nodes and unreliable wireless links. In addition, data transmission through these net-
works is prone to many types of failures and attacks such as node capture, physical
tampering, eavesdropping, denial of service, etc [2, 3, 4]. Some of the previous re-
search efforts in secure routing for WSNs have focused on the design of single-path
routing using low-rate and periodic flooding of events to avoid failed nodes [5]. A
fault-tolerant cluster-based routing scheme has been proposed in [6]. In [24], a multi-
path routing protocol has been presented that uses three paths from the source node to
the sink to provide robustness.
To address the problem of security and efficiency in routing in WSNs, this paper
proposes a scheme that reliably identifies compromised (or faulty) nodes and utilizes
a routing path that avoids these nodes. Essentially, it utilizes a single-path routing
concept and thereby saves energy-consumption. If a malicious node is detected on the
next-hop on the routing path, the node is efficiently bypassed and the packets are
routed around the node to the base station still in a single-path. The proposed proto-
col is a modification of the routing scheme proposed in [7]. However, it is more en-
ergy- efficient and less delay-inducing as observed from the simulation results.
The rest of the paper is organized as follows. Section 2 discusses some related
work in the area of secure routing in WSNs. Section 3 presents the WSN model and
possible types of packet dropping attacks on these networks. It also gives a brief in-
troduction to the LEAP protocol for key management used in the proposed scheme.
Section 4 gives a detailed description of the proposed protocol. Section 5 presents the
simulation results and Section 6 concludes the paper.

2 Related Work
A considerable amount of research effort has been put in design of secure routing
protocols for WSNs. While some of these works have focused on defense against
outsider attacks on the key management protocols being used in the network [8, 9,
10], the others have attempted to secure node-to-node communication [10, 11]. How-
ever, all these schemes fail even if a single node is compromised.
An insider attack on a WSN is even more dangerous, since an attacker in this case
can launch various types of attacks such as dropping of legitimate packets, injecting
of bogus packets and sensing reports, advertising false routing messages, eavesdrop-
ping on the network communication etc. Parno et al have proposed a distributed
detection mechanism [12]. Newsome et al [13] have presented some techniques
for preventing an adversary from launching Sybil attack by arbitrarily creating new
identities.
Deng et al have proposed an intrusion tolerant routing protocol in wireless sensor
networks (INSENS) that adopts a routing-based approach to security in WSNs [14]. It
constructs routing tables in each node, bypassing malicious nodes in the network. The
protocol cannot totally prevent attack on nodes, but it minimizes the damage caused
to the network due to an attack. The scheme has reduced computation, communica-
tion, storage, and bandwidth requirements at the sensor nodes at the cost of greater
computation and communication overhead at the base station. Tanachaiwiwat et al
have proposed a novel secure routing protocol- trust routing for location-aware
 A Secure Routing Protocol for Wireless Sensor Networks 279

sensor networks (TRANS) [15]. It makes use of a loose-time synchronization asym-

metric key cryptographic scheme to ensure message confidentiality. Zhu et al have
proposed a localized encryption and authentication protocol (LEAP) to prevent in-
sider attacks--particularly message eavesdropping and message fabrication attacks
[16]. Compromised or faulty nodes may also drop legitimate packets by launching
selective forwarding attacks [2]. In order to defend against such type of attacks, most
of the existing secure routing protocols in sensor networks are based on multi-path
forwarding scheme [17, 18], or interleaved mesh forwarding scheme [19].
In contrast to multi-path approach, this paper presents a single-path packet-
forwarding scheme in sensor networks where some of the nodes may be malicious
and drop any packets that they receive. The protocol is an extension of the protocol
presented in [7] and is more energy-efficient than the base protocol. The main contri-
bution of the paper is its energy-efficiency and increased reliability, since the packets
in the network are always routed in single-path avoiding costly multi-path approach.
In the next section, the sensor network model and its associated security vulnerabili-
ties are discussed.

3 Network Threats and Key Management

The proposed protocol is suitable for a large and dense sensor network where one or
multiple base stations and a large number of tiny, inexpensive, static and resource-
constrained sensor nodes are deployed over a wide area. Data gathered by the sensor
nodes are sent along multi-hop routes to the base station. It is assumed that each sen-
sor node has the same transmission range and all links are bi-directional.
It is also assumed that a key establishment scheme like LEAP [16] is already de-
ployed in the nodes. A short description of LEAP is given in Section 3.2.

3.1 Threat Model

The goal of the proposed scheme is to defend against packet-dropping attacks that
may be launched by compromised nodes in a WSN. As mentioned in Section 2, the
insider attacks may have devastating effects on the performance of a WSN and suit-
able defense mechanism should be deployed to counter such attacks.
A node in a WSN may drop packets because of two reasons: (i) it may be faulty
due to some problems in its hardware and/or software, or (ii) it may be malicious and
intentionally may not forward any packets. To ensure reliable packet delivery, some
authors [20, 21] have proposed use of acknowledgements (ACK). However, in pres-
ence of malicious nodes, an ACK received from the next-hop node cannot guarantee
that the packet is forwarded by the next-hop node; it only implies that the packet has
been received by the next-hop node. It will not be possible to ensure that the next-hop
node really forwards the packet just by implementing an ACK mechanism.
Due to the broadcast nature of the wireless links in the sensor network, it is possi-
ble for a node to overhear communications to and from its neighbor nodes. Therefore,
a node u after forwarding a packet to its next-hop v can listen on v’s communication
to check whether v really forwards the packet further. The simple monitoring mecha-
nism, however, cannot also guarantee reliable packet delivery either. A malicious
280 J. Sen and A. Ukil

node may forward the packet but it can make its intended next-hop id such that there
is no node in the neighborhood which has the matching id. Therefore, even if the
packet is forwarded by the malicious node it is ultimately lost in the network. This is
called routing disruption attack, and it is impossible to detect such attack by a naïve
neighborhood monitoring scheme.
The objective of the proposed scheme is to address all types of packet-dropping at-
tacks ranging from the most naïve type (e.g., detection of a faulty node) to the most
malicious type (e.g., routing disruption attack).

3.2 Key Establishment Scheme in LEAP

The localized encryption and authentication protocol (LEAP) proposed by Zhu et al

[16] is a key management protocol for WSNs based on symmetric key algorithms. It
uses different keying mechanisms for different types of packets depending on their
security requirements. Four types of keys are established for each node: (i) an indi-
vidual key shared with the base station (pre-distributed), (ii) a group key shared by all
the nodes in the network (pre-distributed), (iii) pair-wise keys shared with its neighbor
node, and (iv) a cluster key shared with multiple neighbor nodes. The pair-wise keys
shared with immediate neighbor nodes are used to protect peer-to-peer communica-
tion and the cluster key is used for local broadcast.
It is assumed that the time required for an attacker to attack a node, Tmin, is greater
than the time interval during which a sensor node can detect all its intermediate
neighbors, Test. This is a reasonable and practical assumption which has also been
made in some of the other research contributions [22].
A newly joined node u in the network executes the following four steps to establish
a pair-wise key with each of its neighbors. The steps are as follows:
1. Key Pre-distribution: each node u is loaded with a common initial key K1 and
derives its master key Ku = fK1 (u), where fK is a pseudo-random function. The
master key depends on the common key and the unique identifier of the node.
2. Neighbor discovery: node u sets up a timer to fire after time Tmin, broadcasts its
id, and waits for each neighbor v’s ACK. The ACK from v is authenticated using
the master key Kv, of node v. Since node u knows K1, it can derive Kv = fK1 (v).
u Æ * : u, Ru
v Æ u : v, MAC (Kv, Ru | v)
3. Pair-wise key establishment: node u computes its pair-wise key (Kuv) with v, as
Kuv = FKv (u). Node v also computes Kuv in the same way. Kuv serves as their
pair-wise key between the nodes u and v.
4. Key revocation: when its timer expires, node u erases K1 and all the master keys
of its neighbors. Every node, however, keeps its own master key, in order to es-
tablish pair-wise keys with nodes which have joined later.
Once K1 is erased, a node will not be able to establish a pair-wise key with any
other nodes that have also erased K1. Without the knowledge of K1, an attacker who
has been able to compromise a node after Tmin, fails to establish pair-wise keys with
any nodes except the neighbors of the compromised node. In such a way, the protocol
is able to localize any possible attack on the network.
 A Secure Routing Protocol for Wireless Sensor Networks 281

The cluster key is established by a node after the establishment of the pair-wise key
establishment. A node generates a cluster key and sends it to each of its neighbors
after encrypting it using the pair-wise shared key. The group key is pre-loaded but is
updated once any compromised node is detected in the network. This could be done
either by base station’s sending the new group key to each node using its individual
key, or a hop-by-hop basis using cluster keys.

4 The Proposed Routing Protocol

This section presents the proposed secure routing protocol for WSNs. The protocol is
based on a robust neighborhood monitoring system (NMS) and is an extension of the
scheme proposed by Lee and Choi [7]. NMS works on promiscuous monitoring of the
neighborhood by a node and detection of any possible malicious packet dropping
attack by a cooperative algorithm using neighbor list checking to be described in
Section 4.1.
The proposed scheme ensures reliable hop-by-hop delivery of packets in a WSN
even in presence of malicious nodes that may launch packet-dropping attack in the
routing path. To defend against packet-dropping attack, most of the proposed algo-
rithms in the literature have exploited the concept of multi-path routing, where a sin-
gle packet is routed through multiple paths from the source to the sink. While this
approach ensures reliable packet delivery, it consumes a high amount of energy for
delivery of each packet. To avoid this problem, the proposed algorithm uses a single-
path routing mechanism. If a malicious node is encountered, the node is avoided and
the packet is routed around it in an efficient manner (discussed later in this section)
still in a single-path mode to the base station. The selection of the new path is based
on some broadcast signaling in the neighborhood of the malicious node.
The proposed algorithm is more efficient than the algorithm proposed in [7], since
it always uses single-path routing in contrast to the occasional multi-path approach
used in [7]. The use of single-path routing makes it more energy-efficient and less
delay-inducing. The algorithm is described below:
1. Neighbor list checking: during the neighbor discovery phase, each node ex-
changes Hello messages with its neighbor nodes to know its 1-hop and 2-hop
neighbors (i.e., neighbors of each of its neighboring nodes). The neighborhood
information is subsequently verified by exchange of neighbor list checking mes-
sages as described in Section 4.1.
2. One-hop packet forwarding: when a node u sends a packet to its neighbor, it first
keeps a copy of the packet in its buffer, and then forwards it to its next-hop node
v before encrypting it with the cluster key of the node u. Sine the cluster key is
shared between the node and all its neighbors, the packet encrypted and sent by
node u to node v can be overheard by all the neighbors of node u.
3. Monitoring nodes selection: as the packet is being forwarded from node u to
node v, the neighbors of node u that are also neighbors of node v receive the
packet and store it in their buffers. These nodes are designated as the secondary
monitoring nodes. For example, in Fig. 1, nodes w and y are the secondary
monitoring nodes for node v. The node u is the primary monitoring node in this
case. The nodes that are not neighbors of node v but have received the packet
282 J. Sen and A. Ukil

because they are neighbors of node u, discard the packet. The primary node
knows the secondary monitoring nodes, since every node has the knowledge of
its 1-hop and 2-hop neighbors.

Fig. 1. Neighbor monitor system (secondary nodes w, y and primary node u)

4. Role of secondary monitoring nodes: secondary monitoring nodes w and y moni-

tor the traffic from node v and compare the outbound packets from node v with
the packets stored in their buffer. The next-hop address of each packet is also
verified to check whether the packet’s intended next-hop is a really a neighbor
of node v, by cross-checking the neighbor list of node v. If both these checks
yield positive results, the secondary monitoring nodes remove the packet from
their buffer and their role of monitoring is complete for that packet. If any
packet is found to remain in the buffer of a secondary monitoring node for more
than a threshold period of time, it first sends a broadcast signal in its neighbor-
hood to inform all its neighbors that it is going to forward the packet to its des-
ignated next-hop so that other neighbors do not forward the same packet. The
secondary monitoring node now forwards the packet to its designated next-hop
after encrypting the packet with the cluster key. The role of the secondary node
now becomes that of the primary node and its neighbors become the secondary
node. This is in contrast to the scheme proposed in [7] in which all the secon-
dary nodes forward the packet in a multi-path mode. The use of a single-path in-
stead of multi-path makes the proposed scheme more energy-efficient and less
delay-inducing compared to the one proposed in [7].
5. Role of primary monitoring node: The role of a primary monitoring node (node
u) is identical to that of secondary monitoring nodes (nodes w and y); the only
difference is that it listens not only on the traffic from node v, but also on the
traffic from the nodes w and y. If the packet is correctly forwarded by any one of
the nodes v, w, y, the node u removes the packet from its buffer. The role of
node u as the primary monitoring node is now complete. If time out occurs for a
packet, the primary monitoring node u forwards the packet (encrypted with its
cluster key) to its next-hop other than node v.

As the packet is routed along a path towards the sink, the above steps of NMS al-
gorithm except the neighbor list checking are executed at each hop so that reliable
packet delivery can happen through a single path. This is in contrast to the previous
schemes proposed in [19, 21, 22]. In these schemes, a node broadcasts a packet with-
out specifying a designated next-hop, and all neighboring nodes with smaller costs
 A Secure Routing Protocol for Wireless Sensor Networks 283

(the cost at a node is the minimum energy required to forward a packet from the node
to the base station) or within a specific geographic region continue forwarding the
packet to the base station. If nodes v, w, and y have smaller costs than node u in Fig.
1, then each of them will forward packets received from node u following the existing
approaches. However, in the proposed scheme, nodes w and y only observe the packet
forwarding activities of node v, instead of actively forwarding the packets. In the
event of no packet drop, the routing to the base station happens in a single-path
thereby making the process highly energy-efficient.
Even in the event of a packet drop, the proposed algorithm works in a single-path
mode. This makes it more efficient than the one proposed in [7]. If the node v in Fig.
1 does not forward the packet it has received from node u, then one of the secondary
monitoring nodes w and y would forward the packet to its next-hop nodes. The node
(either w or y) that forwards the packet to its next-hop neighbors will first send a
broadcast message in its neighborhood so that its other neighbors would not forward
the same packet. This single-path routing approach even in the event of packet drop-
ping makes the proposed scheme extremely energy-efficient and very less delay-
inducing. Fig. 2 shows an example of the application of the scheme, where two mali-
cious (or faulty) nodes are bypassed as the packet is routed to the base station in a
single-path.

Fig. 2. Two malicious nodes identified by secondary monitoring nodes

For the scheme to work, each packet should be encrypted with a cluster key of the
forwarding node so that all the neighbors of the forwarding node can decrypt and
overhear it. If a link-level encryption was applied between each pair of nodes in the
routing path, the scheme would have been more robust, since a compromised node
could decrypt only the packets which were destined to it. However, it would have
made the scheme less resilient to packet dropping attack. Since encryption with a
cluster key provides a reasonable level of robustness to a node compromise, [11] and
also supports local broadcast (i.e. resiliency against packet-dropping) it makes the
algorithm optimum in its performance [11].
To make the scheme robust to routing disruption attack, where a node intentionally
forwards the packets to a spurious address of the next-hop so that the packet is lost in
routing, it is necessary that each node should prove that it really has the claimed
neighbors. Apparently, a node has the knowledge of its direct neighbors by neighbor
discovery and pair-wise key establishment phases discussed earlier. However, in the
284 J. Sen and A. Ukil

case of two-hop neighbors, a malicious node v can inform its neighbor u that it also
has neighbor node x (any possible id in the network) which in fact is not a neighbor of
node v (Fig. 1). Apparently, there is no way node u can detect these false claim of v
since x is not in the neighborhood of u. Therefore, it is absolutely necessary to devise
a scheme by which a node can verify the neighbors of each of its neighboring nodes.
This scheme is explained in the next subsection.

4.1 Neighbor List Checking

To verify the 2-hop neighbors, i.e., the neighbors of the neighbors of each node, the
proposed scheme utilizes a concept of neighbor list checking (NLC). NLC uses the
pair-wise keys established in LEAP. During neighbor discovery, NLC uses three-way
handshaking in order to identify not only the communicating nodes but also their
respective neighbors. There are two distinct types of neighbor discovery in NLC. In
one case, both the nodes involved in the neighbor discovery process are still within
the initial Tmin – both the nodes are referred to as new nodes. In the other case, the
neighbor discovery happens between a newly-deployed node within the initial Tmin
and an existing node over the initial Tmin – the latter node is referred to as an old node.
(i) Neighbor discovery between new nodes: If a new node broadcasts its neighbor
list before the expiry of its initial Tmin, it is verifiable. Therefore, it is essential that
both nodes broadcast their respective neighbor lists before their respective Tmin. The
three-way handshaking neighbor discovery between pure node u and v may be repre-
sented as follows [7]:

u Æ * : u, Ru
v Æ u : v, Tv, Rv, MAC (Kv, Ru | Ku | (u, Tv, Rv))
u Æ v : u, Tu, MAC (Kuv, Rv | (u, Tu))

Tv and Tu are the time remaining to reach Tmin of v and Tmin of u, respectively. Ru is
a random number generated by node u. MAC (K, M1 | M2) denotes message authenti-
cation code of concatenated message of M1 and M2 using key K. Node u broadcasts its
id, and waits for each neighbor v’s ACK. The ACK from every neighbor v is authenti-
cated using the master key Kv of node v. Since node u knows KI, it can derive Kv = fKI
(v). The ACK from node v contains Tv. If Tv is non-zero, node v claims to be a new
node. Ku in MAC proves node v to be a new node, since new node v should know KI
and derive Ku = fKI (u). Node u records T^v (Tv added to the current time of node u) in
the entry for node v in the neighbor information table (NIT). Node u computes its
pair-wise key with v, Kuv = fKv (u). Node u also generates MAC (Kv, v | u) (which
means that v certifies u as an immediate neighbor), and stores it as a certificate.
The ACK from node u also contains Tu. This ACK is authenticated using their pair-
wise key Kuv, which proves node u a new node and u’s identity. Node v then records
T^u (Tu added to the current time of v) in the entry for u in its NIT. It also generates
MAC (Ku, u | v) and stores it as a certificate. This completes the three-way handshak-
ing process.
 A Secure Routing Protocol for Wireless Sensor Networks 285

Every new node u broadcasts its neighbor list just prior to Tmin of u. Each receiving
neighbor v checks whether the receiving time at v is prior to T^u in the NIT. If yes, the
neighbor list of u is now certified by each neighbor v.
(ii) Discovery between a new and an old node: After the bootstrapping phase, new
nodes can join the network. This situation is depicted in Fig. 3, where the new node x
has joined the network. The nodes u, v and w are old nodes. In this scenario, the mes-
sages exchanged during the three-way handshaking are different. The message ex-
change between the new node x and the old node u is as follows [7]:

x Æ * : x, Rx
u Æ x : u, Tu, Ru, v, MAC (Kv, v | u), w, MAC (Kw, w | u), MAC (Ku, Rx | (u, Tu, v,
MAC (Kv, v | u), w, MAC (Kw, w | u))
x Æ u : x, Tx, MAC (Kx, x | u), v, MAC (Kv, x | u), w, MAC (Kw, x | u), MAC (Kxu,
Ru | (x, Tx, MAC (Kx, x | u), v, MAC (Kv, x | u), w, MAC (Kw, x | u))

The node x broadcasts its id, and waits for the ACKs from its neighbors. The ACK
from neighbor u is authenticated using the master key Ku of u. Since x knows K1, it
computes Ku = fKI (u). The ACK from u contains Tu, the time remaining to reach Tmin
of u. As u is an old node in Fig. 3, Tu is zero. Node u sends its neighbor list (v and w)
to x including their respective certificates in the ACK. Node x can verify the list since
it knows KI. If the verification is positive, x computes pair-wise key Kxu = fKu (x), and
also generates MAC (Ku, u | x) and stores it as a certificate.

Fig. 3. Neighbor discover between a new node x and an old node u

The ACK from x also contains Tx, the time remaining to reach Tmin of x. This ACK
is authenticated using Kxu, to prove x’s identity as a new node. Node u computes T^x
(Tx added to the current time of u) and makes an entry for x in the neighbor table.
Node x also sends the certificate for u, MAC (Kx, x | u) in the ACK. It also provides
one-time certificates for each of u’s neighbors (v and w) and completes the three-way
handshaking process.
Node u broadcasts the one-time certificates received from x to its neighbors w and
y so that they are able to verify the authenticity of x. The contents of the one-time
certificates are as follows [7]:

u Æ * : u, x, v, MAC (Kv, x | u), w, MAC (Kw, x | u), KAu, MAC (Kcu, (u, x, v, MAC
(Kv, x | u), w, MAC (Kw, x | u), KAu))
286 J. Sen and A. Ukil

KAu is a local broadcast authentication key in u’s one-way key chain, Kcu is the
cluster key of u. Each neighbor v of u verifies x by checking its one-time certificate,
MAC (Kv, x | u) and x is certified by each neighbor v of u.
Before the expiry of Tmin of x, it broadcasts its neighbor list. Each neighbor u of x
checks whether the receiving time at u is before T^x in the neighbor table. If yes, the
neighbor list of x is certified by u.
Through the above neighbor list checking process, every node knows the certified
neighbors of each of its neighbors. With this knowledge, the monitoring nodes check
the validity of the claims of the next-hop of the target forwarding node.

4.2 Neighbor Information Table

The information obtained through neighbor list checking is stored in the neighbor
information table (NIT) of each node. Table 1 shows the NIT of node u as depicted in
Fig. 3.

Table 1. Neighbor information table of node u

Neighbor ID Certificate Neighbor list

v MAC (Kv, v | u) u, w, t
w MAC (Kw, w | u) u, v, z
x MAC (Kx, x | u) u, r, q

The NIT is used by the monitoring nodes. For example, if node u overhears a
packet sent from w to v, it first checks its NIT and finds that both w and v are its
neighbors. Node u starts listening on v’s traffic as a secondary monitoring node, and if
v does not forward the packet to one of its verified neighbors, u will first broadcast a
message in its neighborhood and then forward the packet to its next-hop other than v.
Node u then starts acting as a primary monitoring node.

5 Simulation Results
To evaluate the performance of the proposed protocol, simulations have been carried
out on network simulator ns-2. To compare the performance of the scheme with that
proposed in [7] identical parameters are chosen. The simulation parameters are pre-
sented in Table 2.
The malicious nodes are configured in such a way that they do not forward any
packets. These nodes are strategically distributed in a square area of 200 m side (from
150m to 350m of each side of the 500 m × 500 m target area), so that they are half-
way between the source and the base station. TinyOs beaconing [23] has been used as
the base routing protocol in the simulation to maintain uniformity with [7]. Each
simulation is repeated with 100 different topologies, and the average value is taken as
the final result.
Fig. 4 and Fig. 5 show the packet delivery ratio (the ratio of the number of packets
successfully reaching the base station to the total number packets sent by the source)
in presence of varying number of malicious nodes (x) with N = 300 and N = 600
 A Secure Routing Protocol for Wireless Sensor Networks 287

Table 2. Simulation parameters

Parameter Values
No. of sensor nodes (N) 300, 600
Area of simulation 500 m * 500 m
Transmission range of each sensor 30 m
Degree of each node 10 (N=300), 20 (N=600)
Position of source (50, 50)
Position of BS (450, 450)
Avg. no of hops from source to BS 18

Fig. 4. Packet delivery ratio vs. no. of malicious nodes (N = 300, x = 0 ~ 50)

Fig. 5. Packet delivery ratio vs. no. of malicious nodes (N = 600, x = 0 ~ 100)

respectively. It is observed that with N = 300, the packet delivery ratio falls with the
increase in number of malicious nodes. The performance of the proposed protocol is
found to be almost identical to that of NWS proposed in [7]. With both these proto-
cols, the packet delivery ratio is on the average more than 80%. Since in NWS some
packets are routed in multi-path, it has a marginally higher of packet delivery ratio in
288 J. Sen and A. Ukil

Fig. 4. As can be observed from Fig. 5, with N = 600, the proposed protocol has same
performance as that of NWS. Even when x = 100, the delivery ratio is as high as 99%.
As has been pointed out in [7], this is because of the higher degree of sensor nodes (d
= 20), more number of malicious nodes can be bypassed.

Fig. 6. No. of forwarding nodes vs. no of malicious nodes (N = 300, x = 0 ~ 50)

Fig. 7. No. of forwarding nodes vs. no of malicious nodes (N = 600, x = 0 ~ 100)

Fig. 6 and Fig. 7 show the number of nodes that has participated in packet forward-
ing for N = 300 and N = 600 respectively. Since the source is 18 hops away from the
base station, the number of relaying nodes in single-path routing case remains con-
stant at 18. The number of forwarding nodes was found to be increasing with x, since
more number of nodes needs to be bypassed. However, in the proposed protocol, the
number of forwarding nodes was found to be much smaller than in the NWS. In
NWS, with N = 600, the number of forwarding nodes increases very fast with the
increase in x, since more number of secondary monitoring nodes participate in rout-
ing. While this ensures reliability, it has an adverse effect on the energy-efficiency of
the network. As the same packet is being routed in multi-path, NWS protocol has
large energy consumption on the average. The proposed protocol while maintaining
 A Secure Routing Protocol for Wireless Sensor Networks 289

the same level of packet delivery ratio (i.e. reliability), involves less number of
forwarding nodes. It, therefore, consumes much less energy than NWS and thereby
enhances the network life-time.

6 Conclusion
This paper has presented a reliable and energy-efficient single-path routing protocol
for WSNs. The proposed protocol works reliably and efficiently even in presence of
malicious nodes on the routing path that drop packets without forwarding them. The
scheme is designed for hop-by-hop reliable forwarding of packets from the source to
the base station while routing every packets in single-path. The simulation results
have shown that the protocol is more energy-efficient than some of the existing rout-
ing protocols for WSNs while maintaining the same level of reliability.

References
1. Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci, E.: A Survey on Sensor Net-
works. IEEE Communications Magazine 40(8), 102–114 (2002)
2. Karlof, C., Wagner, D.: Secure Routing in Wireless Sensor Networks: Attacks and Coun-
termeasures. In: Proc. of 1st IEEE International Workshop on Sensor Network Protocols
and Applications, pp. 113–127 (2005)
3. Perrig, A., Stankovic, J., Wagner, D.: Security in Wireless Sensor Networks. Communica-
tions of the ACM 47(6), 53–57 (2004)
4. Wood, A., Stankovic, J.: Denial of Service in Sensor Networks. IEEE Computer 35, 54–62
(2002)
5. Intanagonwiwat, C., Govindan, R., Estrin, D., Heidemann, J., Silva, F.: Directed Diffusion
for Wireless Sensor Networking. ACM/IEEE Transactions on Networking 11(1), 2–16
(2003)
6. Gupta, G., Younis, M.: Fault-Tolerant Clustering of Wireless Sensor Networks. In: Proc.
of IEEE Wireless Communications and Network Conference (WCNC), pp. 1579–1584
(2003)
7. Lee, S.-B., Choi, Y.-H.: A Resilient Packet-Forwarding Scheme against Maliciously
Packet-Dropping Nodes in Sensor Networks. In: Proc. of the 4th ACM Workshop on Secu-
rity of Ad Hoc and Sensor Networks, pp. 59–70 (2006)
8. Chan, H., Perrig, A., Song, D.: Random Key Pre-Distribution Schemes for Sensor Net-
works. In: Proc. of IEEE Symposium on Security and Privacy, pp. 197–213 (2003)
9. Eschenauer, L., Gligor, V.D.: A Key Management Scheme for Distributed Sensor Net-
works. In: Proc. of the 9th ACM Conference on Computer and Communication Security
(CCS), pp. 41–47 (2002)
10. Perrig, A., Szewczyk, R., Wen, V., Culler, D., Tygar, J.: SPINS: Security Protocols for
Sensor Networks. In: Proc. of ACM MobiCom, pp. 189–199 (2001)
11. Karlof, C., Sastry, N., Wagner, D.: TinySec: A Link Layer Security Architecture for Wire-
less Sensor Networks. In: Proc. of ACM SensSys, pp. 162–175 (2004)
12. Parno, B., Perrig, A., Gligor, V.D.: Distributed Detection of Node Replication Attacks in
Sensor Networks. In: Proc. of IEEE Symposium on Security and Privacy, pp. 49–63
(2005)
290 J. Sen and A. Ukil

13. Newsome, J., Shi, E., Song, D., Perrig, A.: The Sybil Attack in Sensor Networks: Analysis
and Defenses. In: Proc. of IEEE IPSN, pp. 259–268 (2004)
14. Deng, J., Han, R., Mishra, S.: INSENS: Intrusion-Tolerant Routing in Wireless Sensor
Networks. Technical Report CU-CS-939-02, Dept. of Computer Science, University of
Colorado at Boulder (2002)
15. Tanachaiwiwat, S., Dave, P., Bhindwale, R., Helmy, A.: Routing on Trust and Isolating
Compromised Sensors in Location-Aware Sensor Networks. In: Proc. of 1st International
Conference on Embedded Networked Sensor Systems, pp. 324–325 (2003)
16. Zhu, S., Setia, S., Jajodia, S.: LEAP: Efficient Security Mechanisms for Large-Scale Dis-
tributed Sensor Networks. In: Proc. of the 10th ACM Conference on Computer and Com-
munications Security (CCS), pp. 62–72 (2003)
17. Ganesan, D., Govindan, R., Shenker, S., Estrin, D.: Highly Resilient, Energy-Efficient
Multipath Routing in Wireless Sensor Networks. Computing and Communications Review
(MC2R) 1, 11–25 (2002)
18. Deng, J., Han, R., Mishra, S.: A Performance Evaluation of Intrusion-Tolerant Routing in
Wireless Sensor Networks. In: Proc. of the 2nd International Workshop on Information
Processing in Sensor Networks (IPSN), pp. 349–364 (2003)
19. Ye, F., Zhong, G., Lu, S., Zhang, L.: GRAdient Broadcast: A Robust Data Delivery Proto-
col for Large Scale Sensor Networks. ACM Wireless Networks, WINET (2005)
20. Carbunar, B., Ioannidis, I., Nita-Rotaru, C.: JANUS: Towards Robust and Malicious Resil-
ient Routing in Hybrid Wireless Networks. In: Proc. of ACM Workshop on Wireless Secu-
rity, WiSe (2004)
21. Morcos, H., Matta, I., Bestavros, A.: M2RC: Multiplicative-Increase/Additive-Decrease
Multipath Routing Control for Wireless Sensor Networks. ACM SIGBED Review 2
(2005)
22. Yang, H., Ye, F., Yuan, Y., Lu, S., Arbough, W.: Towards Resilient Security in Wireless
Sensor Networks. In: Proc. of ACM MobiHoc, pp. 34–45 (2005)
23. Hill, J., Szewczyk, R., Woo, A., Hollar, S., Culler, D., Pister, K.: Systems Architecture Di-
rections for Networked Sensors. In: ACU ASPLOS IX (2000)
24. Ssu, K.-F., Wu, T.-T., Huang, S.-K., Jiau, H.C.: A Dependable Routing Protocol in Wire-
less Sensor Networks. In: Proc. of ICWMC, pp. 1–6 (2006)
 Efficient Pairwise Key Establishment Scheme Based on
Random Pre-distribution Keys in WSN∗

Hao Wang, Jian Yang, Ping Wang, and Pu Tu

Chongqing University of Posts and Telecommunications,

400065, Chongqing, China
[email protected], [email protected],
[email protected], [email protected]

Abstract. Key establishment and management are the cores of many security
mechanisms and services in wireless sensor networks which are commonly de-
ployed in a hostile environment. Pre-distributing keys into sensor nodes is a
practical option for key management in wireless sensor environment. The nego-
tiation and establishment of pairwise keys are much more important after the
nodes are deployed, so that the whole network will work in expectation and
safely. In this paper, we proposed an efficient pairwise key establishment
scheme used in cluster-based random key pre-distribution protocol, which can
provide perfect connectivity and security for large scale wireless sensor net-
works. Simulation results and performance analysis indicates that it also has
other strongpoint, lower communication and computational overhead, much
more scalable and flexible for the network size, a better balance between the
node storage capacity and communication overhead.

Keywords: Wireless sensor networks, security, key pre-distribution, pairwise

key.

1 Introduction
Wireless Sensor Networks (WSN), the main task is sending data which is collected
from the sensor nodes to the base station in time, consists of a large number of sensor
nodes. And the sensor node, which is powered by the built-in battery and equipped
with a tiny sensor device, has limited capability of data processing and storage, and
can carry out a certain distance wireless communication. WSN is a new research
direction in information technology due to its large-scale of applications, such as in
environmental monitoring, military, homeland security, traffic control, community
security, forest fire prevention, targeting areas and so on. As the sensor nodes are
mostly deployed in unattended or hostile areas, once the sensor network deployed in
enemy controlled areas, the enemy will tap the network communication, palm off the
right nodes ,send malicious information to the network, so that they can destroy the
network, and make it loss of normal function.
∗
The work reported in this paper is supported in part by the national science and technology
major project No. 2009ZX03006-001.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 291–304, 2010.
© Springer-Verlag Berlin Heidelberg 2010
292 H. Wang et al.

Therefore, Sensor network security problem is especially prominent. In order to

ensure the sensor networks works in safe way, when sensor nodes communicate with
each other, the message transmitted between each other should be encrypted, impor-
tant data must also be certified. To achieve this goal, we must first solve the key dis-
tribution problem, establishment of session key. To ensure that communications in
node-to-node or nodes-to-header (such as the cluster head, the base station) is safe, we
should establish a session key, namely, the pairwise key, between two nodes effec-
tively and safely. Focus of this paper is to build the network model and give the estab-
lishment of the pairwise key option.
The main contributions of this paper are summarized in the following:
1) We propose our cluster-based model of the network topology, which is greatly
based on practical application. This model achieves greatly strengthened security with
excellent scalability and increases the connected probability.
2) We present the pairwise key establishment scheme, both with common keys di-
rectly and using different-intermediate-nodes path, which extremely guarantee the key
security and secure communications in the future.
3) We give our analysis of selecting keys randomly and establishing pairwise keys
via intermediate nodes, which excellently balances the storage and the communications.
The remainder of this paper is organized as follows. We briefly describe and dis-
cuss current key pre-distribution schemes in Section 2. We then give a characteriza-
tion of our network topology and model in Section 3. A detailed introduction of our
proposed preferable pairwise key establishment scheme is presented in Section 4.
Section 5 gives analysis and evaluation of our proposed scheme. We conclude the
paper and summarize our results in section 5.

2 Related Work
To ensure secure communications, the sensor nodes need to consult keys; according
to the characteristics of sensor networks and the practical application, secret keys are
pre-loaded into each sensor node by using random deployment method before they
are deployed. There are several key pre-distribution models, the simplest way is that
all nodes share a master key, but this method has poor security; or any two of nodes
share the pairwise key, the security of such programs is good, but poor network ex-
pansion is not suitable for large-scale node network.
E-G program [9] , the basic scheme, with the consideration of security and the space
of storage in nodes, gives the basic random key pre-distribution scheme. This program
is suit for sensor networks and complete the establishment of the keys in three phases,
but did not give a clear way how to determine the pairwise key, or how to establish the
pairwise key by the search path. Chan et al. [3] add the minimum requirements that any
two of nodes share at least q common keys based the scheme [9]. This way increased
security, but with the increase of the compromised nodes, the performance become
relatively poor, meanwhile, it has the same deficiency with scheme [9].
Ren et al. [4] improve the structure of key ring in the basic scheme, give the
method of sub-key pool allocation; Du et al. [5] build a multi-key space based on
matrix theory; Liu et al. [6] use the polynomial instead of the key pool and construct
the application of the safety threshold in the polynomial; Du et al. [7] and Liu et al.
 Efficient Pairwise Key Establishment Scheme 293

[8] further extended the basic scheme using network pre-deployment knowledge. All
of these have a clear way how to determine the pairwise key.
Blom et al. [11] proposed the symmetric matrix operations; in order to the
network has the λ-secure property. Blundo et al. [12] proposed a polynomial-based
key pre-distribution scheme. Scheme [13] [14] also use the symmetric polynomial
f ( x, y ) = f ( y, x) to establish the pairwise keys, either for single base-station or mul-
tiple base-station. All these need much computing and have a poor scalability.
Our goal is to develop and improve the current key pre-distribution schemes. We
give the detailed description how to establish the pairwise keys which provide a per-
fect security and connectivity. The scheme we proposed is not only reduced the keys
stored in nodes but also improve the network’s security.

3 Cluster-Based Deployment Model

3.1 Notation

For clarity, we list the symbols used in the paper below:

P key pool, set of keys randomly chosen from the setup server
|P| size of the key pool, number of keys in the P set
M sub key pool
|M| size of the sub key pool
m number of keys in a node’s key ring
n network size, in nodes
t number of clusters
Ni node whose identifier is Ni
r the expected number of one cluster, r=n/t
p probability
Bu the broadcasting message of node u
kuv pairwise key shared by u and v
N u random value generated by u
k _ ID _ uvw1 common key’s ID between u and v, w1 is the number
Ekuc {M } ciphertext of message M, encrypted by the key kuc

3.2 Network Topology

Sensor network are usually with a large number of nodes, extensive coverage, but a
single node only has a limited short-range radio communication. Except of communi-
cating with the cluster head or base station, ordinary nodes usually communicate
directly with the neighbor nodes. If the entire network is divided into several smaller
regions, we call clusters in this paper; all nodes in a cluster neighboring each other
can communicate with each other directly. At the same time, the key pool is also
divided into the sub-key pools which correspond with the clusters. The sensor nodes
from the corresponding sub-key pool select the keys, and then follow the division of
the region to deploy the nodes. This way not only increases the probability of shard
294 H. Wang et al.

keys within the same cluster of nodes, but also can reduce the number of keys stored
in nodes. Through the overlapping sub-key pool, the nodes in adjacent cluster can be
up to a certain probability, then, when using cluster head mechanism, inter-clusters
communication can be completed through the cluster head. In this model, we use the
head management mechanisms, cluster head products after achieving connectivity in
the cluster; the specific method is not our focus of this paper. After sensor network is
divided into several clusters, when the base station sends control commands, we can
broadcast news according to the requirements of different regions, this has more per-
tinence. When the whole network has requires or control commands to, it broadcasts
to the whole network. If this, it can reduce the numbers of broadcasting information.

Fig. 1. Network topology of our model

The entire network topology model is shown in Fig. 1. Key server (base station) es-
tablishes key pool P by generating random numbers as keys, the size of key pool is |P|.
Suppose that the total number of nodes in the network is n. It has the node clustering
when deploying the networks. Assuming that it is divided into t, the number of nodes
in each cluster is r. Before the deployment of the network, the sub-key pool selected
randomly by the key server for each cluster is M, which has a number of |M|; in each
cluster, for each node we take randomly select m keys from its sub-key pool M.
When the network deployed, it has a certain probability that any two or more of
nodes share the same key within a cluster. When the two nodes have the same key, we
can choose one or more as the pairwise key. As the network takes the cluster head as
manager for each region, each cluster can consider a cluster head using some methods
and establish the pairwise keys with the adjacent clusters. Obviously the different
cluster still has a certain probability to get the same key.

3.3 Probability Model

According to this model of network topology, node connectivity has two kinds of
situation, connectivity between two nodes in one cluster, namely, the probability of
 Efficient Pairwise Key Establishment Scheme 295

existence of the same key in the different clusters. For the same cluster, in the net-
work of Fig. 1 , take C1 for example, the probability that Ni and Nj share at least one
key is p. Ni selects m keys from | M | has a total of C|mM | ways, Nj has also C|mM |
ways. Suppose that nodes Ni, Nj shared w keys, selecting w shared keys has C|wM |
m − w)
methods, Ni and Nj still have 2(m-w) distinct keys, which has a kind of C|2(
M |− w
selection ways. And to partition 2(m-w) different keys to the two nodes equally, we
have C2(m −mw− w) ways. So the probability that two nodes share w keys is
m−w) m− w
C|wM |C|2(
M |− w C2( m − w )
p( w) = (1)
(C|mM | ) 2
The probability that two nodes share at least one key is
w
p = ∑ p (i ) ( w ≤ m) (2)
i =1

According to the probability of the complementary event, the result is

C|mM |C|mM |− m
p1 = 1 − p(0) = 1 − (3)
(C|mM | )2
Similarly, according to the above reason, the probability that two nodes in the differ-
ent clusters shared at least one key is
C||PM| |C||PM|−| |M | C|mM |C|mM |− m
p2 = (1 − )(1 − ) (4)
(C||PM| | ) 2 (C|mM | ) 2

4 Efficient Pairwise Key Establishment and Management

The security connectivity of wireless sensor network is on the basis of communica-
tions connectivity; to ensure that any two nodes within a cluster are able to communi-
cate, it requires that it can establish a secure path at least between any two nodes after
the deployment of networks. Based on random key pre-distribution scheme, the key
server loads the pre-shared keys to each node, that means that it can establish the
communication link after self-organized, and the nodes guarantee the session key by
consultation according to the predetermined rules prior to the establishment of the
link, that is the pairwise key between any arbitrary nodes. Described in detail in this
section about the building process of the pairwise key between the nodes within the
cluster, the entire process of building the key is similar to the literature [9], but im-
prove and perfect it, and make it closer to practical application.

4.1 Establishment of the Pairwise Keys Directly

When the deployment of a whole network is completed, each cluster enters bootstrap-
ping phase. For each cluster, the node find the same key by broadcasting consultation
296 H. Wang et al.

mechanism, and select one of the common keys (if there were multiple) as the pairwise
key through the rules scheduled, the process of looking for shared key is as follows.
The node u broadcasts its own message that contains the node’s ID u, random
number N u and m keys’ IDs,
u − > *: Bu = {u, N u }
U {k _ ID1, k _ ID 2,L , k _ IDm}
When other nodes of the cluster, such as the node v, received the message Bu , it will
compare all the ID with its own stored in the keys’ ring, then reply to u with the same
ID which is in order from small to large. Due to the rules predetermined by the node u
and node v: they take the key of the smallest identifier as their pairwise key, and when
this key is compromised or leaks, the key by followed will be as pairwise key. Node v
sends a message to the node u as follows,

v − > u : M = {v, N u }
U {k _ ID _ uv1, k _ ID _ uv 2,
L , k _ ID _ uvw}
When u received all the nodes’ messages of common key in the cluster, the pairwise
keys will be stored in a form just as Table 1. (1 ≤ wi ≤ wi +1 ≤ m (1 ≤ i ≤ n′) )
The other nodes broadcast their own key messages similarly, and build the pair-
wise key tables with neighbor nodes. If network needs, such as unified management
in cluster, data integration and so on, we can use hierarchical, that’s to say, the estab-
lishment of cluster head for each cluster. The communication between clusters will be
achieved by the cluster head who already had the pairwise keys with the cluster head
of the adjacent clusters. The method establishing the pairwise key between the cluster
heads is the same as the above.

Table 1. Pairwise keys stored in node u

Neighbor
Pairwise key common keys
node
v K _ uv1 K _ uv 2,L , K _ uvw1
a K _ ua1 K _ ua 2,L , K _ uaw2
b K _ ub1 K _ ub 2,L , K _ ubw3
c K _ uc1 K _ uc 2,L, K _ ucw4
…… …… ……

4.2 Establishment of the Pairwise Keys via Intermediate Nodes

In section 4.1, the pairwise key is established when there exist the common keys be-
tween two nodes, so it just needs one session. When the direct establishment of the
 Efficient Pairwise Key Establishment Scheme 297

pairwise key is completed, the node-to-node within the cluster not only has the pair-
wise key, but also forms a table of shared keys. So does with the cluster head-to-head
between the clusters. The table is composed of the neighbor nodes and their security
links which have established their pairwise keys. The physically adjacent nodes which
are not able to find common keys can create it via intermediate nodes. This section
will give an approach about indirectly establishing the pairwise key.

Fig. 2. Indirectly establishing the pairwise key

When there is no common key between two nodes, such as the node u and v, but it
has formed the table of the pairwise key and the shared keys with other neighboring
nodes. Now the node u and v have to establish the pairwise key, and it must be safe,
not known to other nodes in the establishing process. The principle of establishment
is: we choose the node which exists the least shared keys with the sending node as the
forward transmitting node. Shown in Fig. 2, let u stored the shared key with a, b, c, d,
e, it needs to choose two nodes from them, making the both nodes and v have the
maximum probability of the common keys. Assume that a, c have the least numbers
of shared keys in u’s pairwise keys’ table, then u chooses a, c as the forwarding node.
The reason is that there is no shared key between u and v, yet a, c has the least shared
keys with u, then the probability of existing shared keys with v will be the largest.
The specific description of establishing process is as follows.
When node v receives the broadcast message of node u, found that there is no
common key, and then returns the following message,
M = {v, N u , N v } .
Node u receives the message M, stores N v , then selects a, c from the pairwise
keys table to forwarding the key’s materials in two paths according to above principle.
u − > a : Μ = {u , Ekua {v | N u1}}
u − > c : Μ = {u , Ekuc {v | N u 2 }}
When node a, c receive their message M, if they have the pairwise keys with v,
transmit them to v. As the node v receives the message from node a and c, calculates
the pairwise key:
kuv = N v ⊕ N u1 ⊕ N u 2 ,
and node u generates pairwise key using the same method.
298 H. Wang et al.

When either a or c doesn’t have pairwise key with v, u receives the response; it
will choose other nodes in the pairwise key table in the same principle except a and c,
repeats the above steps until successfully establish pairwise key with v. Other nodes
use the same method to establish pairwise keys with the nodes they do not have
shared keys in the cluster, and add the pairwise keys into their own pairwise keys
table. If there is no shared key between the adjacent cluster-heads, they establish the
pairwise key through the base station. In the whole network, each node established
pairwise key among the nodes in the cluster and between the cluster-heads, we
achieved the whole network connectivity.

5 Performance Analysis
In this section, we analysis and evaluate the security and performance of our scheme
in connectivity, security, scalability and node storage. Connectivity: it is on behalf of
the capability of the network security connectivity. It defined as the probability of
establishing secure link when two nodes are within the scope of communication; for
this article, it means that the probability of the two nodes in a cluster storing at least
one common key, as well as the two nodes in different clusters. When nodes store
fewer keys, the probability should be as high as possible. Security: the nodes in the
network should ensure that expected security, which should store the least key and
cost a minimum communication overhead. Node storage: since the node’s limited
storage resource, when the network reaches certain connectivity, the nodes should
store fewer keys.

5.1 Network Connectivity

In the section 3.2, we have got the probability of at least one shared key existed in two
nodes. Suppose the number of key pool is |P|=10000, when |M|=(0.01,0.02,
0.03,0.04)|P|, this probability is equation (3), the relationship between the probability
and the keys stored in node is shown in Fig. 3.

ye
k
en
ot
sa
el
ta
gn
ir
ah
sf
o
yit
ilb
ab
or
P

Fig. 3. Probability of sharing at least one key in two nodes in cluster

 Efficient Pairwise Key Establishment Scheme 299

Table 2. In a certain probability, the number of keys stored in node in cluster

p
0.92~0.95 0.95~0.98 0.98~0.99
|M|
100 15 16,17 18,19
200 22,23 24-26 27,28
300 27,28 29-32 33,34
400 31-33 34-37 38-40

The Fig. 3 shows, when the sub-key pool is constant, the probability increases as
the number of keys stored in nodes increases; nodes store the same amount of keys,
the smaller the sub-key pool, the greater the probability; that’s to say the possibility of
finding pairwise keys will be larger. Even if the two nodes does not exist any same
key, we can establish pairwise keys which is described in section 4.2. Use the method
in section 4.2, we can guarantee the network wide connectivity. Table 2 shows when
asked a certain probability, a single node, should store the number of keys. Compared
to E-G scheme, the new scheme significantly reduces the number of keys.
For the adjacent two clusters, the probability of two nodes, equation (4), simplified
as,

p2 = (1 − f1 ( M , P))(1 − f 2 (m, M ))
( P − M )( P − M − 1)L ( P − 2M + 1)
f1 ( M , P) =
P( P − 1)L ( P − M + 1)
( M − m)( M − m − 1)L ( M − 2m + 1)
f 2 (m, M ) =
M ( M − 1)L ( M − m + 1)
Fig. 4 shows, when sub-key pool is constant, the relationship between the probabil-
ity and the number of keys stored in node.
Fig. 4 and Table 3 shows that sub-key pool is small, such as 100,200, the probabil-
ity of finding common keys between clusters will be very small; when the sub-key
pool increases, the probability significantly increases while the node storage has not
increased much; such as when sub-key pool is 400, the amount of keys stored in dif-
ferent clusters is approximately as 1.4 times as the number of inner cluster. According
to practical application, when we need a large amount of inner-cluster communica-
tion, we may properly increase both the number of sub-key pool and key ring stored
in nodes. Then any node who will be the cluster head has a big probability with oth-
ers’. Even if there is no shared key, we can establish pairwise key through the meth-
ods in 4.2. In actual application, the communication between the nodes in different
clusters is relatively small, if need, we can transmit through the cluster head.
In all, we can see that the scheme has a good performance. When ensure connec-
tivity, we largely reduce the number of key stored in node; and we can pre-configured
key flexibility according to the specific application requirement. When cannot find
shared keys between nodes, we can create pairwise key through a small amount of
communications.
300 H. Wang et al.

Table 3. In a certain probability, the number of keys stored in node in different clusters

p
0.92~0.95 0.95~0.98 0.98~0.99 Maximum p
|M|

100 -- -- -- 0.3895
200 -- -- -- 0.8696
300 38-41 42-49 50-72 0.9900
400 43-46 47-52 53-57 0.9997
1000 69-74 75-85 86-92 1.0000
2000 99-106 107-122 123-132 1.0000
3000 121-131 132-150 151-162 1.0000
4000 140-152 153-173 174-188 1.0000

1.1
1

0.9

0.8 |M|=0.01 |P|

|M|=0.02 |P|
0.7
* |M|=0.03 |P|
0.6 |M|=0.04 |P|

0.5

0.4

0.3

0.2

0.1

0
0 20 40 60 80 100 120 140 160 180 200
size of m
Probability of sharing at least one key

Fig. 4. Probability of sharing at least one key in two nodes in different clusters
 Efficient Pairwise Key Establishment Scheme 301

5.2 Security Analysis

For the random key pre-distribution scheme, the network must establish all the com-
munication keys in shortest time after the distribution of the network nodes. If there is
neither leak of any key nor captured of nodes by attacker, the network is relatively
supposed to be secure and all keys are successfully achieved by each node in the fu-
ture communication. Otherwise, the message encrypted by the pairwise key will lose
confidentiality, and one node compromised by adversary will pose a threat to others
even the network.
The goal of this scheme is to allow sensor nodes to find a common secret key
safely with each of their neighbors after deployment. So every node broadcasts its
key’s identifiers at the network’s initial time to avoid disclose keys. We’ll analyze the
network’s resiliency when one or more nodes are captured and disclose all their keys.
Let the number of captured nodes be x. Since each node contains m keys, the prob-
ability that a given key has not been compromised is
x
⎛ m ⎞
⎜1 − | | ⎟ .
⎝ S ⎠
Therefore, the expected fraction [6,16] of total keys being compromised can be esti-
mated as:
x
⎛ m ⎞
p3 = 1 − ⎜1 − ⎟ (5)
⎝ | S |⎠
When |M|=400, |P|=10000, m=35 and m=55, the results and comparison with existing
key pre-distribution schemes are depicted in Fig. 5, which contains the in-cluster and
global resilience of our scheme, E-G scheme[9] that is same as our global resilience
and Chan-Perrig-Song scheme[3].The Fig. 5 show that our scheme substantially low-
ers the fraction of compromised communication after x nodes are compromised com-
pared with the Chan-Perrig-Song scheme in network-wide. We get that if the x nodes
were compromised in one cluster, the resilience of the network in that local area (in
the cluster) is lower than that of the entire network. We also realize that in reality,
these x nodes might be randomly distributed in the entire region, so the network will
have a higher probability of resilience.
The result is that our scheme is secure and efficient. Even the adversary eaves-
dropped the message, it only get the key’s ID instead of the keys themselves. For
another, other pairwise keys are calculated by the nodes themselves. Because the key
materials are transmitted through different intermediate nodes, the forwarding nodes
only know one of the materials, random value in this method; they can’t evaluate the
pairwise key. So, the key established in this method is absolutely secure compared
to using the secure channels in previous scheme; it is only known the two nodes
who’ll communicate with each other in future instead of any path-node or other third
node.
302 H. Wang et al.

Fraction of communications compromised

Fig. 5. Network Resilience: Comparisons with existing schemes

5.3 Communication and Computational Overhead

After the deployment of all the nodes in the network, to discover the common keys,
each node of the schemes [9] [10] broadcast all of the keys or the key identifiers to
their neighbor nodes. In particular, when the network size is large, the node will store
a large number of keys in its subset. Then key information exchange process, not only
consumes a lot of energy but also wastes considerable radio bandwidth, resulting in
the link load overhead, collision opportunities to grow. Our method, according to
specific requirements, has the strongpoint that nodes only need to store the appropri-
ate number of keys in cluster, broadcast key identifier, to establish a certain number
of pairwise key. The remaining nodes that do not share common keys only take a
small amount of communication and calculation to establish the pairwise keys, which
can reduce the network traffic load and energy costs as much as possible. With our
method, the node transmits its key materials by using its neighbors who have the least
common keys to the destination, which means he/she will find the right intermediate
nodes in the largest probability. This path-materials computation and pairwire key
establishment approach extremely has less CPU power and time consumption than
previous two schemes. We can almost guarantee any two neighbor nodes to establish
a pairwise key and have a safe communication.
In our scheme, they use the very key whose identity is the smallest as pairwise key
in all key’s IDs when the two nodes find their common keys. Otherwise, they only
need four communications to get their key materials to calculate the key via XOR. This
extremely reduces the computation overload for resource constrained sensor nodes
especially their limited computing capability. Meanwhile, if they leak their pairwise
key, they can easily delete it and take the key of the second smallest ID in their key
tables as their later pairwire key while the two nodes have more than one common key.
This is much flexible and secure for updating the pairwise key. Therefore its commu-
nication and computational overhead are much lower than previous schemes.

6 Conclusion
In this paper, we describe our pre-distribution key management model in WSN and
propose a cluster-based key management scheme depending on the application
 Efficient Pairwise Key Establishment Scheme 303

requirements. Since the main communication in the network is point-to-point, estab-

lishing pairwise keys between each two nodes is especially important. Therefore we
present a new efficient pairwise key establishment scheme for large scale wireless
sensor networks. Compared to the E-G scheme and the Chan-Perrig-Song scheme, our
scheme has a number of appealing properties, our scheme can provide the full net-
work connectivity and the substantially best resiliency, as well as the maximum sup-
ported network size for fixed key pool size, which is much more scalable and flexible.
Lower computational and communication overhead are also achieved in our proposed
scheme. Simulation results and performance analysis shows that this network model
not only enhanced communication security but also greatly reduced the number of
keys stored in nodes, and it also achieved a better balance between the node storage
capacity and communication overhead.

References
1. Wang, Y., Attebury, G., Ramamurthy, B.: A survey of security issues in wireless sensor
networks. IEEE Communications Surveys and Tutorials 8(2) (2006)
2. Su, Z., Lin, C., Feng, F.-J., Ren, F.-Y.: Key Management Schemes and Protocols for Wire-
less Sensor Networks. Journal of Software 18(5), 1218–1231 (2007)
3. Chan, H., Perrig, A., Song, D.: Random key pre-distribution schemes for sensor networks.
In: Proceedings of Symposium on Research in Security and Privacy, Oakland, USA, pp.
197–213 (2003)
4. Ren, K., Zeng, K., Lou, W.: A new approach for random key pre-distribution in large-scale
wireless sensor networks. Journal of Wireless Communication and Mobile Computing
(Special Issue on wireless Networks Security) 3(6), 307–318 (2006)
5. Du, W., Deng, J., Han, Y.: A pairwise key pre-distribution scheme for wireless sensor
networks. In: Proceedings of the 10th ACM Conference on Computer and Communica-
tions Security, Washington, DC, USA, pp. 42–51 (2003)
6. Liu, D., Ning, P.: Establishing pairwise keys in distributed sensor networks. In: Proceed-
ings of the 10th ACM Conference on Computer and Communications Security, Washing-
ton, DC, USA, pp. 52–61 (2003)
7. Liu, D., Ning, P.: Location-based pairwise key establishments for static sensor networks.
In: Proceedings of SASN 2003, Fairfax, USA, pp. 72–82 (2003)
8. Du, W., Deng, J., Han, Y.: A key management scheme for wireless sensor networks using
deployment knowledge. In: Proceedings of IEEE INFOCOM 2004, Hong Kong, China,
pp. 586–597 (2004)
9. Eschenauer, L., Gligor, V.D.: A key-management scheme for distributed sensor networks.
In: Proc. CCS 2002, pp. 41–47. ACM Press, New York (2002)
10. Bollobas, B., Fulton, W., Katok, A., et al.: Rand Graphs, 2nd edn., pp. 160–200. Cambridge
University Press, Cambridge (2001)
11. Blom, R.: An optimal class of symmetric key generation systems. In: Beth, T., Cot, N.,
Ingemarsson, I. (eds.) EUROCRYPT 1984. LNCS, vol. 209, pp. 335–338. Springer,
Heidelberg (1985)
12. Blundo, C., Santis, A.D., Herzberg, A., Kutten, S., Vaccaro, U., Yung, M.: Perfectly-
secure key distribution for dynamic conferences. In: Brickell, E.F. (ed.) CRYPTO 1992.
LNCS, vol. 740, pp. 471–486. Springer, Heidelberg (1993)
304 H. Wang et al.

13. Wang, Y., Ramamurthy, B., Xue, Y.: A Key Management Protocol for Wireless Sensor
Networks with Multiple Base Stations. In: IEEE International Conference on In Communi-
cations, ICC 2008, pp. 1625–1629 (2008)
14. Wang, Y., Ramamurthy, B., Zou, X.: KeyRev: An efficient key revocation scheme for
wireless sensor networks. In: Proc. ICC 2007, Glasgow, Scotland, UK (June 2007)
15. Traynor, P., Choi, H., Cao, G., Zhu, S., Porta, T.L.: Establishing pairwise keys in hetero-
geneous sensor networks. In: Proc. INFOCOM 2006 (2006)
16. Ma, C.-g., Shang, Z.-g., Wang, H.-q.: Domain-based key management for heterogeneous
wireless sensor networks. Journal on Communications 0l.30(5) (2009)
17. Chan, A.C.-F.: Distributed symmetric key management for mobile Ad Hoc networks. In:
Proceedings of the INFOCOM 2004, Twenty-third Annual Joint Conference of the IEEE
Computer and Communications Societies, vol. 4, pp. 2414–2424 (2004)
18. Liu, Z.-H., Ma, J.-F., Huang, Q.-P.: Domain-Based Key Management for Wireless Sensor
Networks. Chinese Journal of Computers 29(9) (September 2006)
 Agent Based Approach of Routing Protocol Minimizing
the Number of Hops and Maintaining Connectivity of
Mobile Terminals Which Move One Area to the Other

Kohei Arai1 and Lipur Sugiyanta1,2

1
Department of Information Science,
Faculty of Science and Engineering, Saga University, Japan
2
Department of Electrical Engineering,
Faculty of Engineering, State University of Jakarta
[email protected], [email protected]

Abstract. An ad-hoc network is a special kind of network, where all of the

nodes move in time. So the topology of the network changes as the nodes are in
the proximity of each other. Ad-hoc networks are generally self-configuring,
because of the concept of no stable infrastructure takes a place. The research is
focused in Wireless (Ad-Hoc) Mesh Networks. In this network, each node
should help relaying packets of neighboring nodes using multi-hop routing
mechanism. This deployment is needed to reach far destination nodes to solve
problem of dead communication. This multiple traffic "hops" within a wireless
mesh network caused dilemma. It can extend the coverage area, but the
repeatedly relayed traffic will exhaust the radio resource. Wireless mesh
networks that contain multiple hops become increasingly vulnerable to
problems such as energy degradation and rapid increasing of overhead packets.
Minimizing hop or relay is expected to solve this dilemma.
Keywords: wireless ad-hoc mesh network; multi-hop routing; energy, overhead
packet.

1 Introduction
Wireless networks in today meaning are usually set up with a centralized access point
for each area of connectivity. The access point has knowledge of all devices in its area
and routing to them is done in a table driven manner. The future trend of mobility, it
becomes important to give users the possibility of finding service they need anywhere
in the network. Users would like to obtain access to services automatically, without
reconfiguring their system. Especially with mobile devices dynamic discovery of
services in other than home or corporate network and automatic network
configuration will be very useful.
An ad-hoc network is a special kind of network, where all of the nodes move in
time. So the topology of the network changes as the nodes are in the proximity of
each other. Ad-hoc networks are generally self-configuring, because of the concept of
no stable infrastructure takes a place. Each node in the network must be able to take

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 305–320, 2010.
© Springer-Verlag Berlin Heidelberg 2010
306 K. Arai and L. Sugiyanta

care of routing of the data. So the concept of centralized network with pre-defined
central routing tables could be applied only with the limited nodes changing and
coverage area. The coverage and connectivity of wireless mesh networks decreases as
the end-user becomes further from the other nodes. Multi-hop communications among
mesh nodes can enable long distance communication with acceptable performance
degradation. Multi-hop wireless network is network where communication between
two nodes may go through multiple consecutive wireless links without relying on any
existing, pre-configured network infrastructure or centralized control. These networks
are constrained by the interference of wireless communications, the finite battery
energy and the mobility of nodes.
However, wireless network throughput and multi-hop coverage extension are two
contradictory goals. On one hand, the multi-hop communications can extend the
coverage area. On the other hand, as the number of hops increases, the repeatedly
relayed traffic will exhaust the radio resource. The throughput will sharply degrade
due to the increase of collisions from a large number of users. Therefore, it becomes
an important issue to design a wireless network, so that the coverage of a wireless
network can be extended without sacrificing the overall throughput.
Beside throughput, the ability to transport data from a source to a destination is a
fundamental ability of wireless network. Delay is characterized by their lack of
connectivity, resulting in a lack of instantaneous end-to-end paths. This is due to
protocol routing trying to first establish a complete route and then, after the route has
been established, forward the actual data. Also, wireless clients are mobile nodes and
commonly run on batteries. Thus, power usage of mesh clients should be limited.
These natures of wireless ad hoc networks made them require many improvements
compared to wireless managed networks, before it ready to be implemented.

2 Wireless Mesh Network Architecture

A wireless mesh network is a communications network made up of radio nodes
organized in a mesh topology. It can be seen as a special type of wireless ad-hoc
network. In the IEEE 802.11s network, there is WLAN mesh and is defined as a set of
mesh points interconnected via wireless links with the capabilities of automatic
topology learning and dynamic path selection. Fig. 1 shows an example of IEEE
802.11s WLAN mesh. In the figure, there are two classes of wireless nodes. The mesh
points (MPs) are the nodes supporting wireless mesh services, such as mesh routing
selection and forwarding, while the non mesh nodes are the pure client STAs. In
addition to mesh services, the mesh access point (MAP) also provides wireless access
services. The pure client STAs do not participate in the WLAN mesh, but they can
associate with the mesh APs to connect to the mesh networks. The WLAN mesh can
connect to other networks by the mesh portals (MPPs). Multiple WLAN meshes can
also be connected by the MPP.
The research used that network model with modified scheme. The architecture of
the network is as follow. Nodes are mesh points with capability of routing selection
and forwarding included in. Mesh access points that provide wireless services do not
exist. Services are generated among nodes. Other nodes and new nodes that do not
participate in the mesh network previously, can be associated to the networks by
using a reached node which already joined the network as relay. The wireless mesh
 Agent Based Approach of Routing Protocol Minimizing the Number of Hops 307

can connect to other networks by the mesh portals (MPPs), but this scheme is not
covered in this research. Multiple wireless meshes can also be connected by the MPP.
There will be several wireless areas, each with the nodes inside. Nodes are always
active moving with direction, speed, and distance change randomly. It is possible that
nodes able to move inside wireless area and leaving its home area. During this
activity, nodes always reconfigure their routing path allowing dynamically changing
of route selection. Interaction among nodes became rapidly increased.
When moving, nodes also allowed entering new wireless area, leaving home/
original area. In order to maintain connectivity along this movement, individual
components will need to discover and maintain awareness of surroundings nodes and
also to configure and adapt themselves in response to changing node’s condition.

Fig. 1. The network architecture for the IEEE 802.11s WLAN mesh network

Fig. 2. Network model used in this research

308 K. Arai and L. Sugiyanta

There are special nodes called MPP. A MPP provides the ability to bridge wireless
networks over a mesh. MPP has capacity to find optimum path with minimum hops to
other nodes. Therefore MPP should support traffic propagation to both wireless areas
at where it located. Nodes at both areas able to sent and receive data outside its
home/original mesh and vice versa by using this MPP. It has consequences that nodes
must aware about both their neighbors and also MPP. As long as communication path
between source node and destination node is built inside wireless area, then MPP is
not involved. In this paper, this MPP node is called Agent. The agent node
reorganizes the network considering the available nodes.
Agents are located in such place to facilitate communication between wireless
areas and minimize the number of hops to achieve optimum throughput of mesh
clients which communicate each other. Through minimizing the number of hops
between source and destination, it is hoped that the optimum throughput between
source node and destination can be reached. The optimal tradeoff between network
capacity and throughput for wireless mesh networks need to be investigated.
Simulation programming approach is built to determine the optimal deployment
parameters. This simulation runs with large number of nodes, with high dynamic
parameters and considering mobile devices resources (battery power, bandwidth) and
wireless environment constraints.
Simulation is split into two steps. First has to cover movement among nodes inside
area, not move to another area. Second, simulation covers node’s activities that
include communication among nodes which also move to other areas. During the first
step, route discovery play important role. Source nodes that do route discovery flood
the network with a route request. Its neighbors propagate this route request in the
same manner until it reaches the destination node or a node that already has a route to
the destination. The data analyzed in this paper is intended to this first step.

3 Algorithm
3.1 Establishing the Initial Topology
Initialization topology. Nodes discover themselves and use their maximum trans-
mission power to build the initial topology. We consider the case where all nodes in
the network are similar, i.e., assuming an homogeneous infrastructure. Let us assume a
square area of side l, in which n nodes are deployed uniformly, distributed at random
position. This deployment produces a connected topology under some assumptions;
sometimes a completely connected topology is built. From the simulation, a large
connected component is built quickly using a communications radius considerably
smaller than the radius needed to have the entire network connected.

3.2 Deriving the Dynamic Minimum Hop Tree

The concept used was backbone-based. This backbone approach is to find a subset of
nodes that will be a gateway to guarantee connectivity and communication coverage
while allowing every other node in the network to reach at least one node on this
backbone in a direct way. These nodes provide complete network coverage. Fig. 3
shows an example of this approach in which 35 nodes out of 200 nodes in the network
were selected to build the backbone. (nodes with red color are initiator nodes)
 Agent Based Approach of Routing Protocol Minimizing the Number of Hops 309

Fig. 3. 200 nodes distributed randomly in the area of l2, some nodes become backbone for
others. Captured at initial time (t= 0).

Topology is dynamically changed along as node moving. An easy way to illustrate

the tree construction process is by the minimum spanning tree of a graph. The process
works from the set of nodes that are initiator nodes at time t. Between t and t + 1,
based on certain parameters, the nodes in the set evaluate all adjacent nodes in order
to extend the tree. The process continues until all nodes on the graph are evaluated.
Not every node will be selected to be part of the tree, and those which were not
selected will have a chance until a new tree is requested. In general, the process
usually starts on an initiator node, but if there are more than one initiator nodes, then
several trees may be built in parallel.
The growing a tree algorithm builds connected graph considering the remaining
energy in the nodes and the distance between them. The tree is built using three types
of messages: Hello Message, Parent Reply Message, and Time Out Message. The
building process is started by a predefined node that might be the initiator, right after
the nodes are deployed. The initiator, node 72 in Fig. 3, starts the protocol by sending
an initial Hello Message. This message will allow the neighbors of 72 to know their
“parent”. This message also includes a selection metric that is calculated based on the
signal strength of the received Hello Message and the remaining energy in the node.
The metric will be used later by the parent node to sort the candidates. The parent
node waits a certain amount of time to receive the answers from its neighbors. Once
this timeout expires, the parent node sorts the list in decreasing order according to the
selection metric. The neighbors of 72 that have already received Hello messages, after
send Reply messages, and then prepare to send Hello messages to their neighbors.
The process continues to all reachable nodes. If the neighbors have been already
covered by another node, it ignores the Hello Message. The method that a node needs
to call on in order to send a message is broadcast. The method that a node needs to
send a packet is broadcast.
310 K. Arai and L. Sugiyanta

3.3 Routing Mechanism

Routing algorithms intended to maintain routes to all reachable destinations for
sending data. Reactive ad-hoc routing algorithms use on-demand routing, so the
routes are only requested when they are needed. This approach stores active routes in
memory and no maintaining information for unused routes. Proactive ad-hoc routing
algorithms (table-driven) maintain routes to all nodes in the network, so they rely
according to tables. Nodes use reactive routing algorithm, whereas agents use
proactive routing algorithm.
Main concepts of reactive routing algorithms are route discovery, route caching
and route maintenance. Nodes used this method to discover route path. The first data
(this means that no data has been sent to that destination before) initialize the route
discovery from the source to the destination. A route reply is then sent back the same
way the route request traveled. The route reply contains a list of all the intermediate
nodes between source and destination. When the route reply reaches back to the
source, the source can send the data and cache the new route for future use. Whenever
a node discovers that the link to one of its neighbor nodes has failed, it will send a
route error packet to the source of any cached routes that use the failed link. This
route error will propagate through all intermediate nodes that also have cached this
route. Upon receipt of it, node will update and purge the failed route.
On the other hand, agent nodes have enough knowledge to compute minimal route
(min. hop distance) to any other destination nodes in both areas. In this simulation,
agent nodes create source tree and computes route to the destination by computing the
minimal-hop path. Each regular node keeps a path to agents.

3.4 Energy
Energy is power kept in each node. It was assumed that the radio dissipates Eelec =
50 nJ/bit to run the transmitter or receiver circuitry and ϵamp = 100 pJ/bit/m2 for the
transmit amplifier. Thus, to transmit a k-bit message a distance d using this radio
model, the radio expends:
, (1)
and to receive this message, the radio expends:
(2)
The energy model included in simulation was based on the following formulas, taken
from [9]:
E E ϵ . πr (3)

E E (4)
The energy behaviors of nodes are defined as:

• During the idle time, a node does not spend energy. Even though this
assumption has been proven untrue because being idle might be as costly as
receiving data, this is still an assumption that can be done in most experiments,
since the most important factor is the overhead in terms of message exchange
and its associated cost.
 Agent Based Approach of Routing Protocol Minimizing the Number of Hops 311

• The nodes are assumed to have one radio for general messages. The main
radio is used in all operations when the node is in active mode, and to send and
receive control packets. When this radio is turned off, then no messages will
be received and no energy will be used.
• Energy distribution among nodes can either be constant value, normally
distributed, Poissonly distributed, or uniformly distributed.

4 Properties and Evaluation

Simulation was deployed on the abstract area, which is a rectangle where the nodes of
the network are located. In order to create the deployment area, its width and height
are manually defined.
The wireless mesh network model used in this simulation is based on the following
assumptions:

• The communication range of the nodes is a perfect symmetric unit disk. If dx,y
≤ rx → x and y can see each other.
• A constant bit error rate (BER) is defined for the whole network. It is a simply
implementation of an error model. Whenever a packet is going to be sent, a
random number is generated and compared to the message error rate (that
depends on the size of the message). If the random number is greater, the
message is received, otherwise it is lost. The default value for the BER is 0,
which means there is no packet loss.
• To increase the randomization of the simulation process, simulation introduces
some delay on some common processes in the network, like message
transmission delay, processing delay, time out, etc. so each instance of a
simulation, would produce different results.
• Simulation manages homogeneous networks, in which all nodes have the same
characteristics. The parameters that can be defined in a homogeneous family
of nodes are number of nodes, communication radius, and size of the area of
deployment, position distribution, and energy distribution.

4.1 Initiator Nodes

Initiator node is node that initiates transmission of packet. It can be route discovery or
data transmission. Like other nodes, initiator is always moving with random direction,
speed, and distance. Along moving, initiator node is always sensing its neighbor to
maintain connectivity.

4.2 Node’s Energy Performance

Nodes uses radio channel that is symmetric such that the energy required transmitting
a message from node A to node B is the same as the energy required transmitting a
message from node B to node A for a given SNR. For our experiments, we also
assume that all nodes are sensing the environment at a fixed rate. Energy was used
whenever node sends, receives, and processes message.
The average of remaining node’s energy at the end of simulation (with 200 cycles)
for all nodes is shown in Fig 4.
312 K. Arai and L. Sugiyanta

Node's Energy Performance (in Average)

Hundreds 12

9.799 9.623
10

9.575 8.161
9.323 9.214
8.933 7.551
8
6.800
7.807 7.814
6.145
7.016
6 5.510
5.092
4.417
5.360
4 4.748 4.686 3.516

3.874
3.235
2 2.553
2.340

0 0.658 0.737 0.726138776

10 20 30 40 50 60 70 80 90 100

-1.119
-2
# Nodes

2 Initiator nodes 3 Initiator nodes 4 Initiator nodes

Fig. 4. Node’s energy performance in relationship with number of nodes

16

14

12

10

0
106
113
120
127
134
141
148
155
162
169
176
183
190
197
204
211
218
225
232
239
246
253
260
267
274
281
288
295
302
309
316
323
330
337
344
351
358
365
372
379
386
393
400
1
8
15
22
29
36
43
50
57
64
71
78
85
92
99

Used Energy *100 Speed

(a)
30

25

20

15

10

0
100
109
118
127
136
145
154
163
172
181
190
199
208
217
226
235
244
253
262
271
280
289
298
307
316
325
334
343
352
361
370
379
388
397
1
10
19
28
37
46
55
64
73
82
91

Used Energy *100 Speed

(b)
30

25

20

15

10

0
1

100
109
118
127
136
145
154
163
172
181
190
199
208
217
226
235
244
253
262
271
280
289
298
307
316
325
334
343
352
361
370
379
388
397
10
19
28
37
46
55
64
73
82
91

Used Energy *100 Speed

(c)
Fig. 5. The faster node is moving, the more Energy is be used
 Agent Based Approach of Routing Protocol Minimizing the Number of Hops 313

Each node was given energy of 1000 at the beginning of simulation. Simulation
executed repeatly with number of initiator nodes increment by 10. As number of
initiator node increase, the total energy remained on nodes is decreased. Simulation
with 4 initiator nodes gave result negative average value at # Nodes= 80. It was to
show that some nodes were already dead/OFF.
With node always moving, simulation shows that energy used during cycles is
affected with its speed. More energy is used by node when node moves faster. Fig. 5
below illustrated the effect of speed to energy used by initiator node.

Used Energy

20

18

16
Energy used *100

14

12

10

0
1 13 25 37 49 61 73 85 97 109 121 133 145 157 169 181 193 205 217 229 241 253 265 277 289 301 313 325 337 349 361 373 385 397

Simulation cycle
Energy used (2 In) Energy used (1 In) Energy used (1 In) Energy used (1 In) Energy used (1 In)

Fig. 6. Comparison for energy used to number of initiator node

# Overheads Packets
20,000
Hundreds

18,000 17,818 17,738

17,077 16,959

16,000 16,023 15,930

14,000

12,520 12,412 12,467 12,401

12,000 11,745 11,674

10,000
9,642 9,592
9,057 9,002
8,512 8,462
8,000
7,520 7,476
7,029 6,956

6,000 6,006 5,992

5,032 4,985
4,402 4,410
4,370 4,384
4,000 3,969 3,931
3,038 3,017
2,238 2,213
2,000 1,848 1,829
1,223
1,029 1,202
1,018
815 792 594 580
123
74 116 286
73 136 286
137
0 35 33

10 20 30 40 50 60 70 80 90 100

# Nodes

2 Initiator nodes (sent) 2 Initiator nodes (recv) 3 Initiator nodes (sent)

3 Initiator nodes (recv) 4 Initiator nodes (sent) 4 Initiator nodes (recv)

Fig. 7. Overhead packets increase rapidly as more number of initator nodes

314 K. Arai and L. Sugiyanta

It is clear from Fig. 6, the increasing number of initiator nodes caused nodes to
route a large number of messages. Thus these nodes will die out quickly, causing the
energy required to get the remaining messages to increase and more nodes to die. This
will create a cascading effect that will shorten system lifetime. In addition, as nodes
close to the initiator die, then the environment is no longer being monitored. To prove
this point, we ran simulations using the random 10 – 100 node shown in Fig. 5. After
completed cycle, several nodes were considered dead.
Using Equations 3 and 4 and the random node from 10 until 100, we simulated
transmission of first data from initiator to every node (communication range 100 m).
Fig. 4 showed the total energy remained on nodes. Topology was changed all the time
cycle. With transmission energy is on the same order as receive energy, it can be seen
that the energy-efficient to use depends on the network topology.

25

20

15

10

0
10
19
28
37
46
55
64
73
82
91
100
109
118
127
136
145
154
163
172
181
190
199
208
217
226
235
244
253
262
271
280
289
298
307
316
325
334
343
352
361
370
379
388
397
1

Simulation cycle

Speed # Packet sent # Packet recv

(a)
20
18
16
14
12
10
8
6
4
2
0
10
19
28
37
46
55
64
73
82
91
1

100
109
118
127
136
145
154
163
172
181
190
199
208
217
226
235
244
253
262
271
280
289
298
307
316
325
334
343
352
361
370
379
388
397

Simulation cycle

Speed # Packet sent # Packet recv

(b)
16

14

12

10

0
10
19
28
37
46
55
64
73
82
91
1

100
109
118
127
136
145
154
163
172
181
190
199
208
217
226
235
244
253
262
271
280
289
298
307
316
325
334
343
352
361
370
379
388
397

Simulation cycle

Speed # Packet sent # Packet recv

(c)

Fig. 8. Number of overhead packets is affected with speed of node

 Agent Based Approach of Routing Protocol Minimizing the Number of Hops 315

4.3 Overhead of Topology Discovery

Overhead packet refers to the extra bytes it takes to transmit data on a packet-
switched network. Each packet requires overhead that is format information stored in
the packet header, which, combined with the assembly and disassembly of packets,
reduces the overall transmission speed of the raw data.
In this simulation, TCP/IP will be assumed to be the protocol being used for packet
data. It was because much of the world’s telecommunication networks are forwarding
IP packets to and from at this moment. During the course of an IP packet’s journey
from a node to others, it is encapsulated and de-encapsulated into and out of framing
headers and trailers that define how the packet will make it’s way to the next hop in
the path. This Fig. 8 below is to explain what happens when to overhead packets
when number of nodes increase and so as number of initiator nodes.
The overhead is incurred on a per-packet basis, so the smaller the packets, the more
of them there are, the more overhead. The bigger the frames, the fewer there are, the
less overhead is required to send them. This holds true for any technology that
supports variable length frames, including wireless network. Beside that, multi-hop
routing imply high packet overhead, then more nodes in the network means more hop
available. From [10] shows that the packet overhead of the multi-hop routing is
extremely high compared to single path routing since many nodes near the shortest
path participate in packet forwarding. This additional overhead caused by moving
node can cause congestion in the network. Thus Fig. 8 is simply understood.
Overhead increases almost linearly with speed of node. (red line is for speed data,
whereas other lines are for overhead being sent and received).

4.4 Impact of Multi-hop Transmission

In this section, we validate multi-hop transmission of overhead packet with simulation
result.
• With direct communication protocol, each source node sends its data directly
to the destination node. If the destination is far away, direct communication
will require a large amount of transmit power from each node (since d in
Equation 1 is large). Thus nodes route data through intermediate nodes. In this
simulation, the intermediate nodes are chosen such that the transmit amplifier
energy ϵamp is minimized. Each message must go through n transmits and n
receives. Depending on the relative costs of the transmit amplifier and the
radio electronics, the total energy expended in the system might actually be
greater than direct transmission.
• In multi-hop environment, the nodes closest to the destination will be used to
route a large number of messages to destination. Thus these nodes will die out
quickly. This will create a cascading effect that will shorten system lifetime.
Fig. 5 might prove this point.
• Fig. 9 and Fig. 8 show the packet overhead with various initiator node speeds.
The packet overhead of the multi-hop routing is extremely high compared to
single path routing since many nodes near the shortest path participate in
packet forwarding.
316 K. Arai and L. Sugiyanta

The accumulated transmission delay and increasing number of hops among nodes that
lowered the throughput can be seen from the previous research results. In [7],
variables of evaluation are the number of hops, length of data transmission, and round
trip time between nodes. The data obtained is shown in Fig. 8 below.
ms RTT delay

#HOP affected the #TCP response time #HOP affected TCP throughput
Fig. 9. The effect of multi-hop transmission in Wireless Mesh Network [7]

Those results agreed with Fig. 5 of the simulation results.

The research examined various products from wireless vendors with different
architectures, which is set for the same measurement condition and were measured
separately. One group is wireless LAN devices equipped with ad hoc mode technology
and mesh network routing capability. Meanwhile, other product group is a wireless
LAN product which expected to function as access point.

5 Related Works
Many researchers have evaluated general problems of performance of multi-hop
wireless mesh network.
One branch of related work is concerned with the evaluation of wireless mesh
product considering their applicability into the enterprise network. A second branch of
work considers the evaluation of node behavior that composes multi-hop radio network.
Nemoto Nozomu [7] considered to introduce the technical review of wireless
mesh network products IEEE802.11 standard with fixed wireless mesh network
node. In terms of performance, the research will evaluate the overwhelming use of
outdoors Muni-WiFi with a view to applying the same technology in the corporate
network.
There are many products and technology for wireless LAN. The difference
technology among products includes route, control technology and architecture.
Wireless LAN is not competing with legacy network, otherwise it can be considered as
complement. In this research, wireless LAN and wireless mesh network is compared to
determine the level of performance. Table 1 shows the main contents of evaluation.
 Agent Based Approach of Routing Protocol Minimizing the Number of Hops 317

Table 1. Content of evaluation

Item Test
Installation Ease of installation
Ease of scale expansion
Basic performance Basic transmission performance
Performance of voice and data transmission
QoS control function
Cracking resistance performance
Management functions
Switch failure Switching time when disability
Switching behavior
Stability Load of stability of network behavior

The second research [12] evaluates an autonomous and distributed control

architecture based on the node behavior that composes the multi-hop radio network
and the QoS evaluation of the traffic. The developed multi-hop radio network that uses
this architecture maximize the autonomous node using wireless as a communication
media, monitors the behavior each neighbor nodes and also the traffic in the node that
composes the network. It aims at the establishment the architecture of technology for
the network construction of autonomous decentralized node with high reliability.
The relay node that composed the multi-hop radio network cooperatively executed
and control the autonomous network developed. In a past technology, the behavior of
node was determined so that each relay node reports to a certain manager. This kind
of technique was suitable if there was correct operation of each relay node, and there
was no problem of breakdown and attack. In this research, the behavior of the node is
evaluated by its adjacent nodes. As a result, it becomes possible to evaluate the
behavior of each relay node accurately when the attack and the un-anticipated trouble.
The stabilized operation of the entire network is attempted by excluding the relay
node that not meets a standard. The prototype is like the figure below.

Fig. 10. The prototype of wireless mesh project [12]

318 K. Arai and L. Sugiyanta

6 Consideration for Dissemination to Multi Areas

Wireless mesh network and wireless technology have emerged in recent years. What
type of nodes moving, either inside the home (indoors) electronics appliances or
outdoors ITS (Intelligent Transport Systems) is a major attraction point. The problem
of routing and performance characteristics such as node capacity and power, expected
to be applied to a wide range of subject.
Several other papers provide analytical formulas for calculating performance of
wireless network. The main difference with [7], [12] and related papers is that we are
here focusing on the performance of wireless network (e.g. #hop, throughput, delay,
energy) with nodes which dynamically active. Some results from this simulation tend
to agree with previous research results. To the best of our knowledge, in both of
previous researches, the nodes are static/not moving. The simulation result can be
enhanced through expanding the simulation with multiple areas without violating the
continuous connectivity requirement.

6.1 Agents
The agent nodes are special nodes that are included to receive the information from
all active nodes in both neighborhood wireless areas. They served as bridges and used
to transport the sensed data to its final destination somewhere between the wireless
network areas. They are also in charge of initiating, executing, and/or controlling the
topology construction and maintenance protocols, routing protocols, etc.
Despite the fact that in real life the hardware configuration of the agent nodes may
different compared to a regular wireless device, simulation uses the same data
structure to model both nodes, with the main difference that agent nodes are
considered to have more amount of energy. Agents in the simulation have two main
functions: as service agents and delivery agents.
1. Service agents: takes a place relay between areas which is advertised with service
request from source nodes. They could create paths and provide optimum hop/path
to destination node.
2. Delivery agents: collectors of information received from nodes in both areas.

6.2 Power Efficiency

Power efficiency is defined as the ratio of the total transmission power over the total
maximum transmission power at all nodes and denoted as η. In the worst case that
each node still uses its maximum transmission power after the execution of the
topology control algorithm, the power efficiency is the minimum value, η = 0. The
higher the power efficiency, the more power saved on transmitting in the network.
Theoretically, the upper bound of η is 1. However, practically, η can never be 1 as the
transmission power can not be 0 in order to ensure the network connectivity.

7 Conclusion
In this paper, we evaluate topology of wireless network with multi-hop routing. One
of the main processes for building hop tree is energy-efficient. The algorithm is done
 Agent Based Approach of Routing Protocol Minimizing the Number of Hops 319

in not two-hop information queries, which reduces the amount of overhead.

Transmission energy is set same as receive energy and topology always changed all
the time, and then the energy to use depends on the network topology. Topology is
always changed, so as the initiator nodes. Increasing number of initiator nodes caused
nodes to route a large number of messages. Thus these nodes will die out quickly,
causing the energy required to get the remaining messages to increase and more nodes
to die. This will create a cascading effect that will shorten system lifetime. Multi-hop
routing, however, imply high packet overhead, then more nodes in the network means
more hop available. The packet overhead of the multi-hop routing is extremely high
compared to single path routing since many nodes near the shortest path participate in
packet forwarding. This additional overhead caused by moving node can cause
congestion in the network. This paper doesn’t include agent yet. However, from the
existing result, several parameters must be paid attention. Energy on each node
decreased as number of initiator node increase. It caused by high intensity of nodes to
sense neighbors as number of nodes increases. As hop increases and there exist agents
as bridge, it is estimated that the energy will decrease rapidly. In future work, we plan
to implement agents to minimize hops to achieve optimum transmission while
maintaining connectivity among areas. Further investigation need to be conducted on
dynamic routing advantages and factors which affect routing mode, e.g., flow type,
delay, and etc. We will also further explore the throughput/delay/reliability tradeoffs
between wireless network areas that deploy agents and without agents.

Acknowledgments. The authors would like to thank the anonymous reviewers for the
helpful comments and suggestions.

References
1. Ettus, M.: System Capacity, Latency, and Power Consumption in Multihop-routed SS-
CDMA Wireless Networks. In: Radio and Wireless Conference (RAWCON 1998), August
1998, pp. 55–58 (1998)
2. Rodoplu, V., Meng, T.H.: Minimum energy mobile wireless networks. IEEE Journal on
Selected Areas in Communications 17(8), 1333–1344 (1999)
3. Lin, X., Stojmenovic, I.: Power-Aware Routing in Ad Hoc Wireless Networks. In: SITE,
University of Ottawa, TR-98-11 (Dec. 1998)
4. Meng, T., Volkan, R.: Distributed Network Protocols for Wireless Communication. In:
Proc. IEEEE ISCAS (May 1998)
5. Manfredi, V., Hancock, R., Kurose, J.: Robust routing for dynamicmanets. In: ACITA
2008 (2008)
6. Ye, Z., Krishnamurthy, S., Tripathi, S.: A framework for reliable routing inmobile ad hoc
networks. In: Proceedings of IEEE INFOCOM (2003)
7. Nozomu, N.: Consideration and Evaluation of Wireless Mesh Network. Nomura Research
Institute (NRI) Pacific Advanced Technologies Eng., 70–85 (2006)
8. Jarábek, B.: Advanced Network Approaches for Wireless Environment. In: Bieliková, M.
(ed.) IIT. SRC 2005, April 27, pp. 108–115 (2005)
9. Heinzelman, W., Chandrakasan, A., Balakrishnan, H.: Energy-efficient communication
protocol for wireless microsensor networks. In: Proceedings of the 33rd International
Conference on System Sciences (HICSS), pp. 1–10 (2000)
320 K. Arai and L. Sugiyanta

10. Oh, S.Y., Gerla, M.: Robust MANET Routing using Adaptive Path Redundancy and
Coding. In: COMSNET 2009, Bangalore, India, January 10 (2009)
11. The Dynamic Source Routing Protocol for Mobile Ad Hoc Networks (DSR), RFC Draft
(July 2004),
http://www.ietf.org/internet-drafts/draft-ietf-manet-dsr-10.txt
12. Kyushu University, Kyushu Institute of Technology.: Developing Multi-Hop Wireless
Communication Technology (WiRem), Press Release Project, March 17 (2008)
 Ensuring Basic Security and Preventing Replay Attack
in a Query Processing Application Domain in WSN

Amrita Ghosal1, Subir Halder1, Sanjib Sur2, Avishek Dan2, and Sipra DasBit2
1
Dept. of Comp. Sc. & Engg, Dr. B. C. Roy Engineering College, Durgapur, India
[email protected], [email protected]
2
Dept. of Comp. Sc. & Tech., Bengal Engineering and Science University, Shibpur, India
{sanjib.sur11,avishekdan}@gmail.com, [email protected]

Abstract. Nodes in a wireless sensor network are susceptible to various attacks

primarily due to their nature of deployment. Therefore, providing security to the
network becomes a big challenge. We propose a scheme considering cluster ar-
chitecture based on LEACH protocol to build a security mechanism in a query-
processing paradigm within wireless sensor network. The scheme is capable of
thwarting replay attack while ensuring essential properties of security such as
authentication, data integrity and data freshness. Our scheme is lightweight as it
employs symmetric key cryptography with very short-length key. We simulate
our scheme to show its efficacy of providing basic security to the network as
well as detecting replay attack in the sensor network. Further we compare our
scheme with one of the existing schemes taking packet loss and packet rejection
ratio as performance metrics.

Keywords: replay attack, denial of service attack, authentication, data integrity,

data freshness.

1 Introduction
The advent of efficient short range radio communication and advances in miniaturiza-
tion of computing devices have given rise to strong interest in wireless sensor net-
works (WSNs) [1]. A wireless sensor network (WSN) consists of a large number of
small, battery-powered [2], [3] wireless sensor nodes. The nodes are distributed in ad-
hoc fashion and process sensing tasks. A sensor network may be deployed in hostile
environments where there can be malicious attackers. On the other hand, the WSNs
are used in critical application domain e.g. defense, medical instrument monitoring.
Therefore, securing the activities of WSNs is of utmost importance.
These WSNs are prone to attacks [4], [5] due to their nature of deployment in an
unattended environment and also due to the broadcast nature of the medium. One
such attack is Denial of service (DoS) attack [6] that possesses a great threat to wire-
less sensor network environment. One form of DoS attack attempts to disrupt the
network service, may be by blocking communication between nodes. The other form
of DoS attack is the path-based DoS (PDoS) attack where the network is flooded with
bogus packets along a multi-hop data delivery path [7]. The packets are continuously

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 321–335, 2010.
© Springer-Verlag Berlin Heidelberg 2010
322 A. Ghosal et al.

replayed back to the sink node. This attack is known as replay attack. However, the
objective of both types of such attacks is to eliminate or diminish the network per-
formance and thereby hamper the working of the whole system. In replay attack,
because of the broadcast nature of sensor networks, an adversary eavesdrops on the
traffic, injects new messages and replays/changes old messages.
Many works are so far reported towards the solution of various forms of DoS at-
tack. Deng et al. [7] have considered a type of DoS attack along a multihop data de-
livery path. As WSNs are generally tree structured, so an attack on the nodes of a path
also affects the branches connected to that path. A one way hash chain mechanism has
been proposed by the authors to prevent such path based DoS attacks and protect end
to end communication along a multihop data delivery path. Here an OHC (One way
Hash Chain) number is added with each message packet leading to an extra overhead
of 8 bytes per packet. These 8 bytes of additional overhead is a major constraint for a
resource constrained sensor mode.
Authors in [8] have proposed an authentication protocol capable of resistance
against replay attacks and other DoS attacks. But the scheme uses symmetric keys
where keys are shared by sensor nodes and therefore a compromised node can send
forged messages which may possess a great security threat to the network.
Perrig et al. [9] have used RC5 algorithm for encrypting messages which means a
lot of computations have to be done. Here the µ TESLA protocol is used for secure
broadcast of messages and an 8-byte Message Authentication Code (MAC) is used by
each node for verification along the communication path. However, exchange of huge
authentication information is a real bottleneck for the resource-constrained WSN.
Moreover a node must have prior knowledge of all the nodes along its path.
The authors in [10] have devised a mechanism to prevent replay attacks. Here it is
considered that the packets used for time-synchronization of any two nodes are re-
played. A receiver-receiver model for time synchronization has been considered
where a reference node broadcasts beacon message to all its neighbours. Based on the
arrival time of the beacon message the receiving nodes adjust their clocks. A time
offset is calculated based on the difference between the recording times of the beacon
message by these receiving nodes. The time offsets are exchanged between the receiv-
ing nodes to calculate a threshold value which is the difference between the two time
offsets of two nodes. But the arrival of a beacon message at a node can be delayed by
certain factors leading to gross errors while deriving the time required for synchroni-
zation between the nodes. Moreover, by not using global time synchronization model
a large overhead has been introduced as huge number of time offsets need to be com-
puted by the nodes.
Dong et al. [11] have proposed the use of hash chains in their work where each
node combines the hash value with its own node-id and forwards this combined value
to its next higher hop count node. The receiving node is able to detect replay attacks
by observing the combined value of node-id and hash value. But the computation of
all these values by nodes takes significant amount of time.
Soroush et al. [12] have developed a scheme to defend replay attacks where a
monotonically increasing counter is maintained to keep track of old replayed mes-
sages. But here each node maintains a counter to store the timing information of all
other nodes which requires a large amount of memory leading to a major bottleneck
for memory- constrained sensor nodes.
 Ensuring Basic Security and Preventing Replay Attack 323

In this paper we propose a secured query processing scheme in WSN with a target
to build a security mechanism from within a query-driven application environment.
The proposed security mechanism gives a solution of replay attack while ensuring
authentication, data integrity and data freshness.
The rest of the paper is organized as follows. In section 2 system model along with a
brief description of the present work is given. A detailed description of the scheme is
given in section 3. Section 4 gives the performance evaluation of the proposed scheme
and also the comparative study between the proposed scheme and another scheme.
Concluding remarks and future scope has been stated in section 5.

2 System Model
An attack [6] is defined as an attempt to gain unauthorized access to a service, re-
source, or information, or the attempt to compromise integrity, availability, or confi-
dentiality. DoS is one type of attack which hinders communication among the sensor
nodes. The present work considers replay attack, which is one form of path based
DoS (PDoS) attack. If there is a PDoS attack, the network gets flooded with bogus
packets along a multi-hop delivery path [7]. The occurrence of bogus packets in the
network is due to the replaying of packets by adversaries leading to replay attack. So
PDoS attacks in the network may result in replay attacks.
The system model in the present work considers clustered network architecture
based on LEACH protocol [13]. In this architecture nodes organize themselves into
local clusters with one node acting as the cluster head (CH). LEACH performs local
data fusion to reduce the amount of data sent from the clusters to the base station/
sink. Once all the nodes are organized into clusters, each CH assigns a time slot
(TDMA schedule) for the member nodes in its cluster. The member nodes sense data
and transmit data to the cluster head nodes that are located at one hop distance away
from them. The cluster head nodes transmit the same to the base station after receiv-
ing data from their member nodes.
The present work considers a query-driven application platform where base station
generates query messages and broadcasts the query. Cluster heads receive the broad-
casts query and start the registration process to authenticate their respective member
nodes. Once the registration phase is over, the CHs forward the query messages to
their members. Depending on the nature of queries, specific member nodes send re-
sponse-packets.
The objective of the proposed scheme named as Secured Query Processing Scheme
(SQPS) is to ensure basic security in general and prevent replay attack in particular
while working for a query-driven application. The essential properties [14] of a WSN
required for maintaining basic security within the network are:
• Authentication- A sensor network must be able to differentiate between data
coming from reliable sources and those coming from adversaries.
• Data Integrity- The received data should not be tampered during communica-
tion between the sender and the receiver.
• Data Freshness- The data received at a particular time should be current data
and not old data which may be replayed by adversaries.
324 A. Ghosal et al.

3 The Scheme
In this secured query processing scheme as shown in figure 1, periodically queries are
broadcast from base station. The process begins as soon as the cluster heads receive
queries from base station. The scheme has two phases-
• Registration Phase
• Query Response Phase

3.1 Registration Phase

This phase is used for registering member nodes by the respective cluster head nodes.
The objective of this phase is to register only the authenticated nodes. Moreover the
phase provides a mechanism that will help a CH to ensure data integrity and data
freshness during query processing phase. Upon receiving a query from the base sta-
tion, cluster heads initially broadcast registration packet.
The registration packet contains 8 bits including ‘0’ as MSB. This MSB differenti-
ates between a registration packet and a query packet sent by the cluster heads. As
mentioned in section 2, each member node is allotted a particular time slot which is
used for sending a registration response packet to the cluster head corresponding to
the registration packet and also for receiving identification (node-id) for the member
node from the cluster head for continuing the query processing session. On receiving
the registration packet, the respective member nodes decide the number of bits to be
shifted and accordingly left shift the bits of registration packet. Then the nodes gener-
ate registration response packet including left-shifted registration packet (8 bit), num-
ber of bits left shifted (3 bit), and present time-stamp (12 bit). The key used here is the
number of left shifted bits. Symmetric key cryptography is used as the same key is
used for encryption and decryption of the data packets. Symmetric key cryptography
is beneficial for sensor networks as less computation has to be done and memory
requirement is also minimized. Here, though the key is being sent along with the
registration packet it will not be possible for any adversary who captures or eaves-
drops the contents of the packet to decipher which bits in the packet refer to the key.
Moreover in traditional networks the adversary performs some computations to obtain
the key used in the network; this is not possible in case of sensor networks as the
adversary nodes are also equipped with less computational power. The present time-
stamp indicates the time when a member node sends a reply packet in response to the
registration packet. The time-stamp is represented in minutes and seconds.
Immediately after receiving a registration response packet from one of its member
node, the CH right shifts the reply packet designated number of times. The number of
bits to be right shifted is same as the number of bits the registration packet is left
shifted and it is provided as control data in the registration response packet itself. If
the CH can retrieve the original registration packet after right shifting the received
registration response packet, the member node is considered as an authenticated and
registered node. Upon authenticating a member node, the CH generates a node-id and
sends it to the member node at the same time slot of authenticating and registering the
member node. Once the registration of a member node is successful, the CH stores
this information.
 Ensuring Basic Security and Preventing Replay Attack 325

3.1.1 Encryption at Registration Phase

Let us consider an 8-bit binary registration packet as S= m7…m0, where MSB is m7
and LSB is m0. The registration packet is broadcast by cluster head at a particular time
slot. Time slots are assigned to member nodes according to TDMA schedule. On
receiving the registration packet S, member nodes randomly choose the shift bit B for
left shifting. If B is 100, S is encrypted as m3m2m1m0m7m6m5m4 ( S' ). Accordingly a
3-tuple registration response (RR) packet is formed by the member node as ( S" ,
B, TMN _ present ), where TMN _ present is present time-stamp.

3.1.2 Decryption at Registration Phase

Upon receiving the encrypted registration response packet ( S" , B, TMN _ present ) from
a member node, the cluster head performs the reverse process of encryption to get
back the original registration packet. So S' is right shifted B times and converted
to S" . After decrypting, if S" matches with S, the cluster head accepts the registration
packet and authenticates the corresponding member nodes.

3.1.3 Data Stored by Cluster Head

Once the cluster head authenticates a member node, it generates an identification
(node-id) for the member node and stores a 6-tuple data (node-id, S, S' , B,
TMN _ present , TMN _ previous ) for the member node where S' , B, TMN _ present is RR
(packet sent by the node designated by node-id), S is the original registration packet
and TMN _ previous is a time-stamp which is set to null initially.

3.1.4 Data Stored by Member Node

Once a member node gets its node-id from the cluster head during registration, the
member node stores 3-tuple (node-id, B, TMN _ present ) data.

3.1.5 Algorithm for Registration

Begin
// Action executed by cluster heads (CHs)
1: CH broadcasts 8-bit registration packet (S) at time t
// Actions executed by member nodes (MNs)
2: On receiving S at time (t+1)
3: for i=t+1, i<=t+n, i++
4: left shift S to obtain S'i /* ith member node (MNi) encrypts S by
left shifting it by Bi -bit which is
arbitrarily chosen by MNi to obtain S'i */
5: Generate RR i /* MNi generates 3-tuple registration
response packet RR( S'i , Bi , TMNi _ present ) */
6: Send RR i to CH
// Actions executed by cluster head (CH) on receiving RR i from MNi
326 A. Ghosal et al.

7: right shift S'i to obtain S"i /* CH decrypts S'i by right shifting it by

Bi-bit to obtain S"i */
8: if S"i = S then
9: RR i is authentic /* registration response packet RR i is
authentic */
10: accept RR i /* CH accepts registration response packet
RR i */
11: generate node-id for MNi
12: send node-id to MNi
13: store 6-tuple data /* 6-tuple data: ( node − idi , S, S'i , Bi ,
TMNi _ present , TMNi _ previous )*/
// Action executed by MNi on receiving node − idi
14: store 3-tuple data /*3-tuple data: node − idi , Bi , TMNi _ present */
15: else
16: reject RR i
17: end if
18: end for
End

Node authentication is one of the major parameters of network security. The registra-
tion phase of the present scheme is able to ensure this part of security.

3.2 Query Response Phase

In the event of query processing, the base station broadcasts query. Upon receiving
the query, a CH starts registration phase. Once the registration phase is over (section
3.1), the CH broadcasts the query for its member nodes. Depending on the query,
specific member nodes send response and cluster head nodes forward the responses to
base station. For example, if the query is related to temperature, then member nodes
responsible for sensing temperature provide response to the query.
Query packet is of 8 bits. The MSB bit of the query packet is always 1, to distin-
guish between registration packet and query packet. So, 128 different queries can be
generated with the remaining 7 bits of the query packet.

3.2.1 Encryption at Query-Response Phase

On receiving query packet member nodes respond to the query in the form of a
query response (QR) packet that contains 5-tuple (node-id, m, m ' , TMN _ present ,
TMN _ previous ) where m is an 8-bit response message for the corresponding query and
m' is left shifted (B times) message. Further, TMN _ present and TMN _ previous are
 Ensuring Basic Security and Preventing Replay Attack 327

present time stamps when the member node is sending the response packet for the
query and previous time stamp of last communication (packet exchange) between
member node and CH respectively. To start with, TMN _ present of the 3-tuple data
stored by the member node at registration phase replaces TMN _ previous.

3.2.2 Data Stored by Cluster Head

After receiving query response packet from member node, the cluster head updates
the 6-tuple data stored at registration phase. During registration phase the stored 6-
tuple data was (node-id, S, S' , B, TMN _ present , TMN _ previous ). During query response
phase, S and S' attributes are replaced by m and m' respectively. TMN _ previous which
contained null value during registration phase is now replaced by TMN _ present which
was stored during registration phase and TMN _ present is replaced by the time when the
query response packet is being sent by the member node.

3.2.3 Decryption at Query-Response Phase

Cluster head on receiving the QR packet, finds the node-id of the member node from
the packet. Then to decrypt the encrypted message m', the CH right shifts m ' , B
times where B is found from the stored tuple corresponding to the node-id of the
member node. The m ' is converted to m" after decryption. Once the CH decrypts the
message part of the QR packet, it checks for data integrity, data freshness and replay
attack.

3.2.4 Check for Data Integrity

If it is found that the decrypted message m" is same as the original message m, data
integrity is preserved.

3.2.5 Check for Replay Attack and Data Freshness

In replay attack the malicious node repeats the already sent packets and results in
energy exhaustion of nodes and eventually collapse of the network. With the help of
two attributes TMN _ present and TMN _ previous stored at cluster head corresponding to
every registered/ authenticated member nodes, replay attack is detected. The cluster
head compares the TMN _ present and TMN _ previous of an entry corresponding to a
member node and if TMN _ present stored by it and the TMN _ previous in QR is equal,
then it can be ensured that no replay attack has taken place and data freshness is pre-
served.
If a malicious node attempts to send query response packet posing as an authenti-
cated member node, cluster head rejects the packet. Due to this unauthorized attempt,
an authorized node’s TDMA time slot will be consumed and as a result of which a
packet to be sent by the authorized node is lost. Therefore, there may be two distinct
effects of malicious attempt- one is packet loss and the other is packet rejection.
328 A. Ghosal et al.

3.2.6 Algorithm for Query Response

Begin
// Action executed by cluster head (CH)
1: Broadcasts query packet received from base station at time (t+n+1)
// Actions executed by member nodes (MNs)
2: On receiving query packet at time (t+n+2)
3: for i= t+n+2, i<=t+2n+2, i++
4: send 5-tuple QR i to CH /* ith member node sends query response
packet, QR i ( node − idi , mi , m 'i ( mi left shifted
by Bi bit)), TMNi _ present , TMNi _ previous )to CH */
5: right shift m 'i to obtain m"i /* CH decrypts m 'i by right shifting it by
Bi -bit to obtain m"i */
6: if m"i = mi then /* check for data integrity */
7: accept /* data integrity is maintained */
8: else
9: reject QR i /* data integrity is violated */
10: end if

11: if CH.TMNi _ present =

QR.TMNi _ previous then /* CH compares TMNi _ present
stored in it with TMNi _ previous in QR i */
12: accept QR i /* no replay attack is detected */
13: update its (CH’s) stored data /* replaces S by mi , S'i by m 'i ,
corresponding to MNi TMNi _ previous by TMNi _ present ,
TMNi _ present by TMNi _ present of QR i */

// Action executed by CH on receiving QR i

14: else
15: reject QR i /* CH rejects query response
packet QR i */
16: end if
17: end for
End

The present scheme is based on LEACH protocol which is an internationally

accepted standard protocol. It adds security feature to the LEACH protocol in a
lightweight manner involving shifting of bits. The scheme is implementable as com-
putations for shifting of bits is highly implementable in sensor nodes e.g. mica-2
motes.
 Ensuring Basic Security and Preventing Replay Attack 329

3.3 Diagrammatic Representation of the Scheme

The entire scheme is illustrated with the help of an activity diagram shown in Figure
1.The diagram shows how a cluster with n number of nodes and one cluster head
works. To be more specific, it shows communication among different components of
a wireless sensor network and control flow among various computational modules
within the components of the network. Broadcast communication from cluster head is
shown by lightning symbol whereas unicast communication between cluster head and
member nodes and control flows are shown by solid lines with arrowhead. All com-
munication symbols are labeled to make understand the content of the packet along
with the timestamp at which it is being sent. For example, one broadcast communica-
tion labeled by RBt means registration packet has been broadcast at time t. Time-
stamp helps to know the steps of operations visibly.

Fig. 1. Activity diagram of the proposed scheme (SQPS)

The notations used in Figure 1 has been described below-

RBt – Registration Packet broadcast at time t by cluster head.
RR t +1 – Registration response packet transmitted at time t+1 by member node M1 .
RR t + n – Registration response packet transmitted at time t+n by member node M n .
Node − id M1 – Node-id of member node M1 transmitted at time t+1 by cluster head.
330 A. Ghosal et al.

Node − id M n – Node-id of member node M n transmitted at time t+n by cluster head.

QBt + n +1 – Query packet broadcast at time t+n+1 by cluster head.
QR t + n + 2 – Query response packet transmitted at time t+n+2 by member node M1 .
QR t + 2n + 2 – Query response packet transmitted at time t+2n+2 by member node M n .

4 Performance Evaluation
The effectiveness of the proposed security scheme reported in the earlier section is
evaluated through simulation.

4.1 Simulation Environment

Simulation is performed using MATLAB (version 7.1). We consider 500 nodes in the
network and number of malicious nodes is varied from 25 to 100.
Performance of the scheme is evaluated based on the following two metrics:
Authentication Rate – Number of authenticated nodes / total number of nodes in the
network.
Data freshness (%) – (Number of received packets containing current data / total num-
ber of packets sent) x 100.
The relevant parameters and their associated values are listed below in Table 1-

Table 1. Parameters and their corresponding values used in simulation

Parameters Value
Initial Energy ( Einitial ) 2J
Network area (user input) 25m × 25m to 100 m × 100 m
Communication range of sensor ( R c ) 160 m
Sensing range of sensor ( R s ) 80 m

In presence of replay attack, data freshness is affected. In other words, if replay at-
tack can be considered as a cause, data freshness is an effect. Therefore, the metric,
data freshness is measured to cover analysis on both the security parameters replay
attack and data freshness. Authentication rate is computed to cover another security
parameter i.e. authentication. No separate experiment is carried out to measure data
integrity as violation of data integrity due to presence of malicious node does not
arise.

4.2 Simulation Results

Authentication rate is measured and plotted with time in Figure 2 for varying number
of malicious nodes. Two sets of experiments are conducted to compute authentication
rate.
 Ensuring Basic Security and Preventing Replay Attack 331

TDMA time slots vs Authentication rate (25 malicious

nodes)
1.02

1
Authentication rate
0.98

0.96

0.94
Registration without attack
0.92
Registration with attck
0.9
0 50 100 150 200 250 300
TDMA time slots (sec)

(a) with 25 malicious nodes

TDMA time slots vs Authentication rate (50 malicious

1.05 nodes)

1
Authentication rate

0.95

0.9

0.85

0.8
Registration without attack
0.75
Registration with attack
0.7
0 50 100 150 200 250 300
TDMA time slots (sec)

(b) with 50 malicious nodes

TDMA time slots vs Authentication rate (100 malicious

nodes)
1.2

1
Authentication rate

0.8

0.6

0.4
Registration without attack
0.2
Registration with attack
0
0 50 100 150 200 250 300
TDMA time slots (sec)

(c) with 100 malicious nodes

Fig. 2. Authentication rate over a period of time
332 A. Ghosal et al.

In one set of experiment, we consider that malicious nodes are present and attempt to
participate for sending data but stop participation once it is refused to do so. Results are
plotted in Figure 2 for varying number of malicious nodes and designated as ‘without
attack’. In the other set of experiment, malicious nodes attempt to send data repeatedly
resulting in replay attack. Results of the same are plotted and designated as ‘with attack’.
We observe that in all the cases of Figure 2 ((a), (b), (c)) results ‘with attack’ show
a fall of authentication rate compared to the results of ‘without attack’. This signifies
presence of replay attack from malicious nodes. Further, if we compare results of all
the plots ((a), (b), (c)), it is observed that authentication rate decreases with increase
in number of malicious nodes.
Figure 3 shows percentage of data freshness over a period of time in presence of
malicious nodes. Results in presence of 25 malicious nodes show that average data
freshness is near about 96% whereas its values are 88% and 81% for 50 and 100 ma-
licious nodes respectively. The results indicate that data freshness is inversely related
to the number of malicious nodes present in the network.

TDMA time slots vs Data Freshness (%)

1.2

1
Data Freshness (%)

0.8

0.6

0.4

0.2 Query 25 malicious

Query 50 malicious
Query 100 malicious
0
0 50 100 150 200 250
TDMA time slots (sec)

Fig. 3. Data freshness (%) over a period of time

4.3 Comparative Study

Due to unavailability of an existing suitable work so that our scheme can be compared
based on all the security parameters considered here, we have chosen a work [11] on
secured routing scheme to compare packet loss and data packet rejection ratio as com-
parison metrics as defined below.
Packet loss (%) – Total number of received packets by a CH / total number of packets
sent by the member nodes.
Packet rejection ratio (%) – Total number of received packets by a CH containing
tampered data / total number of packets sent by the member nodes.
Packet loss and packet rejection ratio (%) are computed and plotted for varying num-
ber of malicious nodes in Figure 4 and Figure 5 respectively.
 Ensuring Basic Security and Preventing Replay Attack 333

Number of malicious nodes vs Packet loss (%)

70

60

50
Packet loss (%)

40

30

20
10 M VCS
SQPS
0
0 20 40 60 80 100 120
Number of Malicious nodes

Fig. 4. Packet loss by varying number of malicious nodes

The figures also show the results for secured routing scheme named as MVCS
(mitigating attack against virtual coordinate system) [11]. We observe that packet loss
(Figure 4) increases with the increase of malicious nodes for both the schemes. How-
ever, in our scheme (SQPS), packet loss is about 16% less than MVCS for up to 25
numbers of malicious nodes. For 25 to 85 numbers of malicious nodes packet loss in
SQPS is about 20% less than MVCS and this onwards packet loss is about 16% less
in SQPS. Summarily, it can be said that packet loss in SQPS is less than MVCS for all
sets of values of malicious nodes.

Number of malicious nodes vs Packet Rejection Ratio (%)

70

60
Packet Rejection Ratio (%)

50

40

30

20

10 M VCS
SQPS
0
0 20 40 60 80 100 120
Number of Malicious nodes

Fig. 5. Packet Rejection (%) by varying number of malicious nodes

334 A. Ghosal et al.

Further we observe that packet rejection ratio (Figure 5) increases with the increase
of malicious nodes for both the schemes. But in our scheme (SQPS), packet rejection
is about 10% less than MVCS for up to 25 malicious nodes. For 25 to 100 malicious
nodes packet rejection ratio in SQPS is 20% less than MVCS and this trend continues.
So it can be inferred that packet rejection in SQPS is less than MVCS for all sets of
values of malicious nodes. As all the member nodes’ identities are verified through
registration phase, there is very little chance that a malicious node is able to steal the
identity of a legitimate node and passes through registration phase. That is why SQPS
packet rejection ratio is lower than MVCS.

5 Conclusion
In this paper, we have proposed a scheme to defend replay attacks on nodes of WSN
as well as preserve the essential basic security properties such as authentication, data
integrity and data freshness of such a network. The scheme is designed in such a
manner that no malicious node can take part in actual query processing thereby ensur-
ing authentication.
Moreover, as there is no participating malicious node, violation of data integrity
due to attack has been eliminated. However, a malicious node can attempt to partici-
pate stealing some time slots due to which there may be some packet loss. The merit
of the scheme lies on the fact that simple symmetric key cryptography has been used
to maintain security making the solution very lightweight. We have substantiated our
claims by simulating the scheme in presence of attacks. Finally the scheme is com-
pared with one of the existing routing schemes considering packet loss and data
packet rejection ratio as comparison metrics. Results show our scheme outperforms
the existing one.
As a future extension, the scheme may be made more realistic considering cluster
head nodes are also vulnerable to attack. Further enhancement may be done to make it
applicable for continuous data-flow application domain as well.

References
1. Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci, E.: A survey on sensor networks.
IEEE Communications Magazine 40(8), 102–114 (2002)
2. DasBit, S., Ragupathy, R.: Routing in MANET and Sensor Network- A 3D position based
approach. Journal of Foundation of Computing and Decision Sciences 33(3), 211–239
(2008)
3. Halder, S., Ghosal, A., Sur, S., Dan, A., DasBit, S.: A Lifetime Enhancing Node Deploy-
ment Strategy in WSN. In: Lee, Y.-h., et al. (eds.) FGIT 2009. LNCS, vol. 5899, pp. 296–
308. Springer, Heidelberg (2009)
4. Ghosal, A., Halder, S., DasBit, S.: A Scheme to tolerate Jamming in multiple nodes in
Wireless Sensor Networks. In: Proceedings of Wireless VITAE, pp. 948–951. IEEE Press,
Los Alamitos (2009)
5. Ghosal, A., Halder, S., Chatterjee, S., Sen, J., DasBit, S.: Estimating delay in a data for-
warding scheme for defending jamming attack in wireless sensor network. In: Proceedings
of 3rd International Conference NGMAST, pp. 351–356. IEEE CS Press, Los Alamitos
(2009)
 Ensuring Basic Security and Preventing Replay Attack 335

6. Wood, A.D., Stankovic, J.A.: Denial of Service in Sensor Networks. IEEE Com-
puter 35(10), 54–62 (2002)
7. Deng, J., Han, R., Mishra, S.: Limiting DoS attacks during multihop data delivery in wire-
less sensor networks. Journal of Security and Networks 1(3/4), 167–178 (2006)
8. Liu, D., Ning, P.: Efficient distribution of key chain commitments for broadcast authenti-
cation in distributed sensor networks. In: Proceedings of 10th annual Network and Distrib-
uted System Security Symposium, pp. 263–276 (2003)
9. Perrig, A., Szewczyk, R., Wen, V., Culler, D., Tygar, J.: SPINS: Security Protocols for
Sensor Networks. Wireless Networks Journal (WINET) 8(5), 521–534 (2002)
10. Song, H., Zhu, S., Cao, G.: Attack-Resilient Time Synchronization for Wireless Sensor
Network. In: Proceedings of IEEE International Conference on Mobile Adhoc and Sensor
Systems Conference, vol. (7-7), pp. 772–779 (2005)
11. Dong, J., Ackermann, K.E., Bavar, B., Nita-Rotaru, C.: Mitigating Attacks against Virtual
Coordinate Based Routing in Wireless Sensor Networks. In: Proceedings of 1st ACM con-
ference on Wireless Network Security, pp. 89–99 (2008)
12. Soroush, H., Salajegheh, M., Dimitriou, T.: Providing Transparent Security Services to
Sensor Networks. In: Proceedings of IEEE International Conference on Communications,
pp. 3431–3436 (2007)
13. Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H.: Energy Efficient Communication
protocol for Wireless Microsensor Networks. In: Proceedings of the 33rd Hawaii Interna-
tional Conference on System Sciences, vol. 2, pp. 8020–8029 (2000)
14. Zia, T., Zomaya, A.: Security Issues in Wireless Sensor Networks. In: Proceedings of In-
ternational Conference on Systems and Network Communications, p. 40 (2006)
 A Review of Redundancy Elimination Protocols for
Wireless Sensor Networks

Babak Pazand, Amitava Datta, and Rachel Cardell-Oliver

School of Computer Science & Software Engineering

The University of Western Australia
35 Stirling Highway, Crawley, W.A. 6009, Australia
{babak,datta,rachel}@csse.uwa.edu.au

Abstract. Sensing the same region of a physical environment by many sensor

nodes results in wastage of energy. This behavior conflicts with the most impor-
tant requirement of a wireless sensor network which is power efficiency. Thus,
it is essential to have a mechanism to eliminate high redundancy while preserv-
ing an acceptable level of network coverage. In this paper, we review the pro-
posed protocols in the literature that aim to eliminate the sensing redundancy in
wireless sensor networks.

Keywords: Wireless Sensor Networks, Sensing Redundancy, Node Scheduling.

1 Introduction
Coverage is one of the most important metrics for evaluating the quality of service
(QoS) in wireless sensor networks. A region in a sensor field is considered as a cov-
ered region if every point in this area is in the sensing radius of one or more sensors.
However, it is important to consider how efficiently the sensor field is covered by the
deployed sensors.
Two different aspects of the coverage problem have been discussed in the litera-
ture. The first aspect considers the efficiency of coverage of a deployment region in
terms of the percentage of the total area covered. The second aspect of the coverage
problem is to explore how to reduce redundancy in the coverage. The same area of a
sensor field may be covered by many sensor nodes in a dense deployment. The aim is
to ensure that no more than a few nodes cover a particular area at any time while the
sensor field is fully covered, increasing sensing efficiency. In this paper, we review
the proposed protocols from this sensing efficiency point of view.
We consider three main fields for redundancy elimination protocols. These fields
are node scheduling schemes [1, 2, 3, 4, 5, 6, 7, 8, 9], approaches for solving the cov-
erage problem (redundancy point of view) [10, 11, 12] and solutions for efficient
routing [13, 14]. Although routing protocols proposed in [13, 14] are targeted for
wireless ad-hoc networks. They have the potential to be deployed in wireless sensor
networks.
Redundancy elimination protocols can also be classified into two major groups:
location aware or location free schemes. The former utilizes the precise or relative

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 336–351, 2010.
© Springer-Verlag Berlin Heidelberg 2010
 A Review of Redundancy Elimination Protocols for Wireless Sensor Networks 337

location information of sensor nodes. The latter does not consider any geographical
position information of nodes.
In the following, these protocols are reviewed and a table is presented in the dis-
cussion section to summarize the attributes of each protocol.

2 Location Aware Protocols

Location aware schemes either utilize exact location information or obtain the relative
position of nodes via some recursive trilateration or multilateration techniques [15].
For the first case, each node should be equipped with GPS (Global Positioning Sys-
tem) component or pre-programmed position information. The second case requires
that some anchor nodes with GPS units are deployed in the sensor field.
Because of the knowledge of node’s coordinates, these schemes can eliminate re-
dundancy while preserving 100% network coverage. It means that there are no blind
spots in the sensor field which is the main advantage of location aware schemes.
Moreover, the warm-up stage of these techniques which organizes sensor nodes, usu-
ally results in lower overheads than location free schemes. Also, implementation and
maintenance of these solutions are less complicated.
However, GPS units consume substantial amount of energy and so they decrease the
total energy efficiency of redundancy elimination schemes. Additionally, these units
increase the cost of wireless sensor nodes significantly and they do not work in all
environments. In particular, they assume unobstructed line-of-sight communication
with three or more geostationary satellites which is not available in many situations.
By exploiting range-based localization techniques [16, 17, 18], the relative location
information of nodes are obtained. Nevertheless, we still need a large number of an-
chor nodes equipped with GPS units to reduce the accumulation of localization error
due to iterative multilateration. Although range-free localization methods [19, 34, 20,
21] that utilize connectivity information do not require GPS units, they fail to achieve
practical results with high level of accuracy [22].

2.1 GAF: Geographical Adaptive Fidelity

High redundancy in wireless networks results in several paths between source and
sink. GAF [14] discovers a set of nodes that are equivalent in terms of routing. It then
minimizes the number of paths by turning off the redundant intermediate nodes. A
GAF simulation study confirms that in the presence of representative active nodes, a
stable level of routing fidelity is preserved [14].
The first phase in GAF is to organize nodes in virtual grids. The monitoring area is
divided into a number of grids. Each node is associated with one particular grid based
on its location information. Nodes that are equivalent from the routing point of view
are linked to the same grid. Two grids are adjacent when each node in one grid can
communicate with every node in another gird. Grid size is determined based on the
communication range of nodes.
After assigning a grid to each node, the main phase of GAF begins which determines
and maintains the active node of each grid. In fact, each node has three different states.
These are sleeping, discovery and active. Initially, all nodes are in the discovery mode.
When the discovery timer of a node expires, it broadcasts a discovery message that
338 B. Pazand, A. Datta, and R. Cardell-Oliver

contains the grid id and node id of this node (with this approach equivalent nodes find
each other). If a discovery node does not receive any response from the active node of
its grid, it becomes the active node of this grid. Otherwise, it goes into the sleep state.
The timer of sleeping nodes is set to a random number from the range of [enlt/2, enlt].
Here, enlt is the expected lifetime of the active node. Upon the expiration of this timer, a
sleeping node changes its state to discovery to detect the possible lack of an active node
in its grid. GAF also provides fair energy consumption among sensor nodes. When a
node becomes active, it sets a timer to a pre-defined active duration. When this timer
fires, the active node returns to the discovery state. Now, discovery nodes that recently
have become awake, have the chance to replace the active node of their grid.
GAF has some advantages. First, it sits over any underlying ad hoc routing proto-
col. Second, it dynamically changes the set of active nodes which provides load bal-
ancing. Third, it supports mobility of nodes. However, there are some drawbacks as
well. Although it has been shown in [14] that the percentage of packet loss and route
latency is as good as AODV and DSR routing protocols, it has been investigated un-
der the relatively small number of traffic nodes. GAF does not propose any solution to
preserve the capacity of network after eliminating the redundancy. Therefore, the
performance of GAF can decrease significantly under high data traffic.

2.2 TTS: A Two-Tiered Scheduling Algorithm

The node scheduling presented in [3] poses a novel two-tiered node scheduling archi-
tecture One layer is the coverage-tier which preserves full network coverage and the
layer above is connectivity-tier that maintains a backbone to forward data traffic to the
base station. This approach is targeted to put more nodes in sleep mode by consider-
ing different sleeping behaviors [3]. A lightweight localization method presented in
[23] is utilized for determining the location of sensor nodes. A greedy algorithm is
used to discover a set of sensor nodes for full coverage of the sensor field. The result-
ing cover set forms the coverage-tier. Subsequently, a dominating connectivity set is
calculated so that each node in the cover set can directly communicate with at least
one node of the connectivity set.
Node scheduling begins by considering three different patterns of activity. First,
sensor nodes that neither exist in the coverage-tier nor in the connectivity-tier are put
into sleep mode. Nodes in the coverage-tier wake up periodically to send measured
data to the sink node. Sensor nodes in the connectivity-tier are always awake to for-
ward data to the base station or receive a query from it.
There is an update interval that chooses new coverage and connectivity tiers dy-
namically to balance energy consumption of sensor nodes. At every update interval
the algorithm is targeted to select sensor nodes so that the total remaining energy of
every tier is maximized. A cost function is introduced that calculates the consumed
power per coverage area of each sensor node. Therefore, when a sensor node con-
sumes lower energy to monitor a larger area, it incurs lower cost. Thus, nodes with
lower cost are selected for the next coverage and connectivity tier.
The main advantage of TTS is its tiered architecture. It is argued that having two
different tiers for communication and coverage enables us to devise a wakeup and
sleep pattern for covering nodes instead of keeping the nodes always awake. However,
this architecture fails for the wireless sensor networks that need to monitor the region
of interest constantly and in real time. Also, there is another drawback for TTS which
 A Review of Redundancy Elimination Protocols for Wireless Sensor Networks 339

is its assumption of having the transmission range at least twice the sensing range. This
assumption is necessary to determine the connectivity tier successfully [3].

2.3 K-Coverage Based on Perimeter-Coverage of Sensor Nodes

Authors in [10] propose a solution for the coverage problem. It determines whether
the target area is adequately k-covered or not. Every point in the sensor field should
be covered by at least k sensors. They further propose some applications such as node
scheduling and finding insufficiently covered areas based on their approach.
The main idea of the solution is to determine whether the perimeter of each sensor is
k-covered or not. First, they define that a sensor is k-perimeter-covered if every point on
the perimeter of this sensor is covered by at least k sensors [10]. Then a low-cost algo-
rithm is presented that discovers whether a particular sensor node is k-perimeter-
covered or not. Also, another algorithm is proposed which works with sensor nodes
with different sensing ranges. Finally, the perimeter-coverage of sensor nodes is related
to the coverage property of the monitoring area. It has been proved that when all the
sensor nodes are k-perimeter-covered, then the entire target area is k-covered.
In [10], several applications have been presented based on the proposed approach
for the coverage problem. Determining insufficiently covered areas in the sensor field
is one of the applications. Suppose k is the required value of coverage, first the base
station broadcasts this value to all sensors. Then each sensor communicates with its
neighbours to discover whether it has the segments in its perimeter which are less
than k-covered or not. The base station collects the information about the insuffi-
ciently covered segments to find out the areas that are less than k-covered.
A basic node scheduling protocol is also proposed in [10]. In this approach, each
sensor node S i needs to communicate with its neighbours and ask them to reassess
their perimeter coverage without node. If all the neighbours respond that still they are
k-perimeter-covered, then node S i can turn its radio off.
The proposed algorithm for coverage problem in this research work is quite effi-
cient considering its cost which is Ο(nd log d ) , where n is the number of sensors and
d indicates the highest number of nodes that might intersect a node. An interesting
point about this scheme is its ability to work with irregular sensing ranges based on
polygon approximation. However, its major drawback is its dependency on knowing
the exact coordinates of sensor nodes.

2.4 Integrated Coverage and Connectivity

The basic aim in [5] is to maintain both sensing coverage and network connectivity as
the result of its coverage maintenance protocol. A geometric analysis is presented to
find out the relationship between coverage and connectivity [5]. It has been discov-
ered that full network connectivity is preserved when sensing range is equal or less
than half of the transmission range. As a result, a Coverage Configuration Protocol
(CCP) is devised to eliminate redundancy based on a pre-defined coverage degree. If
the coverage degree requested by the application is k, then the eligibility rule in CCP
is as follows. If all the intersection points inside the sensing range of a particular sen-
sor node are at least k-covered, then the sensor node is not entitled to be an active
node. In order to calculate the intersection points inside the sensing circle, each node
340 B. Pazand, A. Datta, and R. Cardell-Oliver

should maintain a list of its active neighbours located within a distance of twice its
sensing range. Three states have been considered for sensor nodes. These are sleep,
active and listen. Initially all the nodes are in active mode. Then each node checks the
eligibility rule to determine its next state. If the node becomes ineligible it goes to
sleep mode. Then it wakes up periodically and enters into the listen state to re-
evaluate its eligibility. Another solution is also considered in [5] to address the case
when a node’s sensing range is higher than half of the transmission range. SPAN [13]
which is a redundancy eliminator protocol (reviewed later) is integrated with CCP.
The former preserves the full connectivity while the latter maintains the requested
coverage degree for the sensor field. The result of this integration is a new eligibility
algorithm. With this new approach an inactive node executes the eligibility rules of
both SPAN and CCP. If the result of one of them is positive then the inactive node is
eligible to become active. Also, an active node will turn its radio off in case that it
does not comply with the eligibility rule of either SPAN or CCP.
CCP provides very good flexibility in both preserving connectivity and addressing
different levels of coverage. However, there are some shortcomings in this approach.
The computation complexity of CCP is Ο(n 3 ) , where n is the number of active nodes
within twice the sensing range of a node. This complexity is high for wireless sensor
networks and it is not as efficient as other approaches. Moreover, this technique is
infeasible for dense deployment of sensor nodes. Location error can also affect the
performance of CCP. More importantly, the CCP eligibility rule assumes considering
a perfect circular sensing model which is not true in the real world.

2.5 SET K-COVER

Slijepcevic et al. [12] focus on the set k-cover problem as a general form of the cover-
age problem. A heuristic is proposed to solve this problem by determining mutually
exclusive sets of sensor nodes. These sets are called covers that monitor the entire
area of the sensor field. The heuristic has been targeted to maximize the number of
these covers. The proposed solution is based on dividing the deployment area into a
number of fields. Each field is an area of the region of interest in which every point is
covered by the same set of sensor nodes. Suppose A denotes the set of calculated
fields. Then the set k-cover is formulated as follows. There are K disjoint covers, C1,
C2, ….., Ck, in which every member of A is covered by at least one element of every
cover Ci [12]. The first part of the solution organizes points of the deployment area
into a number of fields. A field is a unique area of the sensor field in which points
in this area are covered by the same set of sensors. The second part of the solution
utilizes a greedy algorithm to detect k covers. First, a field with lowest number of
associated sensors is selected. An objective function is then utilized to calculate a
probability value for every node in this field. This function measures the likelihood of
redundant coverage for every sensor node if it is selected. Therefore, a node is se-
lected which has the lowest likelihood. The selected field and its sensors are then
ignored for the next repetition of the heuristic. This process is repeated until the
current cover set attains full coverage. Then the members of calculated cover are
deleted from the fields and the above process is repeated for the next cover set. The
K+1th pass of the above heuristic which is the exit point, occurs when in the begin-
ning of the process there is a field that has not been covered by any of the previously
 A Review of Redundancy Elimination Protocols for Wireless Sensor Networks 341

calculated covers. Calculated covers are responsible to monitor the sensor field. At
every time slice one cover is active while sensors belonging to the other covers are in
sleep mode. The activity period is considered the same for all the covers. However,
the authors have not implemented any scheduling scheme for covers.
The main advantage of this solution is its approach to organize the sensor network
into mutually exclusive sets of nodes. However, the assumption of circular sensing
area is not applicable in the real world. Therefore, the proposed scheme fails to de-
liver the promised full network coverage in real implementations.

2.6 Worst-Case & Best-Case Coverage

Research in [11] specifically manages the coverage problem in wireless sensor net-
works. Two general aspects of the coverage problem are considered. The worst-case
coverage problem identifies the regions in the deployment area that have the lowest
level of coverage compared to other areas. The best-case coverage problem identifies
the regions that have the highest level of coverage by sensor nodes. For each case an
algorithm is presented based on geometric computation and graph theory.
Worst-case coverage is identified by a path with pre-determined starting and end
points in the sensor field. This path is called the Maximal Breach Path. The aim of the
proposed algorithm is to find this path so that for every point on the path the distance
of the point to the closest sensor node is maximized [11]. First, the Voronoi diagram
[24] of the sensor field with its sensor nodes is generated. Then from this diagram a
graph is created so that its nodes are the vertices of the diagram and its edges are the
line segments of the diagram. A weight is also assigned to each edge that is the lowest
distance from the closest sensor in the field. Finally, some graph search algorithms
[11] are utilized to discover the path.
The best-case coverage is identified by a path whose starting point and end point are
initially specified. This path is called the Maximal Support Path. The path is discovered
in a way that the distance of every point on this path to the closest sensor is minimized.
For this case another geometric structure is utilized that is Delaunay triangulation [25].
First, the Delaunay triangulation is computed based on the deployment area and its
sensor nodes. Then a graph is formed such that its vertices are the sensor nodes and its
edges are the lines of the triangles. The weight that is assigned to each edge is the length
of its corresponding edge in the triangle. Finally, a search algorithm with specific search
criteria [11] is applied to the graph to discover the maximal support path.
The most obvious advantages of the proposed scheme are the applications that can
be derived from it. For example, the network administrator is able to discover the
paths in sensor field which suffer from poor coverage. Then, coverage can be easily
improved by deployment of new nodes in those areas. Additionally, by discovering
maximal support paths, it is possible to schedule the redundant nodes of these paths
to sleep. Therefore, the discovery of breach and support paths leads to a balanced
network coverage. There are also some drawbacks in the proposed scheme. It only
considers homogeneous sensor nodes. Moreover, the accuracy of the discovered paths
is strongly related to the precision of location information of sensor nodes. Therefore,
the localization techniques that rely on a few GPS devices can affect the quality of
maximal breach and support paths.
342 B. Pazand, A. Datta, and R. Cardell-Oliver

2.7 Node Scheduling by Sponsoring Nodes

In [4] a node scheduling scheme is proposed in which every node is autonomously re-
sponsible for checking an off-duty eligibility rule. This rule considers the neighbours of
a node to determine whether they can sponsor the node by covering its sensing area. The
proposed solution comprises of two rounds. Round one is a self-scheduling process that
is followed by a typical sensing round. These rounds are repeated periodically.
In the beginning of the self-scheduling round, each node is responsible for discov-
ering its neighbours within its sensing range. This discovery includes location infor-
mation for these nodes. Then, the node determines its eligibility to be an inactive node
by performing some geometric calculations. These calculations include the measure-
ment of the central angles of the sectors that are formed based on intersection of this
node by each of its neighbours.
Every node is responsible for calculating all the central angles that exist between
the node and its neighbours. If the summation of all these angles covers 360°, the
sensing area of the node is fully sponsored by its neighbours and it is eligible to turn
off its radio. In order to avoid any conflict that may occur if nodes sponsor each other
concurrently, a standard back-off technique is utilized. Also, the connectivity between
on-duty nodes is provided by LEACH [26], a cluster-based protocol.
It has been proved in [4] that the requirement of location information can be replaced
by direction information in order to determine sponsor nodes. Additionally, the off-duty
eligibility rule has been further extended to respond to the case that nodes have different
sensing ranges. Nevertheless, there are some major shortcomings. It has been assumed
that the sensing round should be long enough in order to compensate for the overhead of
the self-scheduling round. This assumption can lead to large sensing and communica-
tion gaps in case of node failures. Moreover, in case of dense deployments, there is no
approach to tackle the packet collisions that might occur during neighbourhood discov-
ery. Loss of accuracy of neighbourhood information leads to higher numbers of on-duty
nodes. The last disadvantage is the dependency of the off-duty eligibility rule on the
existence of at least three neighbours for every sensor node.

2.8 Coverage Preserving Redundancy Elimination

Detecting and eliminating redundant sensor nodes while preserving complete network
coverage is the main target of the solution presented in [1, 2]. Moreover, it presents an
approach to discover the coverage-boundary of a sensor field. The proposed methods
are based on Voronoi diagrams generated locally at each sensor node.
A redundant sensor is defined as a node whose sensing disk is completely under cov-
erage of other nodes. This approach utilizes two different Voronoi diagrams. One is the
Voronoi diagram of the network. The second one, constructed for every sensor node, is
the Voronoi diagram of the Voronoi neighbours of that node, called the 2-Voronoi dia-
gram. The vertices of this diagram are called 2-Voronoi Vertices (2-VV). In addition,
the intersection point between the coverage circumcircle of the sensor node and an edge
of its 2-Voronoi diagram is calculated. It is called a 2-Voronoi Intersection Point (2-
VIP). Then, every sensor node is responsible for discovering whether it is redundant or
not by utilizing this information. It has been proved in [1, 2] that a sensor node is
redundant if its 2-VVs and 2-VIPs are under the coverage of its Voronoi neighbours.
Every sensor node is responsible for checking this rule to determine whether it is a
 A Review of Redundancy Elimination Protocols for Wireless Sensor Networks 343

redundant node or not. Redundant nodes that are Voronoi neighbours of each other
should not decide to turn their radio off simultaneously. Since that could cause blind
spots in the sensor field. In order to prevent it, an approach based on discovering a
maximal independent set [27] of redundant nodes is implemented.
The second contribution of the research work presented in [1, 2] is dedicated to the
detection of the coverage-boundary of the deployment area. It is formulated that the
coverage-boundary includes the sensing area of those sensor nodes whose coverage
areas are not fully covered by sensing disks of other nodes. It has been proved that a
node is on the boundary area if its Voronoi cell is not covered by its sensing disk [1,2].
The main advantage of the proposed scheme is its ability to address the heteroge-
neous sensor networks. It is achieved by a modified version of Voronoi diagrams.
Additionally, the proposed scheme utilizes a distributed approach compared to the
centralized Voronoi calculations in [11]. It is capable of recomputing the Voronoi
information locally in order to address node failures and deployment of new nodes.
As a result, it is much more robust in this regard compared to [11]. Nevertheless, this
protocol imposes a large amount of overhead for two reasons. The first issue is the
regular maintenance of Voronoi cells. The second factor is the initial overhead im-
posed when every sensor node is supposed to find its Voronoi neighbours and com-
pute its Voronoi cell. Also, similar to [11], the performance of the scheme is sensitive
to the accuracy of the location information of each node. This is because of the
dependency of these approaches to geometrical calculations.

3 Location Free Protocols

In location free schemes for node scheduling, nodes discover the network topology
and react to its changes by sending control messages. Due to the lack of absolute
position information, the main challenge for these techniques is to maintain a high
level of network coverage after eliminating the redundant nodes. One of the most
important metrics to evaluate the efficiency of these approaches is the number of
active nodes. Obviously, the quality of coverage can be increased by a simple but
wasteful decision of eliminating fewer nodes. However, the issue that determines the
superiority of a proposed scheme is how to have as few as possible active nodes while
preserving a high level of network coverage.
Location free approaches are more cost-effective and more energy efficient than lo-
cation aware protocols because they do not require GPS devices. Therefore, they are
better suited for wireless sensor networks since the major constraint is energy. Loca-
tion independent schemes fail to ensure 100% network coverage and may introduce
high initial packet and energy overhead in the network. However, they extend the life-
time of network considerably without additional cost of special hardware components.

3.1 SPAN: An Energy-Efficient Coordination Algorithm for Topology

Maintenance

Chen et al. [13] propose a protocol to save energy in ad hoc wireless networks by
keeping only the necessary nodes active to forward routing traffic. In SPAN each
node periodically decides whether it is eligible to become a coordinator node and join
the forwarding backbone or it should turn its radio off.
344 B. Pazand, A. Datta, and R. Cardell-Oliver

There is a basic rule which determines the eligibility of a node to be a coordinator

node. The rule is that if a node finds a pair of its neighbours that cannot reach each
other directly or through other coordinator nodes, then the node should be considered
as a coordinator node. However, complying with the eligibility rule does not mean
that the node will surely become an active node because there is the possibility that
multiple nodes in a particular region of the network become eligible at the same time.
This issue is addressed by using random backoff delay for every node that wants to
announce its coordination role. The delay is based on two parameters. The first factor
specifies the number of neighbours that will be connected if this node becomes a
coordinator. The more neighbours get connected, the higher is the priority of this
eligible node. The second factor is based on the amount of remaining energy of the
node, higher priority is given to the node with higher amount of remaining energy.
Each node should announce its willingness to become a coordinator by sending a
HELLO message after the delay time calculated based on the above factors. There-
fore, at the end of the delay time, if a node has not received any HELLO message, it
will broadcast a HELLO packet. Otherwise, the node should reassess its eligibility
based on new conditions of the network. If still the node is eligible it repeats the
above process. Otherwise it will sleep until the next election time.
One advantage of SPAN is its approach to check coordinator nodes regularly to
keep their number as low as possible. A coordinator will be withdrawn if it discovers
that every pair of its neighbours is connected directly or via other coordinators. An-
other useful feature of SPAN is its ability to support load-balancing in the network. A
coordinator will be withdrawn after a period of time, if there is a connection between
every pair of its neighbours directly or via some other neighbours.
In SPAN when a node is asleep the packets that were destined to it will be buffered
at its coordinator. This is achieved based on power-saving feature of 802.11 MAC
layer [28]. The main drawback of SPAN is its dependency on a power-saving feature
of the MAC layer protocol. SPAN will not work with MAC layers that do not provide
this option. Moreover, its performance will be affected according to the challenges
that are related to buffering.

3.2 LUC: Location-Unaware Coverage

Younis et al. [9] propose a solution for node scheduling that relies on four different
conditions. Two tests (RTest-D1 and RTest-D2) are geometrical. They determine that
a node is redundant if its sensing area is covered geometrically by other active nodes.
Two other tests (RTest-H1 and RTest-H2) specify the redundancy of a node if some
conditions based on node density are met [9].
RTest-D1: A node v is redundant if these conditions are satisfied. 1) There are
nodes vj, 1 ≤ j ≤ 3 that lie in the sensing range of node v. 2) vj nodes are active. 3) vj
nodes are pairwise neighbours insides the sensing range. 4) Node v is placed inside
the triangle formed by vj nodes. 5) Boundary of node v is covered by the boundary of
vj nodes.
RTest-D2: A node v is redundant if these conditions are satisfied. 1) There are
nodes vj, 1 ≤ j ≤ 3 that lie in the range 0.618*Rs. Rs is the sensing range of node v. (for
more details please refer to [9]). 2) vj nodes are active. 3) vj nodes are pairwise non-
neighbours within the sensing range of node v.
 A Review of Redundancy Elimination Protocols for Wireless Sensor Networks 345

RTest-H1: A node v is redundant if these conditions are satisfied. 1) There are ac-
tive neighbours vj, j ≥ 4, within distance Rs. 2) Every neighbour of node v is a
neighbour of one or more of vj nodes.
RTest-H2: First these conditions are considered. 1) Node v has at least y active or
undecided neighbours inside its sensing range. 2) Its other neighbours have one or
more neighbours among those y neighbours. Node v is considered redundant if it has
the least remaining energy among those y active neighbours. In [9] the authors as-
sumed that y is 6.
Each node has three states: active, sleep and undecided. A Location-Unaware Cov-
erage (LUC) algorithm utilizes rules D1, D2, H1 and H2 to determine the state of
each node in network. The first phase is started by a neighbourhood discovery when
each node learns its one-hop and two-hop neighbours as well as its distance to its one-
hop neighbours. Then the operational phase begins that determines the active and
asleep nodes. Whenever a node notices that one of its neighbours has become an ac-
tive node, it performs a redundancy check procedure to decide whether it is redundant
or not. This procedure performs the tests in a special order. The order is the same as
we listed above. If the result of any test determines that the node is redundant, it will
quit the procedure and will change its state to sleep. Otherwise, the procedure is trig-
gered by the next active neighbour. However, if a specific timer expires and the node
still remains in an undecided state, the last round of the redundancy checking proce-
dure is performed. Nodes become active if all the tests have a negative result.
There are some shortcomings of this scheme. First and foremost, it has been shown
via the simulations in [9] that the total success ratio of the first two tests is almost
60%. Since the next two tests are exploited for dense networks, the proposed ap-
proach can not detect all the redundant nodes for low density networks. Moreover, the
algorithm is limited by considering the assumption that the transmission range should
be at least two times the sensing range. Last but not least, nodes should be equipped
with special hardware to measure the distance between neighbours.

3.3 Expectancy of Coverage

The proposed solution in [8] comprises of two parts. First, the authors present some
theoretical analysis to estimate the coverage level of the monitoring area based on
some parameters of the wireless sensor network. Second, they devise a location-free
node scheduling scheme.
In order to formulate the coverage expectancy, three factors are considered. These
are sensing range, number of deployed nodes and area of the sensor field. First the
expected coverage of each sensor node is predicted and then the expected coverage
ratio of the entire deployment area is discovered as follows: (more details in [8])
⎡ ⎛ 1 4 4 3 4 3 ⎞ ⎤
n
⎢ ⎜ r − lr − mr + πr 2 ml ⎟ ⎥
E[C n ] = ⎢1 − ⎜1 − 2 3 3 ⎟ ⎥lm (1)
⎢ ⎜ m 2l 2 ⎟ ⎥
⎢ ⎜⎝ ⎟ ⎥
⎠ ⎦
⎣
Here, E[Cn] is the expected coverage level, r is the sensing range, n is the number of
deployed nodes and the deployment area is represented by an l by m rectangle. Also,
346 B. Pazand, A. Datta, and R. Cardell-Oliver

this expression can be utilized to determine the minimum number of sensor nodes
needed to maintain a pre-defined coverage ratio requested by the application.
This analysis for coverage estimation is further utilized to present a node scheduling
scheme. The proposed solution is an uncoordinated scheme in which each node stochas-
tically decides to become active or asleep. This decision is repeated independently. The
duration of time that nodes remain in active or sleep state is based on an exponential
distribution with parameters λ a and λ s , respectively [8]. If we consider that a particular
node has changed its state m times. Then the probability that the node would be active is:
mλ a
Pa = (2)
mλ a + mλ s
Then it can be estimated that the number of active nodes at any given time is nPa.
Now if we put the estimated number of active nodes in equation 1, the expected cov-
erage of the network can be calculated. This calculation can help us to adjust λ a
and λ s to attain the coverage ratio requested by an application.
Although, there are some basic analyses [35, 36] in the literature for estimating the
coverage, this research takes into account the border effect to present a more accurate
approximation. Border effect refers to nodes that are located close to the border of the
monitoring area. Moreover, the proposed node scheduling scheme is an uncoordinated
approach with very low overhead. Nevertheless, the scheduling process is completely
stochastic which might not possess a stable behaviour for some issues such as pre-
serving the network capacity. In addition, the proposed solution for coverage expec-
tancy neither addresses sensor nodes with different sensing ranges nor real-world
irregularity of sensing area.

3.4 PEAS: Probing Environment and Adaptive Sleeping

PEAS [6, 7] is a distributed redundancy eliminator protocol for wireless sensor net-
works. It has been devised to prolong the network lifetime by detecting a group of
working nodes while redundant nodes are put into sleep mode.
PEAS comprises of two main components, probing environment and adaptive
sleeping. The former is responsible for maintaining the necessary set of working
nodes. The latter is responsible for adjusting the sleeping time of inactive nodes.
Three different states have been considered for sensors; sleeping, probing and work-
ing. Initially all the nodes are in sleeping mode. When the sleep timer of a node ex-
pires, it wakes up and broadcasts a PROBE message based on a pre-defined probe
range and waits to receive a REPLY message from each of the neighbouring working
nodes. If the probing node does not receive any REPLY message, it becomes a new
member of the group of working nodes. Otherwise it goes back to the sleep mode. A
probe rate λ is considered for probing nodes.
The most remarkable aspect of PEAS is its feature to adjust the frequency of prob-
ing actions by an adaptive sleeping component. The goal is that every working node
gets probed based on a desired probe rate, λ d . In order to achieve this goal, every
working node is responsible for measuring the aggregate probe rate, λ , that it has
received from its neighbouring probe nodes. Then, probing nodes use the aggregate
probe rate to calculate a new probe rate as follows.
 A Review of Redundancy Elimination Protocols for Wireless Sensor Networks 347

λd
λnew = λ (3)
λ
Subsequently, a new sleeping time is generated according to the equation 4. Where, ts
is the previous sleeping time.

f (t s ) = λ new e − λ
new
ts
(4)
It has been proven in [6, 7] that with this kind of adjustment of probe rate as well as
exponentially distributed sleeping times, the probe rate that every working node re-
ceives is equal to the desired probe rate.
PEAS is a location-free solution for the coverage problem that does not rely on dis-
tance information. Moreover, by simply varying the probe range it can provide differ-
ent coverage levels based on application requirements. Additionally, its proposed
approach of adaptive sleeping can address node failures quite robustly while keeping
the overhead of this action low. Due to the high efficiency of PEAS in terms of han-
dling node failures, we have utilized it as the fault-tolerant module of our node sched-
uling scheme. However, it needs to be mentioned that the main disadvantage of PEAS
is its quality in terms of temporal coverage. Temporal coverage indicates the cover-
age ratio of the sensor field throughout the network lifetime.

3.5 Triangular: Determining Active Sensor Nodes for Complete Coverage by a

Triangular Self-test Mechanism

Wi et al. [29] propose a location-free approach for selecting a set of active sensor
nodes. The main contribution of this research is that in the absence of node coordi-
nates, it can achieve 100% network coverage. However, it assumes that nodes are
equipped with omni-directional antennas to measure the distance between nodes by
received signal strength. The proposed solution comprises of two phases. In phase one
the initial set of active nodes is determined. The second phase is dedicated to detect-
ing blind spots in the sensor field.
First, all the nodes are inactive and each of them runs a back-off timer. Upon the
expiration of this timer, a node wakes up and checks whether there is a working node
within its sensing range or not. If the discovery is not successful, the node becomes a
new working node. Otherwise, the node remains awake for a period of time to build a
Neighbour Table based on some control messages. This table includes node id and
distances of all the one-hop and two-hop neighbours. It is utilized by the second phase
for detection of sensing holes. When this phase is finished an initial set of active
nodes has been discovered.
In the second phase, a triangular self-test is performed by each inactive node to
eliminate any possible blind spots. For every inactive node a triangle is formed based
on the three closest active neighbours. Then relative coordinates are assigned to the
endpoints of the triangle followed by the calculation of the relative coordinates of the
circumcentre of the triangle. Based on calculating the distance from circumcentre to
the endpoints of the triangle and comparing it with the sensing range, the existence of
a blind spot is determined. If the result of this self checking approach is positive the
inactive node becomes active to fill the sensing holes.
348 B. Pazand, A. Datta, and R. Cardell-Oliver

There are two main advantages of this solution. Firstly, a relative coordinate is
assigned to each node based on virtual coordinates of its neighbours. This position
information is not used for other location calculations. Therefore, the scheme is not
affected by propagation of location errors. Secondly, the simulation results presented
in [29] indicate that the proposed scheme is quite scalable so that the packet overhead
does not change when node density increases. However, there are some drawbacks as
well. Although the simulation results show that the set of active nodes covers the
deployment area completely, the scheme achieves that by eliminating the margins of
the sensor field. Also, there are some real-world factors that affect the efficiency of
distance measurement based on received signal strength. This issue decreases the
performance of the triangular self-test mechanism. Moreover, the proposed solution
assumes that the transmission range should be at least three times the sensing range
which limits the applicability of the solution.

4 Discussion
The first and foremost characteristic of a redundancy elimination protocol is its ap-
proach regarding location information. In sections 2 and 3 we discussed the advan-
tages and disadvantages of location-free and location-aware solutions. However, an
important point that needs to be mentioned is that some location-free approaches [29,
9] consider distance information. There are three main drawbacks in this regard. First,
we need special hardware to measure distance. This increases the cost of the wireless
sensor network. Second, environmental conditions affect the accuracy of distance
estimation which might decrease the quality of coverage. The third drawback is re-
lated to the techniques that are used for measuring the distance between nodes. The
RF signal strength approaches might have low performance when the transmission
range is short. Fading, multipath and interferences are the factors that affect the signal
strength when the radio range is short [30]. Moreover, time of flight approaches [31,
32, 33] might deliver inaccurate results mostly because of clock skew and drift be-
tween nodes.
Protocols [10, 12, 4, 5, 29, 9] determine redundant nodes based on some calcula-
tions on either the area of the sensing circle or the perimeter of these circles. These
approaches might not deliver what they have promised to achieve. This is due to the
assumption of perfect circular sensing areas. In the real-world, sensing is not only
irregular but it might vary over time. Although other protocols consider circular sens-
ing areas, they use it only for calculation of the coverage ratio and not as a building
block of their schemes. Therefore, the only shortcoming of these approaches is some
degrees of error in measuring their coverage ratio metric.
All the solutions that we have reviewed in this chapter utilize the perfect circular
radio propagation model. This model does not reflect the real-life behaviour of the
wireless channel. Therefore, the performance of these protocols when they are im-
plemented can be different from what we have seen through simulation results. Al-
though GAF [14] has investigated the effect of the shadowing radio model, this proto-
col is originally proposed for unrealistic radio models.
Elimination of redundant nodes might result in an unconnected set of representa-
tive nodes. In order to tackle this issue, some approaches [3, 29, 9] consider a special
assumption that the transmission range should be at least twice the sensing range.
 A Review of Redundancy Elimination Protocols for Wireless Sensor Networks 349

Therefore, they can not address particular applications where the communication
range might be even lower than sensing range.
Finally, Table 1 compares the main characteristics of the protocols that we have
reviewed in this paper.

Table 1. Comparion of redundancy eliminatro protocols 1

Redundancy Use of
Location Location Distance
Eliminator Distributed Centralized Coordinated Uncoordinated Unrealistic
Free Aware Information
Protocol Radio Model

GAF [14]
- 9 - 9 - 9 - -1

TTS [3]
- 9 - - 9 - 9 9
K-Coverage
based on
Perimeter-
Coverage of - 9 - 9 - 9 - 9
Sensor
Nodes[10]
Integrated
Coverage and
Connectivity - 9 - 9 - 9 - 9
[5]
SET
K-COVER
- 9 - - 9 9 9
[12]
Worst-Case &
Best-Case
- 9 - - 9 N/A - 9
Coverage [11]
Node
Scheduling by
Sponsoring - 9 - 9 - 9 - 9
Nodes [4]
Coverage
Preserving
Redundancy
- 9 - 9 - 9 - 9
Elimination
[1, 2]

SPAN [13] 9 - - 9 - 9 - 9
LUC [9] 9 - 9 9 - 9 - 9
Expectancy of
Coverage [8] 9 - - 9 - 9 9
PEAS [6,7] 9 - - 9 - 9 - 9
Triangular
[29] 9 - 9 9 - 9 - 9

1
GAF is originally proposed for unrealistic radio models. However it has also been investi-
gated under shadowing model.
350 B. Pazand, A. Datta, and R. Cardell-Oliver

References
1. Carbunar, B., Grama, A., VIitek, J., Carbunar, O.: Redundancy and Coverage Detection in
Sensor Networks. ACM Transactions on Sensor Networks 2(1), 94–128 (2006)
2. Carbunar, B., Grama, A., Vitek, J., Carbunar, O.: Coverage Preserving Redundancy Elimina-
tion in Sensor Networks. In: Proceedings of the First IEEE Communications Society Confer-
ence on Sensor and Ad Hoc Communications and Networks (SECON 2004) (October 2004)
3. Tezcan, N., Wang, W.: TTS: A Two-Tiered Scheduling Algorithm for Wireless Sensor
Networks. In: Proceedings of IEEE International Conference on Communications (ICC
2006) (June 2006)
4. Tian, D., Georganas, N.D.: A Coverage-Preserving Node Scheduling Scheme for Large
Wireless Sensor Networks. In: Proceedings of the 1st ACM International Workshop on
Wireless Sensor Networks and Applications (WSNA 2002) (September 2002)
5. Wang, X., Xing, G., Zhang, Y., Lu, C., Pless, R., Gill, C.: Integrated Coverage and Connec-
tivity Configuration in Wireless Sensor Networks. In: Proceedings of the 1st International
Conference on Embedded Networked Sensor Systems (SenSys 2003) (November 2003)
6. Ye, F., Zhang, H., Lu, S., Zhang, L., Hou, J.: A Randomized Energy-Conservation Proto-
col for Resilient Sensor Networks. Wireless Networks 12(5), 637–652 (2006)
7. Ye, F., Zhong, G., Cheng, J., Lu, S., Zhang, L.: PEAS: A Robust Energy Conserving Pro-
tocol for Long-lived Sensor Networks. In: Proceedings of the 23rd International Confer-
ence on Distributed Computing Systems, ICDCS 2003 (May 2003)
8. Yen, L.-H., Yu, C.W., Cheng, Y.-M.: Expected K-Coverage in Wireless Sensor Networks.
Ad Hoc Networks 4(5), 636–650 (2005)
9. Younis, O., Krunz, M., Ramasubramanian, S.: Coverage Without Location Information.
In: Proceedings of the IEEE International Conference on Network Protocols (ICNP 2007)
(October 2007)
10. Huang, C.-F., Tseng, Y.-C.: The Coverage Problem in a Wireless Sensor Network. In:
Proceedings of the 2nd ACM International Conference on Wireless Sensor Networks and
Applications (WSNA 2003) (September 2003)
11. Meguerdichian, S., Koushanfar, F., Potkonjak, M., Srivastava, M.: Coverage Problems in
Wireless Ad-Hoc Sensor Networks. In: Proceedings of the twentieth Annual Joint Confer-
ence of the IEEE Computer and Communications Societies (INFOCOM 2001) (April 2001)
12. Slijepcevic, S., Potkonjak, M.: Power Efficient Organization of Wireless Sensor Networks.
In: Proceedings of the IEEE International Conference on Communications (ICC 2001)
(June 2001)
13. Chen, B., Jamieson, K., Balakrishnan, H., Morris, R.: Span: An Energy-Efficient Coordi-
nation Algorithm for Topology Maintenance in Ad Hoc Wireless Networks. Wireless
Networks 8(5), 481–494 (2002)
14. Xu, Y., Heidemann, J., Estrin, D.: Geography-Informed Energy Conservation for Ad Hoc
Routing. In: Proceedings of the 7th Annual International Conference on Mobile Comput-
ing and Networking (MobiCom 2001) (July 2001)
15. Bulusu, N., Estrin, D., Girod, L., Heidemann, J.: Scalable Coordination for Wireless Sen-
sor Networks: Self-Configuring Localization Systems. In: Proceedings of the 6th Interna-
tional Symposium on Communication Theory and Applications (ISCTA 2001) (July 2001)
16. Priyantha, N.B., Chakraborty, A., Balakrishnan, H.: The Cricket Location-Support System.
In: Proceedings of the 6th Annual International Conference on Mobile Computing and
Networking (MobiCom 2000) (August 2000)
17. Savvides, A., Han, C.-C., Strivastava, M.B.: Dynamic Fine-Grained Localization in Ad-
Hoc Networks of Sensors. In: Proceedings of the 7th Annual International Conference on
Mobile Computing and Networking (MobiCom 2001) (July 2001)
 A Review of Redundancy Elimination Protocols for Wireless Sensor Networks 351

18. Whitehouse, C.D.: The Design of Calamari: An Ad-hoc Localization System for Sensor
Networks. Master Thesis, University of California at Berkeley (2002)
19. He, T., Huang, C., Blum, B.M., Stankovic, J.A., Abdelzaher, T.: Range-Free Localization
Schemes for Large Scale Sensor Networks. In: Proceedings of the 9th Annual International
Conference on Mobile Computing and Networking (MobiCom 2003) (September 2003)
20. Moore, D., Leonard, J., Rus, D., Teller, S.: Robust Distributed Network Localization with
Noisy Range Measurements. In: Proceedings of the 2nd International Conference on Em-
bedded Networked Sensor Systems (SenSys 2004) (November 2004)
21. Niculescu, D., Nath, B.: Ad Hoc Positioning System (APS). In: Proceedings of Global
Telecommunications Conference (GLOBECOM 2001) (November 2001)
22. Zhong, Z., He, T.: MSP: Multi-Sequence Positioning of Wireless Sensor Nodes. In: Pro-
ceedings of the 5th International Conference on Embedded Networked Sensor Systems
(SenSys 2007) (November 2007)
23. Cheng, X., Thaeler, A., Xue, G., Chen, D.: TPS: A Time-Based Positioning Scheme for
Outdoor Wireless Sensor Networks. In: Proceedings of the Twenty-third Annual Joint
Conference of the IEEE Computer and Communications Societies (INFOCOM 2004)
(March 2004)
24. Aurenhammer, F.: Voronoi Diagrams - A Survey of a Fundamental Geometric Data Struc-
ture. ACM Computing Surveys 23(3), 345–405 (1991)
25. Preparata, F.P., Shamos, M.I.: Computational Geometry: An Introduction. Springer-Verlag
New York, Inc., Heidelberg (1985)
26. Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H.: Energy-Efficient Communication
Protocol for Wireless Microsensor Networks. In: Proceedings of the 33rd Hawaii Interna-
tional Conference on System Sciences (HICSS 2000) (January 2000)
27. Luby, M.: A Simple Parallel Algorithm for the Maximal Independent Set Problem. In:
Proceedings of the 7th Annual ACM Symposium on Theory of Computing (STOC 1985)
(May 1985)
28. Wireless LAN Medium Access Control and Physical Layer Specifications. IEEE 802.11
Standard (IEEE Computer Society LAN MAN Standards Committee) (August 1999)
29. Wi, T.-T., Ssu, K.-F.: Determining Active Sensor Nodes for Complete Coverage without
Location Information. International Journal of Ad Hoc and Ubiquitous Computing 1(1-2),
38–46 (2005)
30. Bulusu, N., Heidemann, J., Estrin, D.: GPS-less Low Cost Outdoor Localization for Very
Small Devices. IEEE Personal Communications 7(5), 28–34 (2000)
31. Girod, L., Lukac, M., Trifa, V., Estrin, D.: The Design and Implementation of a Self-
Calibrating Distributed Acoustic Sensing Platform. In: Proceedings of the 4th International
Conference on Embedded Networked Sensor Systems (SenSys 2006) (November 2006)
32. Whitehouse, K., Culler, D.: Calibration as Parameter Estimation in Sensor Networks. In:
Proceedings of the 1st ACM International Workshop on Wireless Sensor Networks and
Applications (WSNA 2002) (September 2002)
33. Youssef, M., Youssef, A., Rieger, C., Shankar, U., Agrawala, A.: PinPoint: An Asynchro-
nous Time-Based Location Determination System. In: Proceedings of the 4th International
Conference on Mobile Systems, Applications and Services (June 2006)
34. Lazos, L., Poovendran, R.: SeRLoc: Secure Range-Independent Localization for Wireless
Sensor Networks. In: Proceedings of the 3rd ACM Workshop on Wireless Security (WiSe
2004) (October 2004)
35. Hall, P.: Introduction to the Theory of Coverage Processes. John Wiley and Sons, Chiches-
ter (1988)
36. Philips, T.K., Panwar, S.S., Tantawi, A.N.: Connectivity Properties of a Packet Radio
Network Model. IEEE Transactions on Information Theory 35(5), 1044–1047 (1989)
 A Light-Weighted Misused Key Detection
in Wireless Sensor Networks

Young-Ju Han1 , Min-Woo Park1, Jong-Myoung Kim2 , and Tai-Myoung Chung1

1
Internet Management Technology Laboratory,
Department of Electrical and Computer Engineering, Sungkyunkwan University,
300 Cheoncheon-dong, Jangan-gu, Suwon-si, Gyeonggi-do, 440-746, Republic of Korea
Tel.: +82-31-290-7222; Fax: +82-31-299-6673
{yjhan,mwpark,tmchung}@imtl.skku.ac.kr
2
Incidents Analysis Team, Korea Internet Security Agency, Republic of Korea
[email protected]

Abstract. In wireless sensor networks, sensor nodes are generally dis-

tributed in hostile environments so the security services such as confi-
dentiality, authentication and integrity are very important. The basis
of these security services is the key management. The majority of key
managements use the random key pre-distribution mechanism based on
the probability model. Thus, an adversary can easily get the keys be-
tween any two non-compromised nodes by compromising some number
of nodes. If the adversary gets the keys, he can modify or insert mes-
sages into the communication channels. These kinds of attacks are very
critical because the adversary may mislead the operation of the net-
works or the application. Recently, Liu and Dong proposed a method
to detect the keys which is misused by an adversary. They suggested
the additional protection for non-compromised sensor nodes even if an
attacker has learned the shared key between them. To detect a misused
key, the traffic overhead is inevitable but their method has some ineffi-
cient aspects by performing a unidirectional misused key detection and
generating unnecessary traffic. In this paper, we enhance the misused
key detecting mechanism by detecting a misused key bidirectionally and
removing the unnecessary traffic while the security functionality of the
Liu and Dong’s scheme is preserved. Our simulation shows that the en-
ergy consumption of our scheme is 36% more efficient than that of the
Liu and Dong’s scheme on average.

1 Introduction
Recent advances in wireless communication, micro system and low-power tech-
nologies have enabled the wireless sensor networks(WSNs). Generally, WSNs
consist of many, low-cost, low-power and small sensor nodes and the base sta-
tion(BS). Once the sensor nodes are deployed at a target area, they collect some
information from the area and report it to the BS. The BS takes the role of
gateway to the traditional networks(:Internet), so we can use the collected in-
formation for specific applications. For example, some applications of WSNs

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 352–367, 2010.

c Springer-Verlag Berlin Heidelberg 2010
 A Light-Weighted Misused Key Detection in WSNs 353

are habitat monitoring, object tracking, fire detection, traffic monitoring and so
on[1,2,7].
The security services in WSNs are very important because the sensor nodes
are often deployed in hostile environments. An adversary can easily overhear
the radio channels, insert or modify some messages, or capture physically some
sensor nodes. To guarantee security service such as the integrity, authentication
and confidentiality in the WSNs, cryptographic methods are essential and the
key management is the most fundamental one. However, the traditional key
management methods such as KDC(Key Distribution Center) and public key
mechanism are the impractical solutions due to the limited resources of sensor
nodes[14].
Many key management techniques in WSNs have been introduced[4,8,9,10,13].
Eschenauer and Gligor proposed the basic random key distribution scheme based
on the probability model to overcome the storage limitation of the sensor node
for storing keys[8]. The main idea is that k keys are randomly selected by each
node out of a large pool of P keys. However, it has the critical drawback that
the attacker can easily threat the network survivability by capturing some nodes.
Chan suggested the q-composite random key distribution scheme[6]. This scheme
makes the resilience against node capture more strong but makes the key sharing
probability between any pair of nodes decreased. Liu and Ning proposed more
enhanced key management scheme using t-degree bivariate polynomials and two-
dimensional grid[3]. It perfectly provides the network resilience until t + 1 nodes
are captured for any specific keys and reduces the overhead on establishing
pairwise keys. An attacker can get the keys between any two non-compromised
nodes because the all of the key management schemes mentioned above are
commonly based on the probability model. If An adversary gets some keys using
physical node capture attack, the attacker might infer some keys between non-
compromised nodes by using the collected keys from compromised nodes, and
then he can attempt some kinds of attacks such as eavesdropping, insertion
and modification of messages. The insertion and modification attacks are more
harmful than the eavesdropping because they can lead the abnormal operations
of the applications in WSNs. Figure 1 describes this situation. Node 3, 6 and

Fig. 1. An example of misused keys

354 Y.-J. Han et al.

7 were compromised by the attacker and the attacker obtained the shared key,
K4,5 , between node 4 and 5. He can insert or modify some messages and finally
threaten the overall functionality of an application.
Recently, Liu and Dong has proposed a misused key detection scheme in WSNs
to prevent these kinds of attacks[5]. According to them, the misused key can be
defined under the following three conditions.
– It should be the key shared between non-compromised nodes.
– The key is exposed to an adversary.
– The key must be used for insertion and modification attacks.
To detect a misused key efficiently in distributed manner, they has introduced
committing values and detecting nodes. Whenever a sender node transfers a mes-
sage to a receiver node, the sender node has to attach committing values to the
message and the receiver node performs the detecting procedure by generating
a report message using the committing values under the predefined probability.
Finally, the detecting nodes and the BS check the report message to determine
whether the key was misused.
However, Liu and Dong’s scheme has four limitations. First, the unused com-
mitting values at the receiver node are big overhead. Second, the reporting
procedure can detect only the unidirectional key misuse from a sender to a
receiver per one report message. Third, it is impractical that the detecting node
lists must be exchanged between sensor nodes because the pairwise key estab-
lishment depends on the routing protocol deployed for the purpose of efficient
operation of specific applications in WSNs. Additionally, the detecting proce-
dure in Liu and Dong’s scheme occurs unbalanced traffic overhead among sensor
nodes.
In this paper, we introduce a light-weighted misused key detection mechanism
based on Liu and Dong’s scheme. Our scheme has better two properties, which
are the reduction of the detecting overhead and the high detecting probability,
than Liu and Dong’s scheme by performing the bidirectional misused key de-
tection, removing of unnecessary message and using the reasonable probability
method for the balanced misused detection among the sensor nodes.
The rest of this paper is organized as follows. The next section briefly re-
views the Liu and Dong’s scheme and discusses about its limitations. Section
3 introduces our scheme: a light-weighted misused key detection mechanism. In
section 4, we will analyze our mechanism from the viewpoints of security and
energy efficiency compared with the Liu and Dong’s method. Finally, section 5
summarizes our paper and discusses about future research.

2 The Liu and Dong’s Work

An adversary can compromise easily a few sensor nodes and may obtain many
shared keys between non-compromised nodes. By using these exposed keys, the
adversary can insert forged or modified messages into the networks. The appli-
cation of WSNs might be misled by the adversary since the modified or forged
 A Light-Weighted Misused Key Detection in WSNs 355

messages make the application choice wrong decisions. To prevent these kinds of
attacks, Liu and Dong suggests the additional protection for non-compromised
sensor nodes even if an attacker has learned the shared keys between them.

2.1 The Assumptions

Before we summarize their method, we briefly review their assumptions as
follows.

– The adversary can eavesdrop, modify, replay, forge or block any network
traffic.
– It is computationally impossible to break the cryptographic primitives.
– The adversary can compromise a few nodes and obtain some shared keys
between non-compromised nodes.

2.2 The Ideas

The main idea of this scheme is based on the committing value(CV) and the
secret key between a sensor node and the BS. The CV is obtained from the
cumulative hashed value about outgoing messages(for a sender) or incoming
messages(for a receiver). If node u and v has established a pairwise key Ku,v ,
each node maintains a variable Cu,v to track the history of the communication
from u to v(Cv,u is also maintained for the inverse direction). The initial value
of the Cu,v is 0. If the node v receives the un-forged, un-modified or un-inserted
messages, the Cu,v of u would be the same as that of v. If the two nodes want
to check whether their pairwise key was misused, they simply calculate the CVs
using the secret key shared with the BS and send it to the BS. Finally, the BS
checks whether the key was misused and updates the key if the key was misused.
In addition to this basic idea, they introduced the detecting node, which are
distributed in WSNs and detects a misused key on behalf of the BS to avoid
the communication overheads resulted from the centralized detection by the BS.
The detecting nodes have the keying materials so they can check the misused
keys without the help of the BS.

2.3 The Operation

The operation of Liu and Dong’s scheme is divided into three parts;initialization
phase, message committing phase and detecting phase.

Initialization Phase. Let m be the total number of detecting nodes and n be

the total number of sensor nodes. Before deployment, each sensor node u pre-
loads with a unique secret key Ku shared with only the BS. Every detecting node
i stores all the hashed values H(Ku , i) for every sensor node,i(i ∈ n). Figure 2(a)
describes this step and shows an example of pre-loaded keys for the sensors and
the detecting nodes where m = 2 and n = 4.
356 Y.-J. Han et al.

(a) Initialization

(b) Committing, sampling and detection

Fig. 2. The operation of Liu and Dong’s scheme

Message Committing Phase. After deployment, each sensor node u makes

the detecting node ID list (u)of the detecting nodes around it. If u and v establish
a pairwise key Ku,v , they exchange their detecting node ID lists((u) and (v))
securely using Ku,v . For every new message M from u to v, u updates the Cu,v =
H(Cu,v , M ) then computes two committing values, V1 = H(Cu,v , H(Ku , i)) and
V2 = H(Cu,v , Ku ), where the i is the ID of the detecting node i in the (v). To
synchronize the detecting node ID, u and v select the detecting node i at the
position 1+(s mod S) in (v), where the s is the sequence number of M and the S
is the size of (v). Finally, u transmits M ||V1 ||V2 with M AC1 which is generated
with the key, Ku,v .

Detecting Phase. When v receives the message M , v updates Cu,v= H(Cu,v , M ).

Depending on the probability p, v makes a report message that contains a two pieces
of information. The first one is u||Cu,v ||V1 for the detecting node i and the other
one is u||v||Cu,v ||V2 for the BS. Figure 2(b) shows these messages. The M AC2 is
generated by using H(Kv , i) and the M AC3 is generated by using Kv .
 A Light-Weighted Misused Key Detection in WSNs 357

When the detecting node i receives the report message, i verifies whether V1
equals H(Cu,v , H(Ku , i)). If the verification fails, i simply forwards
u||v||Cu,v ||V2 ||M AC3 to the BS. Finally, the BS checks if V2 equals H(Cu,v , Ku ).
If not, the BS judges the pairwise key between u and v to be misused and then
distributes the new pairwise key to u and v using the unique secret key Ku and
Kv .

2.4 The Limitations

The redundancy of CVs. Sensor node u transmits the V1 and V2 to node v for
every message M but v uses the CVs only when only the probability p is satisfied
by the receiver-side v. The unused CVs are big overheads since the packet size
of the sensor networks generally is very small. For example, sensor nodes may
report the sensed data such as humidity, temperature and illumination. This
information can be expressed within a few bytes but the size of CV may be
8 bytes which is the general size of hashed value in the WSNs. Unlike in the
traditional network, 8 bytes is considered as a big overhead in WSNs. Besides,
the transmission of unused CVs occurs the unnecessary communication overhead.
This overhead comes from the location where the probability p is applied on.
The receiver-side v sends the report message to the detecting node under the
probability p. Indeed, the sender-side u does not have to send the CVs which
will not be used in the receiver-side v.

Unidirectional detection. The detection of a misused key is only limited to

one direction(u → v or v → u), since the detecting node can check only one
variable(Cu,v or Cv,u ) through one report message. To check the misused key
of the opposite direction, u has to create a new report message containing Cv,u
because Cu,v is the only unidirectional cumulative hashed value from u to v.

Exchanging of the detecting node lists. The selection of detecting node de-
pends on the sequence number of the message and the size of the detecting node
list. The sender node must previously know the detecting node list of receiver
node in secure manner. However, this may be infeasible since the establishment
of pairwise key between the sender node and the receiver node is not predictable.
The reason is that the establishment of a pairwise key not only depends on the
routing or the application of the network but also frequently changes as the
environment varies.

Unbalanced detection among the sensor nodes. The detecting procedure

in Liu and Dong’s scheme occurs the unbalanced traffic overhead among sensor
nodes. To limit the traffic overhead of the misused key detection, they use the
probability p for generating a report message. However, a sensor node u can
generate a new report message although u has reported just before. In this case,
sending the second report message occurs unnecessary traffic overhead. Rather
doing like this, it is efficient that other sensor nodes which did not send the
report message recently have to generate a new report message. For doing so,
358 Y.-J. Han et al.

our scheme will make each sensor node perform the detecting procedure one
per 1/p message transmitting rounds on average for randomized rotation of the
detection procedure among the sensor nodes.

3 The Proposed Scheme

In this section, we present our scheme: a light-weighted misused key detection

mechanism. The main focuses of our scheme are the reduction of overhead caused
by the report message for detecting misused keys and the maximization of the
misused key detection probability while the functionality of the Liu and Dong’s
scheme is preserved.
For convenience, we define some notations like bellow list.

– u and v are non-compromised nodes.

– Ku,v is the pairwise key between u and v.
– fc is the fraction of detecting nodes that are compromised.
– Cu,v is the cumulatively hashed value of the communication from u to v and
is concerned with Ku,v .
– Cv,u is the cumulatively hashed value of the communication from v to u and
is concerned with Ku,v .
– (u) and (v) are the detecting node ID lists of u and v in order.

3.1 Our Ideas

Location of the node requesting the detection procedure. To remove

the overhead resulted from the unused CVs in Liu and Dong’s scheme, we do
not send the CVs for every new messages by applying the probability p at the
sender-side u. This guarantees that the unused CVs do not transmitted and
the traffic overheads are reduced. If T denotes the traffic overhead caused by the
CVs, we can reduce the overhead as T × p.

Bidirectional misused key detection. We will check the two variables(Cu,v

and Cv,u ) with one report message. This idea can reduce the network traffic for
detecting misused keys by half.

No exchange of the detecting node lists According to the section 2.4,

the exchange of the detecting node list between two nodes is impractical. In
our scheme, the selection of the detecting node is performed by the node which
would create a report message. As a result, the detecting node lists do not need
to be exchanged previously.

Balanced rotation of detecting procedure among the sensor nodes.

In the Liu and Dons’s scheme, the misused key detection is unbalanced among
sensor nodes when considering the traffic overhead. To overcome this limitation,
we apply the probability model introduced in LEACH[12]. LEACH uses the
 A Light-Weighted Misused Key Detection in WSNs 359

probability model for the randomized cluster leader election of clustering scheme
in WSNs. Whenever u sends a message M to v, u generates a random number δ
between 0 and 1. If the δ < Ptr (u, r), u performs the detecting procedure which
will be described in section 3.2. Equation (1) defines the Ptr (u, r), where p is
the optimal reporting probability for detecting a misused key of a sensor node,
r is the current message transmission round and G is the set of sensor nodes
that did not perform the reporting procedure during last rmax = 1/p message
transmitting rounds. According to Eq. (1), every sensor node must perform the
reporting procedure only once during rmax message transmission rounds. In this
way, the sensor nodes which did not perform the reporting procedure recently
have a more chance for detecting the misused key than the nodes that have
already performed during last rmax rounds. As a result, our scheme ensures that
each sensor node performs the detecting procedure only one per 1/p message
transmitting rounds on average and it achieves the balanced and fair detection
of a misused key among the sensor nodes.
⎧ p
⎪
⎨   ,if n ∈ G
Ptr (u, r) = 1 − p × rmod 1
p (1)
⎪
⎩
0 ,otherwise

3.2 The Operation

The operation of our scheme consists of four phases;initializing, tracking, report-

ing and detecting phases.

Initializing Phase. Initialization phase is exactly same with the Liu and
Dong’s scheme. Each regular sensor node pre-loads the unique secret key shared
with only the BS. The detecting node i stores H(Ku , i) for every sensor node u as
described in Figure 2(a). After deployment, each regular sensor node u discovers
the detecting nodes around itself and makes the detecting node List (u).

Tracking Phase. In this phase, each sensor node only tracks the history of
messages. Whenever u sends a new message M to v, u updates the Cu,v =
H(Cu,v , M ) where the initial value of Cu,v is zero. v updates the Cu,v with
H(Cu,v , M ) if it receives the message M from u. The Cv,u also is tracked by u
and v for the communication from v to u.

Reporting Phase. Whenever sensor node u sends a message M to v, u appends

the detecting node ID, i, to M with the probability Ptr (u, r), where i is the ID
of detecting node which is randomly selected from (u). As the message ① shown
in Fig. 3, u sends M ||i with M AC1 to v. The M AC1 is generated by using the
pairwise key Ku,v between u and v. Whenever this message is sent, u should set
a timer T1 for detecting the blocking attacks.
Once the sensor node v receives the ID i from u, v updates Cu,v = H(Cu,v , M ).
If there is a message M  to send, v updates Cv,u = H(Cv,u , M  ). Otherwise, v
360 Y.-J. Han et al.

Fig. 3. The operation of the proposed scheme

updates Cv,u = H(Cv,u , M ) and then immediately sends back the committing
message, (M  )||V1 ||V2 with M AC2 as the message ② in Fig. 3. (M  ) denotes the
piggybacked message heading for u if there is a message M  to send. If not, only
the V1 ||V2 is transferred. The V1 is the hashed value with H(Cu,v , Cv,u , H(Kv , i))
and V2 is H(Cu,v , Cv,u , Kv ). The M AC2 is generated by using the pairwise key
Ku,v between u and v.
The sensor node u waits the message ② in Fig. 3 until the T1 expires. If T1 ex-
pires, u reports it to the BS that the message was blocked. Otherwise, u generates
a report message as ③ in Fig. 3, where the M AC3 and M AC4 are generated by
using H(Ku , i) and Ku in order. The first piece of information, v||Cu,v ||Cv,u ||V1 ,
is for the detecting node i and the second one, u||v||Cu,v ||Cv,u ||V2 , is for the BS.

Detecting Phase. When the detecting node i receives the report message ③ in
Fig. 3, it verifies whether the V1 equals H(Cu,v , Cv,u , H(Kv , i)). If it is not same,
it means that the key Ku,v was misused so the detecting node i simply forwards
the second piece of information, u||v||Cu,v ||Cv,u ||V2 ||M AC4 as the message ④
in Fig. 3. Otherwise, the detecting node i sets a timer T2 for detection of the
blocking the report message.
The BS checks whether the V2 equals H(Cu,v , Cv,u , Kv ). If it is not equal,

the BS generates a new pairwise key Ku,v and transmits it to the both u and v
securely using Ku and Kv . The BS concludes a correct verification with a high
probability if the detection procedure at the detecting node was successfully
performed.
 A Light-Weighted Misused Key Detection in WSNs 361

4 Evaluation
To enhance the Liu and Dong’s scheme, we have Lintroduced our ideas in the
previous section. This section compares our scheme with Liu and Dong’s scheme
in the aspects of security and energy efficiency.

4.1 Security
False Key Revocation. Any non-compromised key must not be revoked if the
key is shared between non-compromised regular sensor nodes. There are three
cases that we have to consider.
1. If an adversary sends a forged report message to a detecting node i to revoke
a non-compromised key Ku,v , i will never report a false alarm to the BS.
The adversary cannot generate the M AC2 to authenticate his forged report
message to i since he cannot know H(Ku , i).
2. If a detecting node i is compromised and the adversary tries to revoke a
non-compromised key Ku,v through i, the adversary will never revoke the
key. He cannot create the M AC4 because the attacker does not know the
shared key between u and the BS, Ku . As a result, the forged message by
the compromised detecting node will be always rejected at the BS.
3. If a report message is authenticated by a detecting node i or the BS and the
detecting node i is not compromised, the detection of a misused key always
is guaranteed because the V1 and V2 are protected by the shared key, Kv ,
with only the BS.
Liu and Dong’s scheme has proved that their scheme satisfies these three cases.
Our scheme also satisfies the security requirements as described above.

The Probability of Finding Detecting Node. In the Liu and Dong’s

method, they have introduced a detecting node for the purpose of distributing
the centralized detecting overheads and also analyzed about the number of addi-
tional detecting nodes. According to their analysis, Eq. (2) shows the probability
of finding at least one detecting node in any given regular node’s neighborhood,
where m is the number of detecting nodes, n is the number of sensor nodes and
b is the average number of neighboring regular sensor nodes for every regular
sensor node. When b = 50, only 10% additional detecting sensor nodes make a
regular sensor node find at least one detecting node with a probability 0.99. Our
scheme has adopted the detecting node method so this property is also available
in our scheme.
b m 1 bm
Pcover (m, n) = 1 − (1 − ) ≈1−( ) n (2)
n e

Detection Probability. In the Liu and Dong’s scheme, the probability of a

misuse detection is (1 − (1 − (1 − fc · p))x+1 ) × Pcover (m × (1 − fc ), n)), where
x is the number of messages that v has received since the first misusing of the
Ku,v . The probability of our misused key detection can be inferred like follows.
362 Y.-J. Han et al.

1. we can define the probability of not performing the detecting procedure until
r messages have been transmitted within rmax = 1/p rounds as Eq. (3).


r
Pnotr (u, r) = (1 − Ptr (u, i)) (3)
i=1

2. If we assume that u did not performed the detecting procedure until r mes-
sages have been transmitted within rmax rounds, the probability that u does
not perform the reporting procedure until the additional x messages have
been transmitted from u to v can be calculated from Eq. (4).
⎧
⎪
⎪ 
r+x
⎪
⎪ {1−Pnotr (u, i−1) × Ptr (u, i) × (1−fc)}
⎪
⎪
⎪
⎪
⎪
⎪
i=r+1
⎪
⎪ ,if x + r ≤ rmax
⎪
⎨
Pnotdet|notreport (u, x, r) =
⎪
⎪
⎪
⎪ (x+r)modrmax

⎪
⎪
⎪
⎪ f (x+r)/rmax
× {1−Pnotr (u, i−1)
⎪
⎪ c
⎪
⎪
⎪
⎩
i=1
×Ptr (u, i) × (1 − fc )}] ,otherwise
(4)
3. If we assume that u has performed the detecting procedure when r messages
have been transmitted within rmax rounds, the probability that u does not
perform the reporting procedure until the additional x messages have been
transmitted from from u to v can be calculated from Eq. (5).
⎧
⎪
⎪ 1 ,if x + r ≤ rmax
⎪
⎪
⎪
⎪
⎪
⎨
(x+r)modrmax
Pnotdet|reported (u, x, r) = 
⎪
⎪ f (x+r)/rmax −1
× {1−Pnotr (u, i−1)
⎪
⎪ c
⎪
⎪
⎪
⎩
i=1
×Ptr (u, i)×(1 − fc )} ,otherwise
(5)
4. Since our scheme performs the bidirectional misused key detection, node v
also can perform the detecting procedure for Cu,v and Cv,u . Equation (6) thus
denotes the probability that both u and v does not perform the reporting
procedure until the additional x and y messages have been transmitted.

Pnotdet (u, v, x, y, r1 , r2 )={Pnotdet|notreport (u, x, r1 )×Pnotdet|notreport (v, y, r2 )

+Pnotdet|notreport (u, x, r1 )×Pnotdet|reported (v, y, r2 )
+Pnotdet|reported (u, x, r1 )×Pnotdet|notreport (v, y, r2 )
+Pnotdet|report (u, x, r1 )×Pnotdet|reported (v, y, r2 )}/4
(6)
 A Light-Weighted Misused Key Detection in WSNs 363

(a) fc = 0.1 (b) x = 20

Fig. 4. The probability of detecting a misused key

5. Considering the probability of finding at least one detecting node in any

given regular node’s neighborhood, Eq. (7) is the probability of a misused
key detection.

Pdet (u, v, x, y, r1 , r2 ) = (1 − Pnotdet (u, v, x, y, r1 , r2 )) × Pcover ((1 − fc ) × m, n)

(7)
6. Finally, we can conclude the average probability of a misused key detection
by removing the variables(r1 and r2 ) like following Eq. (8).
r
max r
max  
Pdet (u, v, x, y, i, j)
Pdetavg (u, v, x, y) = 2
(8)
i=1 j=1
rmax

Figure 4 depicts the probability of detecting a misused key under assuming that
the average number of neighbors is b = 50, the total number of detecting nodes is
n
m = 10 and y = x. In Fig. 4(a), the detecting probability of our scheme is higher
1
than that of the Liu and Dong’s scheme under the same probability p(= rmax ).
Moreover, the detecting probability of our scheme with p = 0.1 is almost same
with that of the Liu and Dong’s scheme with p = 0.2. Figure 4(b) shows that
the probability of detecting a misused key is slowly decreased as the fraction of
the compromised detecting nodes is increased. Our performance is more slowly
decreased than the Liu and Dong’s scheme under the same probability p. This is
because our scheme performs the bidirectional misused key detection and uses
the probability model of LEACH.
The p is very important factor since it controls a tradeoff between the traffic
overhead and the detecting probability. The higher p is, the higher the detecting
probability is but the more the traffic overhead is increased. Our scheme increases
the probability of a misused key detection while consuming the lower traffic
overhead compared to the Liu and Dong’s scheme.

No Exchange of Detecting Node List. In the Liu and Dong’s method,

the detecting node lists (u) and (v) have to be exchanged immediately after
364 Y.-J. Han et al.

Table 1. The parameters of the simulation

Parameter Value
Eelec 50nJ
Ef s 10pJ/bit/m2
Emp 0.0013pJ/bit/m4
Size of CV 8 bytes
Size of Cu,v 8 bytes
Size of message 80 bytes
Size of node ID 4 bytes
Size of MAC 8 bytes
Average of d 6m

deployment. However, this is very impractical because establishing a pairwise

key may depend on the routing protocol deployed for the purpose of an efficient
operation of specific applications in WSNs. Especially, in the case of establishing
a path key through another sensor nodes, exchanging the detecting node lists is
very dangerous because a modified or forged detecting node list may frustrate
the whole misused key detection procedure.
In our scheme, we do not have to exchange the detecting node lists. Each
sensor node u just keeps its own detecting node list (u) in local repository. If u
wants to do the detecting procedure, u just randomly selects one ID i from the
(u) and informs it to v. At this time, although an adversary modifies the ID,
this modification can be detected at the detecting node i through the detecting
procedure since the adversary cannot know H(Kv , i) and Kv . Even though the
adversary blocks this message, u can detect this blocking attack using the timer
T1 introduced in the section 3.2.

4.2 The Energy Efficiency

In our scheme, we do not send the CVs per message for energy efficiency unlikely
Liu and Dong’s scheme. In the most general WSNs, the size of message is very
small because it only represents the temperature, humidity, illumination and so
on. So, it can be expressed within a few bytes but the CVs are generally hashed
value which may be about 8 bytes. Sending the CVs per message could cause
a big overhead because the size of CV is very large compared to the size of
message.
To compare the energy efficiency, we have simply simulated the energy con-
sumption of two nodes u and v using Matlab. For the realistic, the first order
radio model in [11,12] is used as a communication model between sensor nodes.
Equation (9) and (10) show the amount of energy consumption in transmitting
and receiving a packet with l bits over d distance according to the first order radio
model. Eelec is the amount of energy consumption per bit to run the transmitter
or receiver circuitry. Ef s and Emp are the amount of energy per bit dissipated
in the RF amplifier. The simulation environments are described in Table 1. For
 A Light-Weighted Misused Key Detection in WSNs 365

simplistic, we have assumed that u and v always have a message to send and
transmit to each other by turns.

Ef s
l × (Eelec + Ef s × d2 ) , if d ≤ Emp
Etrans = { (9)
Ef s
l × (Eelec + Emp × d4 ) , if d > Emp

Erecive = l × Eelec (10)

Fig. 5. Energy consumption of a sensor node

Figure 5 shows that the cumulative energy consumption as the number of

exchanged message is increased. The energy efficiency of our scheme compared
to the Liu and Dong’s scheme is shown in Table 2. As a result, our scheme is
more energy efficient than the Liu and Dong’s scheme.

Table 2. Average Energy Efficiency

Probability p Energy Saved

p = 0.05 29%
p = 0.1 44%
p = 0.2 35%
Average 36%

5 Conclusion and Future Works

In this paper, we have presented an efficient method for detecting a misused key
in the WSNs. We have applied our four ideas into the Liu and Dong’s scheme.
366 Y.-J. Han et al.

Our scheme has better two properties, which are the reduction of the detecting
overhead and the high detecting probability, than Liu and Dong’s scheme by
performing the bidirectional misused key detection, and removing of unnecessary
message and using the reasonable probability method for the balanced misused
detection among the sensor nodes. We can finally achieve an efficient mechanism
with a high detecting probability and low energy consumption. On the average,
our scheme is 36% more energy efficient than the Liu and Dong’s scheme and
the detecting probability of our scheme is higher than that of Liu and Dong’s
under the same reporting probability. The analysis shows that our scheme is
good enough for a misused key detection in the WSNs.
A further direction of this study will be applying our scheme into the existing
sensor networks with some specific routing protocols, key management mecha-
nisms, applications and so on. By applying, we will analyze the effectiveness of
our scheme.

References

1. Cerpa, A., Elson, J., Estrin, D., Girod, L., Hamilton, M., Zhao, J.: Habitat Monitor-
ing: Application Driver for Wireless Communications Technology. In: Proceedings
of the ACM SIGCOMMWorkshop on Data Communications in Latin America and
the Caribbean, San Jose, Costa Rica (2001)
2. Shen, C., Srisathapornphat, C., Jaikaeo, C.: Sensor Information Networking Ar-
chitecture and applications. In: IEEE Personal Communications, August 2001, pp.
52–59 (2001)
3. Liu, D., Ning, P.: Establishing pairwise keys in distributed sensor networks. In: Pro-
ceedings of the 10th ACM conference on Computer and communications security,
pp. 52–61. ACM Press, New York (2003)
4. Liu, D., Ning, P., Du, W.: Group-Based Key Pre-Distribution in Wireless Sensor
Networks. In: Proc. 2005 ACM Wksp. Wireless Security (WiSe 2005), September
2005, pp. 11–20 (2005)
5. Lui, D., Dong, Q.: Detecting Misused Keys in Wireless Sensor Networks. In: Perfor-
mance, Computing, and Communications Conference (IPCCC 2007), April 2007,
pp. 272–280 (2007)
6. Chan, H., Perrig, A., Song, D.: Random Key Predistribution Schemes for Sensor
Networks. In: Proc. IEEE Sec. and Privacy Symp., pp. 197–213 (2003)
7. Akyildiz, I.F., Su, W., Sankasubramaniam, Y., Cayirci, E.: Wireless Sensor Net-
works: A Survey. Computer Networks, 393–422 (2002)
8. Eschenauer, L., Gligor, V.: A Key Management Scheme for Distributed Sensor
Networks. In: Proc. 9th ACM Conf. Comp. and Commun. Sec., November 2002,
pp. 41–47 (2002)
9. Eltoweissy, M., Moharrum, M., Mukkamala, R.: Dynamic Key Management in
Sensor Networks. IEEE Communications Magazine 44, 122–130 (2006)
10. Camtepe, S.A., Yener, B.: Key distribution Mechanisms for Wireless Sensor Net-
works: A Survey, Technical Report TR-05-07, Rensselaer Polytechnic Institute
(March 2005)
11. Mhatre, V., Rosenberg, C.: Design Guidelines for Wireless Sensor Networks: Com-
munication, Clustering and Aggregation. Ad. Hoc. Networks, 45–63 (2004)
 A Light-Weighted Misused Key Detection in WSNs 367

12. Heinzelman, W.B., Chandrakasan, A.P., Balakrishnan, H.: An application-Specific

Protocol Architecture for Wireless Microsensor Networks. IEEE Transaction on
Wireless Communications 1, 660–670 (2002)
13. Du, W., Deng, J., Han, Y.S., Chen, S., Varshney, P.K.: A Key Management Scheme
for Wireless Sensor Networks Using Deployment Knowledge. In: IEEE INFOCOM
2004, vol. 1, pp. 7–11 (March 2004)
14. Xiao, Y., Rayi, V.K., Sun, B., Du, X., Hu, F., Galloway, M.: A survey of key man-
agement schemes in wireless sensor networks. Computer Communications Special
issue on security on wireless ad hoc and sensor networks 30(11-12), 2314–2341
(2007)
 Identifying Mudslide Area and Obtaining Forewarned
Time Using AMI Associated Sensor Network

Cheng-Jen Tang and Miau Ru Dai

The Graduate Institute in Communication Engineering

Tatung University, Taipei, Taiwan
[email protected], [email protected]

Abstract. An accountable disaster prediction and the appropriate forewarned

time are the key issues to reduce the possible damages. Around the globe, land-
slides and mudslides are serious geological hazards affecting people, and cause
significant damages every year. The stability of a slope changed from a stable
to an unstable condition that spawns a landslide or mudslide. In most of mud-
slide-damaged residences, the electricity equipments, especially electricity
poles, are usually tilted or moved. Since the location and status of each electric-
ity pole are usually recorded in AMI (Advanced Metering Infrastructure)
MDMS (Meter Data Management System), AMI communication network is a
substantial candidate for constructing the mudslide detection network. To iden-
tify the possible mudslide areas from the numerous gathered data, this paper
proposes a data analysis method that indicates the severity and a mechanism for
detecting the movement. According the detecting result and the gathered data,
this study calculates the remaining forewarned time when an anomaly happens.

Keywords: Advanced Metering Infrastructure, Mudslide Detection, Sensor

Network, Meter Data Management System, Data Analysis, Disaster Prevention
System.

1 Introduction
Typhoon has a large low-pressure center and numerous thunderstorms that produce
strong winds and heavy rains. The strong winds and heavy rains usually cause land-
slides and mudslides. In Taiwan, there are sometimes over a dozen of typhoons vis-
ited in summer. In fact, the name Typhoon means “the deadly storm from Taiwan.”
On August 8th, 2009, Typhoon Morakot hit southern Taiwan. It is the deadliest ty-
phoon that sweeps Taiwan in recorded history. The storm brought a record-setting
rainfall, nearly 3000mm (almost 10 feet) rainfall accumulated in 72 hours. The rain-
fall spawned mudslide made a devastating damage to several villages and buried
hundreds of lives. There was not any warming mechanism worked when the mudslide
embodied as an advent devil. Threatened by such nature disaster, how good a disaster
warming system performs decides life and death.
A sensor system is one of the key tools to predict a disaster. Several previous stud-
ies proposed some detection techniques. One of the techniques uses the operating
theorem of a mechanical mouse as basic. Another uses the image processing to get the

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 368–379, 2010.
© Springer-Verlag Berlin Heidelberg 2010
 Identifying Mudslide Area and Obtaining Forewarned Time 369

differences between two images. A key issue of a disaster detection system is the
communication network. With the rapidly emerging Advanced Metering Infrastruc-
ture (AMI), AMI communication network is a substantial candidate for constructing
the mudslide detection network. AMI Meter Data Management System (MDMS)
records since the location and status of each electricity pole. With such data, a mud-
slide warming system performs the detection and prediction processes. This study
also quantifies the degree of possible damage as the severity indicator. The mudslide
warming system triggers different alarms according to the given severity indicator. To
obtain a forewarned time, AMI MDMS needs to record the distances information for
each electricity pole and its related residence areas. This study then uses the distance
information to calculate the remaining forewarned time when an anomaly happens.

2 Background

Previous studies developed many detection and identification techniques for the land-
slides and mudslides, including Image enhancement [1, 2, 3], Image differencing [4,
5, 6], Vegetation index differencing [7], Image classification [3, 8] and Image regis-
tration [9]. Image enhancement emphasizes the landslides or mudslides in the satel-
lites images. This technique requires human experience and knowledge of the ob-
served areas for visual interpretation. Image differencing is a straightforward method
and easy to interpret the results. However, this method does not provide a detail
changed matrix. The selection of thresholds affects the correctness of the result.
Vegetation index differencing emphasizes the differences in spectral response. This
method reduces the impacts of topographic effects and illumination. However, this
technique also enhances the random noise or coherent noise. Image classification
minimizes the impact of atmospheric, sensor and environmental differences between
multi-temporal images. However, this method requires a sufficient set of the training
samples for classification. Image registration detects landslide/mudslide movements
with sub-pixel accuracy. The main concern is its requirements of the high computa-
tional cost in terms of CPU time. A simple method of change detection and identifica-
tion uses a local similarity measure based on the mutual information and the image
thresholding [11]. This method uses the mutual information to measure the similarity.
It then detects the landslides and mudslides from different images. However, this
method is not suitable for detecting a small area landslide. This method fails, if the
size of the used imagery is too large in compared with the landslide areas. Table 1
shows the ratio of the approximate size of the landslide over the size of the imagery.
Some studies use the wireless sensor networks (WSN) as the communication infra-
structure of their disaster detection system, such as: Slip Surface Localization in
Wireless Sensor Networks [10]. This method uses a sensor network consisting of a lot
sensor columns. This detection system deploys the sensor columns on hills to find the
early signals preceding a mudslide. This sensor network consists of a collection of
sensor columns placed inside the vertical holes drilled during the network deployment
phase. Installers arrange these sensor columns on a semi-regular grid over the moni-
tored area. The size of the grid depends on the characteristics of the site where the
370 C.-J. Tang and M.R. Dai

Table 1. Comparison on the ratio of the approximation of the landslide area with size of
imagery used. (Reprinted from [11])

sensors are deployed. Each sensor column has two components: the sensing compo-
nent that is buried underground and contains all the instruments, and the computing
component that stays above ground and contains the processor and radio module.

(a) (b)
Fig. 1. (a) Landslide Warning System (b) Sensor Column (Reprinted from [10])

The major problem of this technique is the deployment cost. To obtain a “good”
prediction, the density of the sensors must be high enough. In other words, if there is a
broad area requiring monitoring, the investment must be sky-high. Ironically, the
mudslide usually happens in an area that draws little economical interests. Therefore,
to protect people living in these areas, the detection devices must be inexpensive.
Furthermore, the mounting of these devices must be associated with some public
utilities system to reach where people are living.

3 System Architecture
A disaster warning system conducts the environment sensing, data analysis, and
communication. A database stores the periodically collected sensor data. The pro-
posed system analyses the gathered sensor data from the database that is a part of
AMI system. This disaster warning system uses the tilting degree of an electric pole
to define the severity of the location that the pole is set. The system stores every
 Identifying Mudslide Area and Obtaining Forewarned Time 371

analyzed data in the same database. Using a specialized suffix-tree algorithm, the
warning system sends the forewarning signals according the information. Both data
gathering and message broadcasting use AMI-associated communication network.

3.1 AMI-Associated Sensor Network

Recently, many advanced areas such as Europe, America, and Japan etc. commit
themselves to develop the Smart Grid, SG. SG is an integration of electricity usage
monitoring, generation and distribution automation, meter data management and effi-
ciency improving. SG achieves the energy saving and carbon reduction. SG has func-
tions of energy transmission, energy management and two-way communication,
which involve energy generation, transmission, distribution and the customer site.
Advanced Metering Infrastructure (AMI) is regarded as the fundamental and key
technology of enabling SG.
AMI has a communication network which consists of many advanced metering and
sensing devices, including Smart meters. Smart meters provide interval usage (at least
hourly) and collect at least daily. According to the analysis from Gartner, AMI system
consists of following characteristics: data acquisition, data transform, data cleansing,
data processing, information storage/persistency.

Table 2. AMI Process step and involved technologies [SOURCE: GARTNER]

Process Step Involved Technologies

1. data acquisition Meter device
2. data transform Broadband over Power Lines(BPL),
Wireless, RF Satellite
3. data cleansing validation Editing Estimation (VEE) Tools,
Meter Data Management System(MDMS)
4. data processing MDMS
5. information storage/persistency MDMS
6. information delivery/presentment Portals, Web Services, Electronic Data
Interchange (EDI), MDMS

AMI architecture is shown in Figure 2. The blue lines represent the power lines that
connect the transforming stations, utility devices, and customer sites. Each customer
site has a meter, a meter interface unit (MIU), and a communication module. The in-
formation of power consumption is sent to the data collector through the communica-
tion module. The data collectors also use the same communication channel to send
back messages. Each smart meter is connected to an energy management device. The
smart meter communicates with the control center to manage the energy consumption
at its location. The communication lines at the customer sites are presented in green.
For storing the collected data, there is a meter data management system (MDMS)
that communicates with the customer information portal, the customer information
management system, the dispatch automation system (DAS), and the data collector.
The status of each device within an AMI installation is sent to the data collector,
and then forward to the MDMS.
372 C.-J. Tang and M.R. Dai

This study proposes an inexpensive detection device that is mounted on each elec-
tricity pole. The movement or tilting of an electricity pole triggers the mounted detec-
tion device to send a message back to the data collector through the AMI communica-
tion network. This displacement message then forwards to the MDMS that activates
the mudslide analysis module for further calculations.

Fig. 2. Advanced Metering Infrastructure

3.2 Detection Method

In most of mudslide-damaged residences, the electricity equipments, especially elec-

tricity poles, are usually tilted or moved. In order to obtain the displacement, a
movement detector attaches to a pole. A simple method is to install a camera in a
pole. Each camera photographs the neighbor pole. Comparing different imagery
through image processing techniques identifies whether a movement has happened.

(a) (b)

Fig. 3. (a) Camera photographs neighbor one; (b) Camera adding the laser light

The weather affects the performance of this simple detection method. For example,
the heavy fog hampers the image identification. The unclear images increase the de-
tecting error. Adding a laser light projector in the upside of camera enhances this
situation. Laser light is usually spatially coherent, which means that the light either is
emitted in a narrow or low-divergence beam. Coherent light typically means the
source produces light waves that are in step.
 Identifying Mudslide Area and Obtaining Forewarned Time 373

Another detection method, instead of using an expensive sensor device, the pro-
posed detection technique just uses a portion of a mechanical mouse as the detection
device. Everybody knows that when a mouse is moved, the ball inside rolls. This
motion turns two axles, and each axle spins a slotted wheel. A light-emitting diode
(LED) sends a path of light on one side of each wheel through the slots to a receiving
photo-transistor on the other side. The pattern of light to dark is translated to an elec-
trical signal, which reports the position and the moving speed of the mouse. In a
word, the operation of a mechanical mouse has five steps as the following: i) moving
the mouse turns the ball. ii) X and Y rollers grasp the ball and transfer movement. iii)
Optical encoding disks include light holes. iv) Infrared LEDs shine through the disks.
v) Sensors gather light pulses to convert to X and Y velocities.
To make this device, one needs to attach a small piece of iron or other material that
makes one side of the ball always facing down, and then mounts this “tilt-aware
mouse” on an electricity pole. If the pole is moved or tilted, the ball rolls and the
displacement messages are sent.
According how the rollers turns, the tilting direction is obtained. If the RollerA
turned to the left, the ball is moving forward, as shown in the Figure 4.

Fig. 4. Detection of the Tilted Direction

374 C.-J. Tang and M.R. Dai

To obtain the tilted angle, the ball separates into x-z plane and y-z plane. From the
x-z plane, the radian represents the angle of turning left/right, and the vector is OD .
From the y-z plane, the radian represents the angle of turning forward or backward,
and the vector is OC . The two vectors OD and OC construct a plane ODCP with the
vector sum OP . The angle between OP and z-axis is the tilted angle of the electricity
pole.

Fig. 5. Detection of the Tilted Angle

4 Data Analysis
The displacement information of the tilted electricity pole is sent to an AMI data
collector periodically. The data collector then forwards the information to the MDMS.
MDMS then inserts a displacement record in the Pole_Displacement table. A dis-
placement record has the following fields: pole id, time, tilted direction, tilted angle,
difference of tilted angle, checking bit, and the tilted severity of an electricity pole.
{ t1 , t2 , t3 ,..., tn } denotes the set of recording time of each displacement. an denotes the
tilted angle. The differences between the continued tilted angles are obtained by
an − an −1 = d n and an −1 − an − 2 = d n −1 . d k denotes a difference between two continuous tilted
 Identifying Mudslide Area and Obtaining Forewarned Time 375

angles of a pole. When d k is not zero, the checking bit ck is set and the tilted severity
is also calculated according to d k . If d k ≤ 2m , the tilted severity sk is m. If ck-1 is set, ck
is also set. If ck is set and dk is zero, sk is set to sk-1.
The pairs of checking bit and tilted severity of all electricity poles constructs the
alphabets in the proposed mudslide analysis system. The neighboring poles are as-
signed with adjacent numbers. The pair ci , si of time ti of all numbered electricity
poles constructs the Si according to the pole numbers. The mudslide analysis system
then builds a suffix tree STi of two adjacent Si and Si+1. The longest repeating pattern
in the STi of highest tilted severity indicates the most dangerous area that needs to put
most attentions.
Suffix tree is one of the data structure which uses to process a stream data. For ex-
ample, there is a string “mississippi”.

Fig. 6. The suffix tree of “mississippi”

The main reason of using Suffix Tree as the analysis model is for its well-known
computation complexity of Ο(m) .
Several substrings have the same suffix such as “i”, “m”, “p”, and “s”. The sub-
strings use ‘is’ as the suffix that also use ‘issi’ as the suffix. In this situation, these
substrings share a node in the suffix tree. In the proposed mudslide analysis, this
study assumes that the repeating patterns with ci been set indicate where mudslide
will probably happen.
A table Pole_Displacement in the MDMS stores the status of all poles. The status
of a pole includes its id, time, tilt direction, tilt angle, the difference of tile angle,
checking bit, and the degree.
Field “pole_id” is a unique value that indicates an electricity pole in the system.
Field “time” uses the UNIX timestamp format, and denotes the time to store data.
Figure 4 shows the detection of tilted direction. Through the detection of two rollers
A and B, a detector finds the tilted direction. The roller A stays at X-axle for detecting
376 C.-J. Tang and M.R. Dai

Fig. 7. The fields of table Pole_Displacement

a back-forth movement. The roller B stays at Y-axle for detecting the left-right
movement. Field “t_dir” is a 2-bit value that is for recording the back-forth, and left-
right movement. Field “t_angle” stores the tilted angle. Field “diff_angle” records the
differences of “t_angle”. The field “checkbit” denotes whether a pole is tilted or not.
A tilted pole marks “1”, otherwise “0”. Field “degree” denotes the severity degree of
a pole. diff _ angle ≤ 2 m , and the value m is the degree of severity. An example is
shown in Figure 7.

Fig. 8. The fields of table Distance_Ri_Di

The MDMS stores the distances between each pole and the related residences to
obtain a forewarning time. Table Distance_Ri_Di stores the distances. There are two
fields in this table. Field “pole_id” is the unique id of a pole. The other one is
“Pairs_of_Ri_and_Di”. A pair (Ri, Di) denotes the distance Di between a pole Ek and a
residence Ri. An example is shown in Figure 8.

4.1 Forewarning Time Calculation

Obtaining enough forewarning time is the main purpose of a disaster prevention

system. The database of the proposed system stores the possible titling direction of
each electricity pole. With these tilting directions, the database stores the affecting
scope of each pole. The system gives each residential area a unique identification.
 Identifying Mudslide Area and Obtaining Forewarned Time 377

Fig. 9. The tilted directions of poles

Each affecting scope of an electricity pole than contains a set of residence ID. This
system associates each electricity pole and every residence ID of this pole with an
offset that indicates the distance between the residence area and the pole.
{E1 , E 2 ,..., E k } denotes all poles. {R1 , R2 ,..., Ri } denotes the residence IDs of an
electricity pole k. Dik denotes the distance between Rik and Eik .Assume the probability
of landslide is P(x ) .

⎧1 , landslide happened
P( x) = ⎨
⎩< 1 , otherwise
At any given time after Ek has tilted, the first recorded time is t1, and t0 is subtracting
Δt from t1. Δt is the period time for retrieving data. The tilted displacement for pole Ek
378 C.-J. Tang and M.R. Dai

is m that is hk sin Δθ . Therefore, the tilted velocity for pole Ek is Vk. t 0k is the time
that pole hasn’t tilted for the pole Ek.
ck × hk sin Δθ
Vk = (1)
tnow − t0 k

Fig. 10. The tilted displacement of the pole Ek

The forewarned time of the pole Ek alerting to the residence Ri denotes Ri,AlertTime.
Di
Ri , AlertTime = (2)
ck × hk sin Δθ tnow − t0 k

5 Conclusion
The mudslides and landslides have consumed many lives. AMI communication net-
work is able to reach every residence that makes a substantial candidate for construct-
ing a mudslide detection network. The location and status of each electricity pole are
recorded in AMI MDMS. With an extra movement detector attached to each pole, an
AMI-associated sensor network is able to identify the possible mudslide areas from
the numerous gathered data. To obtain a “good” prediction, the density of the sensors
must be high enough. In other words, if there is a broad area requiring monitoring, the
investment must be sky-high. Ironically, the mudslide usually happens in an area that
draws little economical interests. Therefore, to protect people living in these areas, the
detection devices must be inexpensive. Furthermore, the mounting of these devices
must be associated with the public utilities system to reach where people are living.
Therefore, this paper proposes an inexpensive detection device associated with AMI
network, and a mudslide analysis method that indicates the severity and urgency of a
mudslide.
 Identifying Mudslide Area and Obtaining Forewarned Time 379

This proposed device and method is still a concept rather than a realization. The fu-
ture work is to install such system in an AMI pilot installation to prove this concept.

References
1. Whitworth, M.C.Z., Giles, D.P., Murphy, W.: Airbone remote sensing for landslide hazard
assessment: a case study on the Jurassic escarpment slopes of Worcestershire, UK. The
Quarterly Journal of Engineering Geology and Hydrogeology 38(2), 197–213 (2005)
2. Ostir, K., Veljanovski, T., Podobnikar, T., Stancic, Z.: Application of satellite remote sens-
ing in natural hazard management: the Mount Mangart landslide case study. International
Journal of Remote Sensing 24(20), 3983–4002 (2003)
3. Nichol, J., Wong, M.S.: Satellite remote sensing for detailed landslide inventories using
change detection and image fusion. International Journal of Remote Sensing 9, 1913–1926
(2005)
4. Cheng, K.S., Wei, C., Chang, S.C.: Locating landslides using multi-temporal satellite im-
ages. Advances in Space Research 33, 296–301 (2004)
5. Hervas, J., Barredo, J.I., Rosin, P.L., Pasuto, A., Mantovani, F., Silvano, S.: Monitoring
landslides from optical remotely sensed imagery: the case history of Tessina landslide, It-
aly. Geomorphology 1346, 1–13 (2003)
6. Rosin, P.L., Hervas, J.: Remote sensing image thresholding methods for determining land-
slide activity. International Journal of Remote Sensing 26, 1075–1092 (2005)
7. Lin, W.T., Chou, W.C., Lin, C.Y., Huang, P.H., Tsai, J.S.: Vegetation recovery monitoring
and assessment at landslides caused by earthquake in Central Taiwan. Forest Ecology and
Management 210, 55–66 (2005)
8. Nichol, J., Wong, M.S.: Detection and interpretation of landslides using satellite images.
Land Degradation and Development 16, 243–255 (2005)
9. Yamaguchi, Y., Tanaka, S., Odajima, T., Kamai, T., Tsuchida, S.: Detection of a landslide
movement as geometric misregistration in image matching of SPOT HRV data of two dif-
ferent dates. Int. J. Remote Sensing, preview article:1 12 (2002)
10. Terzis, A., Anandarajah, A., Moore, K., Wang, I.-J.: Slip Surface Localization in Wireless
Sensor Networks for Landslide Prediction. In: IPSN 2006, Nashville, Tennessee, USA,
April 19-21 (2006)
11. Khairunniza-Bejo, S., Petrou, M., Ganas, A.: Landslide Detection Using a Local Similarity
Measure. In: Proceedings of the 7th Nordic Signal Processing Symposium, NORSIG 2006
(2006)
12. Mouse (computing) Wikipedia. Wikipedia (September 24, 2009),
http://en.wikipedia.org/wiki/Mechanical_mouse#Mechanical_mice
 Utilization of Ontology in Health for Archetypes
Constraint Enforcement

Anny Kartika Sari1,2, Wenny Rahayu1, and Dennis Wollersheim1

1
Department of Computer Science and Computer Engineering,
La Trobe University, Victoria 3086, Australia
[email protected],
{w.rahayu,d.wollersheim}@latrobe.edu.au
2
Department of Computer Science, Gadjah Mada University,
Yogyakarta 55281, Indonesia

Abstract. Most existing works in ontology deployment within the health indus-
try are mainly focusing on the standardization and interoperability goals. In this
paper, we propose the utilization of an ontology to apply a new constraint in
health archetypes, i.e. the slot filling constraint. An archetype is a model that
represents functional health concept such as admission record. It can reuse
other existing archetypes through a slot. The name of a slot represents a more
specific health concept such as head. The slot filling constraint restricts the se-
lection of archetypes to fill in that specific slot so that only relevant archetypes
are chosen from the available ones. Ontology is used to enforce this constraint.
An approach on how to apply the constraint is presented based on the semantic
similarity/relevance concept. The evaluation shows that the approach is a better
alternative to the current slot filling process which depends on manual decision
by the archetype author.

Keywords: ontology, health, archetype, slot filling constraint, semantic

relevance.

1 Introduction
Ontologies have been used widely in many areas today. It is primarily employed in
the applications where formal specifications of concepts are needed. One of the fields
that utilize ontology is health. There are even some ontologies related to health which
have been defined and become standards such as SNOMED CT (Systematized No-
menclature of Medicine-Clinical Terms)1 and LOINC (Logical Observation Identifi-
ers Names and Codes)2. However, the main use of ontologies in health so far is still
limited to the effort of achieving the uniform semanticity of health terms, in which the
clinical ontologies become standards. Another common aim of ontology usage in
health is to achieve interoperability between the different frameworks of electronic
health records (EHR), which are presented in e.g. [1], [2], and [3].
1
See: http://www.ihtsdo.org/
2
See: http://loinc.org/

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 380–394, 2010.
© Springer-Verlag Berlin Heidelberg 2010
 Utilization of Ontology in Health for Archetypes Constraint Enforcement 381

This paper will present another kind of ontology employment in health, specifically
in the area related to EHR. Our focus is the EHR architecture proposed by openEHR3,
in which archetype lays as the basis of the model. The archetype concept was first
mentioned as part of the Australian project GEHR (Good Electronic Health Record)
[4]. The development of the archetype concept is mainly motivated by the inability of
the existing medical information system to keep up with the dynamic changing of
health knowledge. In the existing framework, which is commonly referred as “single-
level” methodologies, both informational and knowledge concepts are built into one
level of object and data models [5]. This methodology will become obsolete soon
after there is knowledge updating, and then should be replaced. This produces high
maintenance cost. The archetype concept is proposed to cope with the shortcoming of
the existing medical information model. The key feature of an archetype-based sys-
tem is that it is made up of a two-level architecture: a RIM (Reference Information
Model), consisting of a set of reference types, and a domain or knowledge model,
consisting of a set of archetypes [4]. The reference model is the repository of the
patient clinical data. As clinical knowledge is stored completely separately in
archetypes, resulting EHR systems are more flexible as changes in the clinical knowl-
edge can be embraced by modifications in archetypes, without compromising the
integrity of information in the repository [5]. An archetype itself formally expresses a
distinct health concept such as blood pressure. The concept is expressed in the form
of constraints on data whose instances conform to some RIM [6].
In this work, we are proposing a new archetype constraint, named as the archetype
slot filling constraint, which is defined as the complement of the existing constraints
specified in the archetype specifications. The proposal of the new constraint is moti-
vated by an example of the data inconsistency on the absence of this constraint. With
the enforcement of the constraint, we believe that new archetypes can be defined
more accurately and semantic inconsistency can be avoided. At the end, this will
improve the quality of the whole clinical systems in which archetypes are utilized.
As the constraint is not specified in the archetype specifications, a method will also
be proposed to make the constraint feasible to be applied without influencing the
internal structure of archetypes. Ontology, in this case clinical ontology, will be em-
ployed for the application of the approach. The method also uses the concept of se-
mantic similarity/relevance of an ontology. Of course, the concept needs to be ad-
justed to fit in the clinical ontology to be used.
Actually, the concept of ontology has been used in an archetype as one of its sec-
tions, but the usage is limited to terminologies (clinical ontology) linkings and bind-
ings. The binding is aimed at the uniformity in the meaning of terms contained in the
archetype, even though there is no special mechanism to check the uniformity. The
ontology is simply viewed as collection of terms, not as a ‘real’ ontology. This is
different to the approach in this work in which ontology will be utilized as the integra-
tion of all its concepts, relationships, and attributes.
The rest of the paper will be organized as follows. Section 2 will present the moti-
vating example on the proposal of the new constraint in archetype concept. Related
works is presented in Section 3. The approach to enforce the application of the

3
See: http://www.openehr.org/
382 A.K. Sari, W. Rahayu, and D. Wollersheim

constraint, as well as the related algorithms, is then explained in Section 4. The ap-
proach is evaluated in Section 5. Section 6 closes this paper with some conclusion and
future woks.

2 Motivating Example and the Slot Filling Constraint

According to [7], the term archetype is used to denote knowledge level models which
define valid information structures. The domain concepts defined in the models are
expressed using constraints. Constraints are used to govern the variability that may
occur in the concepts definition. This idea leads to the definition of archetype as a
model defining some domain concept, expressed using constraints on instance struc-
tures of an underlying reference model [7]. From this definition, it can be inferred that
basically archetype contains constraints. Some important built-in constraints are re-
lated to the occurrence of an object node, the existence of an attribute, the cardinality
and ordering of a container attribute, and the containment of an archetype inside an-
other archetype. The syntax used to represent the constraints is cADL (constraint
ADL), while dADL (data ADL) syntax is deployed to express instance data.
In the archetype definition using cADL, constraint in an archetype that is being de-
veloped can be specified by reusing the existing archetypes. This mechanism is
known as an archetype ‘slot’ or ‘chaining point’ [8]. Basically, a slot is the means for
an archetype to link to another archetype and reuse the entire constraints defined in
the linked archetype. The definition of an archetype slot uses the keyword
allow_archetype and utilizes two types of assertion statement: include and
exclude. The assertion statements are the constraints in the slot filling. Archetypes
contained in the include statement can be chosen as the slot fillers, while those
contained in the exclude statement must not be chosen as the slot fillers. The arche-
type slots are available in two of the archetype classes, i.e. the Composition arche-
types and Section archetypes. The classes of archetype that can fill in the slots
are Section, Entry, Observation, Evaluation, Action, and Administration Entry
archetypes.
The wrong choice of the slot fillers may raise semantic inconsistency. An
example of the problem is described here. Consider the openEHR archetype
openEHR.EHR.SECTION-admission_record.v1. Fig. 1. shows the view of that arche-
type as displayed by the Archetype Editor tool. The archetype contains Physical Ex-
amination sub section which has Regional sub section. One of the sub sections of
Regional is Head. Head sub section includes Observation archetypes as its slot. In this
slot, the inclusion and exclusion assertion are not defined. Thus, all archetypes of
Observation class can be used to fill in the archetype slot. However, some archetypes
like openEHR.EHR.OBSERVATION-autopsy.v1, openEHR.EHR. OBSERVATION-
menstruation. v1, and openEHR.EHR. OBSERVATION-urinalysis.v1 are not relevant
to the sub section since they are not semantically related to the term ‘head’. Since
there is no semantic relation between the name of the sub section (slot) with some of
the slot fillers, semantic inconsistency happens in this situation. At the end, this can
lead to the semantic inconsistency of in the data level which is captured using the
archetype.
 Utilization of Ontology in Health for Archetypes Constraint Enforcement 383

Fig. 1. The openEHR.EHR.SECTION-admission_record.v1 and its slot in the Head sub-section

The semantic violation problem occurs because the user, i.e. the archetype author
should fill in the archetype slots without the guidance on which archetypes can be
chosen and which ones should be excluded due to their irrelevancy to the sub section
name. In this example, user should inspect the list of the fifty Observation archetype
names and decide by himself which names can fill in the slots. Since archetype con-
cept is expected to be used widely in the future, and thus, more and more new arche-
types will be created, it will be more complicated for the user to choose the relevant
archetypes when there will be many more than fifty Observation archetypes.
This problem could be prevented if there is a constraint in filling in the slot for ar-
chetypes. This constraint should determine the dependency of the archetype slot fill-
ing to the sub section it is related to or to the archetype itself. We refer this constraint
as archetype slot filling constraint. A method that utilizes ontology can be applied to
enable the slot filling constraint. This method will provide the user with the relevant
archetypes to be filled in a particular slot. The relevancy is measured based on a par-
ticular clinical ontology. However, since the archetype should neutral with respect to
terminologies [6], the utilization of terminology/ontology should be conducted out-
side the archetype definition. The detail of the method is discussed in Section 4.

3 Related Works
There are not many works which utilize ontologies for the works related to arche-
types. However, to the best of our knowledge, there is no existing paper which
discusses the employment of ontology for the validation of archetype constraints.
384 A.K. Sari, W. Rahayu, and D. Wollersheim

Rather, most of the works use ontology as a mediator to achieve semantic interop-
erability between the different frameworks of EHR. For example, [1] shows that the
interoperability between archetypes based web service messages could be accom-
plished by using OWL (Web Ontology Language) as the representation of both ar-
chetype definitions and instances to semantically annotate the web service messages.
Similarly, in [2] one of clinical ontologies i.e. SNOMED CT is used as the standard
into which archetypes and messages are bound. The aim is to allow different systems
to interoperate. And finally, Fernandez-Breis, et al., in [3], presents an ontological
approach to promote interoperability among CEN(European Committee for Stan-
dardization)/TC 2514 and openEHR compliant information systems by facilitating
the construction of interoperable clinical archetypes. This is achieved by construct-
ing a one to one mapping between the models, utilizing OWL as the ontological
representation.
The concept of semantic relevance is also employed in this paper. It is used in the
approach to apply the new constraint in the archetypes. Some works which are related
to the relevance semantic issue have been done. In most of them, e.g. [9], [10], [11],
[12], [13], and [14], the semantic relevance term is related to, or based on, or even
used interchangeably with the terms semantic similarity and/or semantic distance.
Hence, in this work we will also base the semantic relevance measurement on seman-
tic similarity measurement.
There have been many papers related to semantic similarity. According to [15],
based on the structures and input considered, there are three kinds of similarity meas-
ures: edge-based, information content-based, and feature-based approaches. Edge-
based approach is used in a graph representation of an ontology. This approach is
used in e.g. [16] and [17]. Information content-based method, which is applied in
[18], [19], uses text corpus to evaluate the semantic similarity. The feature-based
technique utilizes the properties of concepts for assessing the similarity and is used in
[20]. Some works, such as [21], [22], and [23], use the combination of those methods
or compare them.
To the best of our knowledge, none of the papers related to semantic similarity dis-
cuss its application to clinical ontology, which is the focus of this paper. Clinical
ontology, in this case SNOMED CT, has some special characteristics. First, the mean-
ing of attribute is not the same as attribute in common ontologies. In SNOMED,
attribute is a kind of relationship which relates a concept to another. This different
meaning should be taken into account when feature-based similarity techniques are
used, as they commonly base the measurement on attribute. The next characteristic is
that in SNOMED, the hierarchy is based only on the IS-A relationships. Other types
of relationships are not considered. This will influence the calculation when the edge-
based similarity approach, which is based on all types of relationships, is used. Thus,
there should be adjustment when the approach is applied for SNOMED CT clinical
ontology.

4
CEN/TC 251 is standard for health informatics. The complete published standards, see:
http://www.cen.eu/CENORM/Sectors/TechnicalCommitteesWorkshops/CENTechnicalCom
mittees/CENTechnicalCommittees.asp?param=6232&title=CEN/TC%20251
 Utilization of Ontology in Health for Archetypes Constraint Enforcement 385

4 The Approach for the Slot Filling Constraint Application

This section describes the approach to enforce the slot filling constraint application in
the development of new archetypes. The following sub sections explain the approach
in detail.

4.1 Description of the Method

As previously mentioned, we propose the ontology based approach to apply the slot
filling constraint to archetype. We believe that ontology is suitable to be utilized to
force this constraint for the following reason. The constraint is much related to the
semantic data inconsistency. As ontology is a formal specification, it can maintain
semantic consistency of the concepts it specifies. As the result, the problem of data
inconsistency can be avoided.
In this work, SNOMED CT is chosen as the ontology for the approach. The choice
is based on some reasons. Firstly, SNOMED CT is considered to be the most compre-
hensive, multilingual clinical healthcare ontology in the world [24]. It covers areas
that are closely related to EHR, the main concern of archetypes. Secondly, some
works such as [2] and [25] use SNOMED CT as the ontology for archetype related
works. This shows that SNOMED CT fits well to the concepts defined in archetypes.
Since the slot filling constraint involves only particular classes of archetypes, i.e.
Composition, Section, and Entry, this paper only considers those types of archetypes
as well. For the purpose of clearness of the method used, we categorize those classes
of archetypes into three class layers: the bottom layer, the middle layer, and the top
layer. The bottom layer is the classes of archetypes which can only be the filler of
slots. They are not able to have any slot to be filled in by another archetype. All the
entry classes, i.e. the Instruction class, the Observation class, the Action class, the
Evaluation class, and the Admin Entry class, are included in this layer. The top layer
is the class of archetypes which can only have slots inside them. They themselves
cannot fill in another archetype’s slot. This layer includes the Composition class. The
middle layer comprises the Section class. This layer of archetypes can have slots to be
filled in by other archetypes as well as be the filler of other archetypes. In the rest of
this paper, we will use the layers categorization to refer to the classes of archetypes.
Table 1. summarizes this categorization.

Table 1. The layer categorization of archetype classes used in this work

Name of Layer Included Classes of Archetypes

Bottom Instruction, Observation, Action, Evaluation, Admin Entry
Middle Section
Bottom Composition

The ontology based approach to force the constraint application is explained intui-
tively in this following scenario. An archetype author needs to fill in the slots with the
include and exclude assertion. The archetypes to be included or excluded in the
slot are chosen from some groups of a specific type of archetypes which can be one of
the bottom or middle layer class. The archetype type is chosen by the author. Each
386 A.K. Sari, W. Rahayu, and D. Wollersheim

archetype of the determined type is a candidate for the include assertion. Since
there can be many candidates, it is essential to assure that the chosen archetypes are
indeed relevant to the sub section or the Section/Composition archetype that contains
the slot. For this purpose, ontology is employed. It is used to annotate each archetype
with the concepts it specifies which are related to the archetype. The annotation proc-
ess is performed directly after an archetype of bottom or middle layer is built. The
archetypes of upper layer, i.e. the Composition archetypes, do not need to be anno-
tated because they are never used as slot fillers.
When an archetype is included as an archetype candidate to fill in a slot, its annota-
tion will be used in the filtering process. This process filters the archetypes in the
candidate list which are relevant to the slot. The relevancy is measured by inspecting
the relevance of each annotation of each archetype candidate with the concept in the
ontology which is related to the keyword. The user chooses the keyword from the
name of the sub section where the slot is contained, the names of all the ancestral sub
sections, the name of the archetype, or the combination of them which then become a
set of keywords. Each archetype candidate is then evaluated whether it is considered
relevant or not to fill in the slot. If it is, then the candidate is included in the list of the
relevant archetypes for the slot, otherwise it is excluded. The list is then used to force
the slot filling constraint to the slot filling activity by the user.
From the above description of the method, we determine two technical steps in the
approach: the preliminary process and the filtering process. In this paper, we focus on
the filtering process, thus we will describe the preliminary process briefly, while the
filtering process is described in detail in Section 5.2.
The preliminary process refers to the archetype annotation with respect to the
SNOMED CT ontology. Each archetype is annotated with some terms which repre-
sent the concepts in SNOMED CT which the archetype relates to. The annotation is
applied to any archetype of bottom and middle layer upon its development. Thus, it is
not limited to the archetypes in the candidate list. The process involves two main
steps: the ontology extraction and the archetypes annotation. The ontology extraction
is performed to obtain only the relevant part of the whole ontology, which is called
sub-ontology or ontology view. We reuse the ontology extraction proposed in [26] for
this work. The annotation is used in this work to annotate a specific archetype docu-
ment with the concepts specified in SNOMED CT which are related to its content.
Some annotation tools, such as SMORE [27] and Ontomat [28] can be utilized for the
purpose.

4.2 The Filtering Process

This is the main process of the method. In this process, the relevant archetypes for the
slot to be filled in are determined using a specific evaluation method. First, the arche-
type candidate list is prepared. This is based on the specific type of archetype chosen
by the user, i.e. the archetype author, for the archetype slot to be filled in. then, the
evaluation of relevant archetypes is conducted. Two steps need to be conducted here,
i.e. the determination of the set of keywords and the relevance calculation, which
result in the list of relevant archetypes. For the user, choosing archetypes not included
in the list will violate the archetype slot filling constraint.
 Utilization of Ontology in Health for Archetypes Constraint Enforcement 387

The method for the relevance evaluation will be elaborated. The relevance of a
concept included in the annotation of a specific archetype to the keyword is evaluated
in a simple way. Algorithm 1 below describes the evaluation method.

The relevance evaluation is calculated for each keyword. The keywords are chosen
by the user based on the name of archetypes and the name of sub sections and all its
descendants where the slot is placed. For each keyword, each archetype is evaluated its
relevance to the keyword using its annotations. Each annotation is inspected whether it
is considered relevant to the keyword or not. Once it finds that an annotation of the
388 A.K. Sari, W. Rahayu, and D. Wollersheim

archetype candidate is considered relevant to the keyword based on the ontology, it will
add that archetype to list R, which includes the relevant archetypes for the slot. Thus,
the evaluation of the rest of the annotations is not needed. Moreover, the archetype will
be deleted from list P (list of candidate archetypes). It means that for the relevance
evaluation of the next keyword, it will not be considered anymore.
For evaluating the relevance between the keyword k with concept Pij in the arche-
type annotation, first one or more concepts which have the names similar to k are
searched in SNOMED. Each concept found is then evaluated its relevance to Pij based
on SNOMED CT ontology. Since the basis is an ontology, certainly the word rele-
vance here refers to the semantic relevance, which means the closeness of the rela-
tionship between two terms which in this work refers to the concepts contained in
SNOMED CT ontology. As previously mentioned in Section 3, semantic relevance is
much related to semantic similarity. Thus, some previous works on semantic similar-
ity are reused here with some adjustment.
To measure the semantic relevance between two concepts in SNOMED CT, we
first determine some features which influence the assessment. The factors are based
on the edge-based and information content based approaches of similarity calculation.
They are described below.
1. The distance between the concepts
In edge-based approach, the shortest path between two concepts in the semantic
net is commonly used to evaluate their similarity. The shortest path of concept a
and b is the path where the number of edges in the path is minimum among all
the paths connecting a and b in the nets. Rada et.al. [16] propose a basic semantic
distance measure by only using taxonomic links and computing the shortest path
between the concepts. A variant of the work is [29] in which distance is substi-
tuted by similarity. Thus, we consider the equation is more suitable to our work
and it is reused in this paper. By adopting the equation in [29], we define the
function for semantic relevance as follows:

(1)

where rele(a,b) denotes the edge-based relevance between concept a and concept
b, dist(a, b) represents the number of edges of the shortest path between concept
a and b, and Max is the maximum depth of the taxonomy.
2. The link type connecting the concepts
There are 4 types of relationships in SNOMED CT: defining characteristics (IS-A
relationships and defining attributes), qualifying characteristics (non-defining,
qualifying attributes), historical relationships (relate inactive concepts to active
concepts), and additional relationships (other non-defining characteristics such as
PART OF) [30]. The type of relationship that form the hierarchy of SNOMED
CT is the IS-A relationship. The other types of relationships are not represented
as links in the hierarchy. Two special cases occur here and we consider that the
relevance between two concepts is absolute if at least one of them happens. The
first case is when the two concepts are actually the same concept; the other is
when one of the concepts or one of their descendants becomes an attribute of the
 Utilization of Ontology in Health for Archetypes Constraint Enforcement 389

other concept. The second case is based on the consideration that if the annota-
tion (which is a SNOMED CT concept) of an archetype becomes an attribute of
the keyword (which is also a SNOMED CT concept) of a slot, then the archetype
should be relevant for the slot. It is unlikely that the attribute-based relationships
form a path. Thus, we only consider their existence, not their path length. On the
other hand, the PART-OF relationships often form a path. Thus, the PART-OF
path distance between concept a and b needs to be considered. To differentiate
the IS-A and PART-OF relationships, we define the equations below which are
based on (1):

(2)

(3)

where simIS (a, b) denotes similarity between concepts a and b by using the IS-A
hierarchy, distIS (a,b) is the shortest path between concepts a and b based on the
IS-A hierarchy, MaxIS is the maximum depth of the IS-A taxonomy, simPART (a,
b) represents the similarity between concepts a and b based on the PART-OF re-
lationship, distPART (a,b) is the shortest distance between concepts a and b on the
PART-OF path, and LenIS (a, b) is the maximum length of PART-OF path which
contains a and b.
Then, we need to combine the equations and the consideration of attribute re-
lationship to get the relevance value between concepts a and b, as follows:

(4)

3. The information content of the concepts

The value of the information content of a concept is shown by the probability of
occurrence of this class in a large text corpus [31]. Since we work with the col-
lection of archetypes, we consider that it is more suitable to use the collection of
archetypes as text corpus. As presented in [32], the functions for this factor is de-
fined as follows.

(5)

where relI (a, b) is the information content-based relevance between concepts a

and b, IC(a) = - log P(a) is the information content of concept a and LCS(a,b) is
the least common subsumer of concept a and b. P(a) is the probability of en-
countering an instance of concept a in the corpora.

To get the final relevance equation, we combine (4) and (5). The following is the final
equation.

(6)
390 A.K. Sari, W. Rahayu, and D. Wollersheim

Rel(a, b) is the relevance between concepts a and b, while α1 and α2 are adjustable
factors, α1 + α2 = 1. α1 and α2 can be adjusted to the requirement. For example, when
the information about the text corpus is missing, α2 can be set to 0 to eliminate the
information content-based relevance from the equation.
Finally, equation (6) is connected with Algorithm 1 to determine the value of vari-
able relv. Equation (9) is used for the purpose.

(7)

β is the value which is determined by the user or the archetype author.

5 Evaluation and Discussion

To evaluate the method, we should apply it to the process of filling in some slots.
We choose two types of slots. The first type includes the slots which are contained in
sub section(s). These slots can have more than one keyword to be chosen. On the
other hand, the second type, which is not contained in any sub section, can only have
one keyword, i.e. the archetype’s name. We will try to find out whether the number
of keywords influences the result. Four slots from each type are picked from the
archetypes collection of the Archetype Editor version 1. Not many archetypes have
sub sections; even, we only find the SECTION.admission_record.v1 archetype which
has more than one level of sub section. Thus, we choose four slots from it as the
representation of the first slot type. The slot names are Injury, Sensitivities and aller-
gies, Childhood illness, and History. For the second type, we choose 4 archetypes
which does not has sub section: SECTION.family_history, SECTION.encounter,
SECTION.summary, and SECTION. antenatal_check archetypes. For the filtering
process, we set α1 = 0.95 and α2 = 0.05 as we consider that the number of
archetype collection, as the source of text corpus, is still limited. β is set to be
0.55.
Table 2 shows the comparison on some statistical metrics between the slot filling
result using the method and the current slot filling in which the method is not used.
The precision, recall, accuracy, and F-measure are calculated based on the manual
decision of the slot filling process. As the number of archetypes is still low at
this moment, we believe that manual decision can be used as the valid basis for the
calculation.
From the table, we can see that the precision value of the current slot fillings tend
to be higher than the slot filling result using the method, except for the Encounter and
Antenatal check slots. This can be understood as in the other slots, the percentage of
archetypes to be included is much lower than the actual relevant archetypes. For in-
stance, the current number of included archetypes in the Injuries slot is only 2, com-
pared to 15 actual relevant archetypes for that particular slot. For the Encounter and
Antenatal check slots, all Evaluation archetypes are included to fill in the slots. That
is why the precision is lower than the result on using the method, as not all archetypes
are actually relevant to those slots.
 Utilization of Ontology in Health for Archetypes Constraint Enforcement 391

Table 2. The comparison on some statistical metrics between the result using the method and

Precision Recall Accuracy F-Measure

Number
Slot Names of Using Without Using Without Using Without Using Without
Keywords the the the the the the the the
Method Method Method Method Method Method Method Method
Injury 4 0.77 1.00 1.00 0.13 0.77 0.50 0.83 0.24
Sensitivities
and 4 0.85 1.00 1.00 0.06 0.85 0.38 0.89 0.11
allergies
Childhood
4 0.82 1.00 0.88 0.06 0.81 0.42 0.85 0.12
illness
History 2 1.00 1.00 0.65 0.04 0.69 0.15 0.79 0.08
Family
history 1 0.60 0.80 0.75 0.33 0.65 0.65 0.67 0.47
(archetype)
Encounter
1 0.85 0.23 0.79 1.00 0.81 0.23 0.81 0.38
(archetype)
Summary
1 0.75 0.83 0.60 0.45 0.77 0.73 0.67 0.59
(archetype)
Antenatal
check 1 0.71 0.50 0.83 1.00 0.77 0.50 0.77 0.67
(archetype)

On the other hand, the recall value of the result using the method is generally
higher than the current slot filling. Again, the exception is for two slots, i.e. the En-
counter and the Antenatal check slots, as those two slots included all Evaluation ar-
chetypes. This makes the recall value to be maximum, because the inclusion of all
archetypes results in the inclusion of all relevant archetypes as well.
The accuracy and F-measure metric calculations normalize the opposite results of
the recall and precision metrics. The method generally has higher accuracy than the
current slot filling condition, which means that the method produces the closer results
to the actual relevant archetypes for each specific slot. The result of the F-measure
calculation is also higher. This shows that the method is more effective in retrieving
relevant archetypes for a specific slot than the current slot filling. From the result of
the two metrics, the slot filling process using the method seems to be a better alterna-
tive for the current slot filling condition.
The number of keywords does not seem influencing much on the result. For the
slots with multiple keywords, the accuracy value ranges from 0.69 to 0.85, while
that of the slots with single values ranges from 0.65 to 0.81. It shows that the num-
ber of keywords does not influence the accuracy. A slightly difference measure
happens to the F-measure metric, in which the slots with multiple keywords have a
little bit higher F-measure value. This means that the method is more effective to be
applied to the slots which are contained in a sub section rather than directly in an
archetype.
392 A.K. Sari, W. Rahayu, and D. Wollersheim

6 Conclusion
In this paper, we have presented the new constraint, i.e. the slot filling constraint as
the complement of the existing archetypes’ built-in constraints. The constraint is im-
portant to maintain semantic consistency between the archetype or sub section names
containing the slot and the filling archetypes. An ontology-based approach has been
proposed to validate the constraint on the development of new archetypes. This ap-
proach also shows a different way of utilization of ontology in health from most of the
existing ontology employment, which goal is for interoperability purpose.
The evaluation result shows that the method is a better alternative to the current
slot fillings condition. Based on the manual decision of the slots filling, the precision,
the recall, the accuracy, and the F-measure of the slot filling results using the method
are calculated. Compared to the current slot filling condition, the accuracy and F-
measure of the slot filling using the method is generally better. The evaluation also
shows that the F-measure for the slots contained in a sub section is better than those
contained directly in an archetype. However, the accuracy of the method is not much
different.
While there are two classes of archetypes, i.e. Section and Composition, can have
slots, in the evaluation the method is only applied to some Section archetypes. Further
application to all other archetypes, including the Composition archetypes, can be done
as part of our future works. Another interesting issue is the application of the ap-
proach to another existing clinical ontology, as some archetypes have the binding to
them as well.

Acknowledgement
This work is partially supported by the Ministry of National Education of the Repub-
lic of Indonesia through the scholarship granted to the first author.

References
1. Bicer, V., Kilic, O., Dogac, A., Laleci, G.B.: Archetype Based Semantic Interoperability of
Web Service Message in the Health Care Domain. International Journal on Semantic Web
and Information Systems 1, 1–23 (2005)
2. Qamar, R., Rector, A.: Semantic Issues in Integrating Data from Different Models to
Achieve Data Interoperability. In: MEDINFO 2007, Brisbane (2007)
3. Fernández-Breis, J.T., Menárguez-Tortosa, M., Moner, D., Valencia-García, R.,
Maldonado, J.A., Vivancos-Vicente, P.J., Miranda-Mena, T.G., Martínez-Béjar, R.: An
Ontological Infrastructure for the Semantic Integration of Clinical Archetypes. In: Hoff-
mann, A., Kang, B.-h., Richards, D., Tsumoto, S. (eds.) PKAW 2006. LNCS (LNAI),
vol. 4303, pp. 156–167. Springer, Heidelberg (2006)
4. Wollersheim, D., Sari, A., Rahayu, W.: Archetype-based Electronic Health Records: a Lit-
erature Review and Evaluation of their Applicability to Health Data Interoperability and
Access. Health Information Management Journal 38, 7–17 (2009)
5. Leslie, H., Heard, S.: Archetypes 101. In: Westbrook, J., Callen, J. (eds.) HIC 2006, Syd-
ney (2006)
 Utilization of Ontology in Health for Archetypes Constraint Enforcement 393

6. Beale, T., Heard, S.: Archetypes Definitions and Principles, The openEHR Foundation
(2007)
7. Beale, T.: Archetypes: Constraint-based Domain Models for Future-proof Information
Systems. In: Baclawski, K., Kilov, H. (eds.) The International Conference on Object Ori-
ented Programming, Systems, Languages and Applications 2002, Seattle, Washington, pp.
1–18 (2002)
8. Beale, T., Heard, S.: Archetype Definition Language ADL 1.4, The openEHR Foundation
(2007)
9. Ruotsalo, T., Hyvönen, E.: A Method for Determining Ontology-Based Semantic Rele-
vance. In: Wagner, R., Revell, N., Pernul, G. (eds.) DEXA 2007. LNCS, vol. 4653, pp.
680–688. Springer, Heidelberg (2007)
10. Lombardi, L., Sartori, G.: Concept Similarity: An Abstract Relevance Classes Approach.
In: Fum, D., Missier, F.D., Stocco, A. (eds.) The 7th International Conference on Cogni-
tive Modeling, Trieste, pp. 190–195 (2006)
11. Sartori, G., Lombardi, L.: Semantic Relevance and Semantic Disorders. Journal of Cogni-
tive Neuroscience 16, 439–452 (2004)
12. Zhang, G., Yu, C., Cai, D., Song, Y., Sun, J.: Research on Concept-Sememe Tree and Se-
mantic Relevance Computation. In: The 20th Pacific Asia Conference on Language, In-
formation and Computation, vol. 20, pp. 398–402 (2006)
13. Rhee, S.K., Lee, J., Park, M.-W.: Ontology-based Semantic Relevance Measure. In: The
First International Workshop on Semantic Web and Web 2.0 in Architectural, Product and
Engineering Design, Korea (2007)
14. Ricklefs, M., Blomqvist, E.: Ontology-Based Relevance Assessment: An Evaluation of
Different Semantic Similarity Measures. In: Meersman, R., Tari, Z. (eds.) OTM 2008, Part
II. LNCS, vol. 5332, pp. 1235–1252. Springer, Heidelberg (2008)
15. Raftopoulou, P., Petrakis, E.: Semantic Similarity Measures: a Comparison Study. Techni-
cal University of Crete, Department of Electronic and Computer Engineering (2005)
16. Rada, R., Mili, H., Bicknell, E., Blettner, M.: Development and Application of a Metric on
Semantric Nets. IEEE Transactions on Systems, Man, and Cybernetics 19, 17–30 (1989)
17. Lee, J.H., Kim, M.H., Lee, Y.J.: Information Retrieval Based on Conceptual Distance in
IS-A Hierarchies. Journal of Documentation 49, 188–207 (1993)
18. Resnik, P.: Semantic Similarity in a Taxonomy: An Information-Based Measure and Its
Application to Problems of Ambiguity in Natural Language. Journal of Articial Intelli-
gence Research 11, 95–130 (1999)
19. Ljubešić, N., Boras, D., Bakarić, N., Njavro, J.: Comparing Measures of Semantic Similar-
ity. In: 30th International Conference on Information Technology Interfaces, Cavtat (2008)
20. Rodriguez, M.A., Egenhofer, M.J.: Comparing Geospatial Entity Classes: An Asymmetric
and Context-Dependent Similarity Measure. International Journal of Geographical Infor-
mation Science 18, 229–256 (2004)
21. Liu, M., Shen, W., Hao, Q., Yan, J.: A Weighted Ontology-based Semantic Similarity Al-
gorithm for Web Service. Expert Systems with Applications 36, 12480–12490 (2009)
22. Guisheng, Y., Qiuyan, S.: Research on Ontology-Based Measuring Semantic Similarity.
In: International Conference on Internet Computing in Science and Engineering, Harbin,
pp. 250–253 (2008)
23. Xu, X.-h., Huang, J.-l., Wan, J., Jiang, C.-f.: A Method for Measuring Semantic Similarity
of Concepts in the Same Ontology. In: 2008 International Multi-symposiums on Computer
and Computational Sciences, Washington DC, pp. 207–213 (2008)
24. IHTSDO: SNOMED-CT, http://www.ihtsdo.org/snomed-ct/
394 A.K. Sari, W. Rahayu, and D. Wollersheim

25. Sundvall, E., Qamar, R., Nyström, M., Forss, M., Petersson, H., Åhlfeldt, H., Rector, A.:
Integration of Tools for Binding Archetypes to SNOMED CT. In: SMCS 2006, Copenha-
gen, pp. 64–68 (2006)
26. Wouters, C.: A Formalization and Application of Ontology Extraction. PhD Thesis. De-
partment of Computer Science and Computer Engineering, La Trobe University, Mel-
bourne (2005)
27. Mindswap: SMORE - Create OWL Markup for HTML Web Pages,
http://www.mindswap.org/2005/SMORE/
28. Ontomat Homepage - Annotation Portal,
http://annotation.semanticweb.org/ontomat/index.html
29. Leacock, C., Chodorow, M.: Combining Local Context with WordNet Similarity for Word
Sense Identification. In: WordNet: A Lexical Reference System and its Application, pp.
265–283. MIT Press, Cambridge (1998)
30. IHTSDO: SNOMED Clinical Terms User Guide,
http://www.ihtsdo.org/fileadmin/user_upload/Docs_01/
Technical_Docs/SNOMED_CT_User_Guide_20080731.pdf
31. Jiang, J.J., Conrath, D.W.: Semantic Similarity Based on Corpus Statistics and Lexical
Taxonomy. In: International Conference Research on Computational Linguistics X, Tai-
wan (1997)
32. Mihalcea, R., Corley, C., Strapparava, C.: Corpus-based and Knowledge-based Measures
of Text Semantic Similarity. In: AAAI 2006 Conference, pp. 775–780 (2006)
 Time-Decaying Bloom Filters for Efficient
Middle-Tier Data Management

Kai Cheng

Faculty of Information Science, Kyushu Sangyo University

3–1 Matsukadai, 2-chome, Higashi-ku, Fukuoka, Japan
[email protected]

Abstract. Distributed enterprise applications are typically based on a

multiple–tier client–server architecture where large volume of data is
transferred between tiers frequently. When the amount and frequency
of data to be transferred become large, network bandwidth will become
a bottleneck and efficient middle–tier data management is critical. In
this paper, we propose a semi–persistence model to capture the evolving
nature of data in a middle tier data management system. We also propose
to use Bloom Filters (BF) as an efficient data structure to maintain
the time-sensitive frequency profile of the underlying data items. We
first extend the standard Bloom Filters by replacing the bit-vector with
an array of counters. We then optimize it by allocating lowest space
necessary for each counter to store its value. The preliminary experiments
show that the optimized BF achieves considerable improvement on space
usage while providing the same results of frequency profile.

1 Introduction

Distributed enterprise applications are commonly adopted in almost every busi-

ness. large or small. In such applications a multi–tier client–server architecture is
typically used to provide adaptability, interoperability for enterprise computing
[1]. Because data in this architecture are frequently transferred between tiers,
netowork bandwidth will become a performance bottleneck, especially in large–
scale distributed enterprise applications. To alleviate this problem, a middle tier
is often used to reduce the data transferring, by storing frequently used data
nearly to the data consumers. A middle tier dyanamically maintains a set of
data items in response to the changing access patterns. Thus data in the middle
tier is not necessarily persistent.
According to persistence requirement, one can classify data in current data
management researches into two categories. On the one hand, persistent data is
the main topic of traditional database research. Database management systems
are designed to guarantee all data can be stored persistently and consistently all
the time. Persistent data will be there at any time unless it has been explicitly
moved to somewhere else or removed completely. In other words, persistent data
is data with ∞ lifespan. On the other hand, stream data has recently gained

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 395–404, 2010.

c Springer-Verlag Berlin Heidelberg 2010
396 K. Cheng

most importance. In many applications from IP network management to tele-

phone fraud detection, data arrives in high-speed streams, and queries over those
streams need to be processed in an online fashion to enable real–time responses.
Data streams pose a serious challenge for data management systems as the tra-
ditional DBMS paradigm of set–oriented processing of disk–resident tuples does
not apply. Stream data implies data with no persistence requirement and it is
with 0 lifespan.
In middle–tier data management systems, data with a finite life bound, or
(0, ∞) lifespan, we identify such data as semi–persistent data. Efficient manage-
ment of semi–persistent data requires new framework and technique. Particu-
larly, it is important maintain frequency of access and other profiles of each data
item because semi–persistent data management system manages such profiles as
first class objects. This means that profiles are explicitly materialized, and man-
aged in the same way that data items are managed. This allows the system to
answer trend–report query such as finding the most frequently used data
items on ’mining unstructured data’.
In this paper, we first propose a semi-persistent data model for modeling data
in the middle tier. We also propose to use Bloom Filters (BF) as an efficient data
structure to maintain the time-sensitive frequency profile of the underlying data
items. We then extend the standard Bloom Filters by replacing the bit-vector
with an array of counters and optimize it by allocating lowest space necessary
for each counter to store its value. The preliminary experiments show that the
optimized BF achieves considerable improvement on space usage while providing
the same results of frequency profile.

2 Related Work

As the multi-tier client/server architecture becomes increasingly common in the

web, mid-tier data management, i.e. cache with database management capa-
bilities has recently gained importance [2,3]. In such a multi-tier architecture,
application servers implementing most process logic connect to a backend (cen-
tral) DBMS, and the latter often becomes the bottleneck of performance. We
can improve the backend DBMS by storing frequently used results in the at-
tached cache. This technique can be regarded as “materialized view” of DBMSs.
Such kind of application of cache will not contribute to the reduction of network
traffic. Thus the objectives are different.
Another relevant work is multi-level store of persistent objects. M. Stonebraker
[4] proposed the extension of disk-resident database to include multiple storage
levels, where time -critical objects reside in main memory, other objects are disk
resident, and the remainder occupy tertiary memory. It is possible that more
than three levels will be present and that some of these levels will be on remote
hardware. Distribution of objects in a storage hierarchy is based on semantic
criteria. For example,
main memory representation: EMP where age >= 30
and age < 60
 Time-Decaying Bloom Filters for Efficient Middle-Tier Data Management 397

disk representation: age < 30

archive representation age >= 60
The distribution criteria can be changed dynamically by either an application
or a database administrator. A specific application can temporarily redistribute
instances by temporary redistribution criteria prior to execution. A database
administrator can permanently change the distribution by defining permanent
distribution criteria. A special program called vacuum cleaner is dedicated to
enforcement of the distribution criteria as well as management of buffers for data
from lower levels.

3 Semi–persistence Model for Middle–Tier Data

Management
In this section, we describe the semi–persistence model. After giving a few moti-
vation examples, we shall analyze the insufficiency of the cache model and show
the features that a semi–persistence model should have.
To motivate the necessity of a semi–persistence model for the middle–tier
data management, we suppose there is a corpus of data items in a middle tier
downloaded from remote servers. We also suppose we have an extended SQL–like
query language with the capability to query profiles maintained for each data
item. For example,
SELECT title, author, abstract
FROM Docuement
WHERE title LIKE ’%Ajax%’
ORDER BY ::profile(frequency) DESC
OFFSET 0 Limit 10
Here, we ask the system to retrieve the title, author and abstract of the top 10
most frequently accessed documents about ’Ajax’. The ORDER BY clause is
extended to include query for profiles and :: prof ile(· · ·) is a function provided
by the system. The function :: prof ile(· · ·) relies on the efficient management of
statistics about the underlying data items. We will discuss this in Section 4.

3.1 A Semi–persistence Model

We propose semi–persistence model to model data managed in a middle–tier
context. In A data item is associated with a vector of statistics (fi , ri , ai , vi ).
– fi : frequency of reference
– ri : recency of reference
– ai : age of the data item
– vi : time to live of the data item
The value fi for frequency of reference is obtained by a algorithm for efficient
monitoring frequency counts. The ri is the timestamp when the most recent ref-
erence occurs. The age ai represents time since last synchronization. If ai ≤ vi
398 K. Cheng

we say data is living, otherwise we say it is dead. Later we will discuss in details
how to set the value of vi ’s adaptively.
The recency statistic can be obtained by timestamp when the data item was
referenced. As users may not actually download a data item from the system,
instead they may go directly to the origin site. Thus we have to clarify the
meaning of data item reference.

4 The Space–Optimized Counting Bloom Filters (SBF)

As so far discussed, the semi–persistence model is dependent heavily on the
scalable approach to management of various statistics. In this section, we propose
a space–efficient scheme for this purpose. Our scheme is based on a randomized
data structure, called Bloom Filters.
The characteristics of a semi–persistence model can be described as follows:
(1) Relaxed persistence requirement. Evolving data has relaxed persistence
requirement, where data may have a life cycle and may have multiple versions.
(2) Autonomous life cycle management. Each data item has a life cycle. Any
changes during its life cycle may not be initiated by the user operation. Instead
it is managed by the system autonomously in terms of the lifespan of the data
item. This is essential for management of the evolving data.
(3) Retrievable metadata. To capture the trends of data access, statistics
about the data should be maintained efficiently. Furthermore, in order to sup-
port trend–report queries, the statistics should be managed as first–level citizen
in a semi–persistent data management system.
(4) Trend–report queries. One of the most appealing features of the
semi–persistence model is that it supports queries for the trends of data evolution
as well as the reference trends by the user. Although data warehouse or OLAP sys-
tems also support some forms of trend–report queries, the difference lies in the fact
that the data under those systems are non-volatile, static, read-only and once com-
mitted never over-written or deleted. Whereas data in a semi–persistence model
takes the dynamics of the underlying data as useful trends on its own right.

4.1 Bloom Filters for Membership Query

A Bloom Filter is a space–efficient data structures that maintain a very compact
inventory of the underlying data, supporting membership queries over a given
data set [5]. The space requirements of Bloom filters fall significantly below the
information theoretic lower bounds for error-free data structures. This efficiency
is at the cost of a small false positive rate (items not in the set have a small
constant probability of being listed as in the set), but have no false negatives
(items in the set are always recognized as being in the set). Bloom filters are
widely used in practice when storage is at a premium and an occasional false
positive is tolerable.
The standard Bloom Filters use a bit–vector to hold information about the
underlying data set. Initially, all bits in the bit–vector are turned off. Each item
 Time-Decaying Bloom Filters for Efficient Middle-Tier Data Management 399

in the set is hashed into several locations of the bit–vector using different hash
functions. Bits at these locations are then turned on. In order to find whether
an item is a member of the set, we first compute the locations to which the
item is mapped according to the same hash functions. We then check if all
bits at these locations are on. If so, the answer is yes; otherwise answer no. As
a bit can be set by other items due to the hash collisions, false positives are
possible.
Standard Bloom Filters have several limitations when extended to deploy
in other applications. First, Bloom Filter does not support deletes as simply
turning off the corresponding bits may introduce false negative errors (some bits
of other items, although still in the set, may be turned off). Second, it is note
suitable for dealing with multiset (set with duplicates), where multiplicities of
items should be reported. To extend standard Bloom Filters several variants
have been developed [6,7,8,9].

4.2 Time-Decaying Counters

In a previous work, we have proposed Time–decaying Bloom Filters (TBF),
an improvement of Counting Bloom Filters[9] by employing the exponentially
aging model [10]. This is done by introducing time-decaying counter or decay-
ing counter for short, based on some form of decay functions. In [7], a num-
ber of decay functions are given, such as exponential decay, sliding window
decay, polynomial decay, poly–exponential decay and choral and polygonal
decay.
A time-decaying counter or simply decaying counteris a counter whose value
decays periodically. The fashion of how a counter decays with time is determined
by a special non–increasing, non–negative function φ(t), called time–decaying
function (tdf ). A tdf function should satisfy the following conditions:
1. φ(0) = 1
2. φ(t) is non–increasing
3. 0 ≤ φ(t) ≤ 1 for all t ≥ 0
(i)
Let fe be the frequency count of item e in S up to the current time, tn . Let fe be
 (i)
the frequency count of item e up to the time ti since ti−1 . That is, fe = i fe .
Then, the decayed value of the counter for item e is

n
fe∗ = φ(ti ) · fe(i)
i=1

Cohen, E. et al [11] give a variety of tdf functions. In this paper, we adopt

exponential tdf, one of the most widely used tdf’s in practice.

φ(t) = λt/T , 0 ≤ λ < 1, T >0

400 K. Cheng

h1(x) h2(x) h2(x) hk(x)

......
0000
0010
0000
0111
0000
0012
0000
0101
0000
0000
0110
0000
0000

0000
0000
0001
0100
Fig. 1. Basic Time-Decaying Bloom Filters

T and λ are parameters that control the behavior of how fast the counter
will decay. T is the period between two consecutive applications of φ. In other
words, it controls the granularity of time–sensitivity. A time period of T time
units is referred to as an epoch. The parameter λ ∈ [0, 1], called exponential
decaying factor or simply decaying factor, controls the speed of exponential
decaying. In a time–decaying Bloom Filter, basically we use the exponential
decay since it is most commonly used decay function in practice. An expo-
nential decay is a function of the the form g(x) = λx , λ ∈ [0, 1]. A decaying
counter c is maintained in such a way that its value ct at time t is obtained as
follows:
ct = λ · Δt + (1 − λ) · ct−1 (1)

where Δt is the fluctuation of c between time t − 1 and t. Suppose c is updated

at an time interval of T . The value of cs at time s = kT will be decayed to
be cs · (1 − λ)m−k at time t = mT (k ≤ m) which is as per the definition of
g(x).
Time–decaying counter defined in (1) can not be easily maintained using single
counter, because Δt and ct−1 should be maintained separately. To remedy this,
we proposed the following scheme when λ > 0. Let ct = ct /λ, we obtain the
following equation.

ct = Δt + (1 − λ) · ct−1 , λ ∈ (0, 1] (2)

Using this equation, we need only one counter for each value. We can increment
the counter directly after the decay function was applied, that when a new
(1 − λ)ct−1 was obtained. Whenever, say at time t we need the value of c, the
result of ct can be obtained by the following equation.

ct = λ · ct (3)

Note that when the value of a counter can become very large and the exact
counts are not necessary, some form of probabilistic counting can be adopted
[12,13].
 Time-Decaying Bloom Filters for Efficient Middle-Tier Data Management 401

4.3 The Space–Optimized Counting Bloom Filters

To avoid allocating large counters to small values, we optimize the extended BF
by using allocating minimal size of memory needed by each counter. To do so,
we maintain a free counter pool The data structures include:
1. A standard Bloom Filter, BF (m, k), with a vector of m bits, k pairwise
independent hash functions.
2. A basic SBF (m, k) with k pairwise independent hash functions and m small
counters
3. A free counter pooliP oolj with m/2 free counters
4. A lookup table (Lookup)
First, a standard Bloom Filters is used to enable quick membership query. The
basic SBF with small counters hold frequency counts for most ”cold” data items
with smaller values. For few ”hot” items, however, extra counters can be al-
located dynamically when necessary. When a small counter in the SBF gets
overflowed, we allocate a new counter from the free counter pool (Pool) for the
carried digits. A lookup table (Lookup) maintains the bi–directional links be-
tween the overflowed counters and the extra counters. A counter in the SBF can
extend to a linked counter list to represent much larger values.
To insert a new item q into the optimized SBF filter, we simply insert the
item into SBF, incrementing each of its counters h1 ,i(i = 1, · · · , k) by 1 unless
the counter becomes overflow. If ALL these counters get overflowed, and if there
is an extra counter for this SBF counter, increment that counter by 1; otherwise
allocate an extra counters for the SBF counter. Repeat the process if a extra
counter gets overflow too.
To query for the frequency count of an item, say, q, we first check if q is
recorded in BF. If not, return 0; otherwise, we check the SBF and the lookup
table to construct the whole counter for q from higher significant bits to lower
ones. In Fig.2, q is recorded in the basic BF, and there are 2 extra counters with

BF SBF Pool
0 00
h01 01
1 h11 12 Lookup
02
q h02 0 h12 02
q 05
0 00
1 00
h03 0 h13 00
1 03
0 15

Fig. 2. Space–Optimized Counting Bloom Filters

402 K. Cheng

values of 01 and 02 respectively. Assuming all are 4-bit counters, the frequency
count for q by the SBF counter Ch12 (q) is 256 × 1 + 2 × 16 + 0 = 288.
The usage of web is well known to be quite biased, with a small fraction of
popular sites getting very high hits, while the rest and also the majority are
rarely used. As the values of counters vary significantly from very small values
for most ”cold” pages to thousands of hits for a few ”hot” pages it is not suitable
to allocate the bits to count each of these items.
To avoid allocating large counters for many ”cold” pages, we optimize the
basic SBF as well as TMF using a dynamic representation of large counters.
The optimization is do in the following way:
1. Allocate a small counter to each cell, for example 4 bits or 8 bits.
2. Introduce a free counter pool (Pool) and dynamically allocate to large coun-
ters, each may be 16 or 32 bits. The Pool can also be organized as a hierarchy
with counters of different length.
3. Allocate a flag bit for each small counter to indicate whether it contains
the real frequency value or just a index to a large counter. Its initial value
is on or 1 indicating real value contained. When a small counter becomes
overflow, turn off the flag bit and allocate a free large counter from the Pool
and record its index in the small counter.
The query process will begin with a flag bit check. If it is on, return the value
in small counter as the frequency count. Otherwise use the value as an index to
find the large counter and return value of that counter.

5 Preliminary Experiments
We evaluate the optimization of the proposed Space–Optimized Counting Bloom
Filters. The data set we will use is proxy web access logs of 12 days from June 13
to 24,2003 provided by IRCache 1 . The sanitized cache access logs contain 541.4
MB data in gzipped form. We preprocess the data set by excluding accesses
without a success status code 200 and keep requests of htmltext mime type. The
cleaned data set consists of as many as total 2, 626, 434 requests in total and
1, 027, 679 distinct requests. Fig. 3 shows that it follows Zipf-like distribution.
First we compare the basic TBF, or B–TBF with its optimization called O–
TBF, with different values of λ. The result is shown in Figure 4. With the
increase of λ, the number of large counters in use increase. Thus more memory
space is consumed.
We then evaluate the the basic TBF, or B–TBF with its optimization called
O–TBF, with different window sizes. The result is shown in Figure 5. With the
increase of window size, the number of large counters in use increase and more
memory space is required.
From Figure 4 and Figure 5, we see the optimization by allocating large coun-
ters dynamically on demand is effective.
1
http://www.ircache.net/
Time-Decaying Bloom Filters for Efficient Middle-Tier Data Management 403

㪈㪇㪇㪇㪇㪇

㪈㪇㪇㪇㪇

㩺㩷㫆㪽㩷㫉㪼㪽㪼㫉㪼㫅㪺㪼㫊
㪈㪇㪇㪇

㪈㪇㪇

㪈㪇

㪈
㪈 㪉㪐 㪌㪎 㪏㪌 㪈㪈㪊 㪈㪋㪈 㪈㪍㪐 㪈㪐㪎 㪉㪉㪌 㪉㪌㪊 㪉㪏㪈 㪊㪇㪐 㪊㪊㪎 㪊㪍㪌 㪊㪐㪊 㪋㪉㪈 㪋㪋㪐 㪋㪎㪎 㪌㪇㪌
Frequenct Rankings

Fig. 3. The workload with a Zipf–like distribution

㪈㪍㪇
㪈㪋㪇
㪤㪼㫄㫆㫉㫐㩷㪪㫀㫑㪼㩷㫀㫅㩷㪢㪙

㪈㪉㪇
㪈㪇㪇
㪙㪄㪫㪙㪝
㪏㪇
㪦㪄㪫㪙㪝
㪍㪇
㪋㪇
㪉㪇
㪇
㪈 㪉 㪊 㪋 㪌 㪍 㪎 㪏 㪐 㪈㪇 㪈㪈 㪈㪉 㪈㪊
㫍㪸㫃㫌㪼㫊㩷㫆㪽㩷㫃㪸㫄㪹㪻㪸㩷㫇㪸㫉㪸㫄㪼㫋㪼㫉㩿㩼㪀

Fig. 4. Comparison of the basic TBF and its optimization

㪈㪍㪇
㪈㪋㪇
㪤㪼㫄㫆㫉㫐㩷㪪㫀㫑㪼㩷㫀㫅㩷㪢㪙

㪈㪉㪇
㪈㪇㪇
㪙㪄㪫㪙㪝
㪏㪇
㪦㪄㪫㪙㪝
㪍㪇
㪋㪇
㪉㪇
㪇
㪈 㪉 㪊 㪋 㪌 㪍 㪎 㪏 㪐 㪈㪇 㪈㪈 㪈㪉 㪈㪊

Fig. 5. Comparison of the basic TBF and its optimization

404 K. Cheng

6 Concluding Remarks
To capture the evolving nature of data in middle–tier data management, we
proposed a semi–persistence model, which enhances the cache approach by fea-
tures like trend–report queries as well autonomous life cycle management. The
proposed model relies on efficient management of statistics of the underlying
data over time. We proposed Time-Decaying Bloom Filters as an efficient data
structure to manage time-sensitive access patterns of the underlying data. As
a future direction, a suitable scheme for adaptive management of the time to
live value of vi for each element is importance. Results from cache coherency
maintenance can be used but more semantic factors should be considered.

References
1. Cooper, B.F., Sample, N., Franklin, M.J., Olshansky, J., Shadmon, M.: Middle-tier
extensible data management. World Wide Web 4, 209–230 (2001)
2. Luo, Q., Krishnamurthy, S., Mohan, C., Pirahesh, H., Woo, H., Lindsay, B.G.,
Naughton, J.F.: Middle-tier database caching for e-busines. In: Proceedings ACM
SIGMOD Conference on Management of Data (SIGMOD 2002), Madison, Wiscon-
sin, pp. 600–611 (2002)
3. Team, T.: Mid-tier caching: The timesten approac. In: Proceedings ACM SIGMOD
Conference on Management of Data (SIGMOD 2002), Madison, Wisconsin, pp.
588–593 (2002)
4. Stonebraker, M.: Managing persistent objects in a multi-level store. In: Proceedings
ACM SIGMOD Conference on Management of Data (SIGMOD 1991), Denver, CO,
pp. 2–11 (1991)
5. Bloom, B.H.: Space/time trade-offs in hash coding with allowable errors. Commun.
ACM 13, 422–426 (1970)
6. Bonomi, F., Mitzenmacher, M., Panigrahy, R., Singh, S., Varghese, G.: An im-
proved construction for counting bloom filters. In: Azar, Y., Erlebach, T. (eds.)
ESA 2006. LNCS, vol. 4168, pp. 684–695. Springer, Heidelberg (2006)
7. Cohen, S., Matias, Y.: Spectral bloom filters. In: SIGMOD 2003, pp. 241–252.
ACM Press, New York (2003)
8. Deng, F., Rafiei, D.: Approximately detecting duplicates for streaming data using
stable bloom filters. In: SIGMOD 2006, pp. 25–36. ACM Press, New York (2006)
9. Fan, L., Cao, P., Almeida, J., Broder, A.Z.: Summary cache: a scalable wide-area
web cache sharing protocol. IEEE/ACM Trans. Netw. 8, 281–293 (2000)
10. Cheng, K., Xiang, L.: Efficient web profiling by time-decaying bloom filters. DBSJ
Letters 4, 137–140 (2005)
11. Cohen, E., Strauss, M.: Maintaining time-decaying stream aggregates. In: PODS
2003: Proceedings of the twenty-second ACM SIGMOD-SIGACT-SIGART sym-
posium on Principles of database systems, pp. 223–233. ACM Press, New York
(2003)
12. Morris, R.: Counting large numbers of events in small registers. Commun. ACM 21,
840–842 (1978)
13. Whang, K.Y., Vander-Zanden, B.T., Taylor, H.M.: A linear-time probabilistic
counting algorithm for database applications. ACM Trans. Database Syst. 15, 208–
229 (1990)
 Soft Decision Making for Patients Suspected Influenza

Tutut Herawan1,2 and Mustafa Mat Deris1

1
Faculty of Information Technology and Multimedia
Universiti Tun Hussein Onn Malaysia, Johor, Malaysia
2
Department of Mathematics Education
Universitas Ahmad Dahlan, Yogyakarta, Indonesia
[email protected], [email protected]

Abstract. Computational models of the artificial intelligence such as soft set

theory have several applications. Parameterization reduction under soft set the-
ory can be considered as a technique for medical decision making. One possible
application is the decision making of patients suspected influenza. In this paper,
we present the applicability of soft set theory for decision making of patients
suspected influenza. The proposed technique is based on maximal supported
objects by parameters. At this stage of the research, results are presented and
discussed from a qualitative point of view against recent soft decision making
techniques through an artificial dataset.

Keywords: Soft set theory; Decision making; Influenza artificial dataset.

1 Introduction
The reliance on information obtained from databases is very critical and important.
Almost in every part of our life, there are lots of instances where we have either direct
or indirect dealing with databases. One aspect that database plays an important role is
in the field of decision making. Inputs obtained from data are stored in terms of re-
cords and attributes in databases do contribute a lot in the process of decision making.
To this, one practical problem is faced: for a particular property, whether all the at-
tributes in the attributes set are always necessary to preserve this property [1]. Soft set
theory [2], proposed by Molodtsov in 1999, is a new general method for dealing with
uncertain data. We note that the soft set is designed to replace a Boolean-valued in-
formation system. The research described in this paper is a part of our short term ef-
fort in applying soft set theory in order to make a decision and further grouping of
patients under certain symptoms of influenza. Since soft set theory and soft decision
making techniques are not yet widely known, we start from a tutorial introduction.
The notions are first introduced by means of simple examples and later formalized.
The second part of this research elucidates a soft decision making technique in sig-
nificant detail. We present a problem of decision making through an artificial data of
patient suspected influenza. The data contains 50 patients with 7 symptoms. Using
soft set and maximal symptoms co-occurences in patients, we explore how soft
decision making technique can be used to reduce the number of dispensable
symptoms and further make a correct decision. This technique may potentially

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 405–418, 2010.
© Springer-Verlag Berlin Heidelberg 2010
406 T. Herawan and M.M. Deris

contribute to lowering the complexity of medical decision making without loss of

original information.
The rest of this paper is organized as follows. Section 2 describes the notion of in-
formation system. Section 3 describes the theory of soft set. Section 4 describes soft
Parameterization reduction and decision making under soft set. Section 5 describes
soft decision making using maximal supported objects by parameters. Section 6 de-
scribes an application of soft set theory for decision making and grouping patients
suspected influenza. Finally, the conclusion of this work is described in section 7.

2 Information System
The syntax of information systems is very similar to relations in relational data bases.
Entities in relational databases are also represented by tuples of attribute values. An
{ }
information system is a quadruple S = (U , A, V , f ) , where U = u1 , u 2 ,L , u U is a

{ }
non-empty finite set of objects, A = a1 , a 2 ,L , a A is a non-empty finite set of attrib-
utes, V = Ua∈A Va , Va is the domain (value set) of attribute a, f : U × A → V is an
information function such that f (u, a ) ∈ Va , for every (u, a ) ∈ U × A , called informa-
tion (knowledge) function. An information system is also called a knowledge repre-
sentation systems or an attribute-valued system. In an information system
S = (U , A, V , f ) , if Va = {0,1}, for every a ∈ A , then S is called a Boolean-valued
information system. In many medical information systems, there is an outcome of
classification that is known. This a posteriori knowledge is expressed by one (or
more) distinguished attribute called decision attribute; the process is known as super-
vised learning. An information system of this kind is called a decision system. A de-
cision system is an information system of the form D = (U , A U {d }, V , f ) , where
d ∉ A is the decision attribute. The elements of A are called condition attributes. A
simple example of a decision table can be found in Table 1.

Example 1. A decision system of 6 patients with six symptoms (conditions) and a

decision.

Table 1. An example of a decision system

U Fever Cough Runny Nose Lethargic Tired Headache Flu

p1 yes yes no yes no yes yes
p2 no no yes no yes no no
p3 yes yes no yes no yes yes
p4 yes no no no yes no no
p5 yes yes yes no no yes yes
p6 no no no yes yes no no
 Soft Decision Making for Patients Suspected Influenza 407

3 Soft Set Theory

Throughout this section U refers to an initial universe, E is a set of parameters, P(U )

is the power set of U .

Definition 2. (See [2].) A pair (F , E ) is called a soft set over U, where F is a map-
ping given by
F : E → P(U ) .

In other words, a soft set over universe U is a parameterized family of subsets of uni-
verse U. For e ∈ E , F (e ) may be considered as the set of e-elements of the soft set
(F , E ) or as the set of e-approximate elements of the soft set. Clearly, a soft set is not
a (crisp) set.

Definition 3. (See [3].) The class of all value sets of a soft set (F , E ) is called value-
class of the soft set and is denoted by C ( F , E ) .

Clearly C ( F , E ) ⊆ P(U ) .

Example 4. Let we consider a soft set (F , E ) which describes the “conditions of pa-
tients suspected influenza” that a hospital is considering to make a decision. The six
influenza symptoms, i.e., fever, respiratory, nasal discharges, cough, headache and
sore throat are adopted from [4] and one symptom added is lethargic. Suppose that
there are six patients in the hospital under consideration,
U = {h1 , h2 , h3 , h4 , h5 , h6 } ,
and E is a set of decision parameters
E = {e1 , e2 , e3 , e4 , e5 , e6 , e7 } ,
where
e1 stands for the parameter “fever”,
e2 stands for the parameter “respiratory”,
e3 stands for the parameter “nasal discharges”,
e4 stands for the parameter “cough”,
e5 stands for the parameter “headache”
e6 stands for the parameter “sore throat”
e7 stands for the parameter “lethargic”.

Consider the mapping F : E → P(U ) given by “patients (⋅) ”, where (⋅) is to be filled
in by one of parameters e ∈ E . Suppose that
F (e1 ) = {p1 , p3 , p4 , p5 , p6 } ,
F (e2 ) = {p1 , p2 } ,
408 T. Herawan and M.M. Deris

F (e3 ) = {p1 , p2 , p4 } ,
F (e4 ) = {p1 } ,
F (e5 ) = {p3 , p4 } ,
F (e6 ) = {p2 , p4 } ,
F (e7 ) = {p1 , p3 , p5 , p6 } .

Therefore, F (e1 ) means “patients suffer fever”, whose functional value is the set
{p1 , p3 , p 4 , p5 , p6 } . Thus, we can view the soft set (F , E ) as a collection of approxi-
mations as below
⎧ fever = {p1 , p3 , p 4 , p5 , p6 }, ⎫
⎪ respiratory = {p1 , p 2 }, ⎪
⎪ ⎪
⎪nasal discharges = {p1 , p2 , p 4 },⎪
(F , E ) = ⎪⎨ cough = {p1 },
⎪
⎬
⎪ head ache = {p3 , p4 }, ⎪
⎪ ⎪
⎪ sore throat = {p 2 , p4 }, ⎪
⎪ ⎪
⎩ lethargic = {p1 , p3 , p5 , p6 } ⎭

Table 2. Tabular representation of a soft set in the above example

U e1 e2 e3 e4 e5 e6 e7
p1 1 1 1 1 0 0 1
p2 0 1 1 0 0 1 0
p3 1 0 0 0 1 0 1
p4 1 0 1 0 1 1 0
p5 1 0 0 0 0 0 1
p6 1 0 0 0 0 0 1

Each approximation has two parts, a predicate e and an approximate value set p.
For example, for the approximation “ fever = {p1 , p3 , p4 , p5 , p6 } ”, we have the predi-
cate name of patients with fever and its value set is {p1 , p3 , p 4 , p5 , p6 } .

The relation between a soft set and a Boolean-valued information system is given in
the following proposition.

Proposition 3. If (F, E ) is a soft set over the universe U, then (F, E ) is a binary-
valued information system S = (U , A, V{0,1} , f ) .

Proof. Let (F , E ) be a soft set over the universe U, we define a mapping

F = {f1 , f 2 ,L, f n } ,
 Soft Decision Making for Patients Suspected Influenza 409

where

⎧1, x ∈ F (ei )
f i : U → Vi and f i (x ) = ⎨ , for 1 ≤ i ≤ A .
⎩0, x ∉ F (ei )

Hence, if A = E , V = Ue ∈A Ve , where Ve = {0,1} , then a soft set (F , E ) can be con-

i i

sidered as a binary-valued information system S = (U , A, V{0,1} , f ) .

i

From Proposition 3, it is easily to understand that a binary-valued information system

can be represented as a soft set. Thus, we can make a one-to-one correspondence be-
tween (F , E ) over U and S = (U , A, V{0,1} , f ) .

4 Parameterization Reduction and Decision Making

In this section, we present the existing techniques on soft reduction and decision mak-
ing techniques in [5], [6] and [7]. The purpose of this analysis is to present the com-
parison in the previous techniques and how our proposed technique will provide an
alternative way for soft decision making. Suppose we have a soft set (F, E ) over
universe U with the Boolean representation as displayed in Table 2. Let
f E ( pi ) = ∑ j pij , where hij are the entries of symptoms in the Boolean-table of
(F , E ) . Obviuosly, based on the Table 3 below, p1 will be the patient of first choice
for selecting since it is clearly shown that the total symptoms, f E ( p1 ) = 5 , is the
maximum choice value. In this case, we say that p1 is the optimal decision of patient
having influenza. Patients p4 , ( p 2 , p3 ) and ( p5 , p6 ) then referred as first, second and
third sub-optimal decisions, respectively.

Table 3. An example of Boolean table of representation of a soft set (F, E )

U e1 e2 e3 e4 e5 e6 e7 f E ( pi )
p1 1 1 1 1 0 0 1 5
p2 0 1 1 0 0 1 0 3
p3 1 0 0 0 1 0 1 3
p4 1 0 1 0 1 1 0 4
p5 1 0 0 0 0 0 1 2
p6 1 0 0 0 0 0 1 2

The problem of parameterization reduction and decision making through this view,
then is to reduce the number of parameter (symptoms) which preserved the consis-
tence of optimal and sub-optimal decisions.
410 T. Herawan and M.M. Deris

4.1 Parameterization Reduction of Maji et al. [5]

Maji et al. presented a reduction of soft sets and its applications in a decision making
problem, which can be briefly explained as follows. The most optimal decision
derived by Maji will be only be deduced by identifying the rough set-based reduction
set first. This can by maintained using a partition on U based on the indiscernibility
relation on a set of attributes in rough set theory [8]. As for our example based on
Table 3, the partition induced by the set of all attributes E = {e1 , e 2 , e3 , e4 , e5 , e6 , e7 } ,
denoted by U / E is given by
{{p }, {p }, {p }, {p }, {p , p }} .
1 2 3 4 5 6

Therefore, if any attribute to be deleted, then the partition induced by any subset
Q ⊂ E can only be considered as a reduct if partition induced by Q / E is still
{{p }, {p }, {p }, {p }, {p , p }} . For example,
1 2 3 4 5 6
Q ⊂ E , where Q = {e1 , e5 , e6 , e7 } and Q / E
is still
{{p }, {p }, {p }, {p }, {p , p }} .
1 2 3 4 5 6

Therefore, for Maji’s techniques, the set of attributes, Q can also be considered as a
reduct of E.

Table 4. A Boolean table after reduction of Q ⊂ E

U e1 e5 e6 e7 f E ( pi )
p1 1 0 0 1 2
p2 0 0 1 0 1
p3 1 1 0 1 3
p4 1 1 1 0 3
p5 1 0 0 1 2
p6 1 0 0 1 2

However, for Maji’s techniques, it can be seen from Table 4, that the optimal deci-
sion of patients is no longer p1 . Patient p1 has instead become sub-optimal patient
together with that of p5 and p6 . The maximum value of parameters after reduction,
i.e., f E ( p3 ) = f E ( p4 ) = 3 , is now the maximum choice. Thus, the new optimal decision
of patient in this case will be both p3 and p4 . It is very obvious, selection for opti-
mal patient will yield inconsistent result as shown from one example of reduct
Q ⊂ E , as in the above example. Even if we are to consider the selection of the next
optimal patient, in the case of Q, it will also provide inconsistent result. If for another
example, G ⊂ E , G = {e1 , e3 , e5 , e6 , e7 } and G / E is still producing the partition like
{{p1 }, {p2 }, {p3 }, {p 4 }, {p5 , p6 }} . Therefore G / E can also be considered as a reduct of E.
 Soft Decision Making for Patients Suspected Influenza 411

Table 5. A Boolean table after reduction of G ⊂ E

U e1 e3 e4 e5 e7 f E ( pi )
p1 0 0 1 1 1 3
p2 1 0 0 0 1 2
p3 0 1 1 1 0 3
p4 1 1 0 1 1 4
p5 0 0 1 1 0 2
p6 0 0 1 1 0 2

As can be seen clearly again, now the optimal decision of patients has now
changed to p4 , for this new set of reduct G which is G ⊂ E . Again the issue of
changing the value of the optimal decision of patient due to the change of maximum
value of different set of reduct will is very critical and of our main concern here. This
will not be an ideal technique of Maji’s proposal, since it does not provide consistent
result in terms of different optimal decision of patients that is suggested. Furthermore,
Maji did not discuss the sub-optimal patient that can be derived from the reduced set.
And as we have explored, the sub-optimal patient if addressed by Maji, then again
different set of reduct will yield different sub-optimal patient, thus again not able to
provide consistency in selecting sub-optimal patient. As pointed out earlier in this
paper, that the main intention of reduction is to reduce complexity in decision making
but at the same time, still able to provide consistencies in decision making.

4.2 Parameterization Reduction of Chen et al. [6]

For the sake of comparison results from different proposed techniques, let us again
use the Table 3 as example for analysis purposes. The main purpose of Chen tech-
niques, was to maintain consistency in optimal value of decision thus improving the
∑
process of decision making from Maji. Chen has defined f E ( pi ) = j pij where hij
are the entries in the Boolean-table of (F , E ) and M E denoted for collection of

Table 6. A Boolean table after parameterization reduction

U e1 e2 e3 f E ( pi )
p1 1 1 1 3
p2 0 1 1 2
p3 1 0 0 1
p4 1 0 1 2
p5 1 0 0 1
p6 1 0 0 1
412 T. Herawan and M.M. Deris

objects in U which has the maximum value of f E . Chen has in fact defined a dispen-
sable set A ⊂ E if only if M E \ A = M E . Clearly, parameter reduction of Chen has been
able to provide consistency in optimal object’s decision.
Let assume J, where J ⊆ E and J = {e4 , e5 , e6 , e7 } . Hence, M E \ J = M E is true, and
therefore J is dispensable and we are left with S = {e1 , e2 , e3 } and S ⊆ E . In the Table
6 above, p1 is still the optimal decision of patient, thus parameter reduction thus still
up-hold consistency in selecting optimal patient as opposed to Maji’s attribute reduc-
tion. However, Chen’s parameter reduction is only able to provide in consistency in
selecting optimal decision but not in the sub-optimal decision. As can be seen from the
Table 6 above, that the sub-optimal patients are p2 and p4 . Meanwhile, from Table 3,
p4 is the only patient to be referred clearly as the sub-optimal patient. Therefore pa-
rameter reduction of Chen does provide consistency in selecting optimal patient but
falls short in providing consistency for the selection of sub-optimal patients.

4.3 Normal Parameterization Reduction of Kong et al. [7]

The main patientives of Kong’s normal parameter reduction is to provide consistency

in selecting optimal and sub-optimal objects for any reduced set that conforms to the
original decision provided. Kong has maintained the same the partitions of objects by
defining indiscernibility relation IND( A) , for A ⊂ E as follow
IND( A) = {( pi , p j )∈ U × U : f A ( pi ) = f A ( p j )} .

The decision partition of U generated by IND(E ) is defined as

{
CE = {p1 ,L , pi }f , {pi+1 ,L , p j }f ,L , {pk ,L , pn }f .
1 2 s
}
In the case that for A ⊂ E , if f A ( p1 ) = f A ( p2 ) = ... = f A ( p n ) implies C E = C E − A , then
A is called dispensable set. For this definition, Kong has termed E − A as normal
parameter reduction.
Based on on the Boolean table in Table 3, the decision partition induced will be
C E = {{p1 }5 , {p4 }4 , {p2 , p3 }3 , {p5 , p6 }2 } , and p1 is the optimal decision and p4 will be

Table 7. A Boolean table after normal parameter reduction

U e1 e2 e3 e4 e5 f E ( pi )
p1 1 1 1 1 0 4
p2 0 1 1 0 0 2
p3 1 0 0 0 1 2
p4 1 0 1 0 1 3
p5 1 0 0 0 0 1
p6 1 0 0 0 0 1
 Soft Decision Making for Patients Suspected Influenza 413

the first sub-optimal decision following by p2 , p3 and p5 , p6 as the second and third
sub-optimal patients, respectively. To do a normal parameter reduction from Table 3,
since f {e ,e } (h1 ) = f {e ,e } (h2 ) = ... = f {e ,e } (h6 ) = 1 , thus Z = {e6 , e7 } ⊂ E , is dispensable
6 7 6 7 6 7

since the decision partition has not changed, that is C E = C E \ Z . By deleting parameter
Z = {e1 , e4 } ⊂ E , we will have as what is been displayed in Table 7. And Kong has
successfully shown that the optimal and sub-optimal decisions, p1 and p4 , {p2 , p3 } ,
{p5 , p6 } respectively, thus maintaining consistency in patients decision after the
reduction.

5 Soft Decision Making Using Maximal Supported Objects

Throughout this sub-section the pair (F, E ) refers to the soft set over the universe U
representing a Boolean-valued information system S = (U , A, V{0,1} , f ) .

Definition 4. Let (F, E ) be a soft set over the universe U and u ∈ U . A parameter
co-occurrence set of an object u can be defined as
coo(u ) = {e ∈ E : f (u , e ) = 1} .

Obviously, Coo(u ) = {e ∈ E : F (e ) = 1} .

Definition 5. Let (F, E ) be a soft set over the universe U and u ∈ U . Support of an
object u is defined by
supp(u ) = card({e ∈ E : f (u, e ) = 1}) .

Definition 6. Let (F, E ) be a soft set over the universe U and u ∈ U . An objects u is
said to be maximally supported by a set of all parameters E, denoted by Msupp(u ) if

supp(u ) > supp(v ) , ∀v ∈ U \ {u} .

Based on Definition 6, we can make supported (ranked) ordered objects according

their support value as
U1 > U 2 > L > U n ,

where U i ⊆ U and U i = {u ∈ U : u is i - th maximal supported by E} , for 1 ≤ i ≤ n .

Thus, U i is a collection of objects in U having the same support, i.e., objects of the
same support of are grouped into the same class. Obviously U = U1≤i≤n U i and
U i I U j = φ , for i ≠ j . In other word, a collection of U / E = {U1 , U 2 ,L ,U n } is a deci-
sion partition of U, so called cluster decision of U.

Definition 7. Let (F, E ) be a soft set over the universe U and A ⊂ E . A is said to be
indispensable if U / A = U / E . Otherwise, A is said to be dispensable.
414 T. Herawan and M.M. Deris

Based on Definition 7, we can reduce the number of parameters without changing the
optimal and sub-optimal decisions.

Definition 8. For soft set (F, E ) over the universe U and A ⊆ E . A is reduction of E
if only if A is indispensable and supp A (u ) = A supp(v ) , for every u, v ∈U .

Definition 9. For soft set (F, E ) over the universe U and u ∈ U . An object u will be
the optimal decision if u is maximally supported by E.

Example 10. As for example, the following will be the co-occurrence objects derived
from Table 3.
coo( p1 ) = {e1 , e2 , e3 , e4 , e7 } coo( p2 ) = {e2 , e3 , e5 } , coo( p3 ) = {e1 , e5 , e7 } ,
,
coo( p4 ) = {e1 , e3 , e5 , e7 } , coo( p5 ) = {e1 , e7 } and coo( p6 ) = {e1 , e7 } .
Thus, support of each object is given as follow
supp( p1 ) = 5 > supp( p4 ) = 4 > supp( p2 ) = supp( p3 ) = 3 > supp( p5 ) = supp( p6 ) = 2 .
Based from the Definition 7, the partition contain 4 clusters, i.e.,
{{p }, {p }, {p , p }, {p , p }} ,
1 4 2 3 5 6

where it is arrange in descending order of support value.

As noted that, the first maximal supported patient is p1 , where supp( p1 ) = 5 and
as been defined by Definition 9, p1 is the optimal decision. p4 can be considered
as the second maximal supported patient based on the support which the next
highest, i.e., supp( p4 ) = 4 . And it is also noted that p2 , p3 and p5 , p6 can be the
second and third maximal supported patients, since supp( p2 ) = supp( p3 ) = 3 and
supp( p5 ) = supp( p6 ) = 2 , respectively.
To elaborate Definition 8, let A = {e2 , e3 , e5 , e6 , e7 } . Then we will obtain

U / A = {{p1}, {p4 }, {p2 , p3 }, {p5 , p6 }} = U / E ,

and since E \ A = {e1 , e4 } , then

supp E \ A ( p1 ) = supp E \ A ( p2 ) = L = supp E \ A ( p6 ) = 1 .

Therefore, A is parameter reduction of E and we can now delete attribute A = {e1 , e 4 } .

Note that, by deleting A, we now have U / A = {{p1}, {p4 }, {p2 , p3 }, {p5 , p6 }} , which is
still the same decision partition as in U / E . Also in this case, the maximum sup-
ported objects are still maintained. As in [5] and [6], only the issue of optimal was
addressed, but in our paper, any set of reduct that conforms to our rule of reduct will
still provide the same optimal and sub optimal. By comparing optimal and sub-
optimal decision from our proposed technique with normal parameterization reduction
from [7], also giving the result of the same the optimal and sub-optimal decisions.
 Soft Decision Making for Patients Suspected Influenza 415

Furthermore, our proposed technique confirming that the reduction also provide the
right objects for decision making.

6 Application for Patients Suspected Influenza

Yearly influenza epidemics can seriously affect all age groups, but the highest risk of
complications occur among children younger than age two, adults age 65 or older, and
people of any age with certain medical conditions, such as chronic heart, lung, kidney,
liver, blood or metabolic diseases (such as diabetes), or weakened immune systems [9].

6.1 Artificial Dataset

The proposed technique as proposed in Section 5 is elaborate for patient decision

making from an artificial dataset of patients suspected influenza. The dataset is shown
in Table 8. The table consists of 50 patients with 7 symptoms, i.e., fever (e1 ) , cough
(usually dry) (e2 ) , headache (e3 ) , muscle and joint pain (e4 ) , severe malaise (feeling
unwell) (e5 ) , sore throat (e6 ) and runny nose (e7 ) .

Table 8. An artificial data of patients suspected influenza

Patients e1 e2 e3 e4 e5 e6 e7
1 1 0 1 0 0 0 1
2 1 1 0 1 0 1 1
3 1 1 0 1 0 1 1
4 1 0 1 1 1 0 0
5 1 0 0 0 0 0 1
6 1 0 0 0 0 0 1
7 1 0 1 1 1 0 0
8 1 0 0 0 0 0 1
9 1 0 0 0 0 0 1
10 1 0 0 0 0 0 1
11 0 0 0 1 1 1 0
12 1 0 1 1 1 0 0
13 1 0 1 0 0 0 1
14 1 0 1 0 0 0 1
15 1 1 0 1 0 1 1
16 1 1 0 1 0 1 1
17 1 0 0 0 0 0 1
18 1 1 0 1 0 1 1
19 1 0 1 1 1 0 0
20 1 0 0 0 0 0 1
21 1 0 0 0 0 0 1
22 1 0 0 0 0 0 1
23 0 0 0 1 1 1 0
24 0 0 0 1 1 1 0
25 1 0 0 0 0 0 1
416 T. Herawan and M.M. Deris

Table 8. (continued)

26 0 0 0 1 1 1 0
27 1 0 0 0 0 0 1
28 0 0 0 1 1 1 0
29 1 0 0 0 0 0 1
30 0 0 0 1 1 1 0
31 0 0 1 1 0 0 0
32 1 0 1 0 1 1 0
33 0 0 1 1 0 1 0
34 1 0 0 0 1 0 1
35 0 0 1 1 0 0 0
36 0 0 1 1 0 1 0
37 0 0 1 1 0 0 0
38 1 0 1 0 1 1 0
39 0 0 1 1 0 0 0
40 0 0 1 1 0 0 0
41 0 0 1 1 0 0 0
42 1 0 1 0 1 1 0
43 1 0 1 0 1 1 0
44 1 1 1 1 0 0 1
45 0 0 1 1 0 1 0
46 1 1 1 1 0 0 1
47 1 1 1 1 0 0 1
48 0 0 1 1 0 1 0
49 1 0 0 0 1 0 1
50 0 0 1 1 0 1 0

6.2 Result

After calculating co-occurrences of symptoms in each patient, figure 1 lists the de-
tailed results of dispensable symptoms and reduct obtained. It is clear that symptoms
severe malaise (feeling unwell) (e5 ) and runny nose (e7 ) are indicated as dispensable
symptoms. Therefore, reduct of the symptoms is fever (e1 ) , cough (usually dry) (e2 ) ,
headache (e3 ) , muscle and joint pain (e4 ) , and sore throat (e6 ) .

Dispensable symptoms e5 e7

Reduct e1 e2 e3 e4 e6

Fig. 1. Reduction of symptoms

After reduction, we can make easily make decisions. The decisions are based on
the support symptoms of each patient. Patients having the same support are clustered
in the same class. There are 9 clusters decision we can made and they are given in
 Soft Decision Making for Patients Suspected Influenza 417

Cluster 1 (fever) 5 6 8 9 10 17 20 21 22 25 27 29 34 49

Cluster 2 (fever headache) 1 13 14

Cluster 3 (muscleandjoinpain sorethroat) 11 23 24 26 28 30

Cluster 4 (headache muscleandjoinpain) 31 35 37 39 40 41

Cluster 5 (fever headache) 4 7 12 19

Cluster 6 (headache sorethroat) 32 38 42 43

Cluster 7 (headache muscleandjoinpain sorethroat) 33 36 45 48 50

Cluster 8 (fever cough muscleandjoinpain sorethroat) 2 3 15 16 18

Cluster 9 (fever cough headache muscleandjoinpain) 44 46 47

Fig. 2. Decision clusters

Figure 2. Patient with higher supported symptoms are indicated having flu rather than
lower supported symptoms.

7 Conclusion
In this paper, we have presented an application of soft decision making technique for
decision making of patients suspected influenza. We have proposed an alternative ap-
proach for parameterization reduction and decision making under soft set theory. We
have used the co-occurrence of parameters concept in a Boolean-valued information
system for defining support of each object by parameters. Based on maximal supported
objects, we have defined the notion of indispensable subset and reduct. We have pre-
sented an artificial dataset of 50 patients suspected influenza to show that the proposed
technique can be used to make a correct decision. We have shown that the results ob-
tained are equivalent with that normal parameterization and decision making problem
under soft set theory. For future research, we will investigate the possibility of parame-
terization reduction and decision making problem under a certain criteria of parameter,
e.g., under spatial ordered parameter. With this approach, it is possible that the reduct
and decision obtained are different with the proposed technique in this paper.

Acknowledgement
This work was supported by the FRGS under the Grant No. Vote 0402, Ministry of
Higher Education, Malaysia.

References
1. Zhao, Y., Luo, F., Wong, S.K.M., Yao, Y.Y.: A General Definition of an Attribute Reduct.
In: Yao, J., Lingras, P., Wu, W.-Z., Szczuka, M.S., Cercone, N.J., Ślȩzak, D. (eds.) RSKT
2007. LNCS (LNAI), vol. 4481, pp. 101–108. Springer, Heidelberg (2007)
418 T. Herawan and M.M. Deris

2. Molodtsov, D.: Soft Set Theory-First Results. Computers and Mathematics with Applica-
tions 37, 19–31 (1999)
3. Maji, P.K., Biswas, R., Roy, A.R.: Soft Set Theory. Computers and Mathematics with Ap-
plications 45, 555–562 (2003)
4. Abbas, K., Mikler, A.R., Gatti, R.: Temporal Analysis of Infectious Diseases: Influenza.
In: The Proceeding of ACM Symposium on Applied Computing (SAC), pp. 267–271
(2005)
5. Maji, P.K., Roy, A.R., Biswas, R.: An Application of Soft Sets in a Decision Making Prob-
lem. Compututer and Mathematics with Application 44, 1077–1083 (2002)
6. Chen, D., Tsang, E.C.C., Yeung, D.S., Wang, X.: The Parameterization Reduction of Soft
Sets and its Applications. Computers and Mathematics with Applications 49, 757–763
(2005)
7. Kong, Z., Gao, L., Wang, L., Li, S.: The Normal Parameter Reduction of Soft Sets and Its
Algorithm. Computers and Mathematics with Applications 56, 3029–3037 (2008)
8. Pawlak, Z., Skowron, A.: Rudiments of Rough Sets. Information Sciences 177(1), 3–27
(2007)
9. http://www.who.int/mediacentre/factsheets/fs211/en/index.html
 Personal Identification by EEG Using ICA
and Neural Network

Preecha Tangkraingkij1 , Chidchanok Lursinsap1 , Siripun Sanguansintukul1 ,

and Tayard Desudchit2
1
Advanced Virtual and Intelligent Computing (AVIC) Research Center,
Department of Mathematics, Faculty of Science,
Chulalongkorn University Bangkok 10330, Thailand
2
The Chulalongkorn Comprehensive Epilepsy Program (CCEP),
Faculty of Medicine, Chulalongkorn University, Bangkok 10330, Thailand

Abstract. The problem of identifying a person using biometric data is

interesting. In this paper, the uniqueness of EEG signals of individuals
is used to determine personal identity. EEG signals can be measured
from different locations, but too many signals can degrade the recogni-
tion speed and accuracy. A practical technique combining Independent
Component Analysis (ICA) for signal cleaning and a supervised neural
network for classifying signals is proposed. From 16 EEG different signal
locations, four truly relevant locations F7 , C3 , P3 , and O1 were selected.
This selection can identify a group of 20 persons with high accuracy.

Keywords: Electroencephalogram, Independent Component Analysis,

Neural Network, Pattern-recognition.

1 Introduction
Biometrics such as fingerprints, retinal or iris scanning, face recognition are ac-
tively used for identifications [1]. Cognitive biometrics using brain signals have
become interesting as identification tools. To understand how the brain functions,
many techniques such as electroencephalography (EEG), magnetoencephalogra-
phy (MEG), function magnetic resonance imaging (fMRI), and positron emission
tomography (PET) have been utilized. Each technique has its own strengths and
weaknesses. In this paper EEG has been used to analyze patterns because EEG has
a desirable property for excellent time resolution and low cost of instrumentation.
The objectives of this work strive to acquire locations on the scalp (channels)
that are the most promising locations for personal identification. Furthermore,
the minimum numbers of channel necessary for the identifications would also
be explored. Identifications by EEG using Independence Component Analysis
technique has been used to identify individual signals from different areas of the
brain. Neural network technique was then performed on pattern recognition for
identification. Our study is organized as follows. Section 1 is the introduction,
using EEG for identifications. Section 2 presents the background work; tech-
niques employed in the experiment such as ICA and neural networks and the

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 419–430, 2010.

c Springer-Verlag Berlin Heidelberg 2010
420 P. Tangkraingkij et al.

experimental process. Then, the results are displayed and discussed. Finally, the
conclusion is included in Section 3.

2 Background Methodology and Experiments

2.1 Background
Electroencephalography (EEG). EEG is the measurement of electrical ac-
tivity produced by the brain as recorded from electrodes placed on the scalp.
These signals are small and combined with other signals such as eye tracking
and EMG (electromyography). EEG is surrounded by large electrical potentials
from the environment. Brain waves have been categorized into five basic groups
according to its frequency 1) Delta (1-4Hz), 2) Theta (4-8Hz), 3) Alpha (8-12
Hz), 4) Beta (12-30Hz), and 5) Gamma (30-50Hz) waves. Although none of these
waves are ever emitted alone, the state of consciousness of the individual may
result in one frequency being more pronounced than others. Brain waves were
first recorded in 1874 by Richard Caton, who connected equipment directly to
the cerebral cortex of a rabbit. In 1929, Hans Berger published the first infor-
mation on scalp-recorded brain waves in humans. The differential input was first
amplified by B. H. C. Matthews in 1934 and revolutionized the high-gain am-
plification of biologic electrical signals, including brain waves. In 1935 Frederic
Gibbs, Hallowell Davis, and William Lennox, published the first EEG paper in
English dealing with epilepsy in humans. EEG is now extensively used in diag-
nosing epilepsy and in the study of how the brain functions in both animals and
humans.

Independent Component Analysis (ICA). ICA is a member of a class

of Blind Source Separation (BSS). The aim of source separation is to recover
original signals from known observations where each observation is an unknown
mixture of the original signals [2]. In this paper ICA was used to solve the prob-
lem of separating multi-channels EEG data into independent sources. We tested
the potential usefulness of the ICA algorithm for EEG source decomposition by
applying the algorithm to simulated EEG data.

Neural Networks. Neural network is a process paradigm that mimics the

structures and functions of the human nervous system. Pattern recognition is
an important application which can be implemented using a feed-forward neural
network that has been trained accordingly. During the training, the network
learns to associate output with input patterns. When the network is used, it
identifies the input pattern and tries to output the associated output pattern
similar to the way the human brain works.

2.2 Methodology and Experiments

The experimental methodology in this study consists of five processes as shown
in Figure 1: Collect EEG signal; Feature extraction using ICA 22 algorithms;
 Personal Identification by EEG Using ICA and Neural Network 421

Fig. 1. Experimental diagram process

Find the best ICA algorithms applied with 5,10,20 subjects; Reduce the number
of channels from 16 channels; Test the neural network with 20 subjects.

Collect EEG signal. EEG signals were collected from 20 normal patients (eight
men and twelve women) from the Chulalongkorn Comprehensive Epilepsy Pro-
gram (CCEP). The age range of subjects is between 12 and 40 years. The EEG
signal was recorded while subjects were resting with their eyes open. The elec-
trodes using 16 gold cups were placed on the scalp. Electrode placement are used
to record EEG signals from the locations FP 1 , F7 , T3 , T5 , FP 2 , F8 , T4 , T6 , F3 , C3
P3 , O1 , F4 , C4 , P4 , O2 according to 10-20 system shown in Figure 2. In this pa-
per, channels are labeled as ch1, ch2, ..., ch16 representing EEG from location
FP 1 , F7 , ..., O2 as mentioned earlier.
Recording sessions used monopolar montage with reference at the mastoid
area. The EEG amplifier was Grass model 8 plus. The sampling rate was 200 Hz.
The data was digitized by BMSI board using Stellate Harmony EEG software.
The data was exported as EDF (European Data Format). Electromyography
(EMG) and electrocardiogram (ECG) signals were initially removed from the
EEG signals since they were irrelevant. Finally, there were 3,000 sample data
sets of 16 channels collected from 20 subjects. A sample of the raw EEG signals
is shown in Figure 3(a).

Feature extractions using ICA. The purpose of this step is to identify the
individual signal coming from each channel using 22 different ICA algorithms.
The signal data obtained from the previous step were processed using ICALAB.
The ICALAB for signal processing is the package for MATLAB that implements
422 P. Tangkraingkij et al.

Fig. 2. The locations of electrode placements on the scalp using 10-20 system [3]

Fig. 3. The samples of raw and feature extraction using ICA. (a) Raw EEG signals
from four channels (ch2, ch10, ch11, ch12) of a subject are displayed. The x-axis of
each channel denotes number of samples, whereas y-axis represents the amplitude of
the EEG signal. (b) The EEG signals after being processed by ICA algorithm (ERICA).
 Personal Identification by EEG Using ICA and Neural Network 423

a number of efficient algorithms for ICA [4]. The list of these 22 different ICA
algorithms used in the experiment is shown in Table 1. The sample of raw EEG
signals in Fig. 3(a) after being processed by ICA algorithm for feature extraction
is shown in Fig. 3(b).

Find the best ICA algorithms using a neural network. The objective of
this part is to explore which ICA algorithm performs the best among 22 different
ICA algorithms. The experiments were conducted by using pattern recognition
feature of the neural network on MATLAB 2008a. All EEG data processed from
feature extractions were applied to 5, 10, and 20 subjects, respectively. Then,
the results were compared as shown in Table 1. The first column shows the
name of 22 ICA algorithms. The second column shows accuracy, i.e. the per-
centage of correctly classified cases from training, validation and testing data
sets for five subjects. Note that there are several algorithms achieving 100 per-
cent. To see whether this accuracy is still retained for the other sizes of ex-
perimented subjects, the numbers of study subjects were increased to 10 and
20 as shown in the third and fourth columns, respectively. The experimental
results of 5, 10 and 20 subjects were studied. ERICA, EWASOBI, JADEop,
SIMBEC, SOBI, SOBI-BPF, SOBIRO, and WASOBI algorithm gave high per-
centage of accuracy. Therefore, one of these algorithms was chosen to employ
in the next step to reduce the number of channels. Here, ERICA is the selected
algorithm.

Reduce the number of channels from 16 channels. In order to find the

minimum number of relevant channels for personal identification, the experi-
ments started with a group of two channels. The total number of selected groups
of two channels can be computed by simple binomial coefficient (16 C2 ), which
is equal to 120 combinations. In other words, this is the same as counting the
total number of possible ways of selecting 2 channels from 16 channels. The EEG
data were processed for feature extraction using ERICA algorithm. 500 sample
signals of each subject were used for testing. The identification process is based
on the extracted features using a 2-layer neural network (one hidden and one
output layer) with 20 hidden neurons and some output neurons. The number of
output neurons is set according to the number of subjects. For example, there
are five output neurons in case of five subjects.
The networks were trained with scaled conjugate gradient backpropagation.
The learning algorithm performance was evaluated using mean square error and
confusion matrices. All signals of the selected channels were systematically ar-
ranged and grouped as the input vectors for training, validation , and testing
sets. In the case of two channels, data set consists of two channels of 5, 10 and
20 subjects with each subject using 500 samples. The 500 samples data were di-
vided into 60% (300 samples) for training set, 20% (100 samples) for validation
set, and 20% (100 samples) for the testing set. In case of two selected chan-
nels, each training pattern is a 4-element vector formed by the signal value from
each channel occurring at the same time. The training patterns are generated
424 P. Tangkraingkij et al.

Table 1. The accuracy percentage of each ICA algorithm when applied with 5, 10, 20
subjects. The details of all algorithms can be seen in ICALAB [2], [4].

Algorithms 5 Subjects 10 Subjects 20 Subjects

(%) (%) (%)
AMUSE 91.44 78.54 57.16
ERICA 100.00 100.00 100.00
EVD2 98.52 84.78 55.66
EWASOBI 100.00 100.00 100.00
FAJDC4 100.00 100.00 95.00
FJADE 100.00 100.00 95.00
FOBI-E 96.20 81.02 44.85
JADEop 100.00 100.00 100.00
JADETD 94.85 75.96 51.16
MULCOMBI 100.00 100.00 95.00
POWERICA 100.00 100.00 95.00
QJADE 100.00 100.00 90.00
SAD 90.88 70.34 45.67
SIMBEC 100.00 100.00 100.00
SOBI 100.00 100.00 100.00
SOBI-BPF 100.00 100.00 100.00
SOBIRO 100.00 100.00 100.00
SONS 94.40 74.38 50.43
SYMMETRIC 100.00 100.00 89.62
THINICA 100.00 100.00 95.00
UNICA 100.00 100.00 95.00
WASOBI 100.00 100.00 100.00

by using the first six signal values from each channel. The next two signal values
are used for forming the validation patterns and another next two signal val-
ues next to validation patterns are formed as testing patterns. The network was
trained with five initial weight sets to test whether the training, validation, and
testing sets are properly formed. If they are properly formed, then the network
will always be converged. Otherwise, it may be diverged by some initial weight
sets.
Table 2 displays the results of groups of two channels when applied with 5,
10 and 20 subjects. When 5 subjects are used in experiments, nine combinations
have 100% accuracy percentage. When 10 and 20 subjects are used in the ex-
periment, the best accuracy percentage in the group of Ch2 and Ch7 is dropped
to 91.96% and 61.67%, respectively.
The results in Table 2, suggested that only two channels may not be enough
to use for identification since the accuracy percentage with 20 subjects is about
60%. Therefore, the number of channel in a group of three is investigated to
increase the accuracy percentage. Again, the possible combinations of 3 out of
16 channels are considered (16 C3 ), which is equal to 560 combinations. From the
three-channel experiment, the training set consists of three channels of 5, 10,
 Personal Identification by EEG Using ICA and Neural Network 425

Table 2. The accuracy percentage of 2 channels when applied with 5 and 10 subjects
using ERICA algorithm

Channel Channel 5 subjects 10 subjects 20 subjects

(I) (II) (%) (%) (%)
ch 2 ch 7 100.00 91.96 61.67
ch 2 ch 9 100.00 91.12 59.04
ch 2 ch 14 100.00 72.62 47.76
ch 2 ch 15 100.00 74.28 48.44
ch 5 ch 10 100.00 71.90 39.46
ch 5 ch 15 100.00 83.76 20.34
ch 9 ch 10 100.00 88.42 44.07
ch 10 ch 16 100.00 88.10 54.05
ch 12 ch 15 100.00 64.96 54.62

and 20 subjects with each subject using 500 samples. The ratio of the training
set, validation set and testing set is 60:20:20 as in the case of two channels. The
results of three channels are shown in Table 3.

Table 3. The accuracy percentage of 3 channels when applied with 5,10,20 subjects
using ERICA algorithm

Channel Channel Channel 5 subjects 10 subjects 20 subjects

(I) (II) (III) (%) (%) (%)
ch 4 ch 5 ch 10 100.00 100.00 90.52
ch 4 ch 10 ch 13 100.00 99.96 68.08
ch 4 ch 13 ch 15 100.00 100.00 75.16
ch 5 ch 9 ch 10 100.00 99.98 55.34
ch 9 ch 13 ch 15 100.00 99.98 83.41

Table 3 shows only partial of the three channels combinations. The accuracy
percentages with 5 subjects reach 100%. With 10 subjects, the accuracy per-
centages reach 100% in the groups of Ch4, Ch5, Ch10 and Ch4, Ch13, Ch15.
When numbers of subjects were increased to 20 subjects, the result shows that
the highest percentage is 90.52% in the group of Ch4, Ch5, Ch10.
The accuracy percentage of the experiment using three channels for 20 sub-
jects is still not high enough (the ideal accuracy should get close to 100%). So
the number of channels was increased to a group of four channels to increase the
accuracy percentage of the 20 subjects.
In case of four channels, total combinations of 4 out of 16 channels are con-
sidered (16 C4 ), which is equal to 1,820 combinations. The training set consists of
four channels of 5, 10, and 20 subjects, with each subject employing 500 samples.
The ratio of the training set, validation set and testing set of four channels is the
same as in the case of two and three channels. The results four channels groups
are shown in Table 4.
426 P. Tangkraingkij et al.

Table 4. The accuracy percentage of 4 channels when applied with 5,10,20 subjects
using ERICA algorithm

Channel Channel Channel Channel 5 subjects 10 subjects 20 subjects

(I) (II) (III) (IV) (%) (%) (%)
ch 2 ch 10 ch 11 ch 12 100.00 100.00 98.71
ch 2 ch 4 ch 5 ch 5 100.00 100.00 98.03
ch 4 ch 5 ch 10 ch 15 100.00 100.00 97.34
ch 2 ch 4 ch 7 ch 11 100.00 100.00 96.96
ch 2 ch 4 ch 13 ch 15 100.00 100.00 96.43

Table 4 shows only a part of the four channels combinations. The accuracy
percentages with 5 and 10 subjects reach 100%. When the numbers of subjects
were increased to 20 subjects, the result shows that the highest percentage is
98.71% in the group Ch2, Ch10, Ch11, Ch12. The results of the experiments for
5, 10, and 20 subjects are acceptable. Therefore, four channels are the lowest
possible channels for this experiment.

Table 5. Reducing the number of channels from 16 to 4 channels

Algorithms The best accuracy The best accuracy The best accuracy
for 5 subjects for 10 subjects for 20 subjects
(combination) (combination) (combination)
ERICA 1,502 combinations 185 combinations ch2 ch10 ch11 ch12
reach 100% reach 100% 98.71%
EWASOBI 862 combinations 34 combinations ch4 ch 5 ch 8 ch13
reach 100% reach 100% 91.17%
JADEop 40 combinations 4 combinations ch9 ch11 ch14 ch15
reach 100% reach 100% 93.48%
SIMBEC 110 combinations 14 combinations ch8 ch12 ch13 ch16
reach 100% reach 100% 93.48%
SOBI 110 combinations 4 combinations ch6 ch 7 ch12 ch13
reach 100% reach 100% 96.22%
SOBI-BPF 1,511 combinations 218 combinations ch3 ch 8 ch 9 ch12
reach 100% reach 100% 98.57%
SOBIRO 170 combinations 22 combinations ch5 ch10 ch13 ch15
reach 100% reach 100% 97.87%
WASOBI 1,814 combinations 65 combinations ch3 ch 6 ch 9 ch10
reach 100% reach 100% 94.26%

The result of Table 1 illustrated that 8 algorithms; namely ERICA, EWA-

SOBI, JADEop, SIMBEC, SOBI, SOBI-BPF, SOBIRO, and WASOBI algo-
rithm, give the 100% accuracy. Thus, all these eight algorithms will be applied
with the groups of four channels to further investigate the best algorithm for
 Personal Identification by EEG Using ICA and Neural Network 427

identification. In the experiments using 8 different ICA algorithms, the total

combination of 4 out of 16 channels is equal to 1,820 combinations. The first
algorithm (ERICA) starts with all 1,820 combinations and tests with 5 subjects.
There are 1,502 combinations reaching 100% accuracy. Then, these 1,502 combi-
nations are experimented with 10 subjects. The result shows that there are 185
combinations that give 100% accuracy. These 185 combinations are then applied
with 20 subjects. The final result is the combinations of Ch2, Ch10, Ch11, and
Ch12, which give the highest accuracy percentage (98.71%). The same process
was applied for other algorithms. The results are shown in Table 5.
From the result of Table 5, the best algorithm for 4-channel groups is ER-
ICA. The accuracy percentage is 98.71% in the group of channels ch2, ch10, ch11,
ch12.

Test the neural network with 20 subjects. The purpose of this step is
to test the accuracy of recognition. The neural network that has been trained
earlier using ERICA algorithm on ch2, ch10, ch11, ch12 will be used to test
the accuracy percentage. From Table 5, ERICA algorithm performs the best. In
order to further investigate the performance of this recognition, 3,000 samples
of the EEG signals are classified into 6 groups G1 - G6 (shown in Fig. 4). Each
group contains 500 samples. The first group: G1 consists of sample data from 1
- 500, the second group: G2 consists of sample data from 501-1000, and so on
until the sixth group: G6 is the sample data from 2501-3000. Note that G1 has
been used for training and testing the recognition in the previous step. However,
G2 to G6 are samples that have not been used to train and test the recognition
performance before. Thus, these samples: G2 to G6 are then used to test the
recognition performance on 20 subjects. The results are shown in Table 6.

Fig. 4. The EEG signals are divided into six groups for testing
428 P. Tangkraingkij et al.

Table 6. The accuracy percentage of 20 subjects when test with the best combination
of ERICA algorithm 4 channels 500 samples

Subject G1 G2 G3 G4 G5 G6
(%) (%) (%) (%) (%) (%)
1 100.00 100.00 99.20 98.40 93.40 98.80
2 100.00 100.00 100.00 98.20 100.00 100.00
3 100.00 100.00 100.00 100.00 100.00 100.00
4 100.00 100.00 100.00 100.00 100.00 100.00
5 100.00 99.60 100.00 99.80 100.00 100.00
6 97.00 99.60 99.80 99.60 97.60 99.40
7 99.60 97.80 95.00 99.00 94.60 94.00
8 100.00 99.40 92.40 100.00 100.00 99.60
9 100.00 100.00 100.00 99.80 100.00 100.00
10 97.40 96.40 93.40 97.80 96.20 90.00
11 100.00 99.00 95.60 99.00 98.80 99.60
12 100.00 100.00 100.00 99.20 99.80 95.00
13 99.80 98.60 99.20 100.00 100.00 100.00
14 95.40 81.00 81.80 91.60 82.40 81.20
15 99.40 92.80 95.40 97.00 92.60 94.20
16 96.80 96.20 97.60 94.80 89.00 93.00
17 94.60 92.60 94.60 94.40 98.00 96.40
18 95.80 89.40 93.20 94.80 86.60 90.60
19 98.40 98.40 99.40 98.40 99.60 90.60
20 100.00 99.80 100.00 100.00 99.80 100.00
Average
G(I) 98.71 97.03 96.83 98.09 96.42 96.59

Numbers in Table 6 are the accuracy percentage for each subject in a par-
ticular group. In the first row, for example, the accuracy percentage of subject
1 was evaluated with all groups of data (G1 to G6) . The accuracy percent-
age of subject 1 using G1 data is 100%. The accuracy percentages using G2 -
G6 data are 100%, 99.20%, 98.40%, 93.40%, 98.80%, respectively. The last row
of the Table 6 shows the average accuracy percentage in each test data group.
It can be seen that the average accuracy of these six groups is in the range
of 96% - 99%. The results show that identifications on 20 subjects are very
encouraging.

3 Conclusion and Future Work

In this paper, a practical technique for identifying from standard 16 EEG lo-
cations of EEG signals is proposed. The signals are cleaned first by apply-
ing Independent Component Analysis. Then, a supervised neural network is
used to test the accuracy of the identification process. We found that minimum
 Personal Identification by EEG Using ICA and Neural Network 429

location of EEG signals locations F7 , C3 , P3 , and O1 are relevant channel effi-

ciently used for identification by ERICA algorithm. The results show that EEG
signals can be used for identification for 20 subjects with high accuracy. However,
the number of subjects is considered to be small and all data are collected at
the same time period. It is more interesting to prove that this study still works
even to higher number of subjects and can separate the subject who does not
in the group. In addition, the EEG data which are still obtained from different
time period on the same subjects would still give the high accuracy percentage
for the identifications. This hypothesis will be further studied.

Acknowledgments. This work was supported by The Centre of Excellence in

Mathematics, PERDO, Commission on Higher Education, Thailand and THE
90th ANNIVERSARY OF CHULALONGKORN UNIVERSITY FUND
(Ratchadaphiseksomphot Endowment Fund).

References
1. Jain, A.K., Ross, A., Prabhakar, S.: An Introduction to Biometric Recognition.
IEEE Trans. Circuits Syst. Video Technol. 14(1), 4–20 (2004)
2. Cichocki, A., Amari, S.: Adaptive blind signal and image processing: learning al-
gorithms and applications. Wiley, New York (2003)
3. Niedermeyer, E., da Silva, F.L.: Electroencephalography: Basic Principles. In: Clin-
ical Applications, and Related Fields, pp. 139–141. Lippincott Williams, Wilkins
(2004)
4. Cichocki, A., Amari, S., Siwek, K., Tanaka, T., et al.: ICALAB toolboxes,
http://www.bsp.brain.riken.jp/ICALAB
5. Poulos, M., Rangoussi, M., Alexandries, N.: Neural network based person identifi-
cation using EEG features. In: Proceeding IEEE of the international conference on
Acoustic, Speech and Signal Processing, ICASSP 1999, Arizona, USA, pp. 1117–
1120 (1999)
6. Poulos, M., Rangoussi, M., Kafetzopoulos, E.: Person identification via the EEG
using computational geometry algorithms. In: Proc. Intl. Conf. EUSIPCO 1998,
Rhodes, Greece (September 1998)
7. Marcel, S., Millan, J.: Person authentication using brain-waves (EEG) and maxi-
mum a posteriori model adaptation, Tech. Rep. 81, p. 11, IDIAP Research Report,
Valais, Switzerland (2005)
8. Paranjape, R.B., Mahovsky, J., Benedicenti, L., Koles, Z.: The electroencephalo-
gram as a biometric. In: Proceedings of Canadian Conference on Electrical and
Computer Engineering, Toronto, Canada, May 2001, vol. 2, pp. 1363–1366 (2001)
9. Mohammadi, G., Shoushtari, P., Ardekani, B., Shamsollahi, M.: Person identifica-
tion by using AR model for EEG signals. In: Proceedings of the 9th International
Conference on Bioengineering Technology (ICBT 2006), Czech Republic, vol. 5
(2006)
10. Daly, D.D., Markand, O.N.: Focal brain lesions. In: Daly, D.D., Pedley, T.A. (eds.)
Current Practice of Clinical Electroencephalography, 2nd edn. Raven Press, New
York (1990)
11. Poulos, M., Rangoussi, M., Alexandris, N., Evangelou, A.: On the use of EEG
features towards person identification via neural networks. Medical Informatics
and the Internet in Medicine 26(1), 35–48 (2001)
430 P. Tangkraingkij et al.

12. Poulos, M., Rangoussi, M., Alexandris, N., Evangelou, A.: Person identification
from the EEG using nonlinear signal classification. Methods of Information in
Medicine 41(1), 64–75 (2002)
13. Palaniappan, R., Ravi, K.V.R.: A New Method to Identify Individuals Using Sig-
nals from the Brain. In: Proc. Fourth Int’l Conf. Information Comm. and Signal
Processing, pp. 15–18 (2003)
14. van Drongelen, W.: Signal Processing for Neuroscientists An introduction to the
analysis of physiological signals. Academic Press, London (2007)
15. Saeid Sanei, J.A.: Chambers, EEG Signal Processing. Wiley, Chichester (2007)
 A Formal Concept Analysis-Based Domain-Specific
Thesaurus and Its Application in Document
Representation

Jihn-Chang Jehng1, Shihchieh Chou2, and Chin-Yi Cheng2

1
Institute of Human Resource Management, National Central University,
No.300, Jhongda Rd., Jhongli City, Taoyuan County 32001, Taiwan
[email protected]
2
Department of Information Management, National Central University,
No.300, Jhongda Rd., Jhongli City, Taoyuan County 32001, Taiwan
[email protected],
[email protected]

Abstract. Many techniques in the process of document retrieval and clustering,

based on the vector space model, represent documents by vectors. They ignore
the conceptual relationships of terms such as synonyms, hypernyms and hypo-
nyms and, especially, treat terms as a bag of terms. The application of concep-
tual relationships of terms has been proved by generating improved results for
document clustering in previous studies. For those studies, thesauri like Word-
Net were used to provide the information of relationships between terms. How-
ever, some domain-specific terms like "query expansion" and "document
clustering" cannot be found in these thesauri. These terms are thought of as im-
portant features in domain-specific documents. In this paper, we propose an
automatic domain-specific thesaurus building approach based on Formal Con-
cept Analysis (FCA) dealing with the problem with general thesauri. We also
apply the domain-specific thesaurus as background knowledge to represent
documents by concept dimension vectors. In the evaluation, an improved result
by our method compared to traditional approaches is shown.

Keywords: Information Retrieval, Formal Concept Analysis, Concept Lattice,

Vector Space Model.

1 Introduction
With the growth of internet technology, ever more information is becoming available
online. However, this causes the problem of information overload, in that users have
to make greater efforts to find the information they need from web sites or document
bases. Many techniques, such as information retrieval, personalized portal or docu-
ment management are developed to help users solve this problem, by organizing the
search results or document bases in hierarchical structures, or providing users addi-
tional vocabularies to reformulate their query text.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 431–442, 2010.
© Springer-Verlag Berlin Heidelberg 2010
432 J.-C. Jehng, S. Chou, and C.-Y. Cheng

Most of the techniques in the processing of document retrieval and clustering rep-
resent documents by vectors. Usually, the dimensions of the vectors are the terms
extracted from documents. This approach has been widely applied to information
retrieval and related techniques, such as document retrieval, document clustering, and
document classification.
The main problem with this bag of terms approach is that it ignores the conceptual
similarity of terms [1][2][3]. Suppose there are two documents s1 and s2; term "query
expansion" appears in the topic of s1 and "document clustering" appears in s2. In the
traditional approach, the two documents will be identified irrelevant, because there is
no hypernym like "information retrieval" to make them relevant.
The application of conceptual relationships of terms to document retrieval and
clustering in some cases can improve the precision and recall rate [1][2]. Usually, the
authors of documents use different words to describe the same thing. However, tradi-
tional approaches do not take these relevant words into synonym sets. A thesaurus
like WordNet [4] can provide the information of conceptual relationships of terms
such as synonyms, hypernyms or hyponyms. Previous studies, such as Hotho et al. [2]
and Recupero [1], used WordNet as background knowledge to improve the results of
document clustering.
A thesaurus like WordNet is a general thesaurus, and does not aim at the use of
specific domains. Therefore, some domain-specific words or phrases cannot be found
in WordNet. These words or phrases are always thought of as important features in
domain-specific documents. For example, phrases such as "information retrieval",
"query expansion" and "document clustering" are important features in information
retrieval-related documents; however, they cannot be found in WordNet. To deal with
the problem of general thesauri, some researches construct domain-specific thesaurus
by ontology [3][5]. However, the building of domain-specific ontology always re-
quires the participation of humans and is manual or semi-automatic.
To tackle the problems mentioned above, an automatic domain-specific thesaurus
building methodology is needed. Formal concept analysis (FCA) is mainly used for
data analysis and in particular for extracting inherent relationships between objects in
a given context [6]. This feature helps us to extract the relationships of terms and
form the terms into a thesaurus. Previous researches applied FCA to identify the re-
lated terms in a given document set and used these terms to generate expanding terms
for query expansion [7].
In this paper, we propose an automatic domain-specific thesaurus building ap-
proach by FCA. Each concept of the domain-specific thesaurus consists of related
terms, and the relations between concepts are inherent. This thesaurus provides the
relations of terms for a given specific domain. We also use it as background knowl-
edge to represent documents by concept dimension vectors.
The rest of this paper is organized as follows: In section 2, we introduce the FCA
and its applications in information retrieval. Major definitions of FCA are also de-
scribed. In section 3, the proposed method of constructing the domain-specific
thesaurus by FCA is explained. In this section, we first introduce the procedures of
document preprocessing, then build our domain-specific thesaurus by FCA. Finally,
we explain the method of representing documents by concept dimension vectors. In
section 4, evaluation of our method is conducted. We make a comparison with tradi-
tional approaches to show the improved results of the use of the domain-specific
thesaurus. Finally, we come to the conclusions of this paper.
 A Formal Concept Analysis-Based Domain-Specific Thesaurus 433

2 Formal Concept Analysis

FCA, first introduced by Rudolf Wille in 1982 [8], is a theory of data analysis which
identifies conceptual structures among data sets [9][10]. It is based on the Lattice
Theory [11], which has been widely applied within many different realms. FCA forms
a given application domain in terms of objects (extensions) and the attributes (inten-
sions) that are used to describe those objects, and provides a meaningful, comprehen-
sible interpretation of the given context according to Concept Lattices [10].
FCA has been successfully used in a wide range of applications in information re-
trieval. Groojen and Weide use FCA to discover the number of related term combina-
tions for query expansion [7]. Rajapakse and Denham make up a learning strategy
with FCA to improve the results of document retrieval [12]. Cimiano et al. apply FCA
to generate concept hierarchies from text corpora for certain domain applications [6].
One of the most important features of FCA is that it can discover the inherent rela-
tionships between objects from a given context. In the application of information
retrieval, some hierarchical document clustering techniques have been developed;
these techniques are based on the inherent relationships between documents. For
example, Carpineto and Romano [13] apply a Concept Lattice generated from FCA to
support conceptual document clustering and provide a browsing retrieval by the sys-
tem they implemented named GALOIS. CREDO, another case, is a system that
organizes web search results in a hierarchical structure by Concept Lattice, is also
developed by Carpineto and Romano [14].
In this paper, we apply FCA to generate a domain-specific thesaurus to extract the
relations of terms. The following are the major notions of FCA.

Definition 1
A formal context is defined as a triple (G, M, I), where G and M are two sets and
I ⊆ G × M is a binary relation between G and M. The elements of G are called objects
and those of M are called attributes. I is the relation defined between G and M. To
represent a relation I between object g and attribute m, we write gIm or ( g , m) ∈ I .
In Table 1, we provide an example of a formal context which contains four objects
from Document1 to Document4, and attributes are the words appearing in the given
documents. The elements of Table 1 represent the binary relation between G and M
that indicates whether terms appeared in the documents.

Definition 2
For A ⊆ G , we define
A' = {m ∈ M | ( g , m ) ∈ I for all g ∈ A}
denotes the set attributes common to all the objects in A.

Table 1. An example of formal context

Document/Term Information Query Document Local Global

retrieval expansion clustering analysis analysis
Document 1 X X X
Document 2 X X
Document 3 X X X
Document 4 X X
434 J.-C. Jehng, S. Chou, and C.-Y. Cheng

Dually, for B ⊆ G ,
B' = {g ∈ G | ( g , m) ∈ I for all m ∈ B}

is the set of objects that have all attributes in B.

For example, in our example of formal context in Table 1, A' can be {Document1,
Document3}'={"Information retrieval", "Query expansion"} and B' can be {"Informa-
tion retrieval", " Document clustering "}'={Document2, Document4}.

Definition 3
A formal concept of context (G, M, I) is defined as a pair (A, B)
where A ⊆ G , B ⊆ G , A'=B and B'=A.
For example, {Document1, Document3}={"Query expansion"} is called a formal
concept in our example of formal context. A of the formal concept (A, B) is called
extent and B is called intent. Extent contains all objects belonging to the formal con-
cept (A, B), and intent is the attributes shared by the objects.

Definition 4
Two formal concepts of a given formal context can establish the super-concept and
sub-concept relation. Let (A1, B1) and (A2, B2) be formal concepts of a formal context;
(A1, B1) is a sub-concept of (A2, B2) if and only if A1 ⊆ A2 (or if and only if B2 ⊆ B1);
then (A2, B2) is called a super-concept of(A1, B1). We write (A1, B1) p (A2, B2), where
p denotes a super-concept and sub-concept relation of the formal concepts. Consider
the set of all formal concepts and their super-concept and sub-concept relations of the
formal context; then CL = (G , M , I ) is a complete lattice called Concept Lattice.
A Concept Lattice can be drawn as a diagram with nodes representing formal con-
cepts and edges representing super-concept and sub-concept relations between formal
concepts. Fig. 1 is the Concept Lattice constructed from the formal context of Table 1.
Each node in Fig. 1 represents a formal concept and the edges are their inherent rela-
tions. From Fig. 1, we find a sub-concept inherits all attributes from its super-
concepts and is more specific than its super-concepts. Many methodologies for

Fig. 1. Concept Lattice for the formal context in Table 1

 A Formal Concept Analysis-Based Domain-Specific Thesaurus 435

generating Concept Lattices can be found in [15]. In this paper, we use the tool of
Concept Explorer1 for producing Concept Lattices.

3 Domain-Specific Thesaurus for Document Representation

Traditionally, document representation is based on the use of bag of terms. Given a

document set D, T={t1,t2,…,tn} is the set of different terms appearing in D. Then, a
document di ∈ D can be represented as a term dimension vector
→
d i = (tfidf (d i , t1 ), tfidf ( d i , t 2 ),..., tfidf ( d i , t n )) ,

where tfidf(di,tn) is a weighting schema for measuring the importance of term tn in

document di. The tfidf weighting schema can be defined as follows.

Definition 5: tfidf (term frequency-inverted document frequency) term weighting

schema
D
tfidf ( d i , t n ) = tf ( d i , t n ) × log( )
df (t n )

where tf(di,tn) is the frequency of term tn in document di; df(tn) is the document fre-
quency that indicates the number of documents containing term tn. This weighting
schema is used to measure the importance of terms in a given document. Some further
applications, such as document retrieval, document clustering, or document classifica-
tion, apply Cosine or Jaccard measure to compute the distance between two vectors
which are represented documents or queries.
This very typical method of document representation, called the vector space
model, was first introduced by Salton [16] and has been successfully and widely ap-
plied to information retrieval. However, one of the main drawbacks of this approach
is its ignoring of conceptual relationships of terms. To tackle this problem, we pro-
posed an automatic domain-specific thesaurus building approach based on FCA. This
domain-specific thesaurus is mainly used to represent the relations between terms via
the relations between formal concepts. We also applied this thesaurus as background
knowledge to represent documents by concept dimension vectors.
In our method, FCA is applied to construct the domain-specific thesaurus. The Con-
cept Lattice generated by FCA for a given document set consists of a set of formal
concepts in which related terms are formed. By the definitions introduced in Section 2,
all relations between formal concepts in a Concept Lattice are inherent, and these rela-
tions are called super-concept and sub-concept relations. Therefore, we assume that the
terms in the same formal concept can be regarded as synonyms; and the terms in super-
concept are hypernyms for those terms in their sub-concepts; on the other hand, terms
in sub-concept can be thought as hyponyms for those terms in their super-concepts. In
the following sections, we introduce the procedure for our method in detail.

1
Concept Explorer can be download at http://sourceforge.net/projects/conexp/
436 J.-C. Jehng, S. Chou, and C.-Y. Cheng

3.1 Document Preprocessing

First, all documents are broken down into sentences and then these sentences are part-
of-speech (POS) tagged, using Stanford POS Tagger2. The tags help us to identify the
POS of each word in a sentence, such as nouns, verbs, and adjectives. The POS tags
are also important in extracting the phrases from documents. These extracted phrases,
we thought, were critical features of domain-specific documents. The following is an
example of POS tagged sentence.

Original example sentence:

In this paper, we propose an automatic domain-specific thesaurus
building approach based on Formal Concept Analysis (FCA) dealing with
the problem with general thesauri.

POS tagged example sentence:

In/IN this/DT paper,/NN we/PRP propose/VBP an/DT automatic/JJ
domain-specific/JJ thesaurus/NN building/NN approach/NN based/VBN
on/IN Formal/NNP Concept/NNP Analysis/NNP (FCA)/NNP dealing/VBG
with/IN the/DT problem/NN with/IN general/JJ thesauri./NN

Each token after the slash is the POS tag of words. For example, IN is preposition; DT
is determiner; and NN is noun. A complete list of POS tags can be found in [17].
Second, we keep all nouns, and in order to extract the domain-specific phrases
from documents, we referr to a list of POS patterns constructed by Ou, et al. [18] for
recognizing phrases (See Table 2). Then, stop-words are removed from remained
nouns.

Table 2. Part-of-speech patterns for recognizing phrases

Part-of-speech tag
1 2 3 Example term
NN QE, IR, FCA
JJ NN unexpanded query
NN NN query expansion, document clustering
JJ NN NN interactive query expansion

Third, we stemmed all remained nouns and phrases by Porter stemming algorithm
[19].

3.2 Concept Lattice Generation

After the document preprocessing, we get a set of terms T which consists of single
nouns or phrases. Then, a formal context is constructed from a training document set
D, where the objects of a formal context fck ∈ FC are the documents in the training
set, and the attributes are the terms belong to T; the relation between objects and

2
http://nlp.stanford.edu/software/tagger.shtml
 A Formal Concept Analysis-Based Domain-Specific Thesaurus 437

attributes is the appearance of t ∈ T in d ∈ D . Then, a Concept Lattice can be

generated by FCA, which contains a number of formal concepts within several terms
Tfck and documents Dfck, and super-concept and sub-concept relations between formal
concepts.

3.3 Domain-Specific Thesaurus Generation

We defined our domain-specific thesaurus as a network-like structure, which is used

to represent relations of terms for a given document set. The following notions are
given for our domain-specific thesaurus.

Definition 6
A domain-specific thesaurus is defined as a tuple DST=(C, p )consisting of a set of
concept C and inherent relation p with C. These inherent relations are also called
sub-concept and super-concept relations. For each concept ci ∈ C, it contains a set of
term Tci.

Definition 7
For c1, c2 ∈ C, if c1 p c2, we called c1 is a sub-concept of c2; on the other hand, c2 is
super-concept of c1.If c1 is sub-concept of c2, then Tc2 is a subset of Tc1 (Tc2 ⊆ Tc1).

Definition 8
A concept ci is the root concept in DST, if and only if there is no super-concept for ci.
We write a root concept as cRoot. A root concept ci is the only concept that its Tci may
be empty in DST.
In order to transform the Concept Lattice CL generated in Section 3.2 into a do-
main-specific thesaurus DST defined above, the following steps are executed. First,
all information about objects is removed from a formal concept, while the object
information is no longer needed in the following process. Second, we delete the most
bottom formal concept whose object set is empty and term set contains all terms of T.
Third, we eliminate the formal concept whose set of terms Tci equals the union of sets
of terms of its all direct super-concepts, and the edges connected to this concept.
This pruned Concept Lattice with no circle is used to represent the domain-specific
thesaurus DST defined above. We keep all relations between formal concepts and use
the pruned formal concepts to represent concept ci defined in DST.

3.4 Transform Term Vector to Concept Vector

In order to deal with the problem with traditional bag of terms approaches which do
not take account of the conceptual similarity of terms in their applications, we apply
the domain-specific thesaurus defined in Section 3.3 as background knowledge to
represent a given document by concept dimension vector. A new concept dimension
vector can be represented as
→
d i = (cw(d i , c1 ), cw(d i , c2 ),..., cw(d i , cn )) ,
where cw(di,cn) is the weight of concept ci ∈ C in the document di. For each docu-
ment, all of its concept weights are determined by the following steps.
438 J.-C. Jehng, S. Chou, and C.-Y. Cheng

First, we compute all concept weights for all concepts ci ∈ C. We define a concept
weighting schema for each concept as follows.

Definition 9
For all ci ∈ C,
cw(d i , cn ) = ∑ tf (d i , t )
t∈Tci '

where Tci' is term set in ci excluded the terms appeared in the direct super-concepts of
concept ci.
Second we update concept weights for all concepts, except the concepts at most
bottom level, from its sub-concepts. In our definitions of DTS, a concept ci may in-
herit from multiple super-concepts. This is significantly different to hierarchical struc-
ture. In hierarchical structure, one concept only has one super-concept. Therefore, we
developed a bottom-up approach which updates the weight of concepts by moving the
weight of a given concept up to all its super-concepts.
A recursive algorithm called UpdateConceptWeights (Fig. 2) is used to update the
concept weights for all super-concepts for a given concept. A given concept c and its
weight SubWeight will be the inputs of UpdateConceptWeights. If concepts c is not
root concept, this algorithm will get the summation value NumberOfSuperTerms of
the number of terms of c’s direct super-concepts (lines 5-7). For each direct super-
concept csuper of c, their weight is determined by the maximum function (line 9). Two
values will be the inputs of this maximum function, one is the original concept weight
of csuper and another is distributed value of SubWeight. The distributed value of Sub-
Weight for csuper is defined as
NumberOfTerms (c Super )
×ω ,
NumberOfSuperConceptTerms

Algorithm: UploadConceptWeights
Input: Concept c, SubConceptWeights SubWeight
Output: Concept weights for all concepts in DST
1. if c=cRoot
2. return
3. else
4. NumberOfSuperConceptTerms=0
5. foreach csuper ∈ SuperConcept(c)
6. NumberOfSuperConceptTerms+=NumberOfTerms(csuper)
7. end foreach
8. foreach csuper ∈ SuperConcept(c)
9. cw(di,csuper)=max(cw(di,csuper),
NumberOfTerms (cSuper )
× ω × SubWeights )
NumberOfSuperConceptTerms
10. UploadConceptWeights (csuper,cw(di,csuper))
11. end foreach
12. end if

Fig. 2. Pseudo code of UpdateConceptWeights

 A Formal Concept Analysis-Based Domain-Specific Thesaurus 439

where NumberOfTerms(csuper) represents the number of terms in csuper and ω is a

constant for tuning. The Updated concept weight of csuper will be the maximum value
of the two input values of maximum function at line 9. Finally, all direct super-
concepts of c and their updated concept weight cw(di,csuper) will be the inputs of algo-
rithm UpdateConceptWeights (line 10). This algorithm will end if input concept is
root concept (lines 1-2).
All concepts, except root concepts, with their weight in DTS will be the inputs of
the algorithm UpdateConcpetWeights. The execution sequence of this algorithm is
from the concepts at the most bottom level to concepts at topmost level, level by
level. For each concept it will move its weight up to all its direct and indirect super-
concepts by this algorithm.

4 Evaluation
In this section, we demonstrate our method introduced in Section 3 through a
simple example given as follows: We first generate a domain-specific thesaurus
from an example training document set. Furthermore, this thesaurus is used as
background knowledge for representing the given document set by concept dimen-
sion vectors. Finally, we analyze and compare the results of the concept dimension
vectors with traditional term dimension vectors in the measuring of document
similarities.
The test data set we use in this evaluation are the sentences selected from [20],
which is a research paper mainly focused on the topic of query expansion. We select
sentences, which describe the techniques about local and global query expansion, as
training and test document sets.
A training document set which consists of four documents from TrainDoc1 to Tra-
inDoc4 is given as follows:

TrainDoc1: Existing techniques for automatic query expansion can be

categorized as either global or local.
TrainDoc2: A global technique requires some corpuswide statistics that
take a considerable amount of computer resources to compute.
TrainDoc3: Local technique processes a small number of top-ranked
documents retrieved for a query to expand that query.
TrainDoc4: A local technique may use some global statistics such as the
document frequency of a term.

Highlighted terms, including single nouns and phrases, are selected by the proce-
dure of document preprocessing introduced in Section 3.1. All selected terms are
stemmed by Porter’s algorithm. We use these terms as the attributes and training
documents as objects to form the formal context. Then, according to the method de-
scribed in Section 3.2 and Section 3.3, a domain-specific thesaurus for this context
can be generated as Fig. 3.
440 J.-C. Jehng, S. Chou, and C.-Y. Cheng

Fig. 3. Domain-specific thesaurus for test document set

Three test documents are given as follows: The important terms in documents are
selected and highlighted.
TestDoc1: One of the earliest global techniques is term clustering.
TestDoc2: Information from the top-ranked documents is used to
reestimate the probabilities of query terms in the relevant set
for a query.
TestDoc3: Local techniques expand a query based on the information in
the set of top-ranked documents retrieved for the query.
According to the procedures described in Section 3.4, each concept in domain-
specific thesaurus, as shown in Fig. 3, is the dimension of vectors which will be used to
represent test documents. Then, we compute concept weights for all concepts in do-
main-specific thesaurus for each test document by concept weighting schema defined in
definition 9, and execute the algorithm of UpdateConceptWeights to update the concept
weights for concepts which have sub-concepts. In this example, the value of ω is set to
0.5. The values of concept weights for each test document are used as the dimension
values for vectors. The results of document representation are shown in Table 3.

Table 3. Vectors of test documents

Dimensions
c1 c2 c3 c4 c5 c6 c7
→
d TestDoc1 1 0 1 0 0 0 0
→
dTestDoc2 0.5 1 0 2 0 0 0
→
d TestDoc 3 1 1.5 0 3 0 0 0
 A Formal Concept Analysis-Based Domain-Specific Thesaurus 441

In evaluating the method, we are especially interested that the sub-concept and su-
per-concept relationships between terms extracted by our method may create a chance
to make two documents relevant. These two documents may be identified irrelevant
by traditional approaches. Table 4 shows the document representations by traditional
bag of terms approach with tf-idf weights value for the test documents. After Cosine
measuring, the similarity between test documents TestDoc1 and TestDoc2 is zero,
because, there are no common terms between these two documents. However, accord-
ing to the results of our method shown in Table 3, TestDoc1 and TestDoc2 are
accounted with small similarity by Cosine measuring, because they have a shared
concept in c1. The terms {"top-ranked documents", "query"} in c4 are sub-concept of
term {"local"} in c2, it seems reasonable, because selecting useful expansion terms
from top-ranked documents is the basic idea of local based query expansion. The term
{"local"} in c2 and term {"global"} in c3 inherit root concept c1 which contains term
{"technique"}; in this context, it may mean the technique of query expansion. There-
fore, although there is no term in TestDoc2 matching the terms in concept c1, Test-
Doc2, however, contains terms in c4 which indirectly inherits c1, and therefore c1
becomes a shared concept between TestDoc1 and TestDoc2 and makes these two
documents relevant.

Table 4. tf-idf weights of test documents

Term dimensions
global technique local information top-ranked probability query query terms term
documents cluster
TestDoc1 0.48 0.18 0 0 0 0 0 0 0.48
TestDoc2 0 0 0 0.18 0.18 0.48 0.18 0.48 0
TestDoc3 0 0.18 0.48 0.18 0.18 0 0.35 0 0

5 Conclusion

In this paper, we provide an automatic domain-specific thesaurus building approach,

and use it as background knowledge to represent documents by concept dimension
vectors. This thesaurus built by FCA contains several concepts with the relations of
sub-concept and super-concept, and each concept consists of a set of terms. Accord-
ing to the concept the terms belong to, the relationships between terms, including
synonyms, hypernyms and hyponyms, can be identified. We apply this domain-
specific thesaurus and develop the methodology of concept weighting for document
representation. The demonstration shows the improved results on the measuring of
similarities between documents. Compared to traditional approaches, our method
successfully makes documents with shared concepts relevant. Our work is ongoing
and will be extended to the information retrieval-related techniques, including
document clustering, document classification and query expansion. Furthermore,
given that Concept Lattice building is a computationally difficult task [13][12], we
also plan to develop a learning strategy for constructing domain-specific thesauri
incrementally.
442 J.-C. Jehng, S. Chou, and C.-Y. Cheng

References
1. Recupero, D.R.: A New Unsupervised Method for Document Clustering by Using Word-
Net Lexical and Conceptual Relations. Information Retrieval 10(6), 563–579 (2007)
2. Hotho, A., Staab, S., Stumme, G.: Wordnet Improves Text Document Clustering. In: Proc.
of the SIGIR 2003 Semantic Web Workshop (2003)
3. Hotho, A., Staab, S., Stumme, G.: Ontologies Improve Text Document Clustering. In:
Proc. of the IEEE International Conference on Data Mining (ICDM), pp. 541–544. IEEE
Press, New York (2003)
4. Miller, G.: WordNet: A Lexical Database for English. CACM 38(11), 39–41 (1995)
5. Hotho, A., Staab, S., Maedche, A.: Ontology-based Text Clustering. In: IJCAI 2001,
Workshop on Ontology Learning. Seattle, Washington, USA (2001)
6. Cimiano, P., Hotho, A., Staab, S.: Learning Concept Hierarchies from Text Corpora Using
Formal Concept Analysis. J. of Artificial Intelligence Research 24, 305–339 (2005)
7. Groojten, F.A., Weide, T.P.: Conceptual Query Expansion. Data & Knowledge Engineer-
ing 56(2), 174–193 (2006)
8. Wille, R.: Restructuring Lattice Theory: an Approach Based on Hierarchies of Concepts.
In: Rival, I. (ed.) Ordered sets, pp. 445–470. D. Reidel Publishing Company, Dordrecht
(1982)
9. A Formal Concept Analysis Homepage,
http://www.upriss.org.uk/fca/fca.html
10. Ganter, B., Wille, R.: Formal Concept Analysis - Mathematical Foundations. Springer,
Heidelberg (1999)
11. Brikoff, G.: Lattice Theory. American Mathematical Society, Providence (1967)
12. Rajapakse, R.K., Denham, M.: Text Retrieval with more Realistic Concept Matching and
Reinforcement Learning. Information Processing and Management 42(5), 1260–1275
(2006)
13. Carpineto, C., Romano, G.: A Lattice Conceptual Clustering System and its Application to
Browsing Retrieval. Machine Learning 24(2), 95–122 (1996)
14. Carpineto, C., Romano, G.: Exploiting the Potential of Concept Lattices for Information
Retrieval with CREDO. J. of Universal Computer Science 10(8), 985–1013 (2004)
15. Carpineto, C., Romano, G.: Concept Data Analysis: Theory and Applications. John Wiley
& Sons, Ltd., Hoboken (2004)
16. Salton, G.: A Vector Space Model for Automatic Indexing. CACM 18, 613–620 (1975)
17. Marcus, M.P., Santorini, B., Marcinkiewicz, M.A.: Building a Large Annotated Corpus of
English: The Penn Treebank. Computational Linguistics 19(2), 313–330 (1993)
18. Ou, S., Khoo, C.S., Goh, D.H.: Design and Development of a Concept-Based Multi-
Document Summarization System for Research Abstracts. J. of Information Science 34(3),
308–326 (2008)
19. Porter, M.F.: An algorithm for suffix stripping. Program. 14(3), 130–137 (1980)
20. Xu, J., Croft, W.B.: Improving the Effectiveness of Information Retrieval with Local Con-
text Analysis. ACM Tran. on Information Systems 18(1), 79–112 (2000)
 On the Configuration of the Similarity Search
Data Structure D-Index for High Dimensional
Objects

Arnoldo José Müller-Molina and Takeshi Shinohara

Department of Artificial Intelligence

Kyushu Institute of Technology
Kawazu 680-4 Iizuka 820-8502, Japan
[email protected],
[email protected]

Abstract. Among similarity search indexes, the D-index introduced by

Gennaro et al. in 2001 is regarded as an efficient metric access method.
The performance of this index depends on several parameters, and their
optimal configuration remains an open problem. We study two perfor-
mance issues that occur when the D-index handles high dimensional
objects. To solve these problems, we introduce an optimization that sim-
plifies the D-index. By doing this, we remove two configuration parame-
ters and improve performance.

Keywords: similarity search, digital library, multimedia system, infor-

mation retrieval, spatial index.

1 Introduction

New similarity search applications are constantly being developed, ranging from
music similarity search to open source license violation detectors [1]. To support
such development, metric access methods (MAMs) are valuable tools as they can
be used by researchers to find suitable similarity measures for objects. Among
centralized metric access methods, the D-index [2] is regarded as one of the
fastest MAMs available . The index implements an addressing scheme based on
the relationship between each object and a set of objects called pivots taken
from the database. Pivoting techniques have been traditionally used to estimate
a lower bound of the distance function employed [3,4,5]. The combination of
these techniques is what makes this index unique.
The D-index requires several parameters whose configuration severely affects
performance. The optimal configuration of these parameters still remains an
open problem to date. In this paper, we study the structure of the D-index
in the context of costly distance functions. Our main objective is therefore to
minimize distance computations.
The D-index employs a clustering technique where data is divided into levels
composed of buckets. Additionally, it employs a pivot filtering technique [5] used

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 443–457, 2010.

c Springer-Verlag Berlin Heidelberg 2010
444 A.J. Müller-Molina and T. Shinohara

to reduce dimensionality. When a large number of pivots is employed, one of the

following issues arise:

– The D-index tries to distribute evenly data among a set of buckets however,
in practice, about 80% to 99% of the data stays in one bucket.
– The search algorithm can access an exponential number of buckets with
respect to the cardinality of the pivot set employed.

To correct these problems, we propose a simplification of the D-index that we

call the SD-index (Simple D-index). Our optimization distributes the data into
a larger number of buckets. Additionally, we propose an heuristic that reduces
the number of buckets that need to be accessed.
Our main conclusion is that it is possible to increase performance by reducing
the structure of the D-index to only one level. Additionally, our technique brings
two benefits. First, two parameters are removed, and this makes the D-index
more accessible to a wide range of users. Second, index creation becomes faster
as bucket levels have been removed.
In Section 2, we review key concepts related to the D-index. In Section 3,
our technique is discussed. In Section 4, we experimentally demonstrate the
effectiveness of our method. Finally in Section 5, we point out future research
objectives and the concluding remarks.

2 Background

2.1 Metric Space

Let M = (D, d) be a metric space for a domain of objects D and d : D × D → R,

a total distance function that satisfies the following properties:

∀x, y ∈ D, d(x, y) ≥ 0
∀x, y ∈ D, d(x, y) = d(y, x)
∀x, y ∈ D, d(x, y) = 0 ⇐⇒ x = y
∀x, y, p ∈ D, d(x, y) ≤ d(x, p) + d(p, y)

Given a collection X ⊆ D, a range query for q ∈ D and a range r retrieves all

the objects within distance r, that is the set {x ∈ X|d(x, q) ≤ r}. A nearest
neighbor query returns the k closest elements to the query object q, or the set
R ⊆ X such that |R| = k, for any x ∈ R and any y ∈ X − R : d(q, x) ≤ d(q, y).
In this paper, we focus on queries that combine both range and nearest neighbor
queries.

2.2 Pivot Filtering

It is possible to reduce the number of distance computations executed by em-

ploying the well established pivot filtering technique [5]. This embedding helps
 On the Configuration of the Similarity Search Data Structure D-Index 445

to reduce distance computations by exploiting the triangle inequality. To un-

derstand this, a well known property of the triangle inequality is useful. For a
metric space M = (D, d) and three objects x, y, p ∈ D:

|d(p, x) − d(p, y)| ≤ d(x, y). (1)

The technique selects n pivot objects, and obtains the distances of all the objects
to these pivots. These pre-computed distances can be used to obtain a lower
bound of the real distance. Formally, for a metric space M = (D, d) and a set
of pivots P = {p1 , . . . pn }, the mapping Ψ : (D, d) → (Rn , L∞ ) can be defined:

Ψ (x) = (d(x, p1 ), d(x, p2 ), . . . , d(x, pn )).

The resulting vector is called the pivot vector. A lower bound of d can be obtained
with:
n
L∞ (Ψ (x), Ψ (y)) = max |d(x, pi ) − d(y, pi )| ≤ d(x, y).
i=1

This lower bound can be used in queries to eliminate objects if L∞ (Ψ (x), Ψ (y)) >
r, thus avoiding costly distance computations. Naturally, as the number of piv-
ots increases the pruning effect improves. The technique has been studied by
different authors in the past [5,6,3,7].

2.3 D-Index
The D-index was introduced by [8] and refined by [2]. The index partitions data
into h levels and 2ni − 1 buckets per level i. A function decides in which level
and bucket any given object must be placed. The main characteristic of the
index is that when a range query has a search radius up to some predefined ρ,
only one bucket must be accessed per level. Figure 1 shows how the structure
would look like for a particular example. Each square drawn with solid lines
represents a bucket that is used to store objects. For levels 0 to 2, these buckets
are called separable buckets. Dashed squares represent exclusion buckets that hold
objects that are recursively partitioned to the next level. At the very bottom,
the exclusion bucket E is employed to capture all the objects that cannot be
stored in all the previous levels.
To construct the D-index, for each level i, ni pivots pik (0 ≤ k < ni ) are
selected, and the corresponding median distances mik of all the objects that
entered the level and each pivot pik are calculated. The excluded middle parti-
tioning scheme [9] is employed to split the data. For x ∈ D, the split function
bps implements this partitioning scheme:
⎧
⎨0 if d(x, pik ) ≤ mik − ρ
bpsk,ρ
i (x) = 1 if d(x, pik ) > mik + ρ (2)
⎩
− otherwise

Objects within the interval ]mik − ρ, mik + ρ], belong to the exclusion set. In this
case, the function returns the value “−”. Separable sets are created from objects
446 A.J. Müller-Molina and T. Shinohara

Level 0, n0 = 2
(4 buckets)

Level 1, n1 = 2
(4 buckets)

Level 2, n2 = 1
(2 buckets)

Exclusion bucket

Fig. 1. Example of the structure of the D-Index.

for which the bps function returns 0 or 1. An example of the split function bps
is illustrated in Figure 2. The exclusion set of width 2ρ contains the triangle-
shaped objects. The innermost circle represents one separable set, containing
circle-shaped objects. Finally, the square-shaped objects belong to the second
separable set. Given a range query such that r ≤ ρ, the function bps guarantees
that only one separable set is accessed. This property is called the separable
property.

mik

pik 2ρ

Fig. 2. Example of the bps function

The separation of data into three sets is not sufficient for most applications.
Therefore, a higher order split function can be defined by concatenating the
results bk of each function bpsk,ρ ρ
i . The joint n-split function is denoted by bpsi (x)
and the return value is a string b = (b0 , . . . , bni −1 ) where bk ∈ {0, 1, −}.
To store each bucket, an addressing technique that receives a string and re-
turns a bucket identification number was proposed:
 ni −1 k
k=0 2 bk if ∀k bk = −
b = (3)
2ni otherwise

When no “−” elements are returned by any of the bpsk,ρi functions, a number
smaller than 2ni is returned. By using the function bpsρi and the . operator,
 On the Configuration of the Similarity Search Data Structure D-Index 447

a bucket identification number can be assigned to each object o ∈ D. From the

previous discussion, the following bucket structure is created:

B0,0 , B0,1 , . . . , B0,2n0 −1

B1,0 , B1,1 , . . . , B1,2n1 −1
..
.
Bh−1,0 , Bh−1,1 , . . . , Bh−1,2nh−1 −1
E

Each bucket Bi,j contains a header that holds the pivot vectors for each object
stored in the bucket. The header is sorted by the first pivot, and the objects
can be stored in an alternate location. This bucket layout was called elastic
bucket [2].

Insert. Figure 3 shows the insertion procedure. For every level, the algorithm
attempts to find a bucket. If a bucket is found, the object is inserted (line 4)
and the process terminates. If all the levels are tried without success, the object
is inserted into the exclusion bucket E (line 8).

Range search. Figure 4 shows the range search algorithm. The algorithm eval-
uates each level and reads one or more buckets depending on r. If r is not greater
than ρ (line 4), two possibilities arise. In the case where both the separable set
and the exclusion set are intersected by the query (line 5), one separable bucket is
accessed. Otherwise, no separable bucket is accessed for the current level because
the query region is contained exclusively in the exclusion set of the level.
When the radius is greater than ρ, more than one separable set must be ac-
cessed per level. Recall that when invoking bpsr−ρ
i , a string b = (b0 , b1 , . . . , bni −1 )
is returned. To explore all the possible buckets, we must replace each bk with
value “−” alternately with 0 and 1, and generate all the possible bucket identifi-
cation number combinations. This is accomplished by calling function si in line
10.
In Figure 5, the function si receives a string b initialized with each bk = 0.
This string is used as a template to generate all the possible bucket identification
numbers. The function will process each split function bpsk,r−ρ i until a bucket
identification number can be created. When k < ni , the kth split function’s
value is computed(line 7). If the resulting value is “−” (line 9), the bucket
identification numbers where bk = 0 and bk = 1 are set in b, and the next split
function is recursively computed. Otherwise, the value obtained by bpsk,r−ρ i is
replaced and the next split function k + 1 is processed in line 12. When k = ni ,
a bucket identification number is already in b and bucket Bi,
b can be searched.

2.4 Parameter Configuration Issues

In this Section, we discuss problems related to the parameter configuration of
the D-index. Our discussion assumes that search is CPU bounded, and therefore
448 A.J. Müller-Molina and T. Shinohara

Insertion Algorithm
1: function insert(o ∈ D)
2: for i = 0 to h − 1 do
3: if bpsρi (o) < 2ni then  o is not in E?
4: Store o in bucket Bi,bpsρi (o)
5: exit
6: end if
7: end for
8: Store o in bucket E
9: end function

Fig. 3. Insertion algorithm of the D-index

Q denotes the query region

q Query object, r Range
1: function RangeSearch(q, r)
2: res ← {}  Result set
3: for i = 0 to h − 1 do
4: if r ≤ ρ then  Search radius up to ρ
5: if bpsρ−r
i (q) < 2ni
then
 Not exclusive containment in exclusion bucket
6: res ← res ∪ Q ∩ Bi,bpsρ−r (q)
i
7: end if
8: else  Search radius is greater than ρ
9: b ← (b0 , b1 , . . . , bni −1 ) where bi = 0
10: res ← res ∪ si (q, r, k, b)
11: end if
12: end for
13: return res ∪ (Q ∩ E)
14: end function

Fig. 4. Range search algorithm

our objective is to minimize distance computations. From our experience [10,1],

this is very common when handling trees or other high dimensional objects that
utilize complex distance functions.
Dohnal et al. [2] pointed out that “generally” the number of pivots (or split
functions) per level is assigned in decreasing order. In the following, we assume
this and therefore level i = 0 has the greatest number of pivots. We also assume
that the pivots used in addressing are the pivots using in filtering.
The D-index structure presents two problems when a large number of pivots is
used. First, as the number of pivots increases, the probability that an object will
fall into the exclusion bucket of the next level becomes higher. This is because as
the number of pivots for a level increases, the probability of obtaining a “-” also
increases. This follows from the definition of bpsρi (Section 2.3). In this setting,
the D-index structure degenerates, and the exclusion bucket E tends to contain a
large percentage of the data. The upper layers, with the biggest amount of pivots,
 On the Configuration of the Similarity Search Data Structure D-Index 449

b = (b1 , b2 , . . . , bni ) string

k index to be changed
v the new value for bk
1: function set(b, k, v)
2: Change the kth element of b to v
3: return b
4: end function
i: Current level, q: Query object, Q query region
r Range, k Processing the kth element of b
b String b = (b0 , b1 , . . . , bni −1 ) where bi ∈ {0, 1}
5: function si (q,r,k,b)
6: if k < ni then  Still creating bucket id.
7: bit ← bpsk,r−ρ i (q)
8: if bit == “−” then  Process both bk .
9: return si (q, r, k + 1, set(b, k, 0))∪
10: si (q, r, k + 1, set(b, k, 1))
11: else  Process only one bk .
12: return si (q, r, k + 1, set(b, k, bit)))
13: end if
14: else  Bucket id is complete.
15: return Q ∩ Bi,b  Search bucket b
16: end if
17: end function

Fig. 5. Combination generator for queries r > ρ

store a small subset of the data. This represents a waste, as pivot vectors of the
upper levels are the most helpful in reducing distance computations because the
pivot set size is larger.
To illustrate this problem, in Table 1 the distribution of objects per level for
different ρ and h = 9 in two real data-sets is displayed. The description of the
data-sets is given in Section 4. The percentage of objects that pass to the next
level is shown for each level. In this example, the D-index will search data in
a limited number of buckets. As ρ grows bigger, the percentage of data stored
in the exclusion bucket increases. This behavior can be observed with relatively
large ρ and pivot sets. The large number of objects stored in the exclusion
bucket suggests that D-index behaves like sequential search on pivot vectors [6].
A solution is to reduce the number of pivots. For fast distance functions this is
possible. Nevertheless, as the complexity of a distance function increases, it is
desirable to have more pivots so that the lower bound becomes more precise.
The second problem occurs when ρ is small. In this case, the probability
that an object will fall into the exclusion zone will be also reduced. Recall from
Section 2.3 that when r > ρ, the function si must be called. This implies that
h−1
at most i=0 2ni buckets would have to be accessed. The generation of bucket
identification number combinations by si is an expensive operation.
Matching with a large set of pivots can negatively affect the distribution of buck-
ets. On the other hand, if distribution is improved by changing ρ, exponential I/O
450 A.J. Müller-Molina and T. Shinohara

Table 1. Level distribution for each level h. The column “%” contains the percentage
of objects that fall into the next level i + 1. The last row represents the percentage of
objects stored in the exclusion bucket. The column ni contains the number of pivots
for level i.

MTD ρ = 8 TED ρ = 7
i ni % ni %
0 16 95.06 30 99.99
1 15 93.26 28 99.99
2 13 92.16 26 99.99
3 12 91.65 24 99.99
4 11 90.61 22 99.99
5 10 90.11 20 99.99
6 9 89.68 19 99.99
7 8 88.88 18 99.99
8 7 88.14 17 99.99

occurs. To avoid these problems, we propose a simplification of the D-index that

attempts to alleviate these issues.

3 Proposed Technique

In this Section, we introduce the SD-index, a simplification of the D-index that

attempts to solve the problem described in the previous section. The main ob-
jective of our modification is to avoid the creation of levels, as they reduce the
size of the pivot vectors, while maintaining the bucket addressing scheme used
in the D-index. Both objectives can be achieved by setting ρ to 0.
Let us explain the implications of this change. First, as bpsk,r−ρ
i never returns
“−”, the exclusion bucket disappears. As a consequence, only one level i = 0
exists. The bucket identification number function (3) remains unmodified.
The insert algorithm (Figure 3) can be reduced to only line 4. Likewise, the
range search algorithm of Figure 4 can be reduced to lines 9 and 10. The main
problem our modification introduces is that si must access at most 2n0 buckets.
As mentioned, this issue occurs to the D-Index when r > ρ. To alleviate this, we
implemented two heuristics that try to reduce the amount of buckets that must
be accessed.

3.1 Heuristic 1: Bucket Visiting Order

The first heuristic is designed to guide si to buckets that are more likely to contain
objects close to the query. Our current implementation works only for discrete
numbers but it can be extended for real valued numbers. We pre-compute n0 arrays
fk (0 ≤ k < n0 ) that contain the distance distribution for each pivot pk0 . The
value fk [i] represents the number of objects that are at distance i from pivot pk0 .
For o ∈ D and a range r, the functions lk (o, r) and rk (o, r) are defined as:
 On the Configuration of the Similarity Search Data Structure D-Index 451


m 0k

lk (o, r) = fk [j],
j=d(p0k ,o)−r


d(p0k ,o)+r
rk (o, r) = fk [j].
j=m0k

During block number generation in si , if lk (o, r) ≥ rk (o, r) then the recursion

will proceed first with bk = 0 and then bk = 1 will be executed. Otherwise
bk = 1 will go first followed by bk = 0. With this, we intend to visit first bucket
identification numbers that have more objects intersecting the query rectangle
on a certain “side” of the median.

3.2 Heuristic 2: Empty Bucket Pruning

A total of 2n0 bucket identification numbers can exist. Nevertheless, in practice

only a subset of all the possible buckets is actually used. The second heuristic is
designed to avoid accessing block identification numbers that do not exist. We

i: Current level, q: Query object, Q: query region

r: Range, k: Processing the kth element of b
b: String b = (b0 , b1 , . . . , bni −1 ) where bi ∈ {0, 1, −}
1: function si (q,r,k,b)
2: if not (b0 , . . . , bk ) ∈ prk then
3: return {}  Heuristic 2
4: end if
5: if k < ni then
6: bit ← bpsk,r−ρ
i (q)
7:  Multiple bk to process
8: if bit == “−” then
9:  Heuristic 1
10: if lk (q, r) ≥ rk (q, r) then
11: return si (q, r, k + 1, set(b, k, 0))∪
12: si (q, r, k + 1, set(b, k, 1))
13: else
14: return si (q, r, k + 1, set(b, k, 1))∪
15: si (q, r, k + 1, set(b, k, 0))
16: end if
17: else  One bk to process
18: return si (q, r, k + 1, set(b, k, bit)))
19: end if
20: else
21: return Q ∩ Bi,b  Search bucket b
22: end if
23: end function

Fig. 6. Combination generator for SD-index

452 A.J. Müller-Molina and T. Shinohara

keep n0 sets prk , each of them containing strings of size k + 1 (0 ≤ k < n0 ). At

insertion time, for each string b = (b1 , b2 , . . . , bn0 ) generated by bpsρi , we extract
the sub-strings a0 = (b0 ), a1 = (b0 , b1 ), . . . , an0 −1 = (b0 , . . . , bn0 −1 ). Each ak will
be inserted into prk . In si , for the kth split function, we can verify if the block
prefix we have constructed belongs to prk . If it belongs to prk , then we can
proceed with the given combination, otherwise search should stop at this point.
This heuristic can be implemented with a trie, or by using a bloom filter [11] for
each prk . A better approach would be perhaps to do a sequential scan on the
binary strings and use the hamming distance as an disk read order criteria. This
would guarantee that the number of shared partitions is maximum.
In Figure 6, we rewrite the function si to add the heuristics described above.
Heuristic one is located between lines 10 and 16, and heuristic two is located at
line 2.
If only distance computations are taken into account, the SD-index is optimal
because all the objects are mapped with an equal number of pivots n0 . As n0
is the largest number of pivots for any level, and only a small percent of the
data is stored in the first level, the D-index is likely perform more distance
computations for large ρ. On the other hand, as ρ reaches zero, more data is
storedin the upper levels of the index. In this case, the D-index must access at
most h−1 i=0 2
ni
buckets. The SD-index will access at most 2n0 buckets.

4 Experiments

We evaluated the SD-index on two real data-sets:

– M T D: 350000 trees compared by the mtd [12] tree distance metric (O(n2 )).
The trees are binary program fragments [10] used by an open source license
violation detector. Each tree has up to 500 nodes.
– T ED: 245000 trees compared by the tree edit distance [13] metric (O(n3 )).
This data-set is a subset of M T D and includes trees that have up to 20
nodes.

4.1 Computational Costs

We have chosen distance functions with relatively high complexity that require
many pivots to become feasible to match. As a rough estimation, one distance
computation takes as long as the computation of the L2 distance on a pair of
integer vectors of 6500 dimensions for mtd and 890000 dimensions for ted.
All our experiments present the averaged results of querying the data-set with
an exclusive set of 400 different objects. The maximum range query radii was
selected so that roughly 20% of the data is returned. The distance distributions
between data and query sets are illustrated in Figure 7.
We measured disk access operations, distance computations and total execu-
tion time. We implemented the algorithms in Java and executed them on Intel
Xeon computers running at 2.66GHz and with 4GB of memory on Ubuntu Linux
 On the Configuration of the Similarity Search Data Structure D-Index 453

Data distribution between data and queries

0.1
0.09 MTD
T ED
0.08
0.07

Frequency
0.06
0.05
0.04
0.03
0.02
0.01
0
1 10 100 1000
Distance

Fig. 7. Distance distribution for data-sets T ED and M T D

7.10. The GPL source code along with the data-sets can be downloaded1 . All
the indexes were built on top of the same framework2.

4.2 Indexing

Besides D-index and SD-index, we have employed sequential search as a baseline.

Additionally, we used the P+tree [14] combined with pivot filtering to compare
our technique with a relatively new high dimensional index. After several trials,
the best setting we found for parameter od for P+tree was 12. We could not
execute the D-index for T ED with ρ < 7 because si performed worse than
sequential search. This is due to the exponential number of read operations
required when ρ > r.
By running several trials, we found that the best number of pivots is 16 for
dataset M T D and 30 for dataset T ED. Pivot sets for SD-index and P+tree
are equal. For D-Index, different sets were generated for each level, and nine
levels were created. Table 1 shows the order for each level of D-index. The pivot
selection strategy employed is called incremental selection and it is described
by [15]; the parameters l = 1000 and m = 1000 were used.
Table 2 shows index creation time without considering pivot selection. For dif-
ferent parameters of D-index, we present an average of the running times. P+tree
took less time in dataset T ED because the current implementation avoids du-
plicate computations. With some effort SD-index could also be optimized in the
same way. On the contrary, the D-index cannot be optimized because most of
the time is spent creating the levels, a step that cannot be avoided.
Table 3 shows the distribution of objects on the buckets. Clearly, SD-index is
able to distribute the data better among the buckets. In the case of D-index, as ρ
becomes smaller, buckets have less objects and the standard deviation becomes
smaller.
1
http://www.obsearch.net/prl.tar.gz
2
http://www.obsearch.net
454 A.J. Müller-Molina and T. Shinohara

Table 2. Index creation time in hours

Index M T D T ED
D 0.80 97
SD 0.15 13.75
P + 0.11 7

Table 3. Object Distribution per Bucket. Average, σ, and minimum and maximum
number of objects per bucket. The “Count” column contains the total number of buck-
ets created for SD-index and different ρ values of D-index.

Index Avg. σ Min Max Count

T ED

SD 5 119 1 21 534 46681

D(7) 61250 122495 1 244993 4
SD 55 421 1 15 241 6338
D(1) 85 1607 1 99980 4067
MT D

D(4) 843 13254 1 269502 414

D(8) 2477 25928 1 307854 141
D(12) 4365 35875 1 321122 80

4.3 Performance Results

In Figure 8, the results for M T D show the performance of D-index for different
values of ρ. As expected when ρ = 1, D-index reduces the amount of computa-
tions because the pivot vectors are larger. This is because a greater number of
objects is inserted into the upper levels of the structure. SD-index and P+tree
benefit from an optimal pivot vector size, executing less distance calculations.
Regarding bucket access counts, ρ = 12 and ρ = 8 are accessing data in a small
number of buckets that contain a large percentage of the data. For example, in
Table 1, the exclusion bucket of ρ = 8 holds 88% of the data. Although few
buckets are accessed, note that data read is considerable because the buckets
that are accessed have a large number of pivot vectors. From an IO perspective,
the P+tree is lagging behind D-index and SD-index because it performs many
disk access operations. For D-index, ρ = 12 performs better than ρ = 1 because
r ≤ ρ and this implies that no more than one bucket is being accessed per level.
Recall from Table 1 that in this configuration, D-Index is working like sequential
search on the pivot vectors.
As the distance computation count reflects, ρ = 1 minimizes distance compu-
tations. As expected, smaller ρ increases the number of objects that are stored in
the upper level s of the structure. In those levels, the lower bound l∞ has better
resolution because the pivot set cardinality is the largest. Time-wise, ρ = 1 lags
behind other configurations because most of the time is spent accessing an expo-
nential number of buckets. The heuristics introduced in Section 3 help SD-index
to obtain the best of the two extremes by both reducing distance computations
and execution time at the expense of moderate I/O increase.
 On the Configuration of the Similarity Search Data Structure D-Index 455

Disk Access Count, M T D Distance Computations, M T D Execution Time in Seconds, M T D

00

6
00

0.
Seq = 350000 Seq = 3.575
00

60
10
0

5
00

00

0.
ρ=1
10

50
ρ=4

0
00

4
ρ=8

00

0.
10

40
ρ = 12

0
0

3
SD

00
10

0.
30
P+

0
10

2
00

0.
20
0

1
1

00

0.
10
1

0
,1
,3

,1
,3

,1
,3
0.
1
1
3
1
3
1
3

1
1
3
1
3
1
3

1
1
3
1
3
1
3
12
12

12
12

12
12
0,
3,
3,
6,
6,
9,
9,

0,
3,
3,
6,
6,
9,
9,

0,
3,
3,
6,
6,
9,
9,
Parameters r and k Parameters r and k Parameters r and k

Fig. 8. Results for data-set M T D. Performance comparison among SD-index, P+tree

and different ρ configurations of the D-index.

Disk Access Count, T ED Distance Computations, T ED Execution Time in Seconds, T ED

00

30
00
00

25

Seq = 245000
10

Seq = 418

25
0
00

0
00
10

20

20
00
10

0
00

SD
15
15
0
10

ρ=7
0

P+
10
00
10

10

5
00
1

50

0
1

3
0,

2,

2,

5,

5,

7,

7,
1

0
0.
1

3
0,

2,

2,

5,

5,

7,

7,

0,

2,

2,

5,

5,

7,

7,

Parameters r and k Parameters r and k Parameters r and k

Fig. 9. Results for data-set T ED. Performance comparison among SD-index, P+tree
and a ρ = 7 D-index.

Figure 9 shows the results for data-set T ED. As in data-set M T D, SD-index

outperforms D-index and P+tree in distance computations and execution time.
Although 3x to 4x improvements can be observed for distance computations, the
execution performance is only improved 1.5x. This is partially caused by the over-
head of accessing secondary storage. In general, distance function cost dominates
the execution. D-index performed worse than sequential search for ρ < 7. In fact,
we could not complete any of the experiments in an acceptable time frame because
for ρ < r, a large number of buckets was read by function si .

4.4 Discussion
When ρ is large, D-index resembles a sequential search on the pivot projection.
The fact that the D-index performs well for large ρ is not surprising as it has been
experimentally verified that sequential search based on pivot vectors outperforms
other hierarchical methods under certain circumstances [6]. The P+tree performs
the largest amount of read operations in all the experiments. On the other hand,
SD-index and P+tree are very close to each other in high dimensions because
456 A.J. Müller-Molina and T. Shinohara

their pivot set cardinality is the largest possible. The advantage obtained by
SD-index comes from heuristic one, as we descend first to buckets that are more
likely to have more objects.

5 Conclusion

We have studied the D-index in the context of high dimensional objects. Our
modification, the SD-index, consistently outperforms the D-index in different
configurations. Additionally, it guarantees that fewer distance computations will
be executed as the pivot set cardinality is maximized.
The D-index cannot scale well for small ρ and range r ≥ ρ and when the
number of pivots is large. In this case, the index must access an exponential
number of buckets that depends on the number of pivots. In our experiments
D-Index with 30 pivots was considerably slower than plain sequential search of
a very costly distance function. On the other hand, if we set a large ρ such that
r ≤ ρ, the index degenerates into a sequential search.

References

1. Müller-Molina, A.J., Shinohara, T.: Fast approximate matching of programs for

protecting libre/open source software by using spatial indexes. In: SCAM 2007,
pp. 111–122. IEEE Computer Society, Washington (2007)
2. Dohnal, V., Gennaro, C., Savino, P., Zezula, P.: D-index: Distance searching index
for metric data sets. Multimedia Tools Appl. 21(1), 9–33 (2003)
3. Micó, L., Oncina, J., Vidal, E.: An algorithm for finding nearest neighbours in
constant averagetime with a linear space complexity. In: Recognition Methodology
and Systems, pp. 557–560 (1992)
4. Ruiz, E.V.: An algorithm for finding nearest neighbours in (approximately) con-
stant average time. Pattern Recogn. Lett. 4(3), 145–157 (1986)
5. Shinohara, T., Ishizaka, H.: On dimension reduction mappings for approximate
retrieval of multi-dimensional data. In: Progress in Discovery Science, London,
UK, pp. 224–231. Springer, Heidelberg (2002)
6. Filho, R.F.S., Traina, A.J.M., Caetano Traina, J., Faloutsos, C.: Similarity search
without tears: The omni family of all-purpose access methods. In: Proceedings
of the 17th International Conference on Data Engineering, pp. 623–630. IEEE
Computer Society, Washington (2001)
7. Micó, M.L., Oncina, J., Vidal, E.: A new version of the nearest-neighbour approxi-
mating and eliminating search algorithm (aesa) with linear preprocessing time and
memory requirements. Pattern Recogn. Lett. 15(1), 9–17 (1994)
8. Gennaro, C., Savino, P., Zezula, P.: Similarity search in metric databases through
hashing. In: MIR 2001, pp. 1–5. ACM, New York (2001)
9. Yianilos, P.N.: Excluded middle vantage point forests for nearest neighbor search.
Technical report, NEC Research Institute, Princeton, NJ (1998)
10. Müller-Molina, A.J., Shinohara, T.: On approximate matching of programs for
protecting libre software. In: CASCON 2006, pp. 275–289. ACM Press, New York
(2006)
 On the Configuration of the Similarity Search Data Structure D-Index 457

11. Bloom, B.H.: Space/time trade-offs in hash coding with allowable errors. ACM
Commun., 422–426 (1970)
12. Müller-Molina, A.J., Hirata, K., Shinohara, T.: A tree distance function based on
multi-sets. In: Chawla, S., Washio, T., Minato, S.-i., Tsumoto, S., Onoda, T., Ya-
mada, S., Inokuchi, A. (eds.) PAKDD 2008. LNCS, vol. 5433, pp. 87–98. Springer,
Heidelberg (2009)
13. Demaine, E., Mosez, S., Rossman, B., Weimann, O.: An optimal decomposition
algorithm for tree edit distance. In: ALP, pp. 146–157. Springer, Heidelberg (2007)
14. Zhang, R., Ooi, B.C., Tan, K.L.: Making the pyramid technique robust to query
types and workloads. In: ICDE 2004, p. 313. IEEE Computer Society, Washington
(2004)
15. Zezula, P., Amato, G., Dohnal, V., Batko, M.: Similarity Search: The Metric Space
Approach. Springer, Secaucus (2005)
 Automatic Chinese Text Classification
Using N-Gram Model

Show-Jane Yen1, Yue-Shi Lee1, Yu-Chieh Wu1,

Jia-Ching Ying2, and Vincent S. Tseng2
1
Dept. of Computer Science and Information Engineering, Ming Chuan University
5 The-Ming Rd., Gwei Shan District, Taoyuan County 333, Taiwan
{sjyen,leeys}@mail.mcu.edu.tw
2
Dept. of Computer Science and Information Engineering, National Cheng Kung University
1 University Rd., Tainan City 701, Taiwan
[email protected], [email protected]

Abstract. Automatic Chinese text classification is an important and well-

known research topic in the field of information retrieval and natural language
processing. However, past researches often ignore the problem of word
segmentation and the relationship between words. This paper proposes an N-
gram-based language model for Chinese text classification which considers the
relationship between words. To prevent from the out-of-vocabulary problem, a
novel smoothing method based on logistic regression is also proposed to im-
prove the performance. The experimental result shows that our approach out-
performs the previous N-gram-based classification model above 11% on
micro-average F-measure.

Keywords: Text Classification, N-gram, Feature Selection, Word Segmenta-

tion, Logistic Regression.

1 Introduction
In recent years, there are more and more digital text services on the internet with the
rapid growth of the World Wide Web. Examples include Google, Yahoo, etc. Search
engine does not only provide user with the information, but also manage the informa-
tion effectively. In order to make it easy, many search engine systems classify the
texts in advance to organize their taxonomy. The methods of automatic text classifica-
tion help people to classify the documents. At present, many techniques have been
proposed to deal with the text classification problems. Yang and Liu [21] experiment
on the news text of Reuter. They reported that support vector machine (SVM) and k-
nearest neighbor (KNN) classifiers achieved the best accuracy in comparison to the
other four methods. Sebastian [14] pointed out that SVM had better performance in
general case.
However, support vector machine classifiers still have many problems. First, it
is designed to solve the binary-class classification problem. Therefore it should be
converted to handle multi-class classification problem. Second, the support vector

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 458–471, 2010.
© Springer-Verlag Berlin Heidelberg 2010
 Automatic Chinese Text Classification Using N-Gram Model 459

machine classifiers generally could not generate probability. Most support vector
machine classifiers represent the similarity between the class and the text using the
cosine similarity.
Unlike probability, the cosine similarity is very abstractive to represent the similar-
ity between the class and the text. These problems are solved by Tipping [17]. He
combines the logistic regression in the support vector machine classifiers to generate
the probability. Meanwhile, the multi-class classification problem also can be trans-
formed into several binary-class classification problems. Nevertheless, a known
problem with the Tipping’s approaches is their relative inability to scale with large
problems like text classification. Fortunately, Silva and Ribeiro [15] solved the
problem by the method of dimension reduction.
Nonetheless, these standard approaches to Chinese text categorization has so far
been using a document representation in a word-based “input space” [6], i.e. as a
vector in some high (or trimmed) dimensional Euclidean space where each dimension
corresponds to a word. The advantages of this method are effective and high feasibil-
ity, but it assumes each word is independent with each other. Hence we need to solve
word segmentation first for Chinese text. Most text classifiers treat each feature are
mutually independent, word segmentation in Chinese is a difficult problem and we
could not ensure that every words which is segmented by some word segmentation
process are mutually independent. To solve it, character level N-gram models
[3][4][12][13][16] could be applied. These approaches generally outperform tradi-
tional word-based methods in term of accuracy. Moreover, these approaches can
avoid the word segmentation step and model the word-by-word relations [5][11].
Cavnar and Trenkle [3] proposed a classifier for English categorization, named
N-gram-based text categorization, which works with letters level feature. They used
N-grams as features for traditional feature selection process and then deployed classi-
fiers based on calculating feature-vector similarities. Cavnar and Trenkle’s approach
has a shortcoming. It was intractable when there is a large feature set, and standard
feature selection approaches do not always cope well in such circumstances. For ex-
ample, given sufficient features, the cumulative effect of uncommon features can still
have an important effect on classification accuracy, even though infrequent features
contribute less information than common features individually. Teahan and Harper
[16] used a PPM (prediction by partial matching) model for text classification where
they seek a model that obtains the best compression on a new document. Later, Peng,
et al. [12][13] proposed a classifier for Chinese classification without the word seg-
mentation step which works with letters character feature by means of using N-gram
model. In their experiment, their approach outperforms Cavnar and Trenkle’s ap-
proach [3], Support Vector Machine and naive Bayes in most case. Although Peng, et
al. proposed a novel classification to improve the traditional classifier for Chinese text
classification, the approach also involve in several problems. First, the research of
Peng, et al. entirely does not process the feature selection. If the testing text is too
large, the classifier maybe uses a lot of unnecessary conditional probability to esti-
mate the probability of a document to assign some class. Because the N-gram model
uses the expression (1) to estimate the probability of an event that a sequence of the
words w1…wT occur in a data set.
T

P( w1 ...wT ) = ∏ P( w | w
i =1
i i − n +1 ...wi −1 ) (1)
460 S.-J. Yen et al.

Therefore, if there exists some N-gram w1…wn whose conditional probability

P(wn|w1…wn-1) are extremely similar in each class, the N-gram w1…wn would not
affect the result of classification. But standard methods of feature selection cannot be
used in this kind of approach.
In this paper, we propose a novel feature selection which fits N-gram model for
Chinese text classification. First, we consider the difference of the conditional prob-
ability of some N-gram in each class to select significant features. Second, there are
many kinds of method for N-gram smoothing to solve sparse data problems, but ap-
proaches [12][13] just use the Katz’s back-off estimator [2] without further adjusting.
Because the original purpose of N-gram model is not to classify a text and the
smoothing estimator is the kernel of the model, we believe that the model should be
smoother and more suitable real world event. The logistic regression model is a fa-
mous classification model, it use a sigmoid function to estimate the probability.
Therefore, a novel smoothing estimator using logistic regression is presented in this
paper. Third, past research studies [12][13] treat the whole document as a sequence of
words. The basis ideal of N-gram model is that the probability happening in some
word, wi, relates to the first n-1 continuous words. This means that when we estimate
the probability of some word, we perhaps consult it to the last words of last sentence.
However, we can regard each sentence as an independent event. In accordance with
this point, our approach adjusts the whole model.

2 Logistic-Regression-Based N-Gram Models

The framework of our approach is shown in Figure 1.

Fig. 1. The framework of our approach

The first step is to discover the frequency of every k-grams where 1 ≤ k ≤ n by

scanning the training dataset and estimate each conditional probability of possible n-
gram with maximum likelihood estimation. Next, we adopt chi-square statistic to
 Automatic Chinese Text Classification Using N-Gram Model 461

select important n-grams. In the third step, we build a logistic regression model to
estimate each conditional probability P(wi|wi-n+1…wi-1). According to the third step we
can make sure each conditional probability P(wi|wi-n+1…wi-1) converge within the
interval 0 to 1(i.e. 0<P(wi|wi-n+1…wi-1)<1).

2.1 Significant N-Grams

According to the previous researches for feature selection [1], the chi-square statistic
is better than mutual information and information gain in selecting useful features.
However, the chi-square statistic cannot be applied to N-gram-based model without
amendment. Therefore, we amend it to suit N-gram-based model. First, we use an
example to expound what kind feature is significant in N-gram-based model as fol-
lows: Assume we wish to calculate a probability of a word sequence s=w1w2w3w4
occurred in two class, C1 and C2, using bi-gram model and the distribution of w1w2
and w1 between C1 and C2 is given in Table 1.

Table 1. The distribution of w1w2 and w1 between C1 and C2

#( w1w2) #( w1)
C1 6 12
C2 1 2

PC ( w1w2 w3 w4 ) = PC ( w1 ) × PC ( w2 | w1 ) × PC ( w3 | w2 ) × PC ( w4 | w3 )
1 1 1 1 1

PC ( w1w2 w3 w4 ) = PC ( w1 ) × PC ( w2 | w1 ) × PC ( w3 | w2 ) × PC ( w4 | w3 )
2 2 2 2 2

Because PC ( w2 | w1 ) = PC ( w2 | w1 ) , we believe PC ( w2 | w1 ) and PC ( w2 | w1 ) are un-

1 2 1 2

concerned with the result of classification. Reversely, if there is an n-gram w1w2…wn

whose distribution over all classes is more different than the distribution of (n-1)-
gram w1w2…wn-1 over all classes, we can say that the n-gram w1w2…wn is significant.
Based on this idea, we not only consider the relation and inverse relation between
some n-gram and some class c, but also consider the relation and inverse relation
between the n-gram w1w2…wn and the (n-1)-gram w2…wn.
Because the n-gram w1w2…wn and the (n-1)-gram w2…wn have the distribution
over all classes, their contingency table should be different.

Table 2. Contingency table with n-gram w1w2…wn and class c

Assigning the document d to class c

Total
Assign Not Assign
Occurrence of the Occur A1 B1 A1+B1
n-gram w1w2…wn in
some document d Not Occur P1 Q1 P1+Q1
Total A1+P1 B1+Q1 N1
462 S.-J. Yen et al.

Table 3. Contingency table with (n-1)-gram w1w2…wn-1 and class c

Assigning the document d to class c

Total
Assign Not Assign
Occurrence of the (n- Occur A2 B2 A2+B2
1)-gram w1w2…wn-1 in
some document d Not Occur P2 Q2 P2+Q2
Total A1+P1 A2+P2 B2+Q2

In Tables 2 and 3, Let A1 be the number of times w1w2…wn and c co-occur, let B1
be the number of times w1w2…wn occurs without c, let P1 be the number of times c
occurs without w1w2…wn, let Q1 be the number of times neither w1w2…wn nor c, and
let N1 be the total number of n-gram occurred in corpus.
The chi-square statistic of n-gram w1w2…wn with class c is represented as follows:
N1 × ( A1Q1 − B1 P1 )
2

χ 2 ( w1w2 ...wn , c) =
( A1 + P1 )× (B1 + Q1 )× ( A1 + B1 )× (P1 + Q1 )
Identically, the chi-square statistic of (n-1)-gram w1w2…wn-1 with class c is repre-
sented as follows:
N 2 × ( A2Q2 − B2 P2 )
2

χ 2 ( w1w2 ...wn−1 , c) =
( A2 + P2 )× (B2 + Q2 )× ( A2 + B2 )× (P2 + Q2 )
It is clear that if an n-gram w1w2…wn whose chi-square value is more different than
chi-square value of (n-1)-gram w1w2…wn-1, then their distributions over all classes are
more different. Therefore, we propose a method as follows:

∑ F (w ...w , c)
1
F ( w1...wn , C ) = 1 n
C c∈C

⎧ χ 2 ( w1...wn , c)
⎪ 2 , if χ 2 ( w1...wn , c) ≥ χ 2 ( w1...wn−1 , c)
⎪ χ ( w ...wn −1 , c )
where F ( w1...wn , c) = ⎨ 2
1

⎪ χ ( w ...w , c)
1 n −1
, if χ 2 ( w1...wn , c) < χ 2 ( w1...wn−1 , c)
⎪⎩ χ 2 ( w1...wn , c)

The χ 2 ( w1 ...wn , c) is the chi-square statistic of n-gram w1w2…wn with class c, and the
χ 2 ( w1...wn−1 , c) is the chi-square statistic of (n-1)-gram w1w2…wn-1 with class c. We
can determine them by contingency table as follows: If the significance probability
(i.e. p-value) of the chi-square statistic of some n-gram w1w2…wn is less than the level
of significance, the n-gram is called significant n-gram. In experiment, we set the
level of significance equal to 5%, 10%, 15%, 20%, and 25%. Therefore, we can cal-
culate the significance probability of each n-gram. If the significance probability of
the n-gram is greater than the level of significance we set, we will not consider the n-
gram into the classifier.
 Automatic Chinese Text Classification Using N-Gram Model 463

2.2 Smoothing Estimator for N-Gram Model

N-gram model is a well-known technique in natural language processing domain. As

we know, the goal of language modeling is to predict the probability of natural word
sequences; or more simply, to put high probability on word sequences those actually
occur (and low probability on word sequences that never occur). Because of the heavy
tailed nature of language one is likely to encounter novel n-grams that were never
witnessed during training. Therefore, some mechanism for assigning non-zero prob-
ability to novel n-grams is a central and unavoidable issue [10]. In resent year, the
combining estimation has the best performance in most case. One standard approach
of combining estimation to smoothing probability estimates to cope with sparse data
problems is to use some sort of linear interpolation estimator.
Pli ( wi | wi −n+1...wi −1 ) = λ1 P( wi ) + λ2 P( wi | wi −1 ) + ... + λn P( wi | wi −n+1...wi −1 )

Where 0 < λi < 1 and ∑λ i i

= 1 . This parameter can be set automatically using the
Expectation-Maximization (EM) algorithm [10].
However, the estimation of some probability by some linear way is not perfect
enough. Especially, the occurrence of some n-gram is a kind of response of a human
been. Most statistician believe that logistic regression can solve the problem which
the estimation is a kind of response of a human been. The logistic regression assumes
that there is a sigmoid function, called response function, to represent the probability
of occurrence of a response as follows:
eα +α x +α x +...+α x
0 1 1 2 2 n n

P (r = 1) =
1 + eα +α x +α x +...+α x
0 1 1 2 2 n n

In the following, we take an example to describe the logistic regression modeling to

estimate the conditional probability of bi-grams in some class c. Table 4 represents
the response variable and independent variables.

Table 4. The response variable and independent variables

Response Variable Independent Variables

#(wi-1wi) #(wi-1) PML(wi| wi-1) PML(wi)
n1 r1 x11 x21
n2 r2 x12 x22
… … … …
nj rj x1j x2j

We can treat #(wi-1) as the frequency we observe and #(wi-1wi) as the frequency of the
response which bi-gram is occurred. Assume there is a probability pi to represent the
probability of the response which bi-gram is occurred, the joint probability density
function is
j
⎛ ni ⎞ r
∏ ⎜⎜ ⎟⎟ pi (1 − p i )n − ri i i

i =1 ⎝ ri ⎠
464 S.-J. Yen et al.

1
According to the assumption of the logistic regression pi = , the joint
1 + eα +α x
0 1 1i +α 2 x2 i

probability density function can be written as

ri ni − ri
j
⎛ ni ⎞⎛ eα +α x +α x ⎞ ⎛ 1 ⎞
∏
0 1 1i 2 2i

⎜⎜ ⎟⎟⎜⎜ α +α x +α x ⎟
⎟ ⎜ α +α x +α 2 x2 i
⎟
i =1 ⎝ ri ⎠⎝ 1 + e ⎠ ⎝1+ e ⎠
0 1 1i 2 2i 0 1 1i

⎛ ni ⎞ α +α x +α x r
⎜⎜ ⎟⎟(e ) (1 + eα +α x +α x )
j

∏
− ni
= 0 1 1i 2 2i
i
0 1 1i 2 2i

i =1 ⎝ ri ⎠

Then we estimate the parameters (i.e. α 0 , α 1 , α 2 ) using maximum likelihood estima-

tion and produce the estimator of the conditional probability as follows:
eα +α P ( w |w )+α P ( w )
0 1 ML i i −1 2 ML i

Plo ( wi | wi −1 ) =
1 + eα +α P ( w |w )+α P ( w )
0 1 ML i i −1 2 ML i

Unlike linear interpolation, logistic regression is a sigmoid function. Therefore the

parameters (i.e. α 0 , α 1 , α 2 ) need not to be constrained. Besides, logistic regression
has another advantage that it will not over-evaluate conditional probability which is
estimated to be 0 by ML estimation as follows Figure 2.

Fig. 2. Comparison of logistic regression and linear interpolation

As we see, if the frequency of AB is equal to zero but the frequency of B is

very high, the conditional probability P(B|A) estimated by the linear interpolation is
more over-evaluate than the conditional probability P(B|A) estimated by the logistic
regression.
 Automatic Chinese Text Classification Using N-Gram Model 465

2.3 Language Models as Text Classifiers

Unlike the approach of Peng, et al. [12], our approach considers that a document is a
set of sentences, and a sentence is a word sequence. Because we believe that calculat-
ing the probability of a word considering the end of last sentence is unreasonable.
There is no poof the end of last sentence would affect the probability of a word in the
start of a sentence.
Assume we wish to classify a document d={s1,s2,…,s|d|} into a category
c ∈ C = {c1 ,..., c C } , where si = w1w2 ...w s . A natural choice is to pick the category c
i

that has the largest posterior probability given the text. That is,
c* = arg max Pr(c | d )
c∈C

Using Bayes rule, this can be rewritten as

c * = arg max P(c) Pr(d | c)
c∈C

Considering each sentence is independent of other sentence in a document, this can be

rewritten as
d

c * = arg max P (c)

c∈C ∏ Pr(s | c)
i =1
i

Using N-gram model applying to classification as

d
⎡ si
⎤
c* = arg max P(c)
c∈C ∏ ∏ Pr(w ⎢ ij
| wi ( j −n+1)...wi ( j −1) ∧ c)⎥
i =1 ⎣ j =1 ⎦
As the above equation, our model considers sentence level.

3 Empirical Evaluation
The Chinese data set we used has been previously investigated in [19]. The corpus is
a subset of the TREC-5 People's Daily news corpus published by the Linguistic Data
Consortium (LDC) in 2001. The entire data set contains 164,789 documents on 59
categories, including IT, economy, etc. However, we could not use the entire data for
our experiments, because there many categories include too few documents. There-
fore, if we use the entire data for our experiments, there will be the imbalanced distri-
bution data problem [22] in it. To void the imbalanced distribution data problem, we
just choose particular of the entire data. The entire data set which we choose contains
105,733 documents on 22 categories.

3.1 Experimental Paradigm

Our goal is to classify text into a set of categories, and each document is assigned a
single category. Each n-gram-based classification, such as Peng et al’s approach [12],
466 S.-J. Yen et al.

is generally set the parameter n. For the experiments on Chinese data, we follow
Peng et al’s approach and experiment on uni-gram model, bi-gram model, and tri-
gram model. In each case, we randomly select 90% of documents evenly from
among the remaining categories to form the training data. The testing set contains the
other 10% documents of original data set. The training set and testing set do no over-
lap and do not contain replicated documents. Then we divide the training data into
9 parts uniformly. We pick 5 parts as training data 1, and pick all parts as training
data 2.

3.2 Measuring Classification Performance

In the experiments, we measured classification performance by micro-averaged F-

measure scores. To calculate the micro-averaged score, we formed an aggregate con-
fusion matrix by adding up the individual confusion matrices from each category. The
micro-averaged precision, recall, and F-measure can then be computed based on the
aggregated confusion matrix.
For example, if there are |C| categories included in training data set. For each cate-
gory c ∈ C , we could determine a confusion matrix as follows.

Table 5. Confusion Matrix

Predicted
Text belong to ci Text NOT belong to ci
Actual
Text belong to ci ai pi
Text NOT belong to ci bi qi

The precision, recall, and F-measure of the category c are shown as follows.
ai ai 2 × precision × recall
precision = , recall = , F − measure =
ai + bi ai + pi precision + recall

The aggregated confusion matrix and the micro-averaged precision, recall, and F-
measure of the category c are shown as follows.

Table 6. Aggregated Confusion Matrix

Predicted
Actual Text belong to c ∈ C Text NOT belong to c ∈ C
C C
Text belong to c ∈ C A = ∑ ai P = ∑ pi
i =1 i =1
C C
Text NOT belong to c ∈ C B = ∑ bi Q = ∑ qi
i =1 i =1
 Automatic Chinese Text Classification Using N-Gram Model 467

A A
micro-averaged precision = , micro-averaged recall =
A+ B A+ P
2 × micro-averaged precision × micro-averaged recall
micro-averaged F − measure =
micro-averaged precision + micro-averaged recall

3.3 Experimental Result and Analysis

Figures 3 and 4 give the results of comparisons of our approaches and Peng et al.’s
approach with feature selection on training data set 2 in tri-gram and bi-gram
models. In Figures 3 and 4, we can see that our approach outperform Peng et al.’s
approach in most case. Tri-gram model outperform bi-gram model in most case.
Remarkably, tri-gram model smoothing by logistic regression is more stable than
bi-gram model smoothing by logistic regression, because the difference in F-
measure of tri-gram model smoothing by logistic regression between best case and
worst case is less than 25%, but the difference in F-measure of bi-gram model
smoothing by logistic regression between best case and worst case is greater
than 43%.

Fig. 3. Comparisons of our approaches and Peng et al.’s approach with feature selection on
training data set 2 in tri-gram model

Figures 5 and 6 give the results of comparisons of our approaches and Peng et al.’s
approach with feature selection on training data set 1 in tri-gram and bi-gram models.
In Figures 5 and 6, we can see that our approach outperform Peng et al.’s approach in
most case. There are evidences that the F-measure of back-off smoothing method will
not be improved by feature selection. Therefore, the F-measure of back-off smoothing
method we be deteriorated following the decreasing of the number of features.
468 S.-J. Yen et al.

However, if we just select 80% number of features to train the classifiers, the F-
measure of back-off smoothing method looks few better then the F-measure of logis-
tic regression smoothing method, because there are too many “significant” features
are filtered out in this case. We also can find that the effect of sentence segmentation
is marginally in bi-gram model, because bi-gram model considers that each word is
observed is just related to the last word. In other words, there are rare bi-grams in
head and tail of sentences.

Fig. 4. Comparisons of our approaches and Peng et al.’s approach with feature selection on
training data set 2 in bi-gram model

Fig. 5. Comparisons of our approaches and Peng et al.’s approach with feature selection on
training data set 1 in tri-gram model
 Automatic Chinese Text Classification Using N-Gram Model 469

Fig. 6. Comparisons of our approaches and Peng et al.’s approach with feature selection on
training data set 1 in bi-gram model

In the following, the results of the best cases of logistic regression smoothing
method and back-off smoothing method on each training data set in bi-gram model
and tri-gram model is represented in Table 7.

Table 7. The best cases of logistic regression smoothing method and back-off smoothing
method

Logistic Back-off
bi-gram tri-gram bi-gram tri-gram
training data set 2 0.800323 0.931512 0.596471 0.839820
training data set 1 0.739898 0.912540 0.513090 0.803358

In summary, first, generally our approach is better than Peng et al.’s approach in
most case. Second, feature selection has no help to classifier smoothing by back-off,
but very helpful to classifier smoothing by logistic regression. Third, sentence seg-
mentation impact on F-measure is marginally in classifier smoothing by back-off.
Fourth, logistic regression smoothing method generally outperform than back-off
smoothing method in most case. Fifth, tri-gram model is much better than bi-gram
model.

4 Conclusions and Future Work

In this paper, we have presented an n-gram-based model for Chinese text classifica-
tion. We use the logistic regression to smooth the probability of n-gram. In our ex-
periment, logistic regression smoothing outperform traditional back-off smoothing,
because logistic regression has the ability to process unknown terms and it will not
over-evaluate the conditional probability which originally is zero. Besides, we pro-
posed a novel feature selection method which is suitable to N-gram-based model. In
470 S.-J. Yen et al.

our experiment, we prove that it could improve the F-measure in most case. Third, we
consider a text as a set of sentences. According to our experimental result, this could
improve system performance especially in the case which use tri-gram model or
whose training data set is large enough.
In the future, we plane to find out a method to evaluate the relationship between
sentences. Since we believe that considering a text as a set of sentences is not an op-
timal assumption and there exist some fact could respect the relationship between
sentences. We also try to reduce the consumption of memory and calculative time of
CPU on feature selection.

Acknowledgments. Research on this paper was partially supported by National Sci-

ence Council grant NSC98-2221-E-130-019, NSC98-2622-E-130-001-CC3, and
NSC98-2221-E-130-022.

References
1. Aizawa, A.: Linguistic Techniques to Improve the Performance of Automatic Text Catego-
rization. In: 6th Natural Language Processing Pacific Rim Symposium, pp. 307–314
(2001)
2. Chen, S., Goodman, J.: An Empirical Study of Smoothing Techniques for Language Mod-
eling. In: 34th Annual Meeting of the Association for Computational Linguistics, pp. 310–
318 (1998)
3. Cavnar, W., Trenkle, J.: N-Gram-Based Text Categorization. In: 3rd Annual Symposium
on Document Analysis and Information Retrieval, pp. 161–175 (1994)
4. Damashek, M.: Gauging Similarity with N-Grams: Language-Independent Categorization
of Text. Science 267, 843–848 (1995)
5. Dumais, S., Platt, J., Heckerman, D., Sahami, M.: Inductive Learning Algorithms and Rep-
resentations for Text Categorization. In: 7th International Conference on Information and
Knowledge Management, pp. 148–155 (1998)
6. He, J., Tan, A., Tan, C.: On Machine Learning Methods for Chinese Document Categori-
zation. Applied Intelligence 18, 311–322 (2003)
7. Jiang, E.: Learning to Semantically Classify Email Messages. In: 2nd International Con-
ference on Intelligent Computing, pp. 664–675 (2006)
8. Joachims, T.: Text Categorization with Support Vector Machines: Learning with Many
Relevant Features. In: European Conference on Machine Learning, pp. 137–142 (1998)
9. Lam, W., Ruiz, M., Srinivasan, P.: Automatic Text Categorization and Its Application to
Text Retrieval. IEEE Transactions on Knowledge and Data Engineering 11, 865–879
(1999)
10. Manning, C.D., Schuetze, H.: Fundations of Statistical Natural Language Processing, pp.
191–227. MIT Press, Cambridge (2004)
11. Peng, F., Huang, X., Schuurmans, D., Cercone, N.: Investigating the Relationship of Word
Segmentation Performance and Retrieval Performance in Chinese IR. In: 15th Interna-
tional Conference on Computational Linguistics, pp. 72–78 (2002)
12. Peng, F., Huang, X., Schuurmans, D., Wang, S.: Text Classification in Asian Languages
without Word Segmentation. In: 6th International Workshop on Information Retrieval with
Asian Languages, pp. 41–48 (2003)
 Automatic Chinese Text Classification Using N-Gram Model 471

13. Peng, F., Schuurmans, D.: Combining Naive Bayes and N-Gram Language Models for
Text Classification. In: 25th European Conference on Information Retrieval Research, pp.
335–350 (2003)
14. Sebastian, F.: Machine Learning in Automated Text Categorization. ACM Computing
Surveys 34, 1–47 (2002)
15. Silva, C., Ribeiro, B.: Scaling Text Classification with Relevance Vector Machines. In:
IEEE International Conference on Systems, Man, and Cybernetics, pp. 4186–4191 (2006)
16. Teahan, W., Harper, D.: Using Compression-Based Language Models for Text Categoriza-
tion. In: Workshop on Language Models for Information Retrieval, pp. 83–88 (2001)
17. Tipping, M.: Sparse Bayesian Learning and the Relevance Vector Machine. Journal of
Machine Learning Research 1, 211–214 (2001)
18. Vapnik, V.: The Nature of Statistical Learning Theory. Springer, Heidelberg (1995)
19. Wu, Y.C.: Chinese Text Categorization with Term Clustering. M.S. Thesis, Mining Chuan
University (2003)
20. Yang, Y.: An Evaluation of Statistical Approaches to Text Categorization. Information Re-
trieval Journal 1, 69–90 (1999)
21. Yang, Y., Liu, X.: A Re-examination of Text Categorization Methods. In: 22nd Interna-
tional ACM SIGIR Conference on Research and Development in Information Retrieval,
pp. 42–49 (1999)
22. Yen, S., Lee, Y., Lin, C., Ying, J.: Investigating the Effect of Sampling Methods for Im-
balanced Data Distributions. In: IEEE International Conference on Systems, Man, and Cy-
bernetics, pp. 4163–4168 (2006)
 Genetic Algorithms Evolving Quasigroups with
Good Pseudorandom Properties

Václav Snášel1 , Jiřı́ Dvorský1, Eliška Ochodková1, Pavel Krömer1 ,

Jan Platoš1, and Ajith Abraham2
1
Department of Computer Science
Faculty of Electrical Engineering and Computer Science
VŠB – Technical University of Ostrava
17. listopadu 15, 708 33 Ostrava – Poruba, Czech Republic
{pavel.kromer,vaclav.snasel,jan.platos}@vsb.cz
2
Center of Excellence for Quantifiable
Quality of Service, Norwegian
University of Science and Technology
O.S. Bragstads plass 2E,
N-7491 Trondheim, Norway
[email protected]

Abstract. Quasigroups are a well-known combinatorial design equiva-

lent to more familiar Latin squares. Because all possible elements of a
quasigroup occur with equal probability, it makes it an interesting tool
for the application in computer security and for production of pseudo-
random sequences. Prior implementations of quasigroups were based on
look-up table of the quasigroup, on system of distinct representatives
etc. Such representations are infeasible for large quasigroups. In con-
trast, presented analytic quasigroup can be implemented easily. It allows
the generation of pseudorandom sequences without storing large amount
of data (look-up table). The concept of isotopy enables consideration of
many quasigroups and genetic algorithms allow efficient search for good
ones.

1 Introduction
The need for random and pseudorandom sequences arises in many applications,
e.g. in modeling, simulations, and of course in cryptography. Pseudorandom
sequences are the core of stream ciphers. They are popular due to their high
encryption/decryption speed. Their simple and cheap hardware design is often
preferred in real-world applications. The design goal in stream ciphers is to
efficiently produce pseudorandom sequences - keystreams (i.e. sequences that
possess properties common to truly random sequences and in some sense are
”indistinguishable” from these sequences).
The use of quasigroups and quasigroup string transformations is a recent
but successful tendency in cryptography and coding [17]. With quasigroups in
the hearth of advanced cryptosystems and hash functions, a need to find good
quasigroups becomes hot topic.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 472–482, 2010.

c Springer-Verlag Berlin Heidelberg 2010
 Genetic Algorithms Evolving Quasigroups 473

Quasigroups and its applications in computer security were studied e. g. in [3].

A design of pseudorandom sequence generator (PRSG) based on quasigroup op-
eration was presented in [4]. The authors have performed an extensive analysis of
216 randomly chosen quasigroups of the orders 5, 6, 7, 8, 9 and 10 and concluded
that different quasigroups produce pseudorandom sequences with different pe-
riod (i.e. the number of elements after which the pseudorandom sequence starts
to repeat). They have show that only a small number of quasigroups feature very
large value of coefficient of period growth, a property that significantly affects the
period of generated pseudorandom sequence[4]. This results encourage research
of efficient methods for search for good quasigroups in the field of pseudorandom
generators and cryptography.
Genetic algorithms are probably the most popular and wide spread member of
the class of evolutionary algorithms (EA). EAs build a class of iterative stochas-
tic search and optimization methods based on mimicking successful optimization
strategies observed in nature [7,8,16,21]. The essence of EAs lies in the emula-
tion of Darwinian evolution, utilizing the concepts of Mendelian inheritance for
practical applications in computer science [7].
EAs operate with a population of artificial individuals (chromosomes) encod-
ing potential problem solutions. Encoded individuals are evaluated using a care-
fully selected objective function which assigns a fitness value to each individual.
The fitness value represents the quality (relative ranking) of each individual as
a solution to given problem. Competing individuals explore in a highly parallel
manner problem domain towards an optimal solution [16].

2 Quasigroups
Definition 1. A quasigroup is a pair (Q, ◦), where ◦ is a binary operation on
(finite) set Q such that for all not necessarily distinct a, b ∈ Q, the equations
a ◦ x = b and y ◦ a = b.
have unique solutions.
The fact that the solutions are unique guarantees that no element occurs twice
in any row or column of the table for (◦). However, in general, the operation (◦)
is neither a commutative nor an associative operation.
Quasigroups are equivalent to more familiar Latin squares. The multiplication
table of a quasigroup of order q is a Latin square of order q, and conversely, as
it was indicated in [6,9,23], every Latin square of order q is the multiplication
table of a quasigroup of order q.
Definition 2. Let A = {a1 , a2 , . . . , an } be a finite alphabet, a n× n Latin square
L is a matrix with entries lij ∈ A, i, j = 1, 2, . . . , n, such that each row and each
column consists of different elements of A.
For i, j, k ∈ A the ordered triple (i, j; k) is used to represent the occurrence of
element k in cell (i, j) of the Latin square. So a Latin square may be represented
by the set {(i, j; k)| entry k occurs in cell (i, j) of the Latin square L.}
474 V. Snášel et al.

All reduced Latin squares of order n are enumerated for n ≤ 11 [19]. Let Ln
be the number of Latin squares of order n, and let Rn be the number of reduced
Latin squares of order n. It can be easily seen that
Ln = n!(n − 1)!Rn .
Number of distinct Latin squares of a given order grows exceedingly quickly
with the order and there is no known easily-computable formula for the number
of distinct Latin squares. The problem of classification and exact enumeration
of Latin squares of order greater than 11 probably still remains unsolved. Thus,
there are more than 1090 quasigroups of order 16 and if we take an alphabet L =
{0 . . . 255} (i.e. data are represented by 8 bits) there are at least 256!255! . . . 2! >
1058000 quasigroups.
Multiplication in quasigroups has an important property; it is proven that
each element occurs exactly q times among the products of two elements of Q,
q 2 times among the products of three elements of Q and, generally q t−1 among
the products of t elements of Q. Since there are q t possible ordered products of
t elements of Q, this shows that each element occurs equally often among these
q t products (see [10]).

2.1 Pseudorandom Sequence Generator Based on Quasigroups

The construction of pseudorandom sequence generator based on quasigroup op-
erations was described in [4]. In this paper we use simplified design of PRSG to
verify the proposed quasigroup optimization method.
Definition 3. Let (Q, ◦) be a quasigroup and Q+ be a set of all nonempty words
formed by the elements qi ∈ Q, 1 ≤ i ≤ n.
For a fixed a ∈ Q let the pseudorandom sequence y1 , y2 , . . . , yn be defined as
y1 = a ◦ a
y2 = a ◦ y1
..
.
yi = a ◦ yi−1

2.2 Isotopism of Quasigroups

Definition 4. Let (G, ·), (H, ◦) be two quasigroups. An ordered triple (π, ρ, ω)
of bijections π, ρ, ω of the set G onto set H is called an isotopism of (G, ·) upon
(H, ◦) if ∀u, v ∈ G, π(u) ◦ ρ(v) = ω(u · v). Quasigroups (G, ·), (H, ◦) are said to
be isotopic.

We can imagine an isotopism of quasigroups as a permutation of rows and

columns of quasigroup’s multiplication table.
Example 1. Consider a multiplication table for a quasigroup isotopic to the
quasigroup of modular subtraction, with operation ◦ defined as a ◦ b = (a +
n − b) mod n):
 Genetic Algorithms Evolving Quasigroups 475

◦0123
00321
12103
21032
33210
The table was created from table of modular subtraction. The second and the
third rows were exchanged. Permutations π, ρ were identities and ω = [0213]. A
multiplication in this quasigroup can be illustrated by e.g. 1 ◦ 0 = ω(1) ◦ 0 =
2 ◦ 0 = 2.
Starting with the quasigroup of modular subtraction, we can explore a large class
of quasigroups isotopic to the quasigroup of modular subtraction [11,22]. This
allows us to utilize quasigroups with very large number of elements without the
necessity of their storage in program memory. The multiplication in such isotopic
quasigroup is defined as follows:
a ◦ b = π −1 ((ω(a) + n − ρ(b)) mod n). (1)
We call the quasigroup defined by its multiplication formula and three selected
permutations an analytic quasigroup[15,27].
The notion of analytic quasigroup enables efficient work with large quasi-
groups. Previos studies in this field used mostly quasigroups of small order [14],
or just a small parts of certain quasigroup were utilized mainly as a key for
Message Authentication Code. Such small quasigroups are represented as look-
up tables in main memory. Larger quasigroup of order 2256 is used by NIST’s
SHA-3 competition1 candidate, hash function EdonR [13].
The properties of one analytic quasigroup isotopic to the quasigroup of mod-
ular subtraction were studied in [15]. The quasigroup was created using three
static functions that divided the sequence of n elements of the quasigroup into
several parts. The parts were rotated in various directions and exchanged among
themselves. It was shown that the investigated quasigroup has some faults in its
properties.

2.3 Constructing Quasigroups Isotopic to the Quasigroup of

Modular Subtraction
Consider a quasigroup on the length n defined by multiplication a ◦ b = (a + n −
b) mod n). Then three permutations π, ρ, ω must be chosen in order to implement
isotopic quasigroup, whose multiplication will be defined by (1).
Obviously, there is n! different permutations of n elements. Because three
independent permutations are used to define any isotopic quasigroup, there are
n!n!n! possible choices of π, ρ and ω.
Permutations of elements cannot be sought for an analytic quasigroup directly,
because its elements are not stored in memory. Instead, the permutation needs
to be implemented as a function of an element of Q. One way to achieve this
goal is the use of bit permutation.
1
http://csrc.nist.gov/groups/ST/hash/sha-3/index.html
476 V. Snášel et al.

A quasigroup over a set of n elements requires log2 (n) bits to express each
element. Each permutation of bits in the element representation represents also
a permutation of all elements of the quasigroup (if n is a power of 2). Bit per-
mutation can be implemented easily as a function of q ∈ Q.
The bit permutation is an elegant way of implementing permutations over n
elements of Q. Although it enables us to explore only a fragment (log2 (n)!log2 (n)!
log2 (n)!) of all possible permutation triples over the quasigroup of n elements, it
is useful because it does not require all n elements in main memory and therefore
fits into the framework of analytic quasigroups.
Bit permutations are computationaly more expensive than the static functions
used to implement permutation in [15]. However, there are ongoing efforts to
implement bit permutation instructions in hardware, which would improve the
performance of the proposed algorithm significantly [12].

3 Genetic Algorithms
Genetic algorithms are generic and reusable population-based metaheuristic soft
optimization method [5,16,21]. GAs operate with a population of chromosomes
encoding potential problem solutions. Encoded individuals are evaluated using
a carefully selected domain specific objective function which assigns a fitness
value to each individual. The fitness value represents the quality of each can-
didate solution in context of the given problem. Competing individuals explore
the problem domain towards an optimal solution [16]. The solutions might be
encoded as binary strings, real vectors or more complex, often tree-like, hierar-
chical structures (subject of genetic programming [18]). The encoding selection
is based on the needs of particular application area.
The emulated evolution is driven by iterative application of genetic opera-
tors. Genetic operators algoritmize principles observed in natural evolution. The
crossover operator defines a strategy for the exchange of genetic information
between parents (sexual reproduction of haploid organisms) while the muta-
tion operator introduces the effect of environment and randomness (random
perturbation of genetic information). Other genetic operators define e.g. parent
selection strategy or the strategy to form new population from the current one.
Genetic operators and algorithm termination criteria are the most influential
parameters of every evolutionary algorithm. The operators are subject to do-
main specific modifications and tuning [21]. The basic workflow of the standard
generational GA is shown in Fig. 1.
Many variants of the standard generational GA have been proposed. The dif-
ferences are mostly in particular selection, crossover, mutation and replacement
strategy [16].
In the next section, we present genetic algorithm for the search for good
analytic quasigroups. It is an extended version of the initial GA for quasigroup
evolution introduced in[27]. In this study, we present modified GA for quasigroup
optimization with reengineered fitness function to find quasigroups that produce
good pseudoramdom sequences.
 Genetic Algorithms Evolving Quasigroups 477

1 Define objective (fitness) function and problem encoding;

2 Encode initial population P of possible solutions as fixed length strings;
3 Evaluate chromosomes in initial population using objective function;
4 while Termination criteria not satisfied do
5 Apply selection operator to select parent chromosomes for reproduction:
sel(Pi ) → parent1, sel(Pi ) → parent2;
6 Apply crossover operator on parents with respect to crossover probability to
produce new chromosomes:
cross(pC, parent1, parent2) → {of f spring1, of f spring2};
7 Apply mutation operator on offspring chromosomes with respect to
mutation probability: mut(pM, of f spring1) → of f spring1,
mut(pM, of f spring2) → of f spring2;
8 Create new population from current population and offspring chromosomes:
migrate(of f spring1, of f sprig2, Pi ) → Pi+1 ;
9 end

Fig. 1. A summary of genetic algorithm

4 Genetic Search for Analytic Quasigroups

The genetic algorithm for the search for analytic quasigroup is defined by en-
coding of the candidate solutions and fitness function to evaluate chromosomes.

4.1 Encoding
As noted in section 2.3, any analytic quasigroup isotopic to quasigroup of mod-
ular subtraction is defined by three permutations. Such permutation triple rep-
resents a problem solution and should be mapped to one GA chromosome. Per-
mutations can be for the purpose of genetic algorithms encoded using several
strategies. In this study, we use random key encoding.
Random key (RK) encoding is an encoding strategy available for problems
involving permutation optimization [24]. In random key encoding, the permu-
tation is represented as a string of real numbers (random keys), whose relative
position changes after sorting corresponds to the permutation index. An example
or random key encoding is shown in (2).
 
0.2 0.3 0.1 0.5 0.4
Π5 = (2)
2 3 1 5 4

To encode a quasigroup (isotopic to the quasigroup of modular subtraction)

of the length n = 2l , we use a vector of 3l real numbers v = (v1 , . . . , vl−1 ,
vl , . . . v2l−1 , v2l , . . . , v3l ). The vector is interpreted as three concatenated RK
encoded permutations of the length l.
This encoding allows us to use traditional implementations of genetic opera-
tors, such as n-point crossover and mutation. Crossover was implemented as mu-
tual exchange of genes between selected parents and mutation was implemented
as a replacement of gene with a uniform random number from the interval [0, 1].
478 V. Snášel et al.

4.2 Fitness Function

Fitness function is used to rank candidate solutions among themselves. There

is a number of methods to measure randomness and evaluate pseudorandom
sequences. The DIEHARD battery of tests2 and NIST test battery3 are de-facto
standard tools to detect non-randomness in pseudorandom sequences. They can
be used to find sound statistical evidence that a pseudorandom sequence feature
non-random patterns. Many randomness tests in both mentioned test suites are
based on χ2 test defined by[1]:


n
(Oi − Ei )2
χ2 = (3)
i=1
Ei

where Oi is the observed count of occurences of i−th number and Ei is expected

number of occurences of i − th number.
In this work, we have used the χ2 statistics of the frequency of numbers
produced by the generator described in def. 3 as the goodness of fit measure
of the generated pseudorandom sequence. The χ2 statistics was computed after
10000 numbers were produced by the generator. We expect the same uniform
n
probability of every symbol Ei = no of symbols in the pseudorandom sequence.
The fitness function f was defined as:
1
f= (4)
χ2

so that better (i.e. producing more uniform sequence of symbols) pseudorandom

generator was characterised by higher fitness value.

5 Experimental Optimization

This section summarizes experimental genetic search for a quasigroups isotopic

to the quasigroups of modular subtraction with the dimensions 128, 512 and
2048 respectively. We have implemented genetic algorithm with permutation
encoding and fitness function as defined above. The parameters of the algorithm
(probability of mutation, probability of crossover etc.) were selected after initial
tuning of the algorithm. The paramteres are summarized in Table (1).

Table 1. The settings of genetic algorithm for quasigroup search

Parameter value
Population size 20
Probability of mutation (PM ) 0.02
Probability of recombination (PC ) 0.8
Selection operator elitist
Max number of generations 1000
 Genetic Algorithms Evolving Quasigroups 479

6
max
avg

4
Fitness

0
0 100 200 300 400 500
Generation

Fig. 2. Maximum and average fitness during an optimization of the quasigroup of

dimension 128

0.16
max
avg
0.14

0.12

0.1
Fitness

0.08

0.06

0.04

0.02

0
0 100 200 300 400 500
Generation

Fig. 3. Maximum and average fitness during an optimization of the quasigroup of

dimension 512

The progress of the genetic optimization of analytic quasigroups of the di-

mensions 128, 512 and 2048 is shown in Fig. 2, Fig. 3 and Fig. 4 respectively.
The evolutionary search has found better quasigroups (in terms of used fitness
function) in all experiments. For the quasigroup of length 128, it improved fit-
ness from average 0.00011 in the randomly generated initial population to 5.58.
For quasigroup of the length 512, the average initial fitness of 0.00044 was im-
2
http://www.stat.fsu.edu/pub/diehard/
3
csrc.nist.gov/rng/
480 V. Snášel et al.

0.025
max
avg

0.02

0.015
Fitness

0.01

0.005

0
0 100 200 300 400 500
Generation

Fig. 4. Maximum and average fitness during an optimization of the quasigroup of

dimension 2048

proved to 0.153 and for quasigroup of the length 2048 improved the optimization
procedure fitness from 0.00053 to 0.023. In all cases was the final improvement
achieved before generation 500.
The results demonstrate that genetic algorithms can improve randomly se-
lected quasigroups (isotopic to the quasigroup of modular subtraction) towards
properties defined by used fitness function. The resulting quasigroups are better
PRSG generators in terms of normality of symbol frequency in the generated
pseudo random sequence. Therefore, the optimized quasigroups are more suit-
able for applications in computer security.

6 Conclusions
In this paper was described a genetic algorithm for optimization of an ana-
lytic quasigroup. The genetic algorithm looks for good bit permutations that
are used to construct analytic quasigroups with desired properties. Both, the
analytic quasigroup and bit permutation, do not rely on the lookup table of
the quasigroup stored in memory. Therefore, large quasigroups can be used and
optimized efficiently.
A χ2 based goodness of fit statistics evaluating the randomness of pseudo-
random sequence produced by the quasigroup operation was used as a basis
of the fitness function. When used as a pseudorandom number generator, the
optimized quasigroups generate better pseudorandom sequences than randomly
chosen quasigroups of the same length.
Conducted numerical experiments suggest that the genetic algorithm is a
good tool to find optimized quasigroups. In our future work, we aim to find
more appropriate fitness function that will allow us to find quasigroups suitable
for cryptography and also other application areas.
 Genetic Algorithms Evolving Quasigroups 481

Acknowledgment

This work was supported by the Czech Science Foundation under the grant no.
102/09/1494.

References

1. Knuth, D.E.: The art of computer programming. In: Seminumerical Algorithms,

3rd edn., vol. 2, Addison-Wesley/Longman Publishing Co., Inc. (1997)
2. Marsaglia, G., Tsang, W.W.: Some Difficult-to-pass Tests of Randomness. Journal
of Statistical Software 7(i03)
3. Markovski, S.: Quasigroup String Processing and Applications in Cryptography. In:
Proceedings 1st Conference of Mathematics and Informatics for Industry, Thessa-
loniki, Greece, pp. 278–290 (2003)
4. Dimitrova, V., Markovski, J.: On quasigroup pseudo random sequence generator.
In: Manolopoulos, Y., Spirakis, P. (eds.) Proc. of the 1-st Balkan Conference in
Informatics, Thessaloniki, November 2004, pp. 393–401 (2004)
5. Bäck, T., Hammel, U., Schwefel, H.-P.: Evolutionary computation: comments on
the history and current state. IEEE Transactions on Evolutionary Computation 1,
3–17 (1997)
6. Belousov, V.D.: Osnovi teorii kvazigrup i lup, Nauka, Moscow (1967) (in Russian)
7. Bodenhofer, U.: Genetic Algorithms: Theory and Applications. Lecture Notes,
Fuzzy Logic Laboratorium Linz-Hagenberg (Winter 2003/2004)
8. Dianati, M., Song, I., Treiber, M.: An introduction to genetic algorithms and evolu-
tion strategies, technical report, University of Waterloo, Ontario, N2L 3G1, Canada
(July 2002)
9. Dénes, J., Keedwell, A.: Latin Squares and their Applications. In: Akadémiai Kiadó,
Budapest. Academic Press, New York (1974)
10. Dénes, J., Keedwell, A.: A new authentication scheme based on Latin squares.
Discrete Mathematics (106/107), 157–161 (1992)
11. Dvorský, J., Ochodková, E., Snášel, V.: Hash Functions Based on Large Quasi-
groups. In: Proceedings of Velikonočnı́ kryptologie, Brno, pp. 1–8 (2002)
12. Hilewitz, Y., Shi, Z.J., Lee, R.B.: Comparing fast implementations of bit permuta-
tion instructions. In: Conference Record of the Thirty-Eighth Asilomar Conference
on Signals, Systems and Computers, Pacific Grove, California, USA, November
2004, pp. 1856–1863 (2004)
13. Gligoroski, D., et al.: EdonR cryptographic hash function. Submition to NIST’s
SHA-3 hash function competition (2008),
http://csrc.nist.gov/groups/ST/hash/sha-3/index.html
14. Gligoroski, D., Markovski, S., Kocarev, L., Svein, J.: The Stream Cipher Edon80.
In: Robshaw, M.J.B., Billet, O. (eds.) New Stream Cipher Designs. LNCS,
vol. 4986, pp. 152–169. Springer, Heidelberg (2008)
15. Snášel, V., Abraham, A., Dvorský, J., Krömer, P., Platoš, J.: Hash functions based
on large quasigroups. In: Allen, G., Nabrzyski, J., Seidel, E., van Albada, G.D.,
Dongarra, J., Sloot, P.M.A. (eds.) ICCS 2009, Part I. LNCS, vol. 5544, pp. 521–529.
Springer, Heidelberg (2009)
16. Jones, G.: Genetic and evolutionary algorithms. In: von Rague, P. (ed.) Encyclo-
pedia of Computational Chemistry. John Wiley and Sons, Chichester (1998)
482 V. Snášel et al.

17. Knapskog, S.J.: New cryptographic primitives. In: CISIM 2008: Proceedings of the
2008 7th Computer Information Systems and Industrial Management Applications,
pp. 3–7. IEEE Computer Society, Washington (2008)
18. Koza, J.: Genetic programming: A paradigm for genetically breeding populations of
computer programs to solve problems. Technical Report STAN-CS-90-1314, Dept.
of Computer Science, Stanford University (1990)
19. McKay, B.D., Wanless, I.M.: On the Number of Latin Squares. Journal Annals of
Combinatorics 9(3), 335–344 (2005)
20. Merkle, R.C.: Secrecy, authentication, and public key systems. Stanford Ph.D.
thesis, pp. 13-15 (1979), http://www.merkle.com/papers/Thesis1979.pdf
21. Mitchell, M.: An Introduction to Genetic Algorithms. MIT Press, Cambridge
(1996)
22. Ochodková, E., Snášel, V.: Using Quasigroups for Secure Encoding of File Sys-
tem. In: Proceedings of the International Scientific NATO PfP/PWP Conference
Security and Information Protection 2001, Brno, Czech Republic, May 9-11, pp.
175–181 (2001)
23. Smith, J.D.H.: An introduction to quasigroups and their representations. Chapman
& Hall/CRC (2007)
24. Snyder, L.V., Daskin, M.S.: A random-key genetic algorithm for the generalized
traveling salesman problem. European Journal of Operational Research 174(1),
38–53 (2006)
25. Vojvoda, M.: Cryptanalysis of One Hash Function Based on quasigroup. In: Con-
ference Mikulášská kryptobesı́dka, Praha, pp. 23–28 (2003)
26. TREC Web Corpus: GOV (2009),
http://ir.dcs.gla.ac.uk/test_collections/govinfo.html
27. Snášel, V., Abraham, A., Dvorský, J., Ochodková, E., Platoš, J., Krömer, P.:
Searching for Quasigroups for Hash Functions with Genetic Algorithms. In: Pro-
ceedings of the 2009 World Congress on Nature & Biologically Inspired Computing,
pp. 367–372. IEEE Computer Society, Los Alamitos (2009)
 Software Openness: Evaluating Parameters of
Parametric Modeling Tools to Support Creativity and
Multidisciplinary Design Integration

Flora Dilys Salim and Jane Burry

Spatial Information Architecture Laboratory, RMIT University, Melbourne, Australia

{Flora.Salim,Jane.Burry}@rmit.edu.au

Abstract. The ubiquitous computing era has pushed the Architecture, Engineer-
ing, and Construction (AEC) industry towards new frontiers of digitally enabled
practice. Are these the frontiers originally identified by the pioneers in the
field? Architectural design has progressively shifted from two-dimensional
paper based pencil sketched models to digital models drawn in various Com-
puter-Aided Design (CAD) tools. The recent adoption of parametric modeling
tools from the aerospace industry has been driven by the need for tools that can
assist in rapid flexible modeling. The adaptation of parametric modeling has re-
formed both pedagogy and practice of architectural design. The question
remains if parametric design has answered all the requirements specified by
Steven Anson Coons in his 1963 proposal for a Computer-Aided Design (CAD)
system. Given the growth of computational power and ubiquitous computing,
how has CAD met the visions of its pioneers with respect to the flexibility and
ease of communication with the computer and support of simultaneous design
conversations with many designers working on the same project? This paper
will revisit ideas conceived by the early inventors of CAD, explore the opportu-
nities for advancing parametric modeling with the existing ubiquitous comput-
ing infrastructure, and introduces the notion of software openness to support
creativity and multidisciplinary design integration.

Keywords: CAD, CAE, Parametric modeling, Parametric design, Software

Openness.

1 Introduction
The rise of Computer-Aided Design (CAD) and Computer-Aided Manufacturing
(CAM) has led to widespread adoption of digital modeling in the field of manufactur-
ing and construction. There is an increasing demand for flexible modeling tools that
allow rapid adaptations to design variations to be produced in real time. CAD is
a term originally introduced by Steven Coons [1] and Ivan Sutherland [2] as an
integrated design approach which leverages both the creativity and imagination of
man and the analytical and computational power of the computer. CAD eventually
evolved to tools that truly assist designers. Parametric design software, which was
hitherto used mainly in the aerospace and automotive industry, had been adapted to

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 483–497, 2010.
© Springer-Verlag Berlin Heidelberg 2010
484 F.D. Salim and J. Burry

the Architecture, Engineering and Construction (AEC) industry [3], [4]. We, a com-
puter scientist (the first author) and a practicing architect in academia (the second
author), are involved in various integrated multidisciplinary architecture-engineering
projects which require explorations of existing design tools and techniques.
This paper aims to investigate the current state-of-the-art of digital modeling and
reevaluate parametric design software retrospectively in the light of the aims of early
CAD inventors. The review also seeks to discover the requirements for a truly flexible
and open parametric design tool that can leverage creative design thinking and mul-
tidisciplinary design integration.
The paper is organized as follows. Section 2 presents the summary of the require-
ments for CAD as proposed by the originators. Section 3 reviews the existing para-
metric CAD software in the light of the requirements stated in the previous section
and outlines the research gap that needs to be investigated. Section 4 presents our
proposed notion of software openness given the need to leverage creative design
thinking and multidisciplinary design integration. Section 5 concludes the paper.

2 Reflections on Requirements for CAD

According to Steve Coons, a CAD system needs to support the design process by
assisting creative thinking, increasing productivity, and improving intercommunica-
tion and collaboration through a well-defined human-computer interaction, or “design
conversation” [1]. There are four key requirements of CAD system as stated in the
paper.
Firstly, it must have the ability to accept, infer, and store shape description intro-
duced graphically using graphical means as well as structural abstractions (with sym-
bolic languages that is capable of representing interconnected ideas within the design)
[1].
Secondly, coupled with the visualization, it must also be an analysis tool that is
able to perform all the computations required for the design process. This includes
structural analyses, mechanical services, electrical analyses, and other analytical proc-
esses that are required to validate and optimize the design [1].
Thirdly, it should be a platform that promotes ease of collaboration by enabling the
same model to be accessed and modified by multiple designers in different locations
simultaneously (Coons, 1963). The notion of the internet and online collaboration was
not even close to reality in 1963. ARPANET (Advanced Research Projects Agency
Network), the forerunner of the Internet, was only introduced in 1968 [5]. The idea
described in the paper [1] depicts a situation that was only realizable in the early 90s
when the Internet had reached one-million users worldwide [5].
Finally, a CAD system must be highly generic to accommodate the creative activi-
ties that reside in a trans-disciplinary design domain. The general problems of archi-
tects and engineers need a system that is flexible and adaptable to multiple specific
domains [1].
Apart from Steve Coon’s manifesto, Kasik et al. has reviewed the progress of CAD
departing from Ivan Sutherland’s Sketchpad application and presented ten challenges
of CAD and classified them into three different categories: “computational geometry,
interactive techniques, and scale” [6]. A number of those challenges have been
 Software Openness: Evaluating Parameters of Parametric Modeling Tools 485

answered in recent developments of parametric modeling tools. There are profusion

of CAD software that perform solid modeling but are not parametric. The scope of
this review only covers CAD packages that perform flexible 3D parametric modeling
and are referred as parametric modeling tools from this point forward.

3 Parametric Modeling, Are We There Yet?

Parametric modeling was first introduced as a means for design reuse [7]. Parametric
modeling uses parameterized relationships between components in the design to de-
fine forms [8]. A parametric model comprises variable attributes, which are called
parameters and fixed attributes, which are called constraints [4]. A parametric model
responds to changes of parameter value or definition without erasure of parts of the
model or starting the design model from scratch. Designers alter the value of the pa-
rameters to explore design variations that can be generated from the same model [4].
Parametric design tools, such as GenerativeComponents by Bentley Systems, Digi-
tal Project by Gehry Technologies (an adaptation from the powerful and expensive
CATIA software used by the aerospace industry for the AEC industry), ArchiCAD by
Graphisoft, and the Revit family of products by Autodesk, have been taken up more
widely by the AEC design and design education communities worldwide only during
this decade. In the previous decade, its use was known but through more isolated and
research-led projects in practice and academia. AutoCAD was the main software
utilized in industry but it was not parametric. The rise of the usage of parametric
modeling tools is due to its power to manipulate geometry and generate variations by
simply changing its parameters. Geometry manipulation is not limited to simple ex-
trusion or Boolean operations. Parametric modeling tools .support the construction of
a single model that has many varying geometrical instances. Initially, parameterized
mathematical associations between objects in the model need to be defined. For some
parameters, it may be possible to determine a range, or the maximum and minimum
values based on design criteria. Then by iterative refinement of both the mathematical
associations and the values of the parameters in the model, various design options are
explored. These design options can be compared and analyzed further and criteria
determined for selection of the better and best solutions.
The Sagrada Família church in Barcelona (Fig. 1 – Left), which was started in
1883 and left unfinished by the architect Antoni Gaudí who died in 1926, is an exam-
ple of a building and a design system in which there is a clearly defined parametric
geometry that provides the key to the design process. With the help of parametric
‘recipe’ left by Gaudi in his plaster models and drawings, the ongoing design and
construction of the Sagrada Família relies on Gaudí’s technique of prescribing varia-
tions of complex geometries based on simple forms exploiting the variation of geo-
metric parameters of those forms. One example of the subsequent influential work to
interpret and model Gaudí’s intentions digitally and parametrically is the columnets
model, in which variation of the surface parameters, rotation and Boolean operations
are applied on trimmed hyperbolic paraboloids to generate the forms [9]. Two exam-
ples of the more recent projects using parametric design tools for design exploration
are the Dubai Towers (Fig. 1 – Right), designed by TVS using Generative Compo-
nents by Bentley Systems, and the “Bird’s Nest” Beijing National Stadium (Fig. 2),
designed by Herzon and de Meuron using Digital Project by Gehry Technologies.
486 F.D. Salim and J. Burry

Fig. 1. Left: The clerestory window in Sagrada Familia, Image courtesy of: Mark Burry. Right:
Dubai Towers. The Lagoons Image courtesy of: Thompson, Ventulett, Stainback & Associates
(www.tvsa.com).

Fig. 2. Left: An early digital model of the Bird’s Nest Beijing National Stadium as modeled
inside Digital Project. Right: the final model. Image courtesy of: Gehry Technologies LLP
(www.gehrytechnologies.com).

We perceive two different groupings of parametric design tools given the varying
approach to parametric design. The first group of parametric design tools is based on
associative-geometry, where parameterized mathematical descriptions and associa-
tions between points, curves, surfaces, and solids are possible. The tools belonging to
this group include Bentley GenerativeComponents and Rhino Grasshopper. The sec-
ond group of parametric design tools are focused on BIM, where parametric relation-
ships encapsulate parametric descriptions of components of a building design across
multiple disciplines [10]. BIM stands for Building Information Modeling, which is
regarded as “a data-rich, object-oriented, intelligent and parametric digital represen-
tation of the facility” by the American General Contractors (as quoted in [11]). The
Autodesk Revit 2010 package (Revit Architecture 2010, Revit MEP 2010, Revit
Structure 2010) belongs to this group as well as Gehry Technologies’ Digital Project.
Revit 2010, however, is unable to handle NURBS geometry modeling [12] and
 Software Openness: Evaluating Parameters of Parametric Modeling Tools 487

therefore Revit is not generally used for modeling complex curved geometries. Digital
Project is known to be a very powerful parametric CAD package that handles both
complex parametric geometric associations as well as parametric design representa-
tions referring to diverse libraries across multiple disciplines. Nevertheless, it is also
known to be the most expensive CAD package that is currently around and customiz-
ing model via the script interface is not trivial. The first group of tools can also be
used for parametric representations that link different disciplines, however, BIM high-
level object definitions are not included in the tools by default.
In a reflective mode, we review the existing parametric modeling tools and revisit
the CAD requirements by the CAD visionary and retrospectively raise the question –
are we there yet? The following evaluation is raised in the light of the use of the cur-
rent parametric modeling tools and processes.

3.1 Are They Capable of Accepting, Inferring, and Storing Geometries Using
Graphical and Symbolical Means?

This section evaluates the existing parametric modeling tools’ capabilities to process
and manage the geometries in the design using graphical and symbolical means. The
abstractions of the design in parametric modeling tools are often represented in a
symbolic diagram. These features are now intrinsic to all the existing parametric de-
sign tools. GenerativeComponents (GC) encompasses a 3D interactive view, a sym-
bolic view, and an object view [13]. One distinguishing feature of GC is the way it
displays both the textual and symbolic representation of the parametric associations
between components in the model. This feature is also available in Grasshopper, the
new extension of Rhino. The symbolical representation of a parametric model is
analogous to an object diagram of a software application. In fact, the notion of com-
ponents, parameters, constraints, and associations in a parametric model resemble
those of the object-oriented programming paradigm. The graphical and symbolical
representations undoubtedly assist designers to rethink their design in the light of the
components and their associations in the model as well as the parametric variations
they would like to generate from the model. Both the graphical and symbolical repre-
sentations visualize design intent that is captured within the tool, particularly during
the early design stage when design options need to be generated rapidly.
Revit 2010 uses the more rigid tree structure to represent the higher level of ab-
straction of the model. Geometrical associations between components in the model
are not visualized in the tree structure. Since each component in a Revit model is an
instance of a Revit family type (e.g. an instance of a 36’ flush panel type of the door
family), the tree structure provides the abstraction and the detailed view of the
instances of each type and family that are created in the model. Symbolic representa-
tions are used to view the 3D model as engineering plans. However, there is no exist-
ing symbolical abstraction to represent the parametric associations of all the parts in a
Revit model. Considering that Revit is aimed to be a BIM software, the tree structure
is useful for viewing abstract representations of parts of the model to give a non disci-
pline specific overview.
CATIA/Digital Project (DP) primarily uses a hierarchy of linked files. No single
file need be very large. It also uses simple graphical files that just show the geometry
for representing large extents of the model where it might otherwise take hours or
488 F.D. Salim and J. Burry

days to open the Part and Product files. There is a tree structure in CATIA/DP that is
an organizational and navigational tool and another that directly maps the model’s
parametric relationships.

3.2 Are They Capable of Performing Analysis in Different Performance Areas?

To date, CAD tools (i.e. parametric design tools) and analysis tools have been distinct
and autonomous. Cross-disciplinary analysis is a challenge since current CAD sys-
tems are not integrated well with CAE (Computer Aided Engineering) analysis tools
[Kasik05]. Some tools, such as Revit Architecture 2010 has released a plug-
in/extension for solar incidence and energy analysis to be performed within the soft-
ware. Another example is Google Sketchup, which has an extended capability for a
simple energy analysis when the IES Virtual Environment plug-in is installed with it.
However, there is no existing tool on the market that currently carries the capability to
perform all-in-one design and analysis, where the analysis can respond to the design
variations made in the same model. Some recent work has come close to realizing this
silver bullet [14].
In order for a parametrically-modeled design proposal to be assessed for structural
loads, energy efficiency, solar incidence analysis, lighting analysis, and other types of
analyses, it must undergo “expert translation” since geometrical representation in vari-
ous domains are interpreted differently [15]. Translation errors often arise from two
key sources, which are floating-point arithmetic and tolerances [6]. Approximations
used in numerical calculations reducing floating-point that substitute higher perform-
ance for high precision cause errors in models passed to disciplines that require high
precision analysis [6]. Tolerances are used to control the level of approximation and
need to be managed to produce geometry that can meet different discipline analysis
requirements. If the tolerances are set high, analysis can be performed more quickly
but with greater likelihood of error in the resulting geometry [6]. If the tolerance is too
small, analysis may fail to converge at all [6]. Therefore, selecting a tolerance to be
used requires good comprehension of the engineering analyses to be used, the envi-
ronment of the analyses, and the specific software to be employed [6].

3.3 Are the Models Generic and Adaptable to Various Domains Pertaining to
the Design Process?

This section deals with Coons’ requirement for a CAD system that can support cross-
disciplinary model interoperability. According to a survey by McGraw-Hill construc-
tion study [16], 3% of a project cost corresponds to software non-interoperability. In
the U.S. construction market worth $1.3 trillion in 2007, this corresponds to $36 bil-
lion of productivity waste. In the global market worth $4.6 billion, the figure escalates
to $138 billion [16]. In order to maintain building information efficiently and the vary
representations in a broadly reusable way, the building industry has recently at-
tempted to define and encourage the adoption of Building Information Modeling
(BIM). With too many industry players and major CAD vendors introducing a num-
ber of commercialized BIM tools, however, many have found themselves reinventing
the wheel in order to bridge tools created by one vendor to another tool by another
vendor.
 Software Openness: Evaluating Parameters of Parametric Modeling Tools 489

Industry Foundation Classes (IFC), the accepted industry standard for interopera-
bility and model exchange, are too heavy and complex (IFC consist of 327 data types,
653 entity definitions, and 317 property sets [17]). This poses challenges to perform-
ance, scalability, and ease of use and adaptation into specific sub-domains. Green
Building XML (gbXML), which was introduced as a lighter option to IFC, only en-
compassing a subset of IFCs has become a de facto standard for interoperability of
green building design. However, semantic translations from those standards to BIM
tools and vice versa (import to/export from BIM tools) still generate translation errors.
Lee et al argue that parametric modeling needs to be central to BIM and provide
mechanisms for translation to various domains [7].

3.4 Do They Promote Collaboration and Communication among the Designers

or Users?

Over the top of any provision for collaboration and model sharing and merging within
existing parametric design software packages, there is a need for additional communi-
cation channels that combine graphical, symbol, and language based information.
Currently, separate tools are required Product Lifecycle management tools for manu-
facturing aim to fulfill this role by integrating information and communication with
customers/clients, suppliers, enterprise/planning and systems development informa-
tion and communication.
However, in practice it is a lightweight tool that allows rapid human interaction as
well as capturing the history of informal decision making and rapid change in the
design model that is required in combination with parametric modeling. Autodesk
Design Review and Octopz are examples of visual annotation environments that sup-
port this level of interaction. Versioning software has been integrated with Digital
Project parametric software as the means of team model sharing online, allowing
visual indicators of ownership of different parts of the model, locked in the versioning
software, at any given time within the modeling environment. The concept is familiar
to Archicad Teamwork and Bentley Projectwise users.
There are two fundamental truths that Jerry Laiserin has stated of the current state
of technology in the industry: that while collaboration across design teams is the most
critical factor for successful building information modeling, as model data integration
goes up, flexibility of workflow and performance in collaboration go down [18]. This
is true even for the most straight forward, explicit geometrical representation of build-
ings and currently leaves models shared between large teams of a fundamentally more
formative and flexible nature largely in the aspirational realm even 40 years on.

3.5 Do They Support the Design Process by Communicating Design Intent?

One of Coons’ manifestos is for CAD to promote design conversations between the
computer and the designer. In parametric design tools such as Digital Project and
GenerativeComponents (GC), the history of geometrical associations and parametric
value variations is recorded. In GC, parametric changes are recorded as transactions
and users can go back and forth in the transaction list to undo or redo changes in the
model. Whenever the value of a parameter or the parametric relationship between one
component to another is changed, the value and relationship changes are recorded in
490 F.D. Salim and J. Burry

the transaction file. This is one way for CAD tools to carry “design conversations”,
since designers constantly perform iterative loops in evaluating their design and men-
tal design conversations that are channeled through their modeling activities can be
recorded.
However, difficulties in communicating design intent for multidisciplinary model
sharing still persist. When standards of model interoperability exchange such as IFC
or gbXML are used for model sharing, parametric representations and associations are
not captured either by IFC or gbXML during translation, since the dynamic associa-
tions between the components in the model are replaced with static values represent-
ing the components themselves. The design intent that needs to be maintained in a
parametric model is lost in translation and unrecoverable once the model is imported
or exported to other tools or analysis programs.

3.6 Do They Increase Productivity in the Design Process?

Parametric design tools undoubtedly increase productivity. The efficiency of using

parametric design approach in modeling depends on the number of the iterations to
the design and the extent of the variations in the model [19]. In this respect, a large
part of Coons’ future gazing has come to pass. However, there are many measures of
productivity. Certainly the speeds of documentation and repetitive reuse of data is
massively increased through moving from manual drafting to CAD. More creative
activities, combining analytical and synthetic approaches to building shape, structure
and articulation have benefited in the ways that were both predicted and prototyped in
the 1960s. The size and complexity of building projects has also increased hugely
since the 1960s and we can speculate that the technological means have fed this trend.
However, the continued use of the term ‘documentation’ hints at the continued lag in
the transition of human systems to the new means of production, arguably slowed by
the introduction of personal computing and its disaggregating impact on teams and
projects [20]. With the growth of computing power and the internet in the last decade
plus, the integration of information and reduction in costly design document error and
inconsistency has been a major focus in increasing the contribution of IT to productiv-
ity measured by building cost.
The usability of parametric design tools has increased with the expansion of com-
putational power. In general, the more complex the software is, the lower the usabil-
ity. However, in architecture and engineering disciplines, the usability of the software
also depends on its capacity to host a model that contains multidisciplinary design
representation for multidimensional model editing and real-time visualization. CATIA
is capable of performing complex curves manipulation and visualization of the whole
airplane model as well as the dependent parts of the model.

3.7 Do They Support the Design Process by Assisting Creative Thinking?

Lars Hesselgren stated “the advantage of using GenerativeComponents is that it helps

me think about what I am doing. The disadvantage is that it forces me so to think”
[13]. Parametric modeling tools have forced designers to quantify their design
to components and associations for generating variations that they would like to ex-
plore further. However, when a parametric design grows and the number of potential
 Software Openness: Evaluating Parameters of Parametric Modeling Tools 491

parametric variations grows exponentially, the flexible model poses the threat of in-
flexibility given that the proliferation of backward/forward chaining of changes can
be unmanageable [21]. This is where customization, scripting, or coding comes in to
pull the weight off from visual design. Designers are now becoming more interested
in increasing the efficiency of model building through scripting or coding. To design-
ers who code, coding activity is considered as a channel for creativity and means of
representing design ideas that goes beyond visual or graphical drawing. Current tools
have indeed assisted creative thinking. However, to innovate, designers need to ex-
tend the exploration of their design ideas using complementary creative methods to
the use of digital/visual synthetic geometric modeling aids.
Lawson [22] stated his desire to see CAD becoming an agent for creativity. As an
agent, CAD would act as a subordinate who learns and understands the design inten-
tion of the master and is able to assist creative thinking by crawling the internet and
finding design ideas that can match the master’s interest [22]. Lawson’s desire can
potentially be realized if parametric modeling is integrated with such computational
techniques as machine learning or data mining. The existing work involving machine
learning techniques for the design process includes evolutionary design [23], [24] and
agent based modeling [25].

3.8 Are We There Yet?

Although it has been a long road to general adoption (and we are still clearly on the
uptake curve), the current parametric design tools have to some degree fulfilled the
requirements of CAD as prescribed by Coons. Parametric design tools have improved
productivity through rapid generation of design variations over simple parametric
changes and helped creative thinking in the process of designing. However, there are
still gaps and opportunities for advancing the existing parametric design tools.
Firstly, multidisciplinary design model integration is still very challenging [18].
Given the complexity of the process, analysis, and tools required in various sub-
domains, an all-in-one engineering analysis tools in a CAD package as proposed by
Coons is not feasible in the AEC industry. Cross-CAD and cross-disciplinary model
sharing is still a challenge due to software interoperability issues and translation er-
rors. There is a need to escalate the degree of software openness in order to carry
design intent between designers of one discipline to another. Software openness is
discussed further in the next section.
Secondly, the need for an online multidimensional design review tool for design
collaboration as an additional layer on top of the existing parametric design software
is apparent. In the current cyber digital and ubiquitous computing era, a multidiscipli-
nary team working on the same project might be dispersed in different locations
across the world. Having to share design variations on the same model requires online
collaboration tool and parametric design tool to be integrated.
With the flexibility of parametric modeling, the creative thinking process of the de-
signer can now inform scripts to drive the process of parametric changes. Given the
notion of openness of parametric modeling, parameters can be scripted and their val-
ues driven from any source, the intent of the designer, the results of performance
analysis in the cyber world, or information gathered through sensors, for example, in
the physical space of our world. It opens endless possibilities for supporting both
creativity and productivity.
492 F.D. Salim and J. Burry

The initial need for parametric design tools is the requirement to support flexible
modeling. Flexibility should be an inherent attribute of any parametric design tool. In
the light of the converging physical and digital spaces, architectural form inside the
CAD should keep on transforming and producing emergent results which are reflec-
tive on the physical environmental changes as well as the virtual energy simulation
results. The current state-of-the-art technology in the cyber/digital world, such as
sensing technology, ubiquitous computing, and cloud computing, can potentially be
combined with the power of parametric modeling to enhance the flexibility, fluidity,
and creativity of designing. Therefore, there is an opportunity to leverage the existing
parametric modeling tools by keeping it open to the user, the environment, and other
existing tools in the space. In addition, given the complexity of the process, analysis,
and tools required in various disciplines, there is a need to escalate the degree of soft-
ware openness in order to carry trans-disciplinary design conversations. This is dis-
cussed further in the next section.

4 Software Openness of Parametric Modeling Tools

Hansen [26] quoted Sanford Kwinter that “the world where everything flows seam-
lessly together in real time”, added an argument for the requirement of “soft systems”
which are driven by their softness and openness to the environment and adaptable to
information exchanged with its surroundings. Concurring with Hansen, we propose
the notion of software openness, which is a state of unimpeded ability of 3D paramet-
ric software to be openly customized in order to:
1. exchange information with other software (open interoperability) and with
the physical environment (physical-virtual information integration);
2. modify the exchanged information and the information handling process;
3. mediate, compare, and consolidate the generated design variations for
analysis, simulation, and fabrication.
Software openness is required to support design integration across disciplines and
creative design uptake of the ubiquitous computing opportunities as the physical
realm and virtual worlds become more closely intertwined. This differs from open
source software. Software openness does not imply source code access, instead, it
requires a methodological approach to defining ways for open input, open process,
and open output to parametric modeling tools.
The fundamental of information handling in any software comprises input, process,
and output (IPO). Conventionally, it is a linear progression without any feedback
loop (Fig. 3). Software openness requires open input, open process, and open output
(Fig 4). The input, process, and output parameters of parametric modeling tools need
to be opened for multi-devices, multi-tasking, multi-disciplinary and multidimen-
sional interaction effectively to harness creativity, productivity, and collaboration.
The information handling in such software should no longer be a linear progression,
but a closed loop of IPO that allows real time and open interaction or feedback.
The degree of software openness can be used to assess the suitability of a design
tool in the process of choosing the right sets of tools for a multidisciplinary design
project. The higher the degree of a tool being receptive to open input, open process,
 Software Openness: Evaluating Parameters of Parametric Modeling Tools 493

Fig. 3. Input Process Output

Fig. 4. Open Input, Open Process, Open Output

and open output, the more usable the tool is in the context of a multidisciplinary de-
sign project. Since a parametric design tool do not generally include the full range of
engineering analysis tools, such as building thermal performance analysis and struc-
tural analysis tools, the higher the degree of the software openness, the higher is the
possibility of the software being used in design model integration across disciplines
and design activities. Since performance analysis and software interoperability are
still challenges in multidisciplinary design model integration, the ability to customize
design representation, translation, and input/output from one software to another en-
ables engineers to customize the software for the purpose of the project. The need to
assess the software openness for interoperability is demonstrated in the PM4D project
[27] by Centre for Integrated Facility Engineering (CIFE), Stanford University, where
the input and output from one design, engineering analysis, or visualization software
to another are compared for the best selection of tools required for the integrated
project delivery.
Open Input. Input to a parametric tool can either be models (imported from another
CAD) or textual files containing analysis requirements (which could be a result im-
ported from external analysis tools). In a multidisciplinary design project, it is neces-
sary to analyze the input capabilities of the tools being used and the existing model
translation mechanisms between one software to another. The interoperability gap
between software, where it exists, could then be tackled in two ways. Firstly, the
‘tool-by-tool’ approach which employs a specific tool or plug-in for export and im-
port of a design model between one tool and another, such as demonstrated by Nicho-
las and Burry [15] in translating a model created in Rhino CAD software to a model
understandable by Radiance, a lighting analysis software. Secondly, the ‘bridge tool’
approach which employs an intermediate tool for model translation among various
CAD, analysis, and visualization software. This approach was exemplified in the
development of DesignLink Software Development Kit (SDK) [28], [29], by Arup
and Spatial Information Architecture Laboratory (SIAL), which relies on the transla-
tion of various models into DesignLink XML, an XML schema designed for model
information exchange in the early design stage.
In order to promote design conversations between human designers and the com-
puter, various sensor and haptic devices need to be explored as an input device to pa-
rametric modeling. Handheld devices help designers to converse with the parametric
494 F.D. Salim and J. Burry

models better. For example, a tangible view cube was developed to interface with
Autodesk tools in order to represent the 3D digital model in a physical environment
and promote tangible interaction between the user and the CAD [30]. The ClayTools
system from Senseable [31] is a free-form robot hand that allows model creation, craft-
ing, and detailing straight to the virtual model with a force feedback applied to the
user. The ClayTools can become input to 3DMax, Maya, and Rhino, which are not
parametric modeling software. However, with Grasshopper, Rhino’s parametric plug-
in, there is a potential in connecting Rhino-Grasshopper virtual parametric models with
ClayTools. Another example of a haptic device that can be used for exploring, design-
ing, and generating inputs for parametric models is the Nintendo Wii remote (or Wii-
mote). A Wiimote plug-in for Autodesk Design Review by Autodesk Labs was devel-
oped for designers to navigate through the 3D model and environment using the Wii-
mote [32]. Another existing work on development of haptic input devices to parametric
modeling software is the Wiimote features development for Generative Components
(GC) [33], which allows the Wiimote to interact with the GC to generate forms and
variations based on real-time data input. The Wiimote can be used to draw in 3D space,
modify existing 3D models, or used as a controller of the model (Fig. 5).

Fig. 5. A Wiimote interacts with GenerativeComponents

Open Process. Software openness can also be used to assess the suitability of the
software to be openly customized in order to generate a creative output. Creativity can
be achieved not only through physical or digital modeling, but also through creative
coding [34]. Parametric design and analysis tools need to accommodate for exten-
sions, customization, and open interoperability, in order that additional plug-ins, class
libraries, software components, and scripts can be developed to support creativity and
productivity. Interoperability checking and automation through code can be devel-
oped when the communication channel between software is open. In Digital Project,
customization through codes is available from their internal scripting window, which
uses VBA script or by expert use of the CATIA SDK. Given the nature of VBA as
procedural language, component and object reuse in the code is a challenge since
classes cannot be defined inside the VBA script. The open component based software
architecture adopted by GC and Grasshopper (Rhino) provides a better flexibility
for customization in comparison to non-parametric modeling tools which largely rely
on internal scripting techniques. GC and Grasshopper (Rhino) allow installation of
 Software Openness: Evaluating Parameters of Parametric Modeling Tools 495

custom plug-ins or Dynamic Link Libraries (DLLs) to define new features, objects, or
parametric relationships or to extend/inherit from the existing type libraries. Autodesk
offers an Application Programming Interface (API) to access the classes, methods and
operations provided by Revit programmatically on .NET platform. Commercial and
non-commercial parametric modeling software could open the process of definition,
generation, modification, and deletion of design objects and parametric relationships
among the objects in the model to promote collaboration, interoperability, design
integration, creativity and interactivity with other software, sensors and haptic
devices.

Open Output. Coons’ paper opens with reference to the first automatically controlled
milling machine at MIT in the 1950s [1]. The CAD/CAM axis so far seems to have
proved if anything more vital in the design process than immersive virtual reality.
Designing and building in the physical world entails, as it always has, working from
and capturing a physical context, representing this virtually, whether through meas-
urement, drawing and physical modeling or through digital capture and representa-
tion. Similarly, collective design decision making is still frequently better supported
through interaction around physical models. While the cost and speed of traditional
physical modeling is in some cases uncompetitive with virtual equivalents during the
design process, the opportunity for rapid prototyping allows for the design conversa-
tion to flow back and forth between the virtual and physical embodiment.

5 Conclusion and Recommendations

This paper has reviewed the existing CAD software and the current state-of-the-art of
parametric modeling software. Although parametric modeling has improved produc-
tivity and encouraged creative thinking, the current parametric design tools and proc-
esses have not fully answered the requirements of the CAD visionary nor coped with
the challenges of the digital age.
The first challenge is to answer the multidisciplinary design model integration, par-
ticularly between parametric design and building performance analysis tools. Further,
parametric design software are becoming more prevalent, but are still difficult to
integrate with analysis tools for informed feedback. In support of the collective use of
open parametric design software, there is also a need for multidimensional online
collaboration and communication capabilities to be integrated within parametric
design tools.
To give their creativity full rein, designers need to think outside the box of the digi-
tal screen and digital world in order to hold design conversations that move more
effortlessly between the cerebral, virtual and physical worlds. Coding and open inter-
action with parametric modeling hold the keys to making the transitions between
these environments more seamless.
The proliferation of sensor and physical computing devices in the ubiquitous com-
puting era presents immense potential for designing in the mixed reality, where the
space of designing is no longer constrained to the physical space. Therefore, software
openness can also be measured in the way a parametric design tool is able to cooper-
ate with physical and ubiquitous computing. Designers who model in the digital space
496 F.D. Salim and J. Burry

often require real-time interaction and feedback that can be manifest in the physical
world. When a parametric design software unlocks the means for software openness,
sensors and haptic devices can be linked with parametric design software to achieve
tools that are capable of simulating adaptive and responsive forms of architecture.
Finally, the degree of software openness can be used to measure the flexibility of
the software to be openly customized for interoperability, model sharing, and support-
ing creativity.

Acknowledgment
The authors thank Australian Research Council (ARC) for the funding of the research
in the ARC Discovery project: Challenging the inflexibility of the flexible model. We
also thank the ARC and Queensland State Government’s Project Services for the
funding of the research in the ARC Linkage project: Assimilation of architectural and
services design in early design modeling. We acknowledge John Frazer, Robin Dro-
gemuller, and Bianca Toth from Queensland University of Technology (QUT) and
Mark Burry from RMIT University for their contribution to the research team. The
authors also acknowledge Daniel Davis from RMIT and Ruwan Fernando from QUT
for their assistance.

References
1. Coons, S.A.: An outline of the requirements for a computer-aided design system. In:
AFIPS Spring Joint Computer Conference, pp. 299–304. ACM, New York (1963)
2. Sutherland, I.E.: Sketchpad: a man-machine graphical communication system. In: AFIPS
Spring Joint Computer Conference, pp. 329–346. ACM, New York (1963)
3. Frazer, J.: Computing without computers. Architectural Design 75(2), 34–43 (2005)
4. Hernandez, C.R.B.: Thinking parametric design: introducing parametric Gaudi. Design
Studies 27(3), 309–324 (2006)
5. Dubendorf, V.A.: Wireless Data Technologies. John Wiley & Sons, West Sussex (2003)
6. Kasik, D.J., Buxton, W., Ferguson, D.R.: Ten CAD Challenges. IEEE Computer Graphics
and Applications 25(2), 81–92 (2005)
7. Lee, G., Sacks, R., Eastman, C.M.: Specifying parametric building object behavior (BOB) for
a building information modeling system. Automation in Construction 15(6), 758–776 (2006)
8. Monedero, J.: Parametric design: a review and some experiences. Automation in Construc-
tion 9(4), 369–377 (2000)
9. Burry, M.: Parametric Design and Sagrada Familia. Architectural Research Quarterly 1(4),
70–80 (summer 1996)
10. Drogemuller, R.M., Crawford, J., Egan, S.: Linking early design decisions across multiple
disciplines. In: Proceedings of the Conference on Product and Process Modelling in the
Building and Construction Industry (ECPPM 2004), Istanbul, Turkey (2004)
11. Aranda-Mena, G., Crawford, J., Chevez, A., Froese, T.: Building information modelling
demystified: does it make business sense to adopt BIM? International Journal of Managing
Projects in Business 2(3), 419–434 (2009)
12. Autodesk Inc.: Autodesk Conceptual Design Curriculum 2010 (2009),
http://usa.autodesk.com/adsk/servlet/
item?siteID=123112&id=14038871
 Software Openness: Evaluating Parameters of Parametric Modeling Tools 497

13. Aish, R., Woodbury, R.: Multi-level Interaction in Parametric Design. In: Butz, A., Fisher,
B., Krüger, A., Olivier, P. (eds.) SG 2005. LNCS, vol. 3638, pp. 151–162. Springer, Hei-
delberg (2005)
14. Maher, A., Burry, M.: The Parametric Bridge: Connecting Digital Design Techniques in
Architecture and Engineering. In: The 2003 Annual Conference of the Association for
Computer Aided Design in Architecture, Indianapolis, Indiana, pp. 39–47 (2003)
15. Nicholas, P., Burry, C.M.: Import as: Interpretation and Precision Tools. In: Proc. of the
12th International Conference on Computer Aided Architectural Design Research in Asia,
CAADRIA 2007, Nanjing, China (2007)
16. McGraw-Hill Construction: Interoperability in the Construction Industry (2007),
http://construction.ecnext.com/mcgraw_hill/includes/SMRI.pdf
17. Steel, J., Drogemuller, R.: Model Interoperability in Building Information Modelling. In:
Industrialized Software: KISS workshop @ Code Generation 2009, Cambridge, UK (2009)
18. Laiserin, J.: Next Gen BIM (2009),
http://www.laiserin.com/features/issue25/feature01.pdf
19. Hoffmann, C.M., Joan-Arinyo, R.: Parametric Modeling. In: Farin, G., Hoschek, J., Kim,
M.-S. (eds.) Handbook of Computer Aided Geometric Design, ch. 21. Elsevier B. V, Am-
sterdam (2002)
20. Aish, R.: Migration from an individual to an enterprise computing model and its implica-
tions for AEC Research. In: Berkeley-Stanford CE&M Workshop: Defining A Research
Agenda, Stanford, California (2000)
21. Burry, J., Burry, M.: The Bonds of Spatial Freedom. In: Proc. of the 26th eCAADe Con-
ference - Architecture in Computro, Antwerpen, Belgium, pp. 301–308 (2008)
22. Lawson, B.: Oracles, draughtsmen, and agents: the nature of knowledge and creativity in
design and the role of IT. Automation in Construction 14(3), 383–391 (2005)
23. Frazer, J.: An Evolutionary Architecture. Architectural Association Publications, London
(1995)
24. Janssen, P.: A generative evolutionary design method. Digital Creativity 17(1), 49–63 (2006)
25. Coates, P.S.: Some experiments using agent modelling at CECA. In: Generative Arts 2004,
Milan, Italy (2004)
26. Hansen, M.: Wearable Space. Configurations 10(2), 321–370 (2002)
27. Fischer, M., Kam, C.: Product Model & 4D CAD - Final Report. Stanford University (2002)
28. Khemlani, L.: SmartGeometry 2009 Conference Day (2009),
http://www.aecbytes.com/feature/2009/SmartGeometry2009.html
(accessed April 27, 2009)
29. Holzer, D., Tengono, Y., Downing, S.: Developing a Framework for Linking Design Intel-
ligence from Multiple Professions in the AEC Industry. In: Dong, A., Moere, A.V., Gero,
J.S. (eds.) Computer-Aided Architectural Design Futures (CAADFutures) 2007, Sydney,
Australia. Springer, Netherlands (2007)
30. Babu, S.: Tangible View Cube. CAD Professor (2009),
http://cadprofessor.in/2009/04/tangible-view-cube
31. Senseable, http://www.sensable.com
32. Sheppard, S.: A Wii Bit of Fun. In: Autodesk Labs: It’s Alive in the Lab (2009),
http://labs.blogs.com/its_alive_in_the_lab/2007/12/
a-wii-bit-of-fu.html
33. Salim, F.D., Mulder, H., Burry, J.: A System for Form Fostering: Parametric Modeling of
Responsive Forms in Mixed Reality. In: Proc. of the 15th International Conference on
Computer Aided Architectural Design Research in Asia, New Frontiers. Hong Kong,
China (2010) (accepted for publication)
34. Maeda, J.: Creative Code. Thames & Hudson, London (2004)
 Dynamic and Cyclic Response Simulation
of Shape Memory Alloy Devices

Yutaka Toi and Jie He

Institute of Industrial Science, University of Tokyo,

Komaba 4-6-1, Megruro-ku, Tokyo 153-8505, Japan
[email protected]

Abstract. Since most of the existing simulation researches on shape memory

alloys are focused on their static behavior, the simulation of their dynamic re-
sponse behavior is the blank field waiting for researchers to explore. In this pa-
per, both one-dimensional and three-dimensional dynamic simulations, as well
as an improved simulation model for shape memory alloys’ cyclic response be-
havior will be discussed. The validity of the dynamic and cyclic response
simulation model is illustrated by comparing the calculated results with the ex-
periment data. The damping capacity of shape memory alloys is also demon-
strated in both one- and three-dimensional simulations.

Keywords: shape memory alloys, dynamic response, finite element method,

superelasticity, shape memory effect.

1 Introduction

Since the first large-scale application of shape memory alloys (SMAs) as couplings in
the Grumman F-14 aircraft, the implementations of SMA devices have been spread
from aerospace to medical field, smart structures, robots etc.. The noninvasive charac-
teristic and excellent bio-compatibility make SMA devices widely applied in biomed-
ical fields. The controllable deformation feature of SMA devices has attracted many
attentions in robot researching field. SMAs have been used as actuators and sensors in
several experimental robots, such as robots made by Sugiyama and Hirai [1]. Besides,
the adjustable stiffness feature of SMAs have been used in passive control devices of
vibration controlling, showing a remarkable capability in controlling vibrations with
unknown frequency. Even more, the damping capacity of SMAs has become another
active research field recently.
However, the practical applications are still limited in small scale such as applica-
tions in medical instruments and aerospace devices. One reason is the high price of
SMAs, which has made the applications of SMA devices limited in the fields which
are not sensitive to cost. Another reason is a lack of understanding of SMAs' proper-
ties and simulation tools, which has limited the flexibility when designing SMA de-
vices. The cost problem can be solved by improving the manufacturing process and
mass producing, which are jobs for material scientists and manufacturing engineers.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 498–510, 2010.
© Springer-Verlag Berlin Heidelberg 2010
 Dynamic and Cyclic Response Simulation of Shape Memory Alloy Devices 499

As for the second problem, developing reliable and accurate models for SMA devices
is the most important task for researchers in the field of mechanics.
It is clear that, for most of the potential application fields of SMAs, such as SMA
actuators, SMA sensors and smart structures with SMAs, dynamic behavior simula-
tion is even more important than static behavior simulation. Since there have been
enough static models developed for a variety of SMAs, this paper is focused on mod-
eling the dynamic behavior of NiTi SMA devices. A SMA dynamic model is devel-
oped in this paper, whose feasibility is proved in both one-dimensional scale and
three-dimensional scales. The model which is used in this paper is based on the static
model developed by Brinson [2] and extended by Toi et al. [3]. The dynamic simula-
tion model is inspired by both Seelecke [4], who performed several one-dimensional
SMA dynamic simulations using Achenbach's model [5], and by Toi and Choi [6],
who have done several impact response simulations of SMA devices by considering a
strain rate effect in the previously formulated multi-axial constitutive equation [7].
The present modeling takes into consideration asymmetric tensile and compressive
deformation as well as multi-axial stress state which are not considered in the dynam-
ic analysis by Seelecke [4]. However, the strain rate effect, which is included in the
constitutive modeling by Toi and Choi [6], is neglected in the present modeling, as
the present study discusses relatively slow, vibration problems.
Afterwards, an improvement of SMA dynamic model considering cyclic effect [8-
11] is developed and discussed. The cyclic effect on constitutive behaviors of SMAs
in cyclic loading as well as cyclic deformation conditions, which is important in vi-
bration behaviors, was studied experimentally or analytically by Tobushi et al. [8],
Dolce and Cardone [9], Gall and Maier [10] and Sun and Rajapakse [11]. As for the
constitutive modeling for the cyclic effect, Tobushi et al. [8] and Sun and Rajapakse
[11] proposed the use of critical transformation stresses depending on the number of
cycles and the loading frequency, respectively. In the present study, the critical trans-
formation stresses and the maximum residual strains are assumed to depend on the
accumulated equivalent strain, which can be more conveniently used than the number
of cycles in the general, dynamic analysis. The frequency dependence is not consi-
dered in the present modeling, as the target of the present study is relatively slow
vibration analysis.

2 Description of Shape Memory Alloys

The special features of shape memory alloys are closely related to their internal phase
transformation behavior. Generally, there are two stable phases in SMAs: austenite in
high temperature and martensite in low temperature or under external loading. As
they are shown in Fig. 1, austenite phase is cubic crystal structure, while martensite
phases are monoclinic crystal structures. When we consider the SMA body as a
whole, if an external load is applied in a horizontal direction, every lattice of the body
is subjected to a shear stress. The direction of every lattice changes toward the same
direction. When the stress is beyond a critical value, an observable macroscopic
change occurs with the phase change from austenite or twinned martensite todet-
winned martensite. Detwinned martensite can also be transformed to austenite
500 Y. Toi and J. He

Fig. 1. Phase and crystal structure in SMAs

because of increasing Gibbs free energy when the SMA element is heated in the ab-
sence of applied stress.
Figure 2 shows the mechanical characteristics of SMAs. In Fig. 2(left), solid line
and dotted line show the stress-strain curve of superelastic effect and shape memory
effect respectively. Figure 2(right) is the critical stress-temperature diagram which
can be used to determine the phase concentration in different temperature and stress
settings. M and M indicate the starting and finishing temperature of phase trans-
formation from austenite to martensite when stress is 0; while A and A indicate the
starting and finishing temperature of phase transformation from martensite to auste-
nite in the same setting. The sloping solid line indicates the critical phase transforma-
tion stress in different temperature. The phase transformation from austenite to
martensite occurs when stress is increasing while it is between σ and σ . The
phase transformation from martensite to austenite occurs when stress is decreasing

Fig. 2. Equivalent stress vs. strain curve (left); Mechanical property of SMAs (right)
 Dynamic and Cyclic Response Simulation of Shape Memory Alloy Devices 501

while it is between σ and σ . As examples, the vertical solid line shows a typical
superelastic effect route, while the vertical dotted line shows typical shape memory
effect route, which correspond to the solid line and the dotted line in Fig. 2(left)
respectively.

3 One-Dimensional Dynamic Simulation

3.1 Phase Transformation and Constitutive Equation

The general one-dimensional stress-strain relation of SMAs is written as

σ σ E ε ε Ω ξ ξ θ T T . (1)
where E is Young's modulus; Ω is the transformation coefficient; ξ is the stress-
induced martensite volume fraction; θ is the thermal elastic coefficient; T is the
temperature. The subscript '0' indicates the initial values.
E is expressed as
E E ξ E E . (2)

where E is Young's modulus of austenite phase, E is Young's modulus of marten-

site phase, ξ is the total martensite volume fraction.
The total martensite volume fraction ξ is expressed as
ξ ξ ξ . (3)
where ξ is the temperature-induced martensite volume fraction.
The transformation coefficient Ω is expressed as
Ω ε E . (4)
where εL is the maximum residual strain, which is shown in Fig. 2 as εres. . To con-
sider the asymmetric behavior under tensile and compressive loading, Drucker-Prager
equivalent stress σDP [7] is applied in phase transformation equations as
σ σ β σ σ σ . (5)
The phase transformation equations are expressed as:
Martensite Transformation, when T M and C T M
C T M
ξ cos C T M . (6)

ξ ξ ξ ξ ξ . (7)

ξ ξ ξ . (8)
502 Y. Toi and J. He

Reverse Transformation, when T A and C T A C T A

ξ 1 cos π . (9)

ξ ξ . (10)

ξ ξ . (11)

3.2 Numerical Examples

For simulations of this section, the same SMA bar is used. Material constants of this
SMA bar are listed in Table 1.

Table 1. Material constants of SMA bar

Category Constants(Unit) Value

Moduli E MPa 60000
E MPa 20000
C MPa/ 8.0
C MPa/ 8.0
C MPa/ 13.0
C MPa/ 13.0
Transformation M -77.5
Temperature
M -65.0
A -14.5
A -21.7
Maximum residual ε 0.07
strains
Mass(kg) 0.0087

Simulation of Superelastic Response. For simulations of symmetric superelastic

response (β 0 in eq. (5)), free vibration case is listed. The temperature of SMA
bars is set at 10 , which is higher than austenite phase transformation finish tempera-
ture A at 14.5 . To present simulation results, the simulation time interval is set
at 10 seconds. The horizontal initial velocity of SMA bar is set at 0.17m/s. Ac-
cording to Fig. 3, the upper graph for the displacement time-history shows a clear
damping behavior, the amplitude keep decreasing until phase transformation no long-
er occurs. The lower graph shows a typical stress-strain curve of superelastic effect.
 Dynamic and Cyclic Response Simulation of Shape Memory Alloy Devices 503

Fig. 3. One-dimensional superelastic free vibration response simulation: Displacement vs. time
(upper), Stress vs. displacement (down)

The martensite phase transformation completes only in the first cycle and almost
stops after the 5th cycle.

Simulation of Asymmetric Superelastic Response. For simulations of asymmetric

superelastic response (β 0.15 in eq. (5)), the displacement time-history and the
stress-strain graphs are described in Fig. 4. The material constants, boundary condi-
tion and initial condition are exactly the same as those in the simulation of superelas-
tic response. Lower level of phase transformations can be observed in compressive
504 Y. Toi and J. He

Fig. 4. One-dimensional superelastic free vibration response simulation in asymmetric setting:

Displacement vs. time (upper), Stress vs. displacement (down)

side, which is caused by the higher requirement of stress for the phase transformation
in asymmetric setting.

4 Three-Dimensional Finite Element Dynamic Simulation

4.1 Constitutive Equation

To describe the three-dimensional dynamic behavior of SMA devices, the one-

dimensional stress-strain relation is extended into:
 Dynamic and Cyclic Response Simulation of Shape Memory Alloy Devices 505

σ D ε ξ Ω Tθ . (12)

where σ is the stress vector; D is the stress-strain matrix; ε is the strain

vector; ξ is the stress-induced martensite volume fraction; Ω is the transforma-
tion coefficient vector; T is the temperature; θ is the thermal elastic coefficient
vector.
The transformation coefficient vector Ω is expressed as

Ω D R ε . (13)

where R is the residual strain direction matrix; ε is the maximum residual

strain vector. The thermal elastic coefficient vector θ is expressed as

θ D α . (14)

where α is the coefficient of thermal expansion. By using Euler's formula for nu-
merical time integration, we can derive the stress-strain relation equation as

∆σ D ∆ε ξ D ∆R ε ∆σ ∆t D ε ξ D R ε
(15)
ξ D R ε TD α TD .

or

∆σ X D ∆ε X Y . (16)

or

∆σ D ∆ε Θ . (17)

where

X I ξ D ∆R ε . (18)

Y ∆t D ε ξ D R ε ξ D R ε TD α TD . (19)

By using Newmark's β method, the finite element discretized equation of motion

governed by the stress-strain relation above can be solved efficiently and accurately.

4.2 Numerical Examples

The forced vibration simulation in superelastic setting is described in this section. The
simulation object is listed in Fig. 5. It is a highway-bridge-like concrete structure with
506 Y. Toi and J. He

SMA dampers installed in their upper part (in black). The bottom of the structure is
fixed in all directions. Therefore not only the dynamic behavior but also the damping
capacity of SMAs will be shown in following subsection.

Simulation of Superelastic Forced Vibration Response. To show the three- dimen-

sional SMA device's superelastic dynamic behavior, the temperature is set at 50 ,
which is higher than the austenite phase transformation starting temperature
As 42 . The external force is a function of sine, which is expressed as

F F sin β time . (20)

In this simulation, the maximum external force is set at 4 10 N. The frequency

coefficient β is set at 0.1256; therefore the cycle time is about 50s. The number of
eight-node hexahedron elements is 1, 2, and 13 in x-, y- and z-direction, respectively.
The calculated result is shown in Fig. 6. Similar to the one-dimensional simulation, a
clear stress-strain curve can be observed. But the displacement history is much
more complicated than that in the one-dimensional simulation. The concrete ele-
ments of this structure show an elastic behavior according to its equivalent stress-
strain curve.

Fig. 5. Forced vibration simulation object

 Dynamic and Cyclic Response Simulation of Shape Memory Alloy Devices 507

Fig. 6. Three-dimensional superelastic forced vibration response simulation: top displacement

vs. time (upper) equivalent stress vs. strain of SMA edge (down)

5 Improved Model for Cyclic Behavior of Shape Memory Alloys

5.1 Constitutive Equation

The constitutive models described in section 3 and section 4 are also applicable for
one-dimensional and three-dimensional cyclic dynamic simulation. After careful
analysis of experiment data from Tobushi et al. [8], Dolce and Cardone [9], Gall and
Maier [10] and Sun and Rajapakse [11], it is found that the cyclic behavior of SMA
devices can also be described by making a connection between the accumulative
strain and the martensite phase transformation start and finish temperature M and
508 Y. Toi and J. He

M , the reverse phase transformation start and finish temperature A and A , as well
as the maximum residual strain ε . The accumulative strain is expressed as follows

e ε dt . (21)

where e is the accumulative strain, ε is the equivalent strain rate. The connection
between the accumulative strain e and the martensite phase transformation start and
finish temperature M and M , as well as the reverse phase transformation start and
finish temperature A and A are expressed as

M e M 0 α 1 exp β e . (22)

M e M 0 α 1 exp β e . (23)

A e A 0 α 1 exp β e . (24)

A e A 0 α 1 exp β e . (25)

where M 0 ,M 0 , A 0 and A 0 are the phase transformation critical tempera-

tures. α , α and β , β are the material constants, which have to be determined
by experiment data. The maximum residual strain is divided into two coefficient, ε ,
which is the maximum residual strain during martensite phase transformation, and ε ,
which is the maximum residual strain during reverse phase transformation. ε and
ε are expressed as

ε e ε 0 α 1 exp β e . (26)

ε e ε 0 α 1 exp β e . (27)

where ε 0 is called the initial maximum residual strain; α , α , β and β are the
coefficients which have to be determined by experiment data. By substituting the equa-
tions above into the martensite phase transformation equation which was described in
section 2, we will be able to obtain the cyclic dynamic behavior of SMA devices.

5.2 Numerical Examples

According to the experimental data provided by Tobushi el al. [8], the simulation
object (SMA bar) is similar to the one which is described in the section 2. Unloading
starts as soon as the strain reaches 9%. Detailed simulation results are listed in Fig. 7.
The experimental data shows the stress and strain relationship during No. 1, 2, 5,
10, 20, 50, 100 cycles. The initial maximum residual strain is set at 3.5%, the
temperature is set at 353K. The experimental data shows that critical stresses of start
 Dynamic and Cyclic Response Simulation of Shape Memory Alloy Devices 509

and finish martensite phase transformation as well as the reverse transformation are
getting lower and lower, while the maximum residual strain is getting larger and
larger. The simulation result which is presented in lower graph of Fig. 7 shows a
similar behavior. The critical stresses and the maximum residual strain are changing
faster at the beginning then slowing down, because of the term (1-exp(-X)) in
coefficients revolution function. Only the cycles No. 1, 2, 5, 10, 20, 50 and 100 are
presented in Fig. 7 to show a better comparison with the experimental data.

Fig. 7. Experimental data and simulation results under cyclic loading: simulation data (upper),
experimental data (down)

6 Concluding Remarks

In this paper, the one-dimensional and three-dimensional dynamic response models

for SMA devices have been developed. Besides, model improvement works have
510 Y. Toi and J. He

been done by introducing cyclic effect and asymmetric behavior of SMA devices into
the phase transformation mechanism. According to the simulations performed in this
paper, several conclusions can be obtained as follows:
1. The validity of the dynamic response simulation model as well as the improved
phase transformation mechanism to describe SMAs' superelastic and quasi-plastic
behavior has been illustrated.
2. An improved SMA model by considering the cyclic effect has been developed and
tested.
3. The potential of SMAs being used as damping devices has been demonstrated.
The future works will be focused on enhancing the simulation model's stability,
especially for high frequency vibration simulations in three-dimensional scales.

References

1. Sugiyama, Y., Hirai, S.: Crawling and jumping by a deformable robot. International Jour-
nal of Robotics Research 25(5), 603 (2006)
2. Brinson, L.: One-dimensional constitutive behavior of shape memory alloys: thermome-
chanical derivation with non-constant material functions and redefined martensite internal
variable. Journal of Intelligent Material Systems and Structures 4(2), 229 (1993)
3. Toi, Y., Lee, J., Taya, M.: Finite element analysis of superelastic, large deformation beha-
vior of shape memory alloy helical springs. Computers and Structures 82(20-21), 1685–
1693 (2004)
4. Seelecke, S.: Modeling the dynamic behavior of shape memory alloys. International Jour-
nal of Non-Linear Mechanics 37(8), 1363–1374 (2002)
5. Achenbach, M.: A model for an alloy with shape memory. International Journal of Plastici-
ty 5(4), 371–395 (1989)
6. Toi, Y., Choi, D.: Constitutive Modeling of Porous Shape Memory Alloys Considering
Strain Rate Effect. Journal of Computational Science and Technology 2(4), 511–522
(2008)
7. Toi, Y., Choi, D.: Computational Modeling of Superelastic Behaviors of Shape Memory
Alloy Devices under Combined Stresses. Journal of Computational Science and Technolo-
gy 2(4), 535–546 (2008)
8. Tobushi, H., Iwanaga, H., Tanaka, K., Hori, T., Sawada, T.: Stress-strain-temperature rela-
tionships of TiNi shape memory alloy suitable for thermomechanical cycling. JSME Inter-
national Journal. Series 1, Solid Mechanics, Strenght of Materials 35(3), 271–277 (1992)
9. Dolce, M., Cardone, D.: Mechanical behaviour of shape memory alloys for seismic appli-
cations 2. Austenite NiTi wires subjected to tension. International Journal of Mechanical
Sciences 43(11), 2657–2677 (2001)
10. Gall, K., Maier, H.: Cyclic deformation mechanisms in precipitated NiTi shape memory
alloys. Acta Materialia 50(18), 4643–4657 (2002)
11. Sun, S., Rajapakse, R.: Simulation of pseudoelastic behaviour of SMA under cyclic load-
ing. Computational Materials Science 28(3-4), 663–674 (2003)
 Adaptive Fuzzy Filter for Speech Enhancement

Chih-Chia Yao and Ming-Hsun Tsai

Department of Computer Science and Information Engineering,

Chaoyang University of Technology, Wufong, Taichung 41349, Taiwan
[email protected]

Abstract. In this paper an adaptive fuzzy filter, based on fuzzy system, is pro-
posed for speech signal enhancement and automatic speech recognition accu-
racy. In the past two decades the basic wavelet thresholding-algorithm has been
widely studied and is common applied to filter noise. In the proposed system
adaptive wavelet thresholds are generated and controlled by fuzzy rules con-
cerning the presence of speech in noise. First an amplified voice activity detec-
tor is designed to improve performance on SNR lower than 5dB. Then an
adaptive threshold decision module based on fuzzy inference system is pro-
posed. In this fuzzy inference system overall relations between speech and
noise are summarized into seven fuzzy rules and four linguistic variables, which
are used to detect the state of signals. The adaptive threshold and membership
functions are optimally obtained by particle swarm optimization algorithm so
the SNR of the filter output for training signal data can be maximized. Experi-
mental results reveal that our proposed system effectively increases the SNR
and the recognition rate of speech.

Keywords: speech enhancement; wavelet thresholding; fuzzy; voice activity

detection.

1 Introduction
Speech enhancement remains an important problem within the field of speech and
signal processing, such as mobile communication, speech recognition and hearing
aids [1]. Speech enhancement algorithms have undergone numerous studies through-
out the last two decades yet issues such as distortions to the original speech signal and
residual noise, sometimes in the form of musical tones created by the enhancement
algorithms, remain unsolved [2, 3, 4].
The main objective of speech enhancement is to reduce corrupting noise while pre-
serving as much original clean speech quality as possible. The objectives of speech
enhancement algorithms can be classified into increasing the recognition rate and the
signal-to-noise ratio (SNR) of the distorted speech [5, 6]. These can be roughly classi-
fied as digital signal processing and statistical analysis. The former usually remove an
estimate of the distortion from the noisy features, e.g. spectral subtractive [7]. The
latter utilizes a statistical model to describe predictable structures and expected pat-
terns in the signal process [8, 9].

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 511–525, 2010.
© Springer-Verlag Berlin Heidelberg 2010
512 C.-C. Yao and M.-H. Tsai

The principle of a spectral subtractive algorithm is to obtain the best possible esti-
mations of the short time spectra of a speech signal from a given contaminated
speech. The main attractions of a spectral subtractive algorithm are: (1) Easy and
simple implementation. (2) High flexibility against subtraction parameter's variation.
On the other hand, spectral subtractive algorithms suffer a number of deficiencies
including an imprecise estimation of signal and noise parameters and a mismatch of
the probability distribution models of speech or noise [10].
The wavelet thresholding method is often used to shrink the signal and remove the
noise. Using an adaptive thresholding algorithm could provide better performance
[11, 12]. In this paper we propose an adaptive fuzzy filter, based on wavelet thresh-
olding, to improve performance in speech signal enhancement and automatic speech
recognition accuracy. In this developed adaptive fuzzy filter the signals are classified
into speech and non-speech segments by an amplified voice activity detector. This
amplified voice activity detector employs the full wavelet packet transform (instead of
the perceptual wavelet packet transform) to decompose the input speech signal into
critical sub-band signals [13]. In each critical sub-band signal the mask construction is
obtained by smoothing the Teager energy operator and the entropy of corresponding
wavelet coefficients. Thus, the amplified voice activity detector enhances the ability
of distinction when the SNR is lower than 5dB.
Although some algorithms have been proposed to generate the adaptive threshold
the model between speech and noise is still not clear and the rules judging the exis-
tence of noise is usually expressed in an ambiguous style. A new-type of speech
enhancement system based on fuzzy rules is proposed here. In the design overall
relations between speech and noise are summarized into seven fuzzy rules. Adaptive
thresholds are then inferred by those fuzzy rules. Moreover, this filter is optimally
designed by particle swarm optimization, so that the SNR of the filter output is maxi-
mum; membership functions to represent the rules are not required to be obtained
beforehand.
As is well know, using critical parameters is helpful on detecting the presence of
speech in noisy environment. Some researchers have focused on energy, zero crossing
rate and time duration to find the boundary between the word signal and background
noise. Other parameters have been proposed such as linear prediction coefficient,
linear prediction error energy, pitch information and time-frequency parameter. How-
ever these parameters cannot yet be well adapted to variable-level background noise
even if more complex decision strategies are used. In this paper four parameters used
as linguistic variables are incorporated in order to detect the presence of speech in
noisy environment. The four parameters are energy, zero crossing rate, entropy and
standard deviation. Examples of processing actual speech with eight types of noises
are shown to demonstrate the high performance of this filter.
The remainder of this paper is organized as follows. In Section 2 we review the ba-
sic concept of wavelet transform and wavelet packet decomposition. In Section 3 the
scheme of the speech enhancement system is introduced. In Sections 4 we introduce
the fuzzy inference model for speech enhancement. In Section 5 a performance
evaluation of the proposed system is presented and comparisons with other protocols
are made. Our conclusions are made in Section 6.
 Adaptive Fuzzy Filter for Speech Enhancement 513

2 Basic Concepts

2.1 Wavelet Packet Transformation

Wavelet Transformation is widely used in signal processing because processing sig-

nals in the frequency domain is often easier to implement [13]. In comparison to short
time Fourier transform, wavelet transform allows the use of long-time intervals to
obtain more precise low-frequency information and shorter intervals for high-
frequency information. The continuous wavelet transform (CWT) of a signal x(t) is
given as follows:
1 ∞ ⎛ t −τ ⎞
X CWT (a, τ ) = ∫ x(t )ψ ⎜⎝ a ⎟⎠dt (1)
−∞
a
where τ and a represent the time shift and scale variables, respectively, and ψ (⋅) is
the mother wavelet chosen for the transform.
In the discrete version, the wavelet decomposes the signal with variable frames to
perform multi-resolution analysis in a dyadic form known as discrete wavelet trans-
form (DWT). In DWT the scale and translation parameters of the discrete wavelet
family are given by a = 2 m and τ = n2 m .
Signals transformed by DWT can be regarded as splitting signals by a series of
high-pass and low-pass filters. The high-frequency, low-scale components of speech
signals through a high-pass filter are preserved as details. In the same way, the low-
frequency, high-scale components are preserved as approximations. In discrete wave-
let transform only approximations can be decomposed iteratively.
A further generalization of the DWT is the wavelet packet transform (WPT) that
offers a richer range of possibilities in signal analysis. In WPT the decomposing proc-
ess is iterated on both high and low frequency components rather than continuing only
on low frequency terms as with a standard DWT.
The principle of wavelet packet transform is that, given a signal, a pair of low pass
and high pass filters is used to yield two sequences to capture different frequency sub-
band features of the original signal. The depth of the wavelet packet tree can be varied
over the available frequency range, resulting in configurable decomposition. The two
wavelet orthogonal bases are defined as
∞
ψ 2j +p1 (k ) = ∑ h[n]ψ
n = −∞
p
j ( k − 2 j n) (2)

∞
ψ 2j +p1+1 (k ) = ∑ g[n]ψ
n = −∞
p
j ( k − 2 j n) (3)

where h[n] and g[n] denote the low-pass and high-pass filters, respectively. ψ (n) is
the wavelet function and parameters j and p are the number of decomposition levels
and nodes, respectively.

2.2 Particle Swarm Optimization

In 1995, Kennedy and Eberhart introduced the particle swarm optimization algorithm
(PSO) into the field of social and cognitive behavior [14]. Traditionally the main prob-
lem in designing a neural fuzzy system is training the parameters. Backpropagation
514 C.-C. Yao and M.-H. Tsai

training is commonly adopted to solve this problem. However the steepest descent
approach, commonly used in backpropagation training to minimize the error function,
may reach the local minima very quickly and never find the global solution. Accord-
ingly, a new optimization algorithm, called particle swarm optimization (PSO), appears
to provide better performance than the backpropagation algorithm.
Like other population-based optimization approaches PSO is initialized with a
swarm of random solutions, each swarm consists of many particles. Each particle is
r
characterized by its current position xi = [ x i1 , x i2 , L , xiD ] and current velocity
r
vi = [v i1 , vi2 , L , viD ] , where D stands for the dimensions of the solution space. In the
PSO the trajectory of each particle in the search space is adjusted by dynamically
altering the velocity of each particle. Then the particles rapidly search the solution
space using the moving velocity of each particle. Each of these particle positions is
scored to obtain a fitness value based on how to define the solution of the problem.
During the evolutionary process the velocity and position of particle i are updated as
r r r r
vi = ω × vi + ϕ1 × rand1 ( ) × (NBesti − xi ) + ϕ 2 × rand 2 ( ) × (GBesti − xi ) (4)
r r r
x i = xi + vi (5)

where ω is the inertia weight, ϕ1 and ϕ 2 are the acceleration coefficients, respec-
tively. The second term in Eq. (4), called the cognitive component, reflects the experi-
ence of a particle since it is dependent on the best position of the respective particle.
The third term is referred to as the social component and contains the information of a
social group due to the dependence on the neighborhood best position. The random
numbers rand1 ( ) and rand 2 ( ) are chosen from the interval U(0,1). In Eq. (4), GBest
is the position with the best fitness found so far for the ith particle, and NBest is the
r r
best position in the neighborhood. The term vi is limited to the range ± v max . If the
velocity violates this limit, then it is set to the actual limit. Changing the velocity en-
ables each particle to search around its individual best position and global best posi-
tion. After initialization of the positions and velocities of the particles update equations
are applied to every particle in each iteration until a stopping criterion is fulfilled.

3 Architecture of Adaptive Fuzzy Filter

This section proposes a new framework for speech enhancement called an Adaptive
Fuzzy Filter. This utilizes a fuzzy inference system to determine the threshold used to
suppress noise. The schematic diagram of an adaptive fuzzy filter is shown in Fig. 1. The
structure of an adaptive fuzzy filter comprises wavelet packet transform, a voice activity
detection module, a fuzzy inference module, a learning module and a speech filter.
In the adaptive fuzzy filter wavelet packet transform is applied to each input frame
and a Level-3 wavelet packet decomposition tree with eight frequency bands is ob-
tained. A voice activity detection module is then used to distinguish the presence or
absence of speech. The fuzzy inference module determines the adaptive threshold
used in the speech filter in order to suppress noise. After processing the inverse wave-
let packet transform is applied to provide the enhanced speech. Particle swarm
optimization is used in the learning module to provide the optimal solution.
 Adaptive Fuzzy Filter for Speech Enhancement 515

Fig. 1. The schematic diagram of fuzzy speech enhancement model

3.1 Voice Activity Detection (VAD)

Voice activity detection (VAD) is used to distinguish speech from contaminated

speech signals and is required in a variety of speech communication systems. In pre-
vious studies a Teager energy operator has been adopted to distinguish noise and
speech and has been proven to provide excellent performance in both additive noisy
and real noisy environments [15]. The discrete form of the TEO is given by
T [ y(n)] = y 2 (n) − y(n + 1) y(n − 1) (6)

where T [ y (n)] is called the TEO coefficient of y(n). However a Teager energy opera-
tor is insensitive when the signal-to-noise ratio is low. As an example, when the SNR
is lower than 0dB the difference between noisy energy and speech energy is not obvi-
ous and the performance of TEO is not satisfactory. Fig. 2 shows an example of using
TEO to distinguish speech from noise signals contaminated by -5dB Gaussian white
noise. To overcome this problem, in our proposed module, a Teager energy operator
is cooperated with entropy to improve the ability of distinction. The formula of en-
tropy is shown as follow:
E ( S ) = ∑ p ( si ) log p ( si ) (7)
i

Entropy represents the degree of variation. When the noise model is stationary or
slight non-stationary the corresponding entropy is kept stable or slight changed. Fig. 3
shows the comparison of using TEO and entropy on VAD to distinguish signals con-
taminated by -5dB Gaussian white noise.
The proposed VAD algorithm computes signals wm ( x i ) that have been produced
by the wavelet package transform on each input frame to produce 2J sub-band wavelet
packets, where J is the number of levels for the wavelet packet decomposition tree
and 1 ≤ m ≤ 2 J . Then, a set of T [ wm ( xi )] can be derived from Eq. (6). The scheme of
the voice activity detection is designed as follows:
⎧ 0, if var(T [ wm ( xi )]) < λ J or var( E[ wm ( xi )]) < ς (8)
Vm ( x i ) = ⎨
⎩ xi , otherwise
where ς is user-defined and λ J = σ J 2 log( N J ) , proposed by Johnston and Silverman
[16]. var(⋅) denotes the variance.
516 C.-C. Yao and M.-H. Tsai

Fig. 2. The property of Gaussian white noise with -5db

Fig. 3. The comparison of VAD based on TEO and entropy with -5db Gaussian white noise

4 Adaptive Threshold Decision Module Based on Fuzzy Inference

System
In this section, an adaptive threshold decision module based on fuzzy inference sys-
tem is proposed. In the module the adaptive threshold for each sub-band signal is
 Adaptive Fuzzy Filter for Speech Enhancement 517

generated by the fuzzy inference system and the noise components were removed by
thresholding the wavelet coefficients.
Let s(t) be clean speech with finite length and n(t) be noise, then contaminated
speech y(t) can be expressed as
y(t ) = s(t ) + n(t ) (9)

If W denotes the wavelet transform matrix, Eq. (9) can be written in the wavelet
domain as
Y (t ) = S (t ) + N (t ) (10)

where Y (t ) = W ⋅ y (t ) , S (t ) = W ⋅ s(t ) and N (t ) = W ⋅ n(t ) . The estimated speech sig-

nal Sˆ (t ) can be obtained by using the thresholding function

Sˆ (t ) = FT (Y , T ) (11)

where FT (Y , T ) denotes the thresholding function and T is the threshold. The standard
thresholding function includes the soft thresholding function which is defined as

⎧sign(Y )( Y − T ), Y ≥T (12)
FTS (Y , T ) = ⎨
⎩ 0, Y <T

In this paper an adaptive threshold decision module based on a fuzzy inference system
was proposed. In the fuzzy inference system overall relations between speech and
noise are summarized to seven fuzzy rules. Moreover, four linguistic variables are
used to detect the state of signals. They are energy, zero crossing rate, standard devia-
tion and average residual. The reasons for adopting those four parameters are listed as
follows.
-Energy:
Energy is an effective factor on measuring the degree of noise when the SNR is big-
ger than 0dB. Studies have shown that energy provides excellent performance on
voice activity detection. However, rapid variation of energy in the speech model
causes implementation difficulties.
-Zero crossing rate:
Zero crossing rate (ZCR) is a basic acoustic feature that can be computed easily. It is
equal to the number of zero crossing of the waveform within a given frame. In general
the ZCR of both unvoiced sounds and environment noise are larger than voiced
sounds. ZCR is often used in conjunction with volume for end-point detection. It is
hard to distinguish unvoiced sounds from environment noise by using ZCR alone
since they have similar ZCR values.
-Standard deviation:
Standard deviation is a measure of how wide any given numbers are spread. It is use-
ful in comparing sets of data, which may have the same mean but different range. In a
given frame standard deviation is helpful to distinguish the signal's mode. The
standard deviation of a stable noise is different to voice sound. The observation of
standard deviation is useful in the threshold decision level.
518 C.-C. Yao and M.-H. Tsai

-Average residual:
The statistical property of average residual is similar to standard deviation. In this
paper standard deviation is used to detect the variation within a frame but average
residual is used to compare the difference between frames. By calculating the average
residual across several frames the signal's mode can be distinguished, whether the
variation is temporal or not, in a very short time interval. In this paper the average
residual is calculated based on the standard deviations of previous frame, current
frame and next frame.
In the fuzzy inference system, Eq. (13) is used to obtain the optimal solution.
T (k + 1) = T (k ) + α (k )THS (k ) (13)
Here, α(k) is a step length for threshold’s variation and THS(k) is the kth threshold.
α(k) and THS(k) are decided by above four parameters. First the range of threshold is
decided by energy and zero crossing rate. Studies have shown that energy is an effec-
tive factor in measuring the degree of noise greater than 0dB. However when the
threshold is only decided by energy any rapid variation of energy causes difficulty
and irrational implementation. On top of which, most mechanisms generally use an 8k
sample rate to record the speech signal and, when the frequency of the noise is bigger
than the sample rate, energy is ineffective. To overcome above problem zero crossing
rate is introduced to cooperate with energy on deciding the threshold.
As described in previous statements the ZCR of unvoiced sounds and environment
noise are larger than the ZCR of voiced sounds. In many studies a zero crossing rate is
used on voice activity detection. The signal's mode can be easily distinguished by
using zero crossing rate . When the noise is high frequency and the speech is low
frequency the zero crossing rate can be taken as the auxiliary and be cooperated with
energy. When the frequency of noise is low the threshold decision is highly dependent
on energy. As the frequency of noise becomes higher zero crossing rate is taken into
consideration. In short the importance of zero crossing rate on the threshold decision
is proportional to the frequency of noise.
Standard deviation and average residual are then introduced to control the degree
of threshold variation. The degree of variation on signal's modes between two adja-
cent frames is reflected in the difference between standard deviations. As the differ-
ence between standard deviations increases so the amount of threshold variation in-
creases. However this does not apply when speech is contaminated by instantaneous
impulse signal. An instantaneous impulse signal in a given frame will usually increase
the standard deviation and often causes a violent vibration of threshold. To avoid this
problem, the average residual of any adjacent three frames is calculated to smooth the
threshold's variation. The average residual is calculated from standard deviations of
any adjacent three frames and can effectively reduce the impact of instantaneous
impulse signals. Should the value of the average residual increase there is a high prob-
ability that it is neither temporal nor instantaneous.

4.1 Fuzzy Inference Rules

The threshold can be set by the local characteristics of the input signals but, as previ-
ously mentioned, there are many ambiguous cases where constructing the noise model
 Adaptive Fuzzy Filter for Speech Enhancement 519

Table 1. Linguistic variables - Fuzzy set relational table

Linguistic Variable Fuzzy Sets

Energy Low High
Zero crossing rate Low High
Standard deviation Low High
Average residual Low High
Threshold Low Median High
Step length Low Median High

from the contaminated signal is hard to achieve. In the proposed speech enhancement
system a fuzzy inference system is adopted to overcome these problems and the adap-
tive threshold is generated in order to suppress noise. To begin with various models of
noises and the relation between noise and speech signals are summarized into several
cases. The fuzzy sets of linguistic variables and their corresponding linguistic terms
are then defined according to the analysis. The relations between linguistic variables
and their fuzzy sets are shown in Table 1.
After eliminating the cases where the noise is impossible to distinguish from
speech signal, seven fuzzy rules are concluded, listed as follows.

Rule 1: If Energy is High and Zero Crossing Rate is Low, then Threshold is Medium.
Energy is an effective factor on measuring the degree of noise when the SNR is big-
ger than 0dB. In most cases speech signal has the property of high energy as opposed
to noise. In attempt to retain speech and suppress noise, the threshold can be set lower
than the speech signal. This rule is applicable in places like airports or streets with
heavy traffic. In this model the zero crossing rate is classified to low when the fre-
quency is between 50Hz and 5 KHz.

Rule 2: If Energy is High and Zero Crossing Rate is High, then Threshold is High.
As the zero crossing rate becomes higher, the probability of noise existence increases.
In rule 2 the probability of the existence of noise is higher than in rule 1, so that the
threshold can be set higher than rule 1. This rule is applicable when the energy is
large and the frequency is high such as noise generated by high frequency electric
products or machines. In this model the zero crossing rate is classified to high when
the frequency is bigger than 5K Hz.

Rule 3: If Energy is Low and Zero Crossing Rate is Low, then Threshold is Low.
Low zero crossing rate usually means the probability of noise existence is not high. In
most case the max energy is detected from the speech signal. Therefore, the threshold

…
is set to low when the energy of contaminated speech signal is low. The type of noises
is generated such as by the low frequency electric products (indoor fan, , etc).

Rule 4: If Energy is Low and Zero Crossing Rate is High, then Threshold is Medium.
On the basis of rule 4, high zero crossing rate denotes a high probability of noise. The
threshold should be bigger than in rule 3 but not exceed the energy of speech signal.
This rule is applicable on the environments such as the electromagnetic wave.
520 C.-C. Yao and M.-H. Tsai

Rule 5: If Standard deviation is Low and Average residual is Low, then Step length is
Low.
This rule is applicable when the model of noise is stable, whether in a very short time
interval or in a longer time interval. The property of this noise is that the standard
deviation is small and the average residual calculated from three successive frame
blocks' standard deviations is also small. Since the standard deviation is small, the
low threshold is necessary. The waveform under this situation is shown in Fig. 4 In
Fig. 4, vertical axis represents the amplitude and horizontal axis represents the time.
Besides, in Fig. 4, each frame is bounded by the rectangle. The standard deviation of
each frame is small as the waveform within each frame is stable. The average residual
of the successive three frame's standard deviation is also small since the standard
deviations are all small.

Fig. 4. The property of the noise is low standard deviation and low average residual

Rule 6: If Standard deviation is High and Average residual is Low, then Step length is
Medium.
This rule is applicable when noise is stable but variant in a very short time interval. The
property of this noise is that the standard deviation is high but the residual that is calcu-
lated from current and neighbor frame's standard deviations is small. In rule 6 high
standard deviation means the amplitude of noise is higher than the noise in rule 5.
Hence, the threshold should be set higher than in rule 5. Under the situation with multi-
ple sources of noise, this rule can bring the function into full play. The waveform under
this situation is shown in Fig. 5. In Fig. 5, a big variation of the waveform within the
middle frame causes the standard deviation to be big. The average residual of the suc-
cessive three frame's standard deviation is small since the standard deviations are all big.

Fig. 5. The property of the noise is high standard deviation and low average residual
 Adaptive Fuzzy Filter for Speech Enhancement 521

Rule 7: If Standard deviation is High and Average residual is High, then Step length is
High.
This rule is applicable when noise is unstable, whether in a very short time interval or
in a long time interval. In rule 7 the current frame's standard deviation is high and is
higher than neighbor frames so that the average residual calculated from current and
neighbor frame's standard deviations is high. The waveform under this situation is
shown in Fig. 6.

Fig. 6. The property of the noise is high standard deviation and high average residual

The next step is to design this function “Threshold” and “Step length” for Energy,
Zero crossing rate, Standard deviation and Average residual. In most fuzzy systems
membership function is obtained by fuzzy approximate reasoning using the member-
ship function of the Low and the High of Energy, Zero crossing rate, Standard devia-
tion, Average residual and its relationship to Threshold and Step length. In this paper
the Gaussian type and the sigmoid function are used to approximate the membership
function. The benefit is efficient calculation. In the field of signal processing the goal
is to obtain the optimum filter, in which the output is as close to the original signal as
possible; thus some optimization is required in designing this filter. The particle
swarm optimization algorithm (PSO) is applied to optimize the parameters in these
nonlinear function. As we know, particle swarm optimization algorithm provides
better performance on finding the global solution while optimizing the overall struc-
ture. Experimental results demonstrate that our proposed system provides excellent
performance on speech enhancement.

5 Experimental Results
This section demonstrates the effectiveness of our proposed system. The experimental
results that pertain to our proposed speech enhancement system are compared to spec-
trum subtraction and signal subspace approach.
First, tests for the enhancement and recognition task were carried out on the Aurora
2 databasen [17]. In the database the speech files include recording of the 10 English
digits and 26 English letters spoken by males and females. The Aurora 2 database
does not only provide the uncontaminated speech of 4004 sentences it also includes
the speech of 48048 sentences contaminated with eight types of noise (Subway, Bab-
ble, Car, Exhibition, Restaurant, Street, Airport and Station) and are mixed with -5, 0,
5, 10, 15 and 20dB.
522 C.-C. Yao and M.-H. Tsai

In order to achieve a consistent comparison the HTK speech recognition system is

adopted as the classifier [18]. The HTK speech recognition system was first devel-
oped by the Speech Vision and Robotics Group of Cambridge University. With this
tool, the HMM model, language model, training model and etc. can be established. In
the Aurora 2 database, HTK has been adopted for the training of the model and for
recognition. Therefore, we make use of the same approach to configure and establish
the model, and process for recognition.

(a) Contaminated signal

(b) Enhancement signal

Fig. 7. (a) Speech signal contaminated by 10dB car noise. (b) Enhancement signals denoised by
our proposed system.

(a)

(b)
Fig. 8. (a) The frequency of speech signal contaminated by 10dB car noise. (b) The frequency
of enhancement signals denoised by our proposed system.
 Adaptive Fuzzy Filter for Speech Enhancement 523

The evaluation of the proposed system includes two parts. First the signal-to-noise
ratios on before and after enhanced signals are evaluated. Next applying those en-
hanced speech signals into the HTK classifier compares the recognition rates of our
proposed system with other methods. The experimental results are listed as follows.
Fig. 7 shows the waveform before and after signal denoising by our proposed sys-
tem. The signal is mixed with 10dB car noise. The corresponding time-frequency
diagrams are reveals in Fig. 8.
The comparisons of SNR are included to demonstrate the effectiveness of our pro-
posed system. Tables 2 and 3 show the SNR before and after signal denoising by
our proposed system. The experimental results show that our proposed system can
effectively remove noise.
Finally, the recognition rates of our proposed system with other methods are listed
in Tables 4 - 6. Table 4 shows the rate of recognizing the signal without denoising.
Tables 5 and 6 show the rate of recognizing the signal after denoising by our

Table 2. The SNR of the signals in Aurora 2 database

SNR (without enhancing)

Subway Babble Car Exhib. Rest. Street Airport Station AVE.
20dB 18.5 18.3 18.4 18.6 18.38 18.32 18.45 18.38 18.41
15dB 13.5 13.6 13.4 13.6 13.45 8.91 8.96 13.55 12.37
10dB 8.85 8.9 8.95 8.92 4.97 4.94 4.96 8.93 7.42
5dB 4.95 4.9 4.98 4.96 4.97 4.95 4.92 4.94 4.94
0dB 0.84 0.62 0.08 0.12 0.11 0.3 0.26 0.09 0.30
-5dB -4.81 -4.26 -4.33 -4.13 -5.01 -4.86 -4.97 -5.06 -4.67

Table 3. The SNR of the signals after enhancing by our proposed algorithm

SNR (after enhancing)

Subway Babble Car Exhib. Rest. Street Airport Station AVE.
20dB 20.1 18.9 20.3 19.7 18.8 19.2 19.4 18.7 19.38
15dB 15.5 14.8 13.4 14.9 13.9 14.2 13.6 14.4 14.33
10dB 11.4 10.3 11.6 11.3 9.2 10.6 9.8 10.94 10.64
5dB 7.4 6.1 8.1 6.6 5.7 6.7 6.4 6.5 6.68
0dB 4.45 3.68 5.3 4.3 3.16 3.77 3.63 3.53 3.97
-5dB 1.33 1.1 1.64 1.21 0.88 1.16 0.86 1.18 1.17

Table 4. The recognition rate of the signals without enhancing

Recognition Rate % (withouth enhancing)

Subway Babble Car Exhib. Rest. Street Airport Station AVE.
clean 98.8 98.6 98.8 99.1 99.3 99.25 99.22 99.5 99.07
20dB 97.6 97.8 98.5 97.4 98.4 97.8 98.2 98.95 98.08
15dB 95.1 94.2 95.7 94.9 95.4 94.35 95.3 95.7 95.08
10dB 84.9 81.4 80.5 84.6 85.6 82.15 83.2 82.6 83.11
5dB 63.75 59.4 49.1 56.1 63.3 54.25 60.8 54.9 57.7
0dB 35.13 35.06 22.73 25.15 37.21 28.46 35.8 26.89 30.80
-5dB 15.06 19.17 11.08 12.48 18.33 15.4 18.18 14.57 15.53
524 C.-C. Yao and M.-H. Tsai

Table 5. The comparison of signal subspace with our proposed algorithm

Recognition rate %
Method Subway Babble Car Exhib. Rest. Street Airport Station AVE.
subspace 99.26 99.06 99.14 99.51 99.26 99.24 99.14 99.51 99.26
clean
Our 99.10 98.90 99.16 99.31 99.36 99.26 99.28 99.57 99.24
subspace 99.28 97.40 98.09 97.81 97.97 97.67 97.55 97.55 97.79
20dB
Our 98.10 97.96 98.56 97.57 98.6 97.93 98.26 98.84 98.22
subspace 95.52 92.87 96.69 94.48 94.32 93.23 94.18 94.18 94.43
15dB
Our 95.65 94.20 96.10 95.43 95.24 94.85 95.6 95.91 95.37
subspace 87.69 81.89 90.3 85.34 84.07 83.92 84.46 84.46 85.26
10dB
Our 87.90 83.60 82.9 87.8 88.3 84.95 84.9 86.7 85.88
subspace 74.70 64.06 74.59 62.94 65.18 67.02 67.52 67.52 67.94
5dB
Our 71.23 65.40 72.3 67.1 66.8 64.25 69.82 63.9 67.6
subspace 52.16 39.84 45.45 38.29 41.23 39.02 43.66 43.66 42.91
0dB
Our 53.33 40.21 44.33 39.1 42.3 41.36 45.1 43.89 43.70
subspace 28.31 21.49 20.34 16.11 17.93 16.84 21.47 21.47 20.49
-5dB
Our 30.20 24.60 22.3 16.9 18.3 17.77 20.6 19.6 21.28

Table 6. The performance of contaminated signals denosing by spectrum subtraction

Method Subway Babble Car Exhib. Rest. Street Airport Station AVE.
20dB spectrum 84.81 89.99 89.47 80.61 91.18 91.77 91.43 97.55 93.15
15dB spectrum 77.22 79.11 81.76 67.78 79.34 85.96 84.81 94.18 90.56
10dB spectrum 65.71 55.01 66.86 47.64 56.96 67.84 64.9 84.46 77.85
5dB spectrum 45.52 27.27 38.72 25.09 29.52 43.5 36.08 67.52 47.53

proposed method, the signal subspace approach and spectrum subtraction method,
respectively. Experimental results demonstrate that our proposed method outperforms
other methods.

6 Discussions
In this paper a novel speech enhancement system controlled by fuzzy rules is pro-
posed for removing additive noise in a contaminated speech signal. In this speech
enhancement system the rules about how to set the system's parameters depending on
the local characteristics of signals are expressed in an ambiguous style and the system
is designed optimally by using partial swarm optimization, so that the mean square
error of the system's output can be minimized. In the system four linguistic variables
and seven fuzzy rules are used to infer the adaptive threshold. Different kinds of
noises can be effectively represented and distinguished by the four linguistic vari-
ables. Experimental results demonstrate that our proposed method effectively re-
moves noise and outperform other methods.
Concerning further researches. Searching for more critical parameters to identify
the noise and finding more rules to add to this system is under considered. Although
some studies have demonstrated algorithms can automatically generate inference
rules, inference time increases as fuzzy rules increase. Finding the balance between
precision and time complexity is an open issue.
 Adaptive Fuzzy Filter for Speech Enhancement 525

References
[1] Takahashi, Y., Takatani, T., Osako, K., Saruwatari, H., Shikano, K.: Blind spatial sub-
traction array for speech enhancement in noisy environment. IEEE Trans. on Audio,
Speech, and Language Processing 17(4), 650–664 (2009)
[2] Abramson, A., Cohen, I.: Simultaneous detection and estimation approach for speech
enhancement. IEEE Trans. on Audio, Speech and Language Processing 15(8), 2348–
2359 (2007)
[3] Wu, S.-N., Wang, J.-S.: An adaptive recurrent neuro-fuzzy filter for noisy speech en-
hancement. In: International Joint Conference on Neural Networks, pp. 3083–3088 (2004)
[4] Er, M.J., Li, Z.: Adaptive noise cancellation using online self-enhanced fuzzy filters with
applications to multimedia processing. In: Intelligent Multimedia Processing with Soft
Computing. Springer, Berlin (2006)
[5] Ma, N., Bouchard, M., Goubran, R.A.: Speech enhancement using a masking threshold
constrained kalman filter and its heuristic implementations. IEEE Trans. on Audio,
Speech, and Language Processing 14(1), 19–32 (2006)
[6] Shao, Y., Chang, C.-H.: A generalized time–frequency subtraction method for robust
speech enhancement based on wavelet filter banks modeling of human auditory system.
IEEE Trans. on Systems, Man, and Cybernetic–Part B: Cybernetics 37(4), 877–889 (2007)
[7] Furuya, K., Kataoka, A.: Robust speech dereverberation using multichannel blind decon-
volution with spectral subtraction. IEEE Trans. on Audio, Speech and Language Process-
ing 15(5), 1579–1591 (2007)
[8] Hao, J., Attias, H., Nagarajan, S., Lee, T.-W., Sejnowski, T.J.: Speech enhancement,
gain, and noise spectrum adaptation using approximate bayesian estimation. IEEE Trans.
on Audio, Speech and Language Processing 17(1), 24–37 (2009)
[9] Chang, J.-H., Jo, Q.-H., Kim, D.-K., Kim, N.-S.: Global soft decision employing support
vector machine for speech enhancement. IEEE Signal Processing Letters 16(1), 57–60
(2007)
[10] Hasan, M.K., Salahuddin, S., Khan, M.R.: A modified a priori SNR for speech enhance-
ment using spectral subtraction rules. IEEE Signal Processing Letters 11(4), 450–453
(2004)
[11] Tahmasbi, R., Rezaei, S.: A soft voice activity detection using GARCH filter and vari-
ance gamma distribution. IEEE Trans. on Audio, Speech, and Language Process-
ing 15(4), 1129–1134 (2007)
[12] Davis, A., Nordholm, S., Togneri, R.: Statistical voice activity detection using low-
variance spectrum estimation and an adaptive threshold. IEEE Trans. on Audio, Speech,
and Language Processing 14(2), 412–424 (2006)
[13] Johnson, M.T., Yuan, X., Ren, Y.: Speech signal enhancement through adaptive wavelet
thresholding. Speech Communication 49, 123–133 (2007)
[14] Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proc. IEEE Int. Conf. Neural
Network, pp. 1942–1948 (1995)
[15] Chen, S.-H., Wu, H.-T., Chang, Y., Truong, T.K.: Robust voice activity detection using
perceptual wavelet-packet transform and Teager energy operator. Pattern Recognition
Letters 28, 1327–1332 (2007)
[16] Johnson, I.M., Silverman, B.W.: Wavelet threshold estimators for data with correlated
noise. J. Roy. Statist. Soc., Ser. B 59, 319–351 (1997)
[17] Hirsch, H.G., Pearce, D.: The aurora experimental framework for the performance
evaluations of speech recognition systems under noisy conditions. In: ISCA ITRW ASR
2000: Challenges for the New Millenium, Paris, France (September 2000)
[18] Speech Vision and Robotics Group, http://htk.eng.cam.ac.uk
 Risk Prediction for Postoperative Morbidity of
Endovascular Aneurysm Repair Using Ensemble Model

Nan-Chen Hsieh1, Chien-Hui Chan2, and Hsin-Che Tsai1

1
Department of Information Management, National Taipei College of Nursing
No. 365, Min-Ten Road 11257, Taipei, Taiwan, ROC
[email protected]
2
Department of Computer Science and Information Engineering, Tamkang University
No. 151, Ying-Chuan Road, 25137 Tamsui, Taipei County, Taiwan, ROC

Abstract. Endovascular aneurysm repair (EVAR) is an advanced minimally in-

vasive surgical technology that is helpful for reducing patients’ recovery time
and postoperative morbidity. This study proposes an ensemble model to predict
postoperative morbidity after EVAR. The ensemble model was developed using
a training set of consecutive patients who underwent EVAR between 2000
and 2008. The research outcomes consisted of an ensemble model to predict
postoperative morbidity, the occurrence of postoperative complications pro-
spectively recorded, and the causal-effect decision rules. The probabilities of
complication calculated by the model were compared to the actual occurrence
of complications and a receiver operating characteristic (ROC) curve was used
to evaluate the accuracy of postoperative morbidity prediction. In this series, the
ensemble of BN, NN and SVM models offered satisfactory performance in pre-
dicting postoperative morbidity after EVAR. Moreover, the Markov blankets of
BN allow a natural form of causal-effect feature selection, which provides a
basis for screening decision rules generated by granular computing.

Keywords: Endovascular aneurysm repair (EVAR), postoperative morbidity,

ensemble model, machine learning, Markov blanket

1 Introduction

Cardiac surgery is a complex surgical operation that is indicated for patients with
severe insufficiency in cardiac function. Major cardiac surgical interventions include
coronary artery bypass grafting (CABG), repair of congenital heart defects, surgical
treatment of atrial fibrillation, heart transplantation, repair or replacement of heart
valves, aortic surgery, aneurysm repair or a combination of these surgical procedures.
During the operation and the postoperative stay at the ICU and nursing ward, there is
considerable morbidity for cardiac surgery patients with postoperative complications,
which results in increased hospital mortality and postoperative morbidity. Many
prediction models for cardiac surgical outcome apply logistic or multivariable regres-
sion to assess preoperative risk [1-3]. EuroSCORE is a predominant model [4]. Most
of the risk prediction models in current use were derived for patients undergoing open

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 526–540, 2010.
© Springer-Verlag Berlin Heidelberg 2010
 Risk Prediction for Postoperative Morbidity of EVAR Using Ensemble Model 527

abdominal aortic aneurysm (AAA) repair and appear to lack utility when applied to
EVAR patients. EVAR is an advanced minimally invasive surgical technology that
helps reduce patients’ recovery time as well as postoperative mortality and morbidity;
it is especially helpful in the treatment of patients judged to be high surgical risk for
conventional surgery. EVAR benefits patients with medical co-morbidities, and post-
operative complications highly significantly influence longer-term postoperative out-
comes in EVAR patients.
The use of machine learning models has become widely accepted in medical appli-
cations. Various machine learning models including BNs, NNs, and SVMs have been
tested in a wide variety of clinical and medical applications [5]. Soft-computing, in-
cluding fuzzy set and rough set techniques that work well in descriptive data mining,
is also a promising technique. BN is a probability-based inference model that has a
wide range of applications and is increasingly used medically as a prediction and
knowledge representation modeling technique. Verduijn, M., et al. [6, 7] presented
the prognostic BN as a new type of prognostic model that builds on the BN method-
ology and implements a dynamic, process-oriented view of cardiac surgical progno-
sis. Lin and Haug [8] proposed BN suitable for exploiting missing clinical data for
medical prediction. NNs have featured in a wide range of medical applications, often
with promising results. Rowan et al. [9] used NNs to stratify the length of stay of
cardiac patients based on preoperative and initial postoperative factors. Eom et al.
[10] developed a classifier ensemble-based, including NNs, DTs, and SVMs, clinical
decision support system for predicting cardiovascular disease level. SVMs have been
successfully used in a wide variety of medical applications. Polat and Güne [11] used
a least square support vector machine to assist breast cancer diagnosis. Babaog˘lu et
al. [12] first used principle component analysis method to reduce data features, and
acquired an optimum support vector machine model for the diagnosis of coronary
artery disease. Choi [13] proposed the detection of valvular heart disorder (VHD) by
wavelet packet decomposition and SVM techniques.
Traditional medical data modeling tends to use statistical techniques to provide re-
ports. However, most diagnoses and treatments in medical decision making are im-
precise and are accompanied by errors. RS is able to circumvent this limitation with
its ability to handle imprecise and uncertain data [14]. Theoretically, statistical
clinical models use predetermined hypotheses and a priori assumptions in initial hy-
pothesis test design. These models also usually require a perfectly random statistical
distribution; furthermore, statistical formulae or equations usually prove poor in char-
acterizing medical data and solving for the characteristics or relationships of that data.
The RSs approach takes advantage of a correct proven philosophy to work with medi-
cal data without a strong a priori reasoning. As an example, Mitra et al. [15] imple-
mented a rule-based RS decision system for the development of a disease inference
engine for ECG classification. Podraza et al. [16] presented a complex data analysis
and decision support system for medical staff based on rough set theory. Wakulicz-
Deja and Paszek [17] implemented an application of RS theory to decision making for
diagnosing mitochondrial encephalomyopathies in children.
This study describes the development of an informative ensemble prediction model
consisting of BNs, NNs and SVMs and argument using an RS model for the predic-
tion of postoperative morbidity and the description of causal effects between preop-
erative variables and complication outcomes in EVAR patients. The development of
528 N.-C. Hsieh, C.-H. Chan, and H.-C. Tsai

this prediction model proceeds as follows. First, an informative discretization method

for numerical values, which employs a fuzzy discretization method to guide the cate-
gorized numerical features as linguistic fuzzy terms on the basis of entropy/MDL
algorithm and laboratory surgeon’s knowledge, is developed. Through fuzzy discreti-
zation, numerical values are converted into discrete values with informative fuzzy
trapezoid functions. Second, the proposed ensemble-based architecture focuses on
fusing three types of models (BNs, NNs, and SVMs) with the RS used to augment the
interpretability of the derived knowledge. The model selection scheme, designated a
CV-bagging scheme, is a mixture of bagging and cross-validation that is chosen in
order to improve overall classification by combining classifiers trained on randomly
generated subsets of the entire training set. This strategy increases both the diversity
and accuracy of individual ensemble classifiers. Finally, the Markov blanket (MB)
concept of BN allows a natural form of feature selection, providing a basis concept
for filtering decision rules by RSs. The learned knowledge is presented in multiple
forms, including causal-effect diagrams and decision rules.

Fig. 1. The proposed architecture

 Risk Prediction for Postoperative Morbidity of EVAR Using Ensemble Model 529

Section 2 of this paper begins with an overview of study background and research
materials in general. Section 3 describes the methods and procedures used in this
study. Section 3 also empirically tests the proposed method using an EVAR clinical
dataset. Section 4 discusses the experimental findings and offers observations about
practical applications and directions for future research.

2 Study Background and Materials

Abdominal aneurysm (AAA) is an enlargement that occurs in a weakened area within

the largest artery in the abdomen (http://www.vascularweb.org/patients/NorthPoint/
Abdominal_Aortic_Aneurysm.html). If an AAA is not treated in due time, the pres-
sure generated by heartbeats causes the aneurysm to continuously grow larger, and the
aortic wall continues to weaken. Finally, rupture occurs and massive internal bleeding
occurs. The best way to prevent the high mortality associated with AAA is to find the
lesion before rupture occurs. However, patients with aortic diseases are often elderly
with severe co-morbidities and sometimes devastating morbidity, making them ex-
tremely challenging candidates for surgery. For such patients, EVAR represents a
lower risk approach than conventional open surgery and is associated with shorter
operating times, shorter hospitalizations, more rapid recovery and improved quality of
life during the perioperative period and postoperative follow-up. Although long-term
data on the clinical outcomes of patients who received EVAR are not yet available,
given its importance, building a prediction of postoperative morbidity after EVAR is
critical.
We retrospectively examined 140 consecutive patients who underwent EVAR sur-
gery at Taipei Veteran General Hospital, a teaching hospital in Taiwan, between 2000
and 2008. The dataset contains preoperative patient characteristics, details of the op-
erative information, and pathological and laboratory findings from the emergency
room, operating room and intensive care unit (ICU). The dataset also included length
of ICU stay, variables that describe postoperative complications that frequently occur
in EVAR surgery, death during hospitalization, and time of death for patients who
expired. Postoperative complication was used as the binary outcome variable of the
ensemble model, and types of complications were used as subsidiary outcome vari-
ables. The original dataset contained 137 variables, but included many missing
values. Preliminary inspection of the dataset showed that many variables contained
missing values for at least 50% of the patients; these variables were not included in
further analysis. In order to identify significant variables for use in the ensemble
model, a number of criteria were employed. Variables that were subjective, ambigu-
ous or inadequately defined were excluded; variables that were frequently incomplete
were also excluded from subsequent analysis. Sixty-seven of the 140 patients experi-
enced severe postoperative complications during their stay at hospital. Data collected
included preoperative patient characteristics, risk factors, details of the operative
information, physical characteristics of the aneurysm, postoperative physiological and
laboratory findings, and postoperative complications as the outcome variable.
We employed RS as a preprocessing tool for data reduction. The main advantage
of RS is that it does not require any preliminary or additional information about data
530 N.-C. Hsieh, C.-H. Chan, and H.-C. Tsai

such as probability distribution in statistics, basic probability assignment in the

Dempster-Shafer theory, or membership degree in fuzzy set theory; RS is especially
suitable for dealing with sparse datasets. Based on the indiscernibility relation of RS,
redundant variables were identified and eliminated to reduce the number of variables.
Originally, RS with Johnson’s search algorithm reported nine variables that showed
significant relatedness to the outcome variable. Following suggestions from surgeons
and review of relevant research studies [1-3], 10 more interesting variables were in-
cluded in each patient’s profile. As shown in Table 1, the final dataset contains nine
prediction variables, 10 profiling variables and one outcome variable. Besides its
usefulness for data reduction, RS is also suitable for finding dependencies among
variables, discovering cause-effect relationships, and generating decision rules. Ac-
cordingly, we also used RS as the preliminary variable selection mechanism for build-
ing MB of BN.

Table 1. Variables used to predict the postoperative morbidity

Variable Type Definition Category Use as

Heart disease Categorical Heart diseases, A0: no, A1: yes Patc. Prediction
Smoking Categorical A0: no smoking, A1: smoking, A2:
Patc. Prediction
uncertain
Hypertension Categorical A0: no, A1: yes Patc. Prediction
Hyperlipidemia Categorical Hyperlipidemia, A0: no, A1: yes Patc. Prediction
CRI Categorical Chronic renal insufficiency, A0:
Patc. Prediction
no, A1: yes
COPD Categorical Chronic obstructive pulmonary
Patc. Prediction
disease, A0: no, A1: yes
Seg Numerical Segmented neutrophil count, 0-
Lab. Prediction
100%
AAA Size Numerical Size of abdominal aortic
Pato. Prediction
aneurysm, mm
AAA Shape Categorical Shape of abdominal aortic
aneurysm, A0: Fusiform, A1:
Pato. Prediction
Saccular, A2: Rupture, A12:
Saccular and Rupture
BUN Numerical Blood urea nitrogen count, mg/dl Lab. Profile
Creatine Numerical Creatine count, mg/dl Lab. Profile
HCT Numerical Hematocrit count, 0-100% Lab. Profile
Hb Numerical Hemoglobin count, g/dl Lab. Profile
PLT Numerical Platelet count, /CUMM Lab. Profile
Age Numerical Integer number, 0- 100 years Patc. Profile
Gender Categorical Male/female Patc. Profile
DM Categorical Diabetes, A0: no, A1: yes Patc. Profile
CVA Categorical Cerebral vascular accident, A0: Profile
Patc.
no, A1: yes
AAA site Categorical Site of abdominal aortic aneu- Profile
rysm, A0: Infra renal, A1: Supra Pato.
renal, A2: Juxta renal
Complications Categorical Postoperative complications,
Outcome
yes/no
i.e. Patc.: patient characteristics; Pato.: pathological; Lab.: laboratory findings.
 Risk Prediction for Postoperative Morbidity of EVAR Using Ensemble Model 531

3 Methods and Procedures

3.1 Entropy/MDL-Based Method for the Discretization of Numerical Values

Most studies dealing with cardiac surgery prediction models have applied logistic or
multivariable regression to assess the preoperative risk. Few studies have utilized
machine learning algorithms i.e., decision trees, Bayesian networks or neural net-
works in analyzing clinical data. These state-of-the-art machine learning algorithms
are often informative and can represent more knowledge in the clinical data. Gener-
ally, clinical datasets involve numerical variables. To satisfy the requirements of
prediction models, the employment of a discretization approach is necessary. Discre-
tization is defined as a process that divides numerical values into states of discrete
categorical values, leading to informative expressed categorical values. For example,
the CART model originally was not designed to handle numerical attributes. During
the construction of a CART model, numerical attributes were divided into discrete
categorical values. Aside from CART, discretization techniques were frequently
adopted in other popular learning paradigms, such as C4.5, BNs, NNs, and genetic
algorithms.
This study employed entropy/MDL to discretize numerical domains into intervals.
Incorporation investigates fuzzy discretization; the discretized value is characterized
by a triple of interval (i,m,a), where i denotes interval name, m denotes membership
value of the value, and a denotes affinity of the value. The steps for automatically
finding fuzzy sets from a given dataset are described herein. Assuming that the do-
main of a numerical attribute ranges from v1 to v2, and that {c1, c2,…,ck} denote the k
cut-points obtained by the entropy/MDL algorithm, using these k midpoints, k/2 + 1
linguistic terms or membership functions can be determined for a trapezoidal fuzzy
set. The first membership function is computed as:
⎧ 1 if v1 ≤ x ≤ c1
⎪ .
f first ( x) = ⎨(c2 − x) (c2 − c1 ) if c1 < x < c2
⎪ x ≥ c2
⎩ 0 if

Generally, the middle pth, p =2,.., k/2, membership function is computed as:
⎧ 0 if x ≤ c2 p −3
⎪( x − c ) (c2 p −2 − c2 p −3 ) if c2 p −3 < x < c2 p −2 .
⎪⎪ 2 p −3

f p ( x) = ⎨ 1 if c2 p −2 ≤ x ≤ c2 p −1
⎪ (c − x ) (c2 p − c2 p −1 ) if c2 p −1 < x < c2 p
⎪ 2p
⎪⎩ 0 if x ≥ c2 p

The final membership function is computed as:

⎧ 0 if x ≤ ck −1
⎪
f final ( x) = ⎨(ck − x) (ck − ck −1 ) if ck −1 < x < ck .
⎪ if ck ≤ x ≤ v2
⎩ 1
532 N.-C. Hsieh, C.-H. Chan, and H.-C. Tsai

For example, AAA_Size is a numerical attribute obtained from the dataset. The
domain of AAA_Size ranges from 5.0 to 9.6. Six cut-points, (5.5, 5.7, 5.8, 7.3, 8.1,
9.6), were obtained using the entropy/MDL algorithm. As depicted in Fig. 2, a trape-
zoidal fuzzy set was obtained with four membership functions corresponding to the
six intervals. If AAA_Size={5.65cm}, then it can be transformed to a triple value
[i2,max(round(0.75,0.8),round(0.3,0.4),i3] = [i2,0.8,i3], meaning that this value belongs
to interval i2 with membership value 0.8, and adheres to interval i3. If AAA_Size =
{6.55cm}, it can be transformed to a triple value [i4,0.5,i4]. Within each analysis, the
laboratory system should contain an interval that delimits life-compatible values.
Therefore, if cut points fall within the interval, they should be eliminated. For exam-
ple, one cut point {1.4 mg/dl} of creatine lay within an interval, i.e., [0.75~1.5] for
male, [0.5~1.2] for female; therefore the cut point {1.4 mg/dl} was eliminated.

Fig. 2. The trapezoidal fuzzy set of AAA_Size and Creatine

3.2 Ensemble Model for the Prediction of Postoperative Morbidity

BNs, NNs, SVMs and RSs were chosen because they represented different ap-
proaches, each of which is simple to implement and has been shown to perform well
in medical applications. The rationale of employing these models is that BNs can
easily model complex relationships among variables, NNs are generally superior to
conventional models, SVMs can be used as a benchmark technique, and RSs can be
used to augment the interpretability of the derived knowledge. The detailed configura-
tions of each individual model are as follows: BNs are Bayesian networks with K2
search algorithm, NNs are RBF neural network with radial basis functions as activa-
tion functions, SVMs use John Platt's sequential minimal optimization algorithm with
logistic to the outputs for proper probability estimates, and RS is Pawlak’s model with
Johnson search algorithm for generating the singleton reduce.
In this study, the model selection scheme is a mixture of bagging and cross-
validation (CV-bagging) that aims to improve the classification by combining models
trained on randomly generated subsets of the entire training set. We first applied a
cross validation scheme for model selection on each subset; subsequently, for the sake
of simplicity and so as not to run into over-fitting problems, we combined the selected
models in a uniform weights approach to reach a final decision. The concept of
 Risk Prediction for Postoperative Morbidity of EVAR Using Ensemble Model 533

uniform weights voting is both appealing and simple to implement; it can generally be
applied to any type of classifier without relying on specific interpretations of the out-
put. In order to validate models, we used random subsets as cross-validation folds for
reasons of simplicity. In k-fold cross validation, the dataset is partitioned into k sub-
sets. Of these k subsets, a single subset is retained as the validation dataset, and the
remaining k-1 subsets are used for training datasets. The cross validation process is
then repeated k times with each of the k subsets used only once as the validation data.
The k results from the folds can then be averaged to produce a single estimation. In
cases where specific inside medical knowledge is not available, such a cross valida-
tion method can be used to select a classification method empirically, because it
seems to be obvious that no classification method is uniformly superior. For each
model, we train several individual classifiers and select the best performing model as
final model. If the trained models of each model differently with different initial con-
ditions, we then find proper values for the free parameters of the model. We extend
these results by combining models from different classes in order to increase the
model diversity. The result, a heterogeneous ensemble, allows classification methods
to be used more effectively.
The area under the ROC curve is based on a non-parametric statistical sign test, the
purpose of which is to estimate the probability of survival for each of a pair of pa-
tients, one of whom did not survive. The surviving patient gains greater probability of
survival. In this study, the area under the ROC metric was assessed through stratified
10-fold CV-bagging. The detailed accuracy of individual models and of the ensemble

TP FP Precisi F- ROC
Models Recall
Rate Rate on Measure Area
10 folds CV- bagging weighted average
BN 0.736 0.270 0.737 0.736 0.734 0.834
NN 0.675 0.330 0.657 0.657 0.674 0.755
SVM 0.757 0.247 0.758 0.757 0.756 0.755
Ensemble 0.757 0.249 0.760 0.757 0.755 0.813

Fig. 3. The results of the postoperative morbidity prediction

534 N.-C. Hsieh, C.-H. Chan, and H.-C. Tsai

model is shown in Figure 3. The results of the experiment show that individual mod-
els’ performance and improvements in performance were achieved by applying the
ensemble of models. This indicates that model combination techniques indeed yield
stable performance improvements for ensemble models.

3.3 BN’s Markov Blanket for the Causal Interpretability of the Prediction
Model

BN models have the ability to reason under conditions of uncertainty on the basis of
probabilistic inference mechanisms; therefore, they exhibit more resistance than other
machine learning models to noise that commonly occurs in medical datasets. That is,
if a dataset contains some samples with missing values, it can still be used to train or
test BN models. Because medical datasets tend to have missing values, this property
is useful for intelligent data analysis. Unlike most regression techniques, BN models
can easily model complex relationships among variables without any requirements for
the underlying distributions of variables and are more transparent and intuitive be-
cause relationships among variables are explicitly represented by a directed acyclic
graph (DAG) for modeling experts’ knowledge. The DAG consists of a set of nodes
linked by directed arcs, where the nodes represent the variables and the arcs indicate
the direct causal relationships among the connected attributes. The strengths of these
relationships are expressed by conditional probabilities. By using BNs, we can find a
given network’s structure and its associated parameters; the resulting model can pro-
vide diagnostic, predictive, and inter-causal reasoning. Therefore, the DAG of BN can
serve as a causal analysis tool to help surgeons arrive at valuable decision rules gen-
erated by RS. Through the use of causal-effect analysis, medical causal knowledge
can be exhibited through a causal DAG graph. Such knowledge can help surgeons
stimulate the hidden knowledge behind the medical dataset, communicate complex
causal concepts, and transform an individual’s tacit knowledge into a team’s explicit
knowledge.
A novel idea of BN for significant feature selection is the Markov blanket (MB)
[18], which is defined as the set of input features such that all other features are prob-
abilistically independent of the target features. In this case, based on a general BN
classifier, we can obtain a set of features that are present in the MB of the target fea-
ture, and features provided by the MB are sufficient for perfectly estimating the dis-
tribution of and for classifying the target feature classes. As an example, suppose that
we build a BN classifier with a complete input training data set. The MB of the target
feature then forms the feature selection mechanism and all features outside the MB
are ignored by the BN. The remaining features are then used to induce constrained
decision rules. In this study, RS is used to augment the interpretability of ensemble
model; all decision rules derived by RS are in the form, Y←X, where Y is the target
variable and X is a subset of variables from the MB. The rationale of this method is
that for any feature xi∉{MB∪xj}, xi≠xj and xi, xj∈MB, we have P(xi|MB,xj) =
P(xi|MB). That is, the MB of a variable xi is a minimal set of features MB such that xi
is conditionally independent of any other feature xj in MB. The MB structure to
be used for BN explores not only the relationship between target and predictive
variables, but also the relationships among the predictive variables themselves.
 Risk Prediction for Postoperative Morbidity of EVAR Using Ensemble Model 535

Since the conditional independence assumption of BN is not real, BN with a TAN

search algorithm can therefore be used to offset this disadvantage. The use of this
approach results in significant improvement in terms of classification accuracy, effi-
ciency and model simplicity. Although use of the TAN search algorithm may not
always result in the best classification accuracy, this study adopts the TAN search
algorithm because it can create a causal effect graph in which the postoperative-
complications target variable treated as the supreme parent node is located at the top
in the DAG. After a BN is learned, a heuristic is applied to ensure that all nodes in the
BN are part of the Markov blanket of the target variable.
Figure 4 depicts an MB for the EVAR dataset composed of 10 variables, 9 predic-
tive variables, and one target variable. The 9 predictive variables contributed signifi-
cantly to identification of all the variables in the BN that are needed to classify the
target variable. We should use these variables as RS’s input variables to generate
decision rules, connecting with proper profile variables for interpretation. The struc-
ture of the MB proposed for analyzing medical decision tasks follow s the logical
structure for features defined in a loss causation model.

Fig. 4. A Markov blanket for EVAR

3.4 Granular Computing for the Extraction of Causal-Effect Knowledge

In recent years, RSs and granular computing have been proven to be good soft-
computing techniques when used in various synergistic combinations with fuzzy
logic, neural networks and genetic algorithms. Since RSs do not require prior parame-
ter settings and extract knowledge from the data itself by means of indiscernibility
536 N.-C. Hsieh, C.-H. Chan, and H.-C. Tsai

relations, RSs provide a stronger framework to achieve tractability, robustness, and

interpretable decision-making knowledge. RSs require fewer calculations than do
other soft-computing techniques, and provide an effective means for analysis of data
by synthesizing upper and lower approximations of set concepts from the acquired
dataset [19]. The induced decision rules are concise and valuable and can benefit
surgeons by revealing hidden medical knowledge in the dataset, even when evidence
is sparse. Granular computing refers to computation performed on information gran-
ules. Therefore, RS leads to both data compression and gains in computation time,
thereby finding wide application.
As an example, rough-fuzzy integration [20] can be considered as a way of emulat-
ing the basis of f-granulation in data summarization, where quantifiers have fuzzy
boundaries and granular attribute values. Since granular computing is performed on
groupings of similar objects, rather than on individual data values, computation time
is significantly reduced. Due to its comprehensive logical rules, RS can invariably
generate a high performance prediction model, and each logical rule represents an
information granule. However, unimportant logical rules obstruct the utility of the
induced knowledge. In this study, we proposed to use an RS-BN integration approach
to induce necessary decision rules derived from medical knowledge, to generate a
prediction model for postoperative morbidity, and to find significant MBs as causal-
effect relationships among the variables. MBs from BN are then used as a basis for
filtering significant variables for granular computing of RS. The obtained decision
rules correspond to different important regions of the feature space or data groups, a
feature that is good for interpretation. Due to the suitability of BNs and RSs for han-
dling medical data, this study adopted MBs of BN as the candidate variables selector
and used RSs to augment the interpretability of classifiers. The direct information
provided by MB on the basis of the probability distribution obtained for each variable,
shown in Fig. 4, represents data representing the most typical circumstances arising
among the EVAR cases. However, explanation of the meaning of the probability
distribution is necessarily rather complex. For easy interpretation, the profiles for
postoperative morbidity were established with the help of the RS. We can certainly
extract decision rules with whole or partial causality. From this, surgeons and deci-
sion makers can create more valuable discrimination rules regarding the possibilities
of complications and postoperative morbidity.
Figure 5 depicts a partial causality of BN with probability distributions. We used
variables in MB as RS’s input variables to generate decision rules, connecting with
proper profile variables for interpretation. The structure of the MB proposed for ana-
lyzing medical decision tasks follows the logical structure for features defined in a
loss causation model. In this network, we can trace how each node will influence and
be influenced by other nodes. For example, COPD and Complications will directly
influence the state probability of Seg, AAA_size, and AAA_shape through a top-
down causal-effect. The state probabilities of the input and profile variables are also
included. From the surgeon’s perspective, the constrained association rules are rea-
sonable for outcome description. Suitable linguistic quantifiers can be employed to
add linguistic confidence.
In most circumstances, medical datasets are uncertain and incomplete and cases
are sparse. By conventional rules, judgment standards such as support and confidence
are not properly used to select valuable rules. The fuzzy logic-based calculus [21]
 Risk Prediction for Postoperative Morbidity of EVAR Using Ensemble Model 537

provides interpretation and validation of truth statements involving complex linguistic

quantifiers such as “always”, “medium” and “often”. The truth statement evaluation is
therefore used to validate and select decision rules. The procedure for determining the
truth value of a linguistically quantified statement is as follows. Let “Q {t1,…,tn} are
S” denotes a linguistically quantified statement, and let {t1,…,tn} denotes a set of
decision rules, DS. If the summary S involves a variable A, and ti denotes a rule that
satisfies the summary S, then the membership value of ti to S is given by:

S (ti ) = max ( μ EQ (ak , bk ) ) , for all ak∈ti[A], bk∈S,

∀k

where S(ti) denotes the degree to which ti satisfies the summary S, and the func-
tion μ EQ ( ak , bk ) = 0 if and only if (ak ≠ bk ) ∨ (ak = bk ∧ μ (bk ) = 0) ;
μ EQ (ak , bk ) = 1 − μ (ak ) − μ (bk ) if and only if ( ak = bk ∧ μ (bk ) ≠ 0) . Then, the individ-
ual truth value of “{t1,…,tn} are S” for a variable A over DS is computed as:

Truth({t1,…,tn} are S) = ⎛ 1 ∑ n S (t ) ⎞
⎜ i ⎟
⎝ n i =1 ⎠
Moreover, when the linguistic summaries are distributed over m variables with
ANDed conditions, that is, S = S1 ∧ … ∧ Sm, then the total truth value
Truth({t1,…,tn} are S)= min mj=1 (Truth({t1 ,..., tn } are S j ) . Finally, T=Q(Truth({t1,…,tn}
are S)) denotes the truth value of the linguistically quantified statement “Q {t1,…,tn}
are S” to the fuzzy quantifier Q in agreement. We gave an example for the procedure
in obtaining truth value using linguistic summaries. For example, let COPD be the
observation target; then {COPD, Seg, AAA_Size, AAA_Shape, Complications} is a
MB of partial causality. By using granular computing of RS, several decision rules
were obtained; we list three relevant decision rules as follows, where “support” means
number of cases supporting this rule:

IF THEN SUPPORT
COPD(A0) and Seg(i2,0.9,i3) and AAA_size(i1,1.0,i2) and Complications(Y) 7
AAA_Shape(A0)
COPD(A0) and Seg(i1,0.7,i2) and AAA_size(i1,0.8,i2) and Complications(Y) 3
AAA_Shape(A0)
COPD(A0) and Seg(i2,0.8,i3) and AAA_size(i4,0.6,i3) and Complications(Y) 1
AAA_Shape(A0)

Next, we defined the fuzzy quantifiers, Q, as shown in Fig. 6, which are repre-
sented as 7 membership functions. Then linguistically quantified statement for vali-
dating these decision rules is therefore given by:
“Q of the patients of postoperative morbidity have
COPD(A0) and Seg(i2) and AAA_size(i1) and AAA_Shape(A0).”

method of computing the individual truth value for the variable Seg is described as
follows. Let the linguistic summary, S1(Seg(i2))=(i2,1.0,i2) and a decision rule compo-
nent ti[Seg]=(i2,0.9,i3). Then, the membership value of ti to S1 of Seg is given by
538 N.-C. Hsieh, C.-H. Chan, and H.-C. Tsai

S1(t1[Seg])=max(1,0.9) if and only if they have overlap interval or S1(t1)=0. Similarly,

S1(t2[Seg])=max(1,0.7) and S1(t3[Seg])=max(1,0.8). For n tuple {t1,…,tn}, the individ-
ual truth value “Truth({t1,…,tn}) of Seg to S1” over DS is computed as
⎛1 n ⎞ =1, where n=11. Similarly, for linguistic summary, S2(AAA_Size(i1)),
⎜ ∑ i =1 S (ti ) ⎟
⎝n ⎠
the individual truth value “Truth({t1,…,tn}) of AAA_Size to S2” over DS is 0.85.
Moreover, COPD and AAA_Shape are categorical values, and data values matched
the values in linguistic quantified statement, therefore, their individual truth values are
1. Finally, the linguistic quantified statement is ANDed with linguistic summaries, the
total truth value for this linguistic quantifier statement is min(1,1,0.85,1)=0.85. A
decision rule with truth is obtained in the following form, which denotes a satisfactory
truth degree:
“IF COPD(A0) and Seg(i2) and AAA_size(i1) and AAA_Shape(A0) THEN
Patients have postoperative morbidity, with almost_always truth 0.95 and support
35%.”

C o m p lic a tio n

No

Yes
Seg COPD

i4 0 .0 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 A999
P e rc e n ta g e %
i3 A2

i2 A1

i1 A0

0 .0 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .0 0 .1 0 .2 0 .3 0 .4 0 .5
P e r c e n ta g e % P e rc e n ta g e %
A A A _ S iz e A AA_Shape

i7
i6 A12

i5
A2
i4
i3 A1
i2
A0
i1

0 .0 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0
P e rc e n ta g e % P e rc e n ta g e %

Fig. 5. The partial causality of MB with probability distributions

 Risk Prediction for Postoperative Morbidity of EVAR Using Ensemble Model 539

Fig. 6. Fuzzy quantifiers for linguistic summaries

4 Conclusions
Although several risk prediction systems have been proposed for patients undergoing
open aneurysm repair (Bohm et al., 2008), they basically rely on traditional statistical
methods and provide scant accuracy and utility when applied to EVAR patients. We
have proposed an ensemble model to predict postoperative morbidity after EVAR and
support clinical decision-making. The proposed ensemble model is constructed by
incorporating fuzzy discretization of categorical numerical values; BNs, NNs, and
SVMs were used to augment the ensemble model and the dataset was processed by
CV-bagging, showing moderate performance. The experimental result shows that the
proposed ensemble model predicts postoperative morbidity with relatively satisfactory
accuracy, even when data is missing and/or sparse, showing its usefulness in support of
clinical decision-making. In particular, the Markov blankets allow for a natural form of
causal-effect feature selection, which itself provides a basis for screening decision rules
generated by granular computing and generates supplementary information for risk
prediction. The learned knowledge is represented in multiple forms, including causal-
effect diagrams and causal-effect decision rules. The supplementary nature of multi-
models distinguish the proposed model from existing risk scoring systems that are
based on conventional statistical methods and from various machine learning models.
To summarize, the advantage of using the proposed ensemble model is that it can pro-
vide surgeons with practical, relatively accurate aid in their daily diagnostic tasks and
enable them to extract meaningful relationships among features of medical datasets
through the use of causality graphs and constrained decision rules.

Acknowledgement

This research was partially supported by National Science Council of Taiwan (NSC
97-2410-H-227-002-MY2).

References
1. Barnes, M., Boult, M., Maddern, G., Fitridge, R.: A model to predict outcomes for En-
dovascular aneurysm repair using preoperative variables. European Journal of Vascular
and Endovascular Surgery 35, 571–579 (2008)
540 N.-C. Hsieh, C.-H. Chan, and H.-C. Tsai

2. Bohm, N., Wales, L., Dunckley, M., Morgan, R., Loftus, I., Thompson, M.: Objective risk-
scoring systems for repair of abdominal aortic aneurysms: applicability in Endovascular
repair. European Journal of Vascular and Endovascular Surgery 36, 172–177 (2008)
3. Stijn, C.W., Wouters, L.N., Freek, W.A.V., Rene, M.H.J.B.: Preoperative prediction of
early mortality and morbidity in coronary bypass surgery. Cardiovascular Surgery 10,
500–505 (2002)
4. Roques, F., Michel, P., Goldstone, A.R., Nashef, S.A.: The logistic EuroSCORE. Euro-
pean Heart Journal 24, 1–2 (2003)
5. Bellazzi, R., Zupan, B.: Predictive data mining in clinical medicine: Current issues and
guidelines. International Journal of Medical Informatics 77, 81–97 (2008)
6. Verduijn, M., Peek, N., Rosseel, P.M., de Jonge, E., de Mol, B.A.J.M.: Prognostic Bayes-
ian networks I: Rationale, learning procedure, and clinical use. Journal of Biomedical In-
formatics 40, 609–618 (2007)
7. Verduijn, M., Peek, N., Rosseel, P.M., de Jonge, E., de Mol, B.A.J.M.: Prognostic Bayes-
ian networks II: An application in the domain of cardiac surgery. Journal of Biomedical In-
formatics 40, 619–630 (2007)
8. Lin, J.-H., Haug, J.: Exploiting missing clinical data in Bayesian network modeling for
predicting medical problems. Journal of Biomedical Informatics 41, 1–14 (2008)
9. Rowan, M., Ryan, T., Hegarty, F., O’Hare, N.: The use of artificial neural networks to
stratify the length of stay of cardiac patients based on preoperative and initial postoperative
factors. Artificial Intelligence in Medicine 40, 211–221 (2007)
10. Eom, J.-H., Kim, S.-C., Zhang, B.-T.: AptaCDSS-E: A classifier ensemble-based clinical
decision support system for cardiovascular disease level prediction. Expert Systems with
Applications 34, 2465–2479 (2008)
11. Polat, K., Güne, S.: Breast cancer diagnosis using least square support vector machine.
Digital Signal Processing 17, 694–701 (2007)
12. Babaoğlu, I., Fındık, O., Bayrak, M.: Effects of principle component analysis on assess-
ment of coronary artery diseases using support vector machine. Expert Systems with Ap-
plications (in Press)
13. Choi, S.: Detection of valvular heart disorders using wavelet packet decomposition and
support vector machine. Expert Systems with Applications 35, 1679–1687 (2008)
14. Pattaraintakorn, P., Cercone, N.: Integrating rough set theory and medical applications.
Applied Mathematics Letters 21, 400–403 (2008)
15. Mitra, S., Mitra, M., Chaudhuri, B.B.: A rough-set-based inference engine for ECG classi-
fication. IEEE Transactions on Instrumentation and Measurement 55, 2198–2206 (2006)
16. Podraza, R., Dominik, A., Walkiewicz, M.: Decision support system for medical applica-
tions. In: Hamza, M.H. (ed.) Applied Simulation and Modelling, Marbella, Spain (2003)
17. Wakulicz-Deja, A., Paszek, P.: Applying rough set theory to multi stage medical diagnos-
ing. Fundamenta Informaticae 54, 387–408 (2003)
18. Cheng, J., Greiner, R., Kelly, J., Bell, D.A., Liu, W.: Learning Bayesian networks from
data: an information-theory based approach. The Artificial Intelligence Journal 137, 43–90
(2002)
19. Pawlak, Z.: Rough sets. International Journal of Computer and Information Sciences 11,
341–356 (1982)
20. Pal, S.K.: Soft data mining, computational theory of perceptions, and rough-fuzzy ap-
proach. Information Sciences 163, 5–12 (2004)
21. Yager, R.R.: Database discovery using fuzzy sets. International Journal of Intelligent
Systems 11, 691–712 (1996)
 Further Results on Swarms Solving Graph
Coloring

Manuel Graña, Blanca Cases, Carmen Hernandez, and Alicia D’Anjou

Grupo de Inteligencia Computacional www.ehu.es/ccwintco UPV/EHU

Abstract. We have proposed the mapping of graph coloring problems

into swarm dynamics. Empirical evidence that flock steering behaviors
augmented with the notion of hostility (enmity and friendliness) are
enough to perform efficiently the task of coloring the nodes of graphs
even in the case 3-coloration hard graph topologies. We discuss here
what are the minimal cognitive capabilities that allow the emergent be-
havior of swarms to solve such NP-complete problem without mediating
an explicit knowledge representation.

1 Introduction

Swarm Intelligent Systems are models of spatial behavior of populations whose

individuals (boids) are steered by decisions taken in a neighboring zone around
them. The models of flocking boids of developed by Craig Reynolds [9,10] repre-
senting the steering behaviors of birds that follow each other following a linear
combination of three local rules: separation, coherence and alignment to other
flockmates or to static goals situated in the environment. This model has been
integrated in a more general models of Self-Organizing Particle Systems (SOPS).
Our focus in this work is on the emergent behavior arising when boids have the
cognitive ability to perceive the hostility quality of spatilly close boids, which
are perceived as friend or enemies. This abilities determine the emergence of a
spatial distribution at a global level minimizing the number of cohabitating ene-
mies. When the definition of enemies and friends is determined by an underlying
graph, this behavior produces a spatial distribution that solves precisely and
efficiently the the Graph Coloring Problem (GCP) [2]. The spatial distribution
follows a kind of supervised behavior in the sense that we define some regions of
the space with singular properties (i.e. representing a color), and the boids are
forced to converge to them.
The Supervised Territorial Flock Coloring (SFTC) when taking to its limit
expression, that is with flocking parameters of coherence, separation and align-
ment set to zero, is able to color graphs efficiently. Surprisingly, it improves with
hard graphs for 3-coloration. That means that territoriality is a powerful driv-
ing force by itself. We identify a key parameter that seems to characterize the
problem because it provides a fine prediction of the results that the algoritm will
obtain. We have called Hostility this parameter, which is the ratio of the average
number of friends to the average number enemies per node in the graph.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 541–551, 2010.

c Springer-Verlag Berlin Heidelberg 2010
542 M. Graña et al.

There are some other efforts to find other swarm inspired ways to solve the
GCP. A k-coloration of a undirected graph with vertex set V and E the set of
edges G = (V, E) is a partition of the graph in k classes of equivalence such that
the number of connected nodes belonging to the same class is minimized [6]. A
recent approach by ACO [5,4] to solve the GCP is [7] identifying ants to colors. In
[2] we have shown that our approach improved over some greedy algorithms over
a collection of hard coloring problems. Here we are more focused in identifying
the minimal swarm dynamics that allow to solve the coloring problem.
Section 2 gives some basic definitions of swarms for the GCP. Section 3 in-
troduces the Supervised Territorial Flock Coloring (SFTC). Section 4 reports
the experiments with the extreme version of the SFTC. Finally we gather some
conclusions and directions for further work in section 5.

2 Basic Definitions of Swarms for GCP

The aim of this paper is to prove that GCP solutions emerge from the swarm
dynamics of flocking boids that have hostility as the basic cognitive ability. They
are able to recognize and run away from enemies, and run towards friends. Hos-
tility is defined in terms of the underlying graph to be colored. Boids correspond
to nodes in the graph. Enemies of a boid are the boids whose corresponding
nodes are connected though an edge in the graph. That is, enemies are nodes at
distance 1 over the graph. The friends are boids whose nodes are a distance 2 or
greater over the graph. A flocking boid has a spatial position p = (p1 , p2 ) and
velocity v = (v1 , v2 ), which results from the composition of the velocities induced
by the different forces pulling the boid, which are combined in the following well
knwon linear model of the boid’s velocity, which will be revisited below:

v = Vm (αs vs + αc vc + αa va + αn vn ) (1)

Here s corresponds to separation, c to coherence, a to alignment and n to noise

R
and parameters α are in the range [0, 1]. If we introduce vA which denotes the
Z
Amity velocity calculated in a neighborhood of radius R and vE which is the
Enmity velocity measured inside radius Z around boid’s position, simplifying
R Z
the notation vA = vA and vE = vE , we could introduce these effects into the
boid’s velocity equation (eq. 1) obtaining the following expression:

v = Vm N (αEs vEs + αAc vAc + αEa vEa + αn vn ) (2)

Here p denotes the norm of a position or vector p and Vm is a non-negative

parameter that limits the norm of a vector and N (p) = p p
represents the
normalized position or vector and |S| denotes the cardinal of set S. Terms αEs vEs
and αEa vEa are respectively the velocity components of separation and oppossite
alignment to enemies, αAc vAc is the component of coherence to friends and
αn vn is a noisy component. We take normalized velocities vEs , vAc , vEa , vn and
normalize their linear combination to turn the head of a boid in the direction of
the vector, giving after an step forward of length Vm .
 Further Results on Swarms Solving Graph Coloring 543

2.1 Supervised Territoriality

The defense of territory by birds was modeled mathematically by J. Maynard
Smith, who related in 1973 natural selection to the theory of games through the
concept of evolutionary stable strategies [12,11]. Territories represent feeding and
breeding zones that individuals deffend from predators and from competitors of
the same specie. In previous works [2] we introduced a territorial concept to
model the a k-coloration process. We defined spatial goals associated with node
colors, so that boids converging to a goal represent nodes being colored. This
idea corresponds to a supervised territoriality modelling of the graph coloring
process.
Supervised Territories: for each color 1, . . . , k we create a goal region on the
swarm space. Usually we place Goals as the vertices of a regular polygon, and
flocking boids are attracted by the goals with velocity αg vg being equation
3 the one that governs the model:

v = Vm N (αEs vEs + αAc vAc + αEa vEa + αn vn + αg vg ) (3)

Satisfaction: In supervised territoriality we attach a level of satisfaction to

boids that measures their strees for being out of the desired spatial goals.
Satisfaction grows until it reaches a maximum level whenever the boid is
safe inside a goal without enemies. Cohabitating with enemies in a goal or
being outside a goal decreases the satisfaction until 0 is reached.
Attack: Attack is a behavior triggered by the decrease of their boid satisfaction
level. Attackers are individuals that triger conflict because of their own low
level of satisfaction, and the attacks can be of two types:
– External: Unsatified boids that are flocking outside the goals select a
goal and try to expell enemies from it to replace them there.
– Internal: unsatisfied boids inside a goal try to expell enemies to obtain
an homogeneous population inside the goal region.

2.2 The Scope of This Work

Besides, we find that the Unsupervised Flock Coloring (UFC), without spatial
goals identified in the swarm evolution space, is able to find the chromatic num-
ber of a graph by convergence of the flock to clusters, for a sample of graphs of
different topologies. A class of hard graphs for 3-coloration [8] by Brélaz algo-
rithm [1] is explored. We measure the performance of UFC by indirect means,
to avoid bias due to the interpretation of the spatial clusters.

3 The Rules of Supervised Territorial Flock Coloring

(STFC)
The step from static models to swarms of flocking birds came from the hand of
Craig Reynold’s works [9,10] on computer animation. Rather that a grid, the
544 M. Graña et al.

board becomes continuous, and cells are agents that can change their positions
in the world. Each individual exhibits a very simple behavior that is specified by
a few simple rules that guide them to get along with the collective motion of the
flock. The global behavior of the flock emerges from these individual decisions.
We will stick to the birds metaphor, so that in the following, we call boids to
the agents that compose a flock.

3.1 STFC Scheme

A territorial flock coloring scheme is a tuple F = (B, E, A, G, α, β) where:

– B is a set of nodes that represent a population of boids.

– E ⊆ B × B is a undirected set of edges called Enmity edges. Graph G =
(B, E)is the one we want to color. Ei = {bj ∈ B : (bi , bj ) ∈ E} is the set of
enemies of boid bi .
– A ⊆ C is a set of edges called Amity edges, being the friends of node b a
subset of nodes at graph distance 1 of b, the enemies of the enemies that are
not enemies of b:

C = {(x, y) ∈ B × B : ∃u ((x, u) ∈ E ∧ (u, y) ∈ E) ∧ (x, y) ∈

/ E}

Ai = {bj ∈ B : (bi , bj ) ∈ A} is the set of friends of boid bi .

– G is a set of goals. Each goal g ∈ G is in a position p and all them have a

goal radius Γ .
– α = (αs , αc , αa , αn , αg ) ∈ [0, 1]5 are the flock parameters for equation 3.
– β = (S, R, Z, Γ , ρ , Vm , Σ, next − goal) are the parameters related to the
size of the world: the side S, vision radius R, separation zone Z, goal radius
Γ , angle of vision ρ, length of step Vm , maximum level of satisfaction Σ and
criterion next − goal to calculate the goal velocity.

The angle of vision, denoted ρ, comprises all the positions in the plane p = pi +v 
where γvi − ρ2  γv  γvi + ρ2 . In this article, the value ρ = 360o , meaning that
individuals can even observe what happens behind their backs. Let R be the
radius of vision: the cone of vision is the disk sector of radius R and angle ρ
around the boid. The private radius Z is a zone that separates the boid from
others.  
Each boid is aware of an spatial region around it, its neighborhood. Let ∂iX 
denote the population in the neighborhood of radius X. The set of boids lying
inside the neighborhood of the i-th individual of radius X is denoted:

∂iX = ∂ X (bi ) = {bj : dist (pi , pj ) < X ∧ pj ∈ conei } .

where dist is the euclidean distance, R is the ratio of vision, and conei represents
the population in the cone of vision of the individual i. The population in goal
g is ∂gΓ = ∂ Γ (g) = {bj : dist (pi , pj ) < Γ }.
 Further Results on Swarms Solving Graph Coloring 545

3.2 STFC Steering Rules

The steering basic rules, used in our model presented in equation 3, are the
classical ones of the Reynolds model: alignment, separation, and cohesion. Com-
bining these rules, the flocking birds, called boids in this model, are able to
flight co-coordinately avoiding collisions. The flocking rules for the boidbi are
formalized as follows:
Separation from enemies: steer to avoid crowding local flock enemies inside
a private zone of radius Z.
⎛ ⎞

vsEi = N ⎝− (pj − pi )⎠ (4)
bj ∈∂iZ ∩Ei

Cohesion towards friends: steer to move toward the average position ci of

local flockfriends in vision area R:
⎛ ⎞
1 
vcAi = N ⎝ci − pi where ci =  R  pj ⎠ (5)
∂i ∩ Ai  R bj ∈∂i ∩Ai

Alignment in opposite direction of enemies: steer in opposite direction of

the average heading of local flockenemies.
⎛ ⎞
1 
vaEi = N ⎝−  Z  vj − vi ⎠ (6)
∂i ∩ Ei  Z
bj ∈∂i ∩Ei

3.3 Dynamics of STFC Systems

We present here the equations governing the velocity and position of a Boid. Let
it be pi (0) the initial conditions and vi the model of territorial flock coloring
defined in equation 3.

pi (t + 1) = pi (t) + vi (t + 1) (7)

In a toric geometry, points p are represented

√
as pmodS. The maximum euclidean
distance between two points is hence S 2 2 .
A boid configuration is a tuple bi (t) = (pi (t) , vi (t) , si (t) , gi (t)) where si (t)
is the level of satisfaction and gi (t) ∈ G ∪ {nobody} denotes next goal, being
nobody a default value meaning that no goals are being sought.
⎧
⎪
⎪ si (t) +̇1 in − goal − without − enemies (bi (t))
⎪
⎨s (t) −̇1 in − goal − with − enemies (b (t))
i i
si (t + 1) =
⎪
⎪ ∨ enemies − in − all − goals (b i (t))
⎪
⎩
si (t) otherwise
546 M. Graña et al.

where x+̇y = min {Σ, x + y} and x−̇y = max {0, x − y}. Predicates in− goal −
without − enemies and in − goal − with − enemies are defines as follows:
 
in − goal − without − enemies (bi ) ⇔ ∃g bi ∈ ∂gΓ ∩ Ēi
 
in − goal − with − enemies (bi ) ⇔ ∃g bi ∈ ∂gΓ ∩ Ei
 
enemies − in − all − goals (bi ) ⇔ ∀g ∂gΓ ∩ Ei = ∅
Next-goal is a criterion that allows the selection of the goal toward which a
boid navigates, g (t + 1) = next − goal (bi (t)) and can be for example: current
goal, nearest goal, nearest free goal,...
 A TFC system is a pair (F , c̄) where F is a flock scheme and c̄ = c̄ (0) =
c1 (0) , . . . , c|B| (0) is the vector of initial configurations. A computation of
length k of a TFC system is the sequence c̄ (0) . . . c̄ (k).

4 Supervised Flock Coloring experiments

We implemented in Wolfram Mathematica Mizuno’s [8] generator of hard graphs

for 3-coloratión by Brélaz heuristic [1], which uses 4 colors for that class of 3-
colorable graphs, and producing a sample of 100 hard graphs with population P
between 100 and 112 nodes. We also generated a sample of 100 random graphs
generated by Kuratowski’s theorem [3], for which e ≤ 3P - 6 being P the number
of nodes a e the number of edges. Complete bipartite graphs of 100 nodes were
also analyzed. The side of the world was S = 400 and goals are placed forming
a regular polygon inscribed in a circle of radius S2 .

4.1 Extreme Supervised Territorial Coloring

An Extreme Supervised Territorial Coloring (ESTC) scheme is a STFC

F = (B, E, A, G, α, β) where the parameters α = (αs , αc , αa , αn , αg ) that con-
trol the flock are zero, αs = αc = αa = 0, meaning that the only motors of the
system are goal velocity and attack. The ESTC algorithm depends on next−goal
criterion, that we make equal to nearest goal in the experiments presented here,
that extend the work that we began in [2]. Assume that internal attack expells
enemies from the goal reseting the maximum satisfaction to both contenders.
As cascade coloration strategy can be used to find the chromatic number of a
graph: it is sufficient to start the process of coloring successively the graph with
colors P, P − 1, . . . For 3-coloration, the execution of the program has two stages
: First, the system attempts to find a 4-coloration (maximum 1500 iterations).
Second, eliminate the less populated goal, the individuals newly freed wander
to seek a new goal (maximum 3500 iterations). This procedure of cascading
coloration is based on known works in reaction-diffusion particle systems [13].
An experiment consists in 25 runs of each one graph, registering the maximum
number of well-colored nodes per run. We register the best configuration for each
 Further Results on Swarms Solving Graph Coloring 547

Fig. 1. Evolution of the number of correctly colored nodes in a cascade coloration

process

experiment, the one that has minimum errors as well as the step at which the
minimum is reached. Each run ends either when a 3-coloring solution is reached
(success). or when 5000 iterations are completed in cascade. Figure 1 shows the
typical plot of the number of correctly colored nodes along a cascade coloration
process of a hard graph.

4.2 Mean Results of Experimentation

Table 1 shows the mean results of the experiments. We calculated the mean
number of friends and enemies per node per graph, computing the Hostility
#f riends
quotient H = #enemies . The final iteration step, the percentage of successful
runs per experiment and the percentage of well 3-colored nodes are registered in
table 1.

Table 1. Average succesful runs for the main topological graph classes and graph
topological measures

Mean values Trees Bipartite Hard Random

Number of friends 2.92 65.20 6.26 11.76
Number of enemies 2.00 33.83 3.48 3.54
Hostility 1.46 1.93 1.80 3.33
Final iteration 482.74 428.44 2345.70 1467.10
% successful runs 100.00 100.00 52.20 11.04
% well 3-colored nodes 100.00 100.00 99.21 92.16

In order to determine which independent variable (Number of friends, Num-

ber of friends or the quotient) is more correlated to dependent variables (Final
iteration step, percentage of successful runs per experiment, percentage of well
548 M. Graña et al.

3-colored nodes per run), we calculated the Pearson coefficient over the rows of
1. The result is given in table

Table 2. Correlation of topological measures with success measures

Pearson coefficient r Final iteration % successful runs % well 3 colored

Hostility 0,26 -0,86 -0,97
Number of enemies -0,52 0,50 0,35
Number of friends -0,50 0,42 0,26

The Hostility quotient shows a very high negative correlation to the percentage
of successful runs per experiment (runs that find a 3-coloration) and to the
percentage of well 3-colored nodes per run. From this result we select Hostility
H as the best graph feature that explains tidily the results regardless of the
topology of the graph.

4.3 Percentage of Well 3-Colored Nodes

Figure 2 shows (squares) that 3-coloration reaches the 100% of mean well 3-
colored nodes in the case of trees and bipartite graphs that are 2-colorable.
Hard graps means are over the 97% of well 3-colored nodes. Random graphs

Fig. 2. Distribution of coloring success related to the Hostility quotient

 Further Results on Swarms Solving Graph Coloring 549

show poor results in the case of graphs that are not 3-colorable since not all
planar graphs are 3-colorable. The graphic shows also (dots) the worst cases per
experiment: even in this case, hard graphs obtain results above 90%.

4.4 Percentage of Successful Runs Per Experiment

Respecting to the hits, i.e. the number of runs per experiment that end in a
3-coloration, figure 3 shows that the majority of the cases are over the 40%
and the mean is about a 75% of runs that end with success. Again, 2-colorable
graphs reach the 100% of hits rate. The sample of hard graphs exhibit a wider
variability, but 1 of two runs ends with success as can be observed in table 1.
Random graphs, remember that not all of them are 3-colorable, obtain quite
poor results, but only 4-colorable random graphs reach the 0% of hits success.

Fig. 3. Number of succesfull runs as a function of the Hostility quotient

4.5 On the Time Needed to Color a Node

To have a measure of computing performance, we estimate the minimal time cost
needed to color correctly one node dividing the minimal number of iterations of
one run in one experiment by the maximum number of well 3-colored nodes in
that experiment. Figure 4 shows the distribution of this value for the experiments
performed, identifying the graph topology class. Lowest costs correspond again
to 2-colorable graphs, in spite of the proportion of friends over enemies of trees
and bipartite graphs are very different. Highest costs correspond to hard graphs;
3-colorable random graphs are divided between those that obtain intermediate
time cost values and those that get inmediatly a 3-coloration due to previous
cascade 4-coloration left the problem almost solved.
550 M. Graña et al.

Fig. 4. Number of iterations versus the Hostility quotient

5 Conclusions
We propose in this paper a framework to study GCP in terms of flocking birds
that show territorial behavior, calling the algorithm Supervised Territorial Flock
Coloring (STFC). We studied separately territoriality and flocking, finding that
Extreme Territorial Flock Coloring (ETFC), based in goal velocity and attack
is able to solve GCP. ETFC algorithm is able to color Mizuno’s hard graphs for
3-coloration while Brélaz algorithm fail for all them.
We found that a global graph parameter, the ratio of nodes at distance one
(enemies) to nodes at distance greater than one (friends), which we have called
hostility, is a good characterization of the difficulty of solving GCP for each
graph. It has negative strong correlation with the success and time to solve the
problem.

References
1. Brélaz, D.: New methods to color vertices of a graph. Communications of ACM 22,
251–256 (1979)
2. Cases, B., Hernandez, C., Graña, M., D’anjou, A.: On the ability of swarms to
compute the 3-coloring of graphs. In: Bullock, S., Noble, J., Watson, R., Bedau,
M.A. (eds.) Artificial Life XI: Proceedings of the Eleventh International Confer-
ence on the Simulation and Synthesis of Living Systems, pp. 102–109. MIT Press,
Cambridge (2008)
 Further Results on Swarms Solving Graph Coloring 551

3. Diestel, R.: Graph Theory, electronic edn. Springer, USA (2005)

4. Dorigo, M., Blum, C.: Ant Colony Optimization theory: a survey. Theorical Com-
puter Science 344, 243–278 (2005)
5. Dorigo, M., Gambardella, L.M.: Ant Colonies for the Traveling Salesman Problem.
BioSystems 43, 73–81 (1997)
6. Galinier, P., Hertz, A.: A survey of local search methods for graph coloring. Com-
put. Oper. Res. 33, 2547–2562 (2006)
7. Galinier, P., Hertz, A., Zufferey, N.: An adaptive memory algorithm for the k-
coloring problem. Discrete Applied Mathematics 156(2), 267–279 (2008)
8. Mizuno, K., Nishihara, S.: Constructive generation of very hard 3-colorability in-
stances. Discrete Applied Mathematics 156(2), 218–229 (2008)
9. Reynolds, C.W.: Flocks, herds, and schools: A distributed behavioral model. Com-
puter Graphics 21, 25–34 (1987)
10. Reynolds, C.W.: Steering Behaviors for Autonomous Characters (1999),
http://www.red3d.com/cwr/papers/1999/gdc99steer.html
11. Smith, J.M.: Evolution and the theory of games. Cambridge University Press,
Cambridge (1992)
12. Smith, J.M., Price, G.R.: The logic of animal conflict. Nature 246 (November)
13. Turk, G.: Generating textures on arbitrary surfaces using reaction-diffusion. SIG-
GRAPH Computers Graphics 25(4), 289–298 (1991)
Data Collection System for the Navigation of Wheelchair
Users: A Preliminary Report

Yasuaki Sumida, Kazuaki Goshi, and Katsuya Matsunaga

Kyushu Sangyo University, Graduate School of Information Science,

2-3-1 Matsukadai, Hgashi-ku, Fukuoka 813-8503, Japan
[email protected],{goshi,matsnaga}@is.kyusan-u.ac.jp

Abstract. In Japan, population of aged people is becoming greater. According

to it, number of the wheelchair user would become larger. The government has
being improved to lessen the barriers for these people. But depending on the
physical power of wheel chair users, it was pointed out that there could be the
places where they could not move by their physical power. This is a reason that
many wheelchair users do not move by their alone. Therefore we tried to de-
velop the navigation system to eliminate the problem. At the beginning we have
developed the system to measure the required force to move the wheelchair, the
width of the space of aisle where they could turn back the direction, and the po-
sition information. By the data base, we planned to develop the navigation sys-
tem for wheelchair users who could move alone to working places or other
places where they want to go. In this paper, we report the system developed to
collect data for navigation of wheelchair users.

Keywords: wheelchair, navigation, GPS, QR-code.

1 Introduction
In Japan, the population of aged people is increasing [1]. Accordingly, wheelchair
users are expected to increase. Government regulation has lessened barriers that hin-
der them. However, depending on the physical power of wheelchair users, there are
places where they are unable to move using their own physical power. Consequently,
many wheelchair users do not move around by themselves.
For wheelchair users to live independently, they should be supported to move
anywhere. For example, the system which helps movement of wheelchair [5] and the
system which guides a movable course of the wheelchair [2, 6] would be necessary.
However, almost all such studies investigate systems to show routes. No system has
been designed to collect information for a navigation database.
This research is undertaken to develop a system that navigates wheelchair users
who want to go to the destination and give them the possibility to venture freely out-
doors. Regarding the research, we first developed a system to measure the force to
move a wheelchair, the size of the space in which a wheelchair might turn back, and
position information. These data of forces for wheelchair movement and space sizes
can be matched with building plans and roads. As described, we report this system
developed to collect data for wheelchair user navigation.

D. Taniar et al. (Eds.): ICCSA 2010, Part III, LNCS 6018, pp. 552–563, 2010.
© Springer-Verlag Berlin Heidelberg 2010
Data Collection System for the Navigation of Wheelchair Users: A Preliminary Report 553

2 Overview of the Developed System

Fig. 1 presents an overview of the system we intend to develop. Using the data collec-
tor, information for the navigation of wheelchair users will be collected and transmit-
ted to the database server. The physical power of wheelchair users to move a wheel-
chair is measured. Using the internet, wheelchair users can find a way to a destination
prior to using the database in the server. Moreover, wheelchair users will be navigated
in real time. Herein, we report the data collection system.

Fig. 1. Overview of the developed system

3 Data Collection Method

The necessary data for the system to navigate wheelchair users is location information
of consecutive measuring points, the necessary force to move a wheelchair to go to
the destination, information about space around the wheelchair for checking whether
it is possible to turn back or not, and the distance and time to the destination.

3.1 Measurement System to Collect Location Data

Force to move the wheelchair from the place of departure to the destination should be
corresponding to the map of the building and the road. The wheelchair positions are
identifiable using a Global Positioning System (GPS). We use the RMC format of GPS
to obtain the position information. The RMC format includes time, latitude, longitude,
ground speed, and direction of movement: such location data are sent as serial data.
The serial data are input to a built-in PC board via an AVR microcontroller with other
measured data. It is thereby possible to record all measured data with location informa-
tion. However, it is difficult to identify places within the buildings using GPS. We
attempted to identify the places in the buildings by the moving distances measured by
frequencies of wheelchair rotation from anchor points differentiated using QR codes
assigned to floor plans.
554 Y. Sumida, K. Goshi, and K. Matsunaga

The direction in which the wheelchair moves is identified using the QR code. In
Fig. 2, sheets on which QR codes are printed are attached on the ceiling at crossing
points of the passageways. The QR codes can indicate the wheelchair’s location and
direction of movement. The wheelchair position in the building is obtained according
to the moving direction identified by QR codes and moving distance measured using
the frequencies of rotation of the wheel from the position of certain QR codes at-
tached on the ceiling. The QR code includes position information in the building. In
Fig. 3, an example of QR code is shown. This QR code contains information of
"33.669835,130.444867,11" which means the degrees of latitude, the degree of longi-
tude and the altitude. The QR code includes position information in the building. In
Fig. 3, a QR code is shown: it contains information of "33.669835,130.444867,11",
denoting the degrees of latitude, the degree of longitude and the altitude. It is thereby
possible to obtain information about direction within buildings.

Ceiling at cross point of corridor

Fig. 2. Places where image charts of QR codes are attached

Fig. 3. A QR code

3.2 System to Measure Forces Moving the Wheelchair

Wheelchair users are not always able to move on any path. For example, it might be
difficult to move on steep slopes, roads with different levels, and pathways covered
with deep-pile carpets. However, no map shows road surface states in detail. There-
fore, the force necessary to move wheelchair must be inferred.
The force necessary to move the wheelchair is determined by the current to the mo-
tor of the electric wheelchair. The current to the motor is obtained by conversion of
voltage by the resistor connected between the motor and the batteries. The voltage
between edges of the resistor is amplified and converted into a digital value by the
microcontroller’s (AVR, ATMEGA644P: Atmel Corp.) AD converter. The obtained
value is calibrated into the value of the drive torque and into the force.
Data Collection System for the Navigation of Wheelchair Users: A Preliminary Report 555

3.3 Space Width Measurement

Wheelchair users want to know the width of spaces on the way to the destination.
They cannot turn back in narrow spaces. Therefore, for route navigation, space infor-
mation must be given that the space is sufficiently wide or not. To turn back, the pas-
sage width should be more than 3.6 m on many hand-driven wheelchairs and many
electricity-driven wheelchairs [4]. Width is measured using distance measuring de-
vices of ultrasonic type (EZ0 and EZ1, LV-MaxSonar: MaxBotix Inc.). The distance
measurement range is 0–6.5 m. The ultrasonic device outputs a pulse width that is
proportional to the distance to the object. The distance is obtainable by measuring the
pulse width by the counter of the AVR microcontroller. Equation (1) is used for com-
putation to make one count equal to 1 cm.
Count cycle = {1 [s] / (0.01 [m] / acoustic velocity [m/s])} /2 (1)

The acoustic velocity is altered by temperature, humidity, and air density. For that rea-
son, it is necessary to calibrate it according to these values.

3.4 Movement Distance and Velocity Measurement

As described previously, it is difficult for the GPS to obtain location data in build-
ings. Therefore, moving distances are measured using the wheelchair wheel’s
rate of rotation. It is obtained by counting by a device of counter a circular
pattern of monochrome alternation inside the wheel. The counter using a photo-
graph interrupter outputs pulses when it detects a white pattern. The rate of wheel
rotation was obtained by counting the standing up of the pulse every second using
an AVR microcontroller. The movement distance was obtained using eq. (2). The
movement velocity is calculable according to the movement distance and movement
time.
Movement distance each one second [m] = {wheel circumference [m] / (2)
number of white pattern} × pulse count each second

4 Design and Implementation

4.1 Hardware and Software Based on ASSIST

We developed the Assistant System for Safe Driving by Informative Supervision

and Training (ASSIST) [3], which records driving parameters including velocity,
headway distance, and position: it sends the data to a supervisor at a remote place to
educate drivers about safe driving. We decided to develop the data collection sys-
tem based on ASSIST because the ASSIST hardware and software for recording
and transmitting are applicable to collection of routing information for wheelchair
users.
556 Y. Sumida, K. Goshi, and K. Matsunaga

4.2 Hardware Design

4.2.1 Camera to Obtain Front View Images

People can easily find their way if landmarks are given. We designed the system to
get pictures using a video camera while the probe car was collecting data. The video
camera was attached on the front of the wheelchair.

4.2.2 Overall Design of the Hardware (W-ASSIST Hardware)

Fig. 4 shows a block diagram of the hardware with an embedded computer board, an
interface, video cameras, distance sensors, GPS, and a rotation sensor. The interface
was designed with an AVR microcontroller (ATMEGA644P: Atmel Corp.). The
interface hands over the data from the sensors to the embedded PC. The embedded PC
board was connected with cameras to obtain front view images and QR codes, and a
mobile phone for data transmission to the data server and the web server.
Fig. 5 shows the electric wheelchair with our measurement system.

Fig. 4. Overview of the hardware

Fig. 5. The electric wheelchair used for measurements

4.3 Software(W-ASSIST Software)

4.3.1 Design of the Software in the Interface and the Embedded PC Board
Data from all sensors and GPS are input into the AVR microcontroller in the interface
board. Then the inputted data are related with the GPS location data and sent to an
embedded PC board as serial data. The data transmission is done every second when
the data of GPS are complete. The data are transmitted to the data server and the web
server. Fig. 6 portrays a block diagram of the system showing the data flow.
Data Collection System for the Navigation of Wheelchair Users: A Preliminary Report 557

In-car unit

Interface GPS Each Sensor

(Aisle width,
AVR microcontroller
Distance and
Speed, Drive
Measurement
torque)
data
Serial data

Built-in PC board
Picture Image
Serial data
with Picture
Mobile Phone

Memory

Video
Camera

Fig. 6. Block diagram of the system and the flow of data

4.3.2 Software Design

It is necessary to monitor the system to verify its function. We developed software to
check the system visually and display graphical images on the screen to confirm the
logged data and to monitor the system working in real time. The software was made
by revising the ASSIST system software.

Fig. 7. Design of the log-viewer software

Fig. 8. Design of the live-monitoring software

558 Y. Sumida, K. Goshi, and K. Matsunaga

We designed and developed two software, log-viewer software for checking log
data, and the live monitor software for a rear-time check. Fig. 7 shows the design of
the log-viewer software. Fig. 8 shows the design of the live-monitoring software.

5 Results
5.1 Results of System Performance

The measured data is stored in a flash memory by the character string by the recorder
program. The log-viewer software obtains each data by dividing this character string.
The stored data can be checked by the log-viewer software. Fig. 9 shows a method of
data processing in log-viewer software. The data is stored in order of the position
information from the GPS, measurement data of a microcomputer, data for debug-
ging. The Log-viewer software displays texts and graphs by reading this character
string, dividing it, and obtaining each data. Fig. 10 presents a screen display drawn
from data logged in the W-ASSIST hardware using the log-viewer software. The
hardware and the software of the system worked well to get signals from the GPS,
the camera, the current measurement device, the distance measurement devices, and
the wheel rotation-counting device. The log-viewer software is developed using Java.
The live-monitoring software consists of a recording program, a transmitting pro-
gram, a W-ASSIST server program and a web interface. Fig. 11 shows the flow of
information of the live-monitoring software. The recording program sends data to
summing program through UDP socket. The transmitting program sends live data to
the server of ASSIST through TCP socket on the mobile network. The live data con-
sists of times, wheelchair’s IDs, measurer’s IDs, latitudes, longitudes, measurement
data and front view images. The graphs are drawn using measured data for 20seconds

Fig. 9. Method of data processing in log-viewer software

Data Collection System for the Navigation of Wheelchair Users: A Preliminary Report 559

Fig. 10. Screen image displayed by the live-monitoring software through the web

Fig. 11. Block diagram of the live-monitoring software

Fig. 12. Screen image displayed by the live-monitoring software through the web
560 Y. Sumida, K. Goshi, and K. Matsunaga

each. Fig. 12 shows a screen display drawn from the data obtained by the W-ASSIST
hardware using the live monitor software. Data were transmitted to the servers using
the mobile phone system via the internet system. Software for live monitoring of the
driving behavior worked well (Fig. 12). The live-monitoring software is developed
using PHP and javascript.

5.2 Measurement Accuracy

y The method of experiment

In order to verify about the reproducibility of the measurements, the measurements
were done five times by running wheel chair (Fig. 13). In fig. 15, A, B, C and D mean
the measurement points.

Fig. 13. Experimental environment

y The result of experiment

Table 1 shows the result of moving mileages measured. In Table 1, moving distances
were calculated by multiplying the length of every one pulse by the pulse number
total. The measured distances have the errors of about 3 meters at most. However, the
differences of the totals of pulse numbers are small enough. The cause of the errors of
the distance measurements is thought to be derived from the distortion of the tires
while wheelchair running.

Table 1. The result of measured mileage

Fig. 14 shows the results of measured aisle width which were gotten by the dis-
tance measuring devices. The differences could be thought to be small.
Fig. 15 shows the result of the measured voltage to analyze the currents to the mo-
tors. Although there are some differences, the reproducibility could be thought to be
high enough.
Data Collection System for the Navigation of Wheelchair Users: A Preliminary Report 561

Fig. 14. Result of measured aisle width

Fig. 15. The result of measured voltage to analyze the driving force

5.2.1 Verification of the Relation between Drive Torque and Actual Power

To confirm the correlation of the voltages and the forces to move wheelchair.
y The method of experiment
The measurements were done five times by running the wheelchair on the slopes of
the same asphalt pavement. Those slope's gradient are different. And the forces to
move wheelchair were measured by the spring balance.
y The result of experiment
Fig. 16 shows the result of the measured forces to move wheelchair using the spring
balance and the voltage while the wheelchair movement for the measurement. Both
inclinations could be thought to be proportional to the slope angles. In addition, we
conducted multiple linear regression analysis about the starting powers and the run-
ning powers, and got the multiple regression equations (eq. (3, 4)).
562 Y. Sumida, K. Goshi, and K. Matsunaga

Fig. 16. Result of experiment

The needed starting power [kg] = 5.2054 + 0.0032 * Max voltage[mV] (3)

The needed running power[kg] =

(4)
-0.037 + 0.00192 * Average driving torque voltage[mV]
The result of the predicted value computed based on these equations is shown in Ta-
ble 1. It is almost the same value, and if some errors would be permitted, the fore to
move wheelchair could be converted into the driving torque by these multiple regres-
sion equations.

Table 2. Prediction result

5.3 Result of a QR Code Reading System

We developed the QRcode reading system using the opencv1.1 library and the libde-
codeqr library by C language. QR Codes are read in the following procedure. First,
the still pictures were gotten using camera. If a QR code in a still picture image was
found, it was clipped out. Next, it’s image was expanded into double and the QR code
was recognized. Finally, the location information in the QR code was analyzed. Fig.
19 shows an example of the screen display where the system was executed actually.

Fig. 17. The examples of execution of the QR code reading system

Data Collection System for the Navigation of Wheelchair Users: A Preliminary Report 563

y The method of experiment

The QR code readings were done using the wheelchair for the measurement under the
five QR codes which were posted to the ceiling. We succeeded in the reading of three
of the QR cords in five pieces.

5 Conclusion and Future Studies

As described herein, we reported a system developed to collect data to navigate
wheelchair users. Results show the possibility of an automatic data collecting system
used for the navigation system database. A measurement function will be added to
collect more data including route height information and ground surface properties.

References
1. Statistics Bureau in Japan,
http://www.stat.go.jp/data/jinsui/tsuki/index.htm
2. Yang, A., Mackworth, A.K.:
http://people.cs.ubc.ca/~mack/Publications/CCAI07.pdf
3. Goshi, K., Matsunaga, K., Kuroki, D., Shidoji, K., Matsuki, Y.: Educational Intelligent
Transport System ASSIST. In: Proceedings of the Fourth IASTED International Confer-
ence Computers and Advanced Technology in Education, Banff, pp. 150–154 (2001)
4. Epoch House, http://www.epochhouse.com/care05.htm
5. Wang, H., Salatin, B., Grindle, G., Ding, D., Cooper, R.: Real-time model based electrical
powered wheelchair control. Medical Engineering & Physics 31(10), 1244–1254
6. Kasemsuppakorn, P., Karimi, H.A.: Personalised routing for wheelchair navigation. Jour-
nal of Location Based Services 3(1), 24–54 (2009)
 Author Index

Abawajy, Jemal H. III-201 Cao, Lu IV-427

Abraham, Ajith III-472 Cao, Yanzhao IV-409
Ahmadian, Kushan I-574 Cardell-Oliver, Rachel III-336
Alarcon, Vladimir J. I-491, I-501 Carlini, Maurizio II-177, II-206
Amorim, Ronan M. IV-395 Casas, Giuseppe Las I-62
Anders, Frauke I-152 Cases, Blanca III-541
Andriamasinoro, Fenintsoa I-476 Cassard, Daniel I-476
An, Hong IV-427 Castellucci, Sonia II-177
Anuar, Mohd Hafiz I-331 Cattani, Carlo II-155, II-164, II-215,
Aoki, Takaaki IV-252 II-225
Arai, Kohei II-71, II-87, II-336, III-305 Cattrysse, Dirk I-414
Areerachakul, Sirilak III-215 Cecchini, Arnaldo I-166
Arikan, Yüksel Deniz II-544 Chai, Kevin II-351
Aritsugi, Masayoshi II-412 Chan, Chien-Hui III-526
Asche, Hartmut I-346, I-515 Chang, Maw-Shang II-314
Atman, Nilüfer II-544 Checiches, Gabriela II-215
Chen, Changqian IV-296
Baba, Kensuke IV-236, IV-273 Cheng, Chin-Yi III-431
Badea, Radu II-215 Cheng, Kai III-395, IV-418
Bagstad, Kenneth J. I-238 Chen, Gong-liang II-14
Bai, Songnan IV-520 Chen, Jiming IV-296
Barros, Diego Martins Vieira I-430 Cheon, Sung Moon I-182
Basuki, Achmad II-87 Che, Z.H. II-533
Battaglia, Francesco I-1 Che, Zhen-Guo II-533
Benigni, Gladys II-422 Chiang, C.J. II-533
Benkner, Siegfried IV-13 Chiang, Tzu-An II-533
Beristain, Andoni I-610 Chi, Hongmei IV-409
Bernardi, Milena II-206 Choi, Bum-Gon III-85
Bhuruth, Muddun II-570 Choi, Hyung-Jin III-118
Bi, Zhongqin IV-482 Choi, Wook III-129
Blecic, Ivan I-166 Chong, Dahae III-21, III-31
Boizumault, Patrice II-432 Choo, Hyunseung III-85, III-96,
Boo, Chang-Jin II-99, II-110 III-129, III-158
Boojhawon, Ravindra II-570 Chou, Shihchieh III-431
Borruso, Giuseppe I-1 Cho, Yongyun III-258, III-269
Bramley, Randall II-503 Chuang, Chun-Ling III-178
Brankovic, Ljiljana II-586 Chu, Hone-Jay I-116
Brooks, Christopher I-501 Chung, Min Young III-63, III-85, III-96
Bucur, Razvan II-215 Chung, Tai-Myoung III-142, III-352
Burry, Jane III-483 Chyan, Shyh-Haw I-293
Ciloglugil, Birol II-556
Cacheiro, Javier Lopez IV-41 Coluzzi, Rosa I-361
Cafer, Ferid II-301 Cong, Ming IV-427
Canters, Frank I-89 Costantini, Alessandro IV-29, IV-41
566 Author Index

Cracknell, Arthur P. I-545 Gavrilova, Marina I-574

Crémilleux, Bruno II-432 Gensel, Jérôme I-445
Crisan, Diana II-215 Gervasi, Osvaldo II-422, IV-41
Crisan, Maria II-215 Ghosal, Amrita III-321
Ghosh, S.K. I-309
Dai, Miau Ru III-368 Gizzi, Fabrizio I-320
Dan, Avishek III-321 Glorio, Octavio I-461
Danese, Maria I-320 Goi, Bok-Min IV-188
D’Anjou, Alicia III-541 Goldfeld, Paulo II-475, IV-395
Daněk, Josef IV-62 Goshi, Kazuaki III-552, IV-497
DasBit, Sipra III-321 Gotoh, Yusuke II-324
Datta, Amitava III-336 Graña, Manuel I-610, III-541
Davoine, Paule-Annick I-445 Gutierrez, Eduardo IV-41
de Doncker, Elise II-139
Delgado-Mohatar, Oscar II-586 Hadden, John IV-358
Deris, Mustafa Mat III-201, Halder, Subir III-321
III-405, IV-175 Hamaguchi, Nobuyuki II-139
Desudchit, Tayard III-419 Han, Yi IV-263
Dickstein, Flavio II-475, IV-395 Han, Young-Ju III-142, III-352
Di Donato, Pasquale I-528 Hara, Hideki IV-119
Dohi, Tadashi IV-441 Harsono, Tri II-71
Dong, Fei I-131 Hasegawa, Hidehiko II-60
Dong, Wei II-463 Hashim, Mazlan I-331, I-545
Dookhitram, Kumar II-570 Hatzichristos, Thomas I-140
dos Santos Amorim, Elisa Portes II-475, Hayashi, Masaki IV-497
IV-395 Hayati, Pedram II-351, II-400
dos Santos, Rodrigo Weber II-475, Hedar, Abdel-Rahman IV-457
IV-395 He, Jie III-498
Dostálová, Taťjána IV-62 Heng, Swee-Huay IV-188
Dvorský, Jiřı́ III-472 Herawan, Tutut III-201, III-405, IV-175
Hernandez, Carmen III-541
Edwards Jr., David II-1
Hernández, Constanza II-361
Emanuele, Strano I-32
Hirata, Kazuhiro IV-497
Embong, Abdullah IV-83
Hirose, Hideo IV-199
Engelen, Guy I-89
Hliňáková, Petra IV-62
Ervin, Gary I-501
Hong, Youngshin III-52
Esposto, Stefano II-206
Hope, Martin III-228
Faisal, Zaman IV-199 Hsieh, Nan-Chen III-526
FanJiang, Yong-Yi II-257 Huang, Wong-Rong II-257
Fidalgo, Robson do Nascimento I-430 Huang, Zequn IV-520
Firoozeh, Nazanin II-370 Hung, Ruo-Wei II-314
Firuzeh, Nazanin II-400
Fischer, Edward II-285 Ikeda, Daisuke IV-236
Florea, Mira II-215 Im, Se-bin III-118
Fujimoto, Junpei II-139 Inceoglu, Mustafa Murat II-556
Fujita, Shigeru IV-119, IV-128 İnceoğlu, Mustafa Murat II-544
Fuster-Sabater, Amparo II-586 Inenaga, Shunsuke IV-236
Ishikawa, Tadashi II-139
Gansterer, Wilfried N. IV-13 Ishiwata, Emiko II-60
Garcia, Ernesto IV-1 Ismail, Rashad IV-457
 Author Index 567

Itokawa, Tsuyoshi II-412 Kim, Tai-Hoon II-422

Ito, Taishi IV-138 Kinoshita, Tetsuo IV-107, IV-138,
Iwane, Masahiko II-488 IV-152, IV-164
Izumi, Satoru IV-152 Kitagata, Gen IV-164
Kitasuka, Teruaki II-412
Jamel, Sapiee IV-175 Kobayashi, Yusuke IV-152
Janciak, Ivan IV-13 Koehler, Martin IV-13
Jang, Jun-Hee III-118 Köhler, Hermann I-152
Jang, Myungjun I-262 König, Reinhard I-152
Jazyah, Yahia Hasan III-228 Konno, Susumu IV-107, IV-119
Jehng, Jihn-Chang III-431 Krömer, Pavel III-472
Jeong, Seungmyeong III-72 Kudreyko, Aleksey II-155
Jeong, Yeonjune III-158 Kuo, Jong-Yih II-257
Jeung, Jaemin III-72 Kurihara, Yoshimasa II-139
Jia, Xiaoqi IV-468 Kusuda, Tetsuya IV-336
Jing, Jiwu IV-468 Kwak, Ho-Young III-11
Ji, Yindong II-463 Kwong, Kim-hung I-374, I-389
Jo, Heasuk IV-510 Kwon, Young Min III-63
Johnson, Gary W. I-238
Joo, Yongjin I-105 Laganà, Antonio IV-1, IV-41
Jun, Chulmin I-105 Lago, Noelia Faginas IV-29
Jung, Jaeil IV-520 Lai, Poh-chin I-374, I-389
Jung, Soon-Young IV-376 Lanorte, Antonio I-361
Ju, Shiguang IV-296 Lasaponara, Rosa I-254, I-361
Laserra, Ettore II-225
Kaio, Naoto IV-441 Le, Thuy Thi Thu I-401
Kalisch, Dominik I-152 Lee, Chang H. IV-370
Kaneko, Kunihiko III-189 Lee, Cheng-Chi I-599
Kang, Ji-Ae III-11 Lee, Eunseok IV-385
Kang, Min-Jae II-99, II-110, III-11 Lee, Im Hack I-182
Kang, Seung Goo III-21 Lee, Jin-Kyung I-271
Kawato, Akifumi IV-164 Lee, Junghoon III-1, III-11, III-52
Khiari, Mehdi II-432 Lee, Ju Yong III-63, III-85
Kim, Byung-Sung III-158 Lee, Kwang Y. II-99
Kim, Chang Seup III-85 Lee, Myungsoo III-31
Kim, Choel Min III-1 Lee, Ok Kyung III-63
Kim, Dong In III-42 Lee, Saebyeok IV-376
Kim, Ho-Chan II-99, II-110 Lee, Sang Joon III-52
Kim, Hyeon-Cheol IV-376 Lee, Seungil I-271
Kim, Hyunduk III-158 Lee, Tae-Jin III-85, III-96
Kim, Jae-Yearn II-119 Lee, WonGye IV-376
Kim, Jingyu III-42 Lee, Youngpo III-21, III-31
Kim, Jong-Myoung III-352 Lee, Youngyoon III-21
Kim, Junhwan III-31 Lee, Yue-Shi III-458
Kim, Kyu-Il I-271 Li, Chun-Ta I-599
Kim, Kyungill IV-370 Li, Jian-hua II-14
Kim, Sang-Wook III-1 Li, Ming II-191
Kim, Seungjoo IV-510 Lim, HeuiSeok IV-370, IV-376
Kim, Shin Do I-182 Lim, Jaesung III-72
Kim, Taeyoung III-129 Lin, Feng-Tyan I-77, I-293
568 Author Index

Lin, Jingqiang IV-468 Nagy, Miroslav IV-62

Lin, Rong-Ho III-178 Nakamura, Toru IV-236
Lin, Yu-Pin I-116, I-224 Namatame, Akira IV-321
Liou, William W. II-25 Nedoma, Jiřı́ IV-62
Li, Peng IV-427 Nickerson, Bradford G. I-401
Lischka, Hans IV-13 Nicolas, Lachance-Bernard I-32
Li, Tiancheng II-44 Ninomiya, Daisuke IV-252
Liu, Dong IV-427 Nishida, Akira II-448
Liu, Fang II-503 Nissen, Volker IV-346
Liu, Hsiao-Lan I-293 Niyogi, Rajdeep IV-309
Liu, Liang I-590 Nomura, Yoshinari II-324
Liu, Peng IV-468 Nomura, Yusuke II-324
Liu, Yuan IV-427
Li, Yin II-14 Ochodková, Eliška III-472
Li, Yu I-590 Ogi, Tetsuro IV-336
Lobosco, Marcelo IV-395 O‘Hara, Charles G. I-491
Lursinsap, Chidchanok III-419 Oh, Chang-Yeong III-96
Lu, Tianbo IV-263 Okamoto, Kouta II-324
Lu, Wenjie I-590 Orshoven, Jos Van I-414
Osada, Toshiaki IV-164
Maggio, Grazia I-210
Pallottelli, Simonetta IV-29
Mahmud, Mohd Rizaludin I-331
Pannacci, Nicola IV-29
Manabe, Yusuke IV-119
Park, Gyung-Leen III-1, III-11, III-52,
Mancini, Francesco I-210
III-107
Mantelas, Lefteris A. I-140 Park, Min-Woo III-142, III-352
Mardiyanto, Ronny II-336 Park, Soohong I-105
Marghany, Maged I-331, I-545 Passeri, Francesco Luca II-422
Martel-Jantin, Bruno I-476 Pazand, Babak III-336
Maruyama, Katsumi II-324 Pecci, Francesco I-46
Ma, Shang-Pin II-257 Perchinunno, Paola I-17
Masini, Nicola I-254, I-320, I-361 Phinitkar, Pattira IV-209
Matsunaga, Katsuya III-552, IV-497 Pirani, Fernando IV-1
Mazón, Jose-Norberto I-461 Platoš, Jan III-472
McAnally, William I-501 Plumejeaud, Christine I-445
Mekhedov, Ivan I-557 Pontarollo, Nicola I-46
Mestetskiy, Leonid I-557 Porceddu, Andrea I-1
Meza, Ricardo Rafael Quintero II-241 Potdar, Vidyasagar II-351, II-370,
Milani, Alfredo IV-309 II-383, II-400
Misra, A.K. II-273 Potenza, Maria Rosaria I-320
Misra, Sanjay II-301 Prastacos, Poulicos I-140
Mitrea, Delia II-215 Přečková, Petra IV-62
Mitrea, Paulina II-215 Prud’homme, Julie I-445
Montrone, Silvestro I-17 Purnami, Santi Wulan IV-83
Moon, Hyun-joo III-269
Moon, Jongbae III-258, III-269 Quintero, Ricardo II-361
Mukherjee, Indira I-309
Müller-Molina, Arnoldo José III-443, Raba, Nikita II-130
IV-252 Rahayu, Wenny III-380
Murgante, Beniamino I-62, I-320 Rampino, Sergio IV-1
 Author Index 569

Ridzuan, Farida II-383, II-400 Suh, Woonsuk IV-385

Robinson, Ian II-44 Sumida, Yasuaki III-552
Röcker, Carsten IV-93 Sur, Sanjib III-321
Rodriguez, Aurelio IV-41
Rotondo, Francesco I-283 Takahashi, Hideyuki IV-138, IV-152
Ruckenbauer, Matthias IV-13 Takahata, Kazuo III-168
Takaoka, Tadao II-519
Saft, Danilo IV-346 Takashita, Taiki II-412
Saito, Tsubasa II-60 Talevski, Alex II-351, II-383, II-400
Sakatoku, Akira IV-164 Tang, Cheng-Jen III-368
Salim, Flora Dilys III-483 Tangkraingkij, Preecha III-419
Sánchez, Leopoldo Z. II-241, II-361 Tangman, Desire Yannick II-570
Sanguansintukul, Siripun III-215, Taniar, David I-574
III-419 Taniguchi, Hideo II-324
Sarencheh, Saeed II-370, II-400 Tan, Syh-Yuan IV-188
Sari, Anny Kartika III-380 Tasso, Sergio IV-29
Scardaccione, Grazia I-62 Tilio, Lucia I-320
Schleupner, Christine I-193 Timothée, Produit I-32
Scorza, Francesco I-62 Tiwari, Ashutosh IV-358
Selicato, Francesco I-210
Toi, Yutaka III-498
Selmane, Schehrazad IV-72
Tokuhisa, Soichiro III-189
Sen, Jaydip III-246, III-277
Torre, Carmelo Maria I-17
Sergio, Porta I-32
Trujillo, Juan I-461
Serikawa, Seiichi II-488
Trunfio, Giuseppe A. I-166
Shan, Pei-Wei II-191
Tsai, Hsin-Che III-526
Shen, Liyong IV-482
Tsai, Ming-Hsun III-511
Shibata, Yoshitaka III-168
Tseng, Vincent S. III-458
Shimizu, Yoshimitsu II-139
Tseng, Wan-Yu I-77
Shin, In-Hye III-1, III-11, III-52, III-107
Tseng, Yuh-Min IV-225
Shinohara, Takeshi III-443, IV-252
Tu, Pu III-291
Shiratori, Norio IV-138, IV-152, IV-164
Turner, Chris IV-358
Shon, Minhan III-129
Shukla, Ruchi II-273
Şimşek, Ömer II-544 Uchida, Noriki III-168
Singh, Kuldip IV-309 Uchiya, Takahiro IV-107
Snapp, Robert R. I-238 Uddin, Mohammad Mesbah IV-199
Snášel, Václav III-472 Uemura, Toshikazu IV-441
Song, Chonghan III-21, III-31 Ufuktepe, Ünal IV-53
Song, MoonBae III-129 Ukey, Nilesh IV-309
Sophatsathit, Peraphon IV-209 Ukil, Arijit III-277
Sosonkina, Masha II-503 Uljee, Inge I-89
Stankova, Elena II-130 Ushijima, Kazuo IV-418
Stankute, Silvija I-515 Ushioda, Tatsuya IV-128
Steinhöfel, Jens I-152
Stéphane, Joost I-32 van der Kwast, Johannes I-89
Suganuma, Takuo IV-138, IV-152 Van de Voorde, Tim I-89
Sugawara, Kenji IV-119 Vanegas, Pablo I-414
Suga, Yuji IV-284 Villa, Ferdinando I-238
Sugiyanta, Lipur III-305 Villarini, Mauro II-206
Suh, Soon-Tak I-262 Villecco, Francesco I-590
570 Author Index

Wang, Cheng-Long I-224 Yang, Yang II-25

Wang, Hao III-291 Yao, Chih-Chia III-511
Wang, Lian-Jun I-599 Yasuura, Hiroto IV-236
Wang, Ping III-291 Yeganeh, Elham Afsari II-370, II-400
Wang, Shuai II-463 Yen, Show-Jane III-458
Wang, Tao IV-427 Yılmaz, Buket IV-53
Wang, Ye IV-263 Ying, Jia-Ching III-458
Wang, Yung-Chieh I-224 Yoe, Hyun III-258
Wolff, Markus I-346 Yokoyama, Kazutoshi II-324
Wollersheim, Dennis III-380 Yoon, Seokho III-21, III-31
Won, Dongho IV-510 Yoo, Yoong-Seok II-119
Won, Kyunghoon III-118 Yuasa, Fukuko II-139
Wu, Bin IV-482 Yu, Hsiao-Hsuan I-224
Wu, Chen-Fa I-116 Yu, Zhi-hong IV-427
Wu, Tsu-Yang IV-225
Wu, Yu-Chieh III-458 Zain, Jasni Mohamad IV-83
Zazueta, Liliana Vega II-241
Xavier, Carolina Ribeiro II-475, IV-395 Zeng, Zhenbing IV-482
Xiang, Limin IV-418 Zha, Daren IV-468
Zhai, Yuyi I-590
Yamamoto, Toshihiko IV-321 Zhang, Hong I-131
Yamawaki, Akira II-488 Zhao, Yaolong I-131
Yang, Jian III-291 Zotta, Cinzia I-320
Yang, Shiliang I-590 Zou, Zhiwen IV-296
Yang, Shiyuan II-463 Zurada, Jacek M. II-110