Lakhmi C. Jain, Vasile Palade and Dipti Srinivasan (Eds.) Advances in Evolutionary Computing for System Design
Studies...

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Lakhmi C. Jain, Vasile Palade and Dipti Srinivasan (Eds.) Advances in Evolutionary Computing for System Design

Studies in Computational Intelligence, Volume 66 Editor-in-chief Prof. Janusz Kacprzyk Systems Research Institute Polish Academy of Sciences ul. Newelska 6 01-447 Warsaw Poland E-mail: [email protected] Further volumes of this series can be found on our homepage: springer.com

Vol. 55. Xianyi Zeng, Yi Li, Da Ruan and Ludovic Koehl (Eds.) Computational Textile, 2007 ISBN 978-3-540-70656-4

Vol. 44. Nadia Nedjah, Luiza de Macedo Mourelle, Mario Neto Borges, Nival Nunes de Almeida (Eds.) Intelligent Educational Machines, 2007 ISBN 978-3-540-44920-1

Vol. 56. Akira Namatame, Satoshi Kurihara and Hideyuki Nakashima (Eds.) Emergent Intelligence of Networked Agents, 2007 ISBN 978-3-540-71073-8

Vol. 45. Vladimir G. Ivancevic, Tijana T. Ivancevic Neuro-Fuzzy Associative Machinery for Comprehensive Brain and Cognition Modeling, 2007 ISBN 978-3-540-47463-0

Vol. 57. Nadia Nedjah, Ajith Abraham and Luiza de Macedo Mourella (Eds.) Computational Intelligence in Information Assurance and Security, 2007 ISBN 978-3-540-71077-6

Vol. 46. Valentina Zharkova, Lakhmi C. Jain Artificial Intelligence in Recognition and Classification of Astrophysical and Medical Images, 2007 ISBN 978-3-540-47511-8

Vol. 58. Jeng-Shyang Pan, Hsiang-Cheh Huang, Lakhmi C. Jain and Wai-Chi Fang (Eds.) Intelligent Multimedia Data Hiding, 2007 ISBN 978-3-540-71168-1

Vol. 47. S. Sumathi, S. Esakkirajan Fundamentals of Relational Database Management Systems, 2007 ISBN 978-3-540-48397-7

Vol. 59. Andrzej P. Wierzbicki and Yoshiteru Nakamori (Eds.) Creative Environments, 2007 ISBN 978-3-540-71466-8

Vol. 48. H. Yoshida (Ed.) Advanced Computational Intelligence Paradigms in Healthcare-1, 2007 ISBN 978-3-540-47523-1

Vol. 60. Vladimir G. Ivancevic and Tijana T. Ivacevic Computational Mind: A Complex Dynamics Perspective, 2007 ISBN 978-3-540-71465-1

Vol. 49. Keshav P. Dahal, Kay Chen Tan, Peter I. Cowling (Eds.) Evolutionary Scheduling, 2007 ISBN 978-3-540-48582-7

Vol. 61. Jacques Teller, John R. Lee and Catherine Roussey (Eds.) Ontologies for Urban Development, 2007 ISBN 978-3-540-71975-5

Vol. 50. Nadia Nedjah, Leandro dos Santos Coelho, Luiza de Macedo Mourelle (Eds.) Mobile Robots: The Evolutionary Approach, 2007 ISBN 978-3-540-49719-6

Vol. 62. Lakhmi C. Jain, Raymond A. Tedman and Debra K. Tedman (Eds.) Evolution of Teaching and Learning Paradigms in Intelligent Environment, 2007 ISBN 978-3-540-71973-1

Vol. 51. Shengxiang Yang, Yew Soon Ong, Yaochu Jin Honda (Eds.) Evolutionary Computation in Dynamic and Uncertain Environment, 2007 ISBN 978-3-540-49772-1 Vol. 52. Abraham Kandel, Horst Bunke, Mark Last (Eds.) Applied Graph Theory in Computer Vision and Pattern Recognition, 2007 ISBN 978-3-540-68019-2 Vol. 53. Huajin Tang, Kay Chen Tan, Zhang Yi Neural Networks: Computational Models and Applications, 2007 ISBN 978-3-540-69225-6 Vol. 54. Fernando G. Lobo, Cl´audio F. Lima and Zbigniew Michalewicz (Eds.) Parameter Setting in Evolutionary Algorithms, 2007 ISBN 978-3-540-69431-1

Vol. 63. Wlodzislaw Duch and Jacek Ma´ndziuk (Eds.) Challenges for Computational Intelligence, 2007 ISBN 978-3-540-71983-0 Vol. 64. Lorenzo Magnani and Ping Li (Eds.) Model-Based Reasoning in Science, Technology, and Medicine, 2007 ISBN 978-3-540-71985-4 Vol. 65. S. Vaidya, L.C. Jain and Hiro Yoshida (Eds.) Advanced Computational Intelligence Paradigms in Healthcare-2, 2007 ISBN 978-3-540-72374-5 Vol. 66. Lakhmi C. Jain, Vasile Palade and Dipti Srinivasan (Eds.) Advances in Evolutionary Computing for System Design, 2007 ISBN 978-3-540-72376-9

Lakhmi C. Jain Vasile Palade Dipti Srinivasan (Eds.)

Advances in Evolutionary Computing for System Design With 134 Figures and 55 Tables

123

Prof. Lakhmi C. Jain

Dr. Dipti Srinivasan

KES Centre University of South Australia Adelaide, Mawson Lakes Campus South Australia SA 5095 Australia E-mail:- [email protected]

Department of Electrical and Computer Engineering National University of Singapore Block E4-06-08 4 Engineering Drive 3 Singapore 117576 E-mail:- [email protected]

Dr. Vasile Palade Oxford University Computing Laboratory Wolfson Building Parks Road Oxford, OX1 3QD England E-mail:- [email protected]

Library of Congress Control Number: 2007926313 ISSN print edition: 1860-949X ISSN electronic edition: 1860-9503 ISBN 978-3-540-72376-9 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, reuse of illustrations, recitation, broadcasting, reproduction on microfilm 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-Verlag. Violations are liable to prosecution under the German Copyright Law. Springer is a part of Springer Science+Business Media springer.com c Springer-Verlag Berlin Heidelberg 2007 ° The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Cover design: deblik, Berlin Typesetting by the SPi using a Springer LATEX macro package Printed on acid-free paper SPIN: 11551430 89/SPi

543210

Preface

In the last two decades, there has been a huge trend towards building intelligent systems that are inspired by biological systems from nature. Many nature-inspired algorithms and approaches have been developed and have become popular for designing intelligent systems with application to various real-world problems. Evolutionary computing paradigms, which are powerful search techniques inspired from natural genetics, proved to be at the forefront of this endeavour. Undoubtedly, many authored and edited books have already revealed the ﬁeld of evolutionary computing from many facets, which include deep theoretical analyses as well as more application-oriented studies. Edited volumes containing conference papers on the theory and application of evolutionary computing have also been very popular. In this book, the editors tried to put together a collection of extended and consistent chapters presenting original and latest research work from leading researchers on evolutionary computing. The focus of the book was to present challenging real-world applications of evolutionary computing paradigms as well as to provide an up-to-date reference manual for someone interested to use evolutionary techniques to system design. This edited book consists of thirteen chapters covering a wide area of topics on evolutionary computing and applications. The application areas are diverse and include system control, bioinformatics, hardware optimization, traﬃc grooming, games theory, grid computing, and so on. Challenging and captivating theoretical issues are also introduced and discussed. These refer especially to the use of evolutionary computing techniques in conjunction with other machine learning approaches, and include: evolutionary neurofuzzy systems, fuzzy systems and genetic algorithms, evolutionary algorithms and immune learning for neural network design, evolutionary techniques for multiple classiﬁer design, hybrid multi-objective evolutionary algorithms, evolutionary particle swarms. Our hope is that this book will serve as a reference to researchers in evolutionary computing and to system designers and practitioners working

VI

Preface

in various application domains who are interested in using evolutionary computing paradigms in system design and implementation. This book can also be used by students and lecturers as an advanced reading material for courses on evolutionary computing. The editors are grateful to the authors for their excellent contributions to the book. Thanks are due to the reviewers for providing precious feedback. The editorial support provided by Springer-Verlag is acknowledged. We hope that this book on evolutionary computing in system design will prove valuable to its readers. Lakhmi C. Jain University of South Australia Australia Vasile Palade Oxford University UK Dipti Srinivasan National University of Singapore Singapore

Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V 1 Introduction to Evolutionary Computing in System Design Lakhmi C. Jain, Shing Chiang Tan, Chee Peng Lim . . . . . . . . . . . . . . . . . .

1

2 Evolutionary Neuro-Fuzzy Systems and Applications G. Castellano, C. Castiello, A.M. Fanelli, L. Jain . . . . . . . . . . . . . . . . . . . . 11 3 Evolution of Fuzzy Controllers and Applications Dilip Kumar Pratihar, Nirmal Baran Hui . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 4 A Neuro-Genetic Framework for Multi-Classiﬁer Design: An Application to Promoter Recognition in DNA Sequences Romesh Ranawana and Vasile Palade . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 5 Evolutionary Grooming of Traﬃc in WDM Optical Networks Yong Xu and Kunhong Liu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 6 EPSO: Evolutionary Particle Swarms V. Miranda, Hrvoje Keko, Alvaro Jaramillo . . . . . . . . . . . . . . . . . . . . . . . . . 139 7 Design of Type-Reduction Strategies for Type-2 Fuzzy Logic Systems using Genetic Algorithms Woei-Wan Tan, Dongrui Wu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 8 Designing a Recurrent Neural Network-based Controller for Gyro-Mirror Line-of-Sight Stabilization System using an Artiﬁcial Immune Algorithm Ji Hua Ang, Chi Keong Goh, Eu Jin Teoh, and Kay Chen Tan . . . . . . . . 189 9 Distributed Problem Solving using Evolutionary Learning in Multi-Agent Systems Dipti Srinivasan, Min Chee Choy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

VIII

Contents

10 Evolutionary Computing within Grid Environment Ashutosh Tiwari, Gokop Goteng, and Rajkumar Roy . . . . . . . . . . . . . . . . . 229 11 Application of Evolutionary Game Theory to Wireless Mesh Networks Athanasios Vasilakos, Markos Anastasopoulos . . . . . . . . . . . . . . . . . . . . . . . 249 12 Applying Hybrid Multiobjective Evolutionary Algorithms to the Sailor Assignment Problem Deon Garrett, Dipankar Dasgupta, Joseph Vannucci, James Simien . . . . 269 13 Evolutionary Techniques Applied to Hardware Optimization Problems: Test and Veriﬁcation of Advanced Processors Ernesto Sanchez and Giovanni Squillero . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303

1 Introduction to Evolutionary Computing in System Design Lakhmi C. Jain1 , Shing Chiang Tan2 , Chee Peng Lim3 1

School of Electrical and Information Engineering University of South Australia, Australia 2 Faculty of Information Science and Technology Multimedia University, Malaysia 3 School of Electrical and Electronic Engineering University of Science Malaysia, Malaysia Summary. In this chapter, an introduction on the use of evolutionary computing techniques, which are considered as global optimization and search techniques inspired from biological evolutions, in the domain of system design is presented. A variety of evolutionary computing techniques are ﬁrst explained, and the motivations of using evolutionary computing techniques in tackling system design tasks are then discussed. In addition, a number of successful applications of evolutionary computing to system design tasks are described.

1.1 Introduction Evolutionary Computing (EC) [1] techniques have gained a lot of interests from researchers and practitioners in various ﬁelds as they provide robust and powerful adaptive search mechanisms that are useful for system design. In addition, the underlying biological concepts on which EC techniques are based on also contribute to their attractiveness in tackling system design problems. Although many diﬀerent EC techniques have been proposed over the years, they, in general, share a common idea, i.e., evolving a population of potential solutions through the process of selection, recombination, and mutation. The search process of EC can be interpreted as a collective learning process in a population. Such interpretation is in agreement with the claims of some evolutionary models, such as the Baldwin evolution [2], that put emphasis on the role of the evolutionary learning process in a population. On the other hand, the emergence of systems plays an important role in our daily life. According to [3], a system is deﬁned as a set of entities, real or abstract, comprising a whole where each component interacts with or is related to at least one other component and they all serve a common objective. Researches in various types of systems are widespread and interdisciplinary, ranging from information and computer science, social and cognitive sciences, L.C. Jain et al.: Introduction to Evolutionary Computing in System Design, Studies in Computational Intelligence (SCI) 66, 1–9 (2007) c Springer-Verlag Berlin Heidelberg 2007 www.springerlink.com

2

L.C. Jain et al.

to engineering and management research. For example, as stated in [3], in information and computer science, a system could also be a method or an algorithm. From the perspective of engineering, system engineering deals with the application of engineering techniques to form a structured development process that covers various stages of the process, from concept to production to operation and disposal. In social and cognitive sciences, systems comprise human brain functions and human mental processes as well as social/cultural behavioral patterns; while in management research, operations research, and organizational development, human organizations are viewed as conceptual systems of interacting components. In this chapter, an introduction to EC in system designs is presented. In sections 1.2 and 1.3, what is EC, and why EC is useful in system design tasks are explained. In section 1.4, some successful applications of EC to a variety of system design problems are described. A description on each chapter in this book is given in Section 1.5, and a summary of this chapter is included in Section 1.6.

1.2 What is Evolutionary Computing Evolutionary computing is referred to as computer methods that are inspired by biological evolutions [1]. From the literature, there are generally ﬁve basic types of EC, viz., genetic algorithms (GA), evolutionary programming (EP), evolution strategies (ES), genetic programming (GP), and learning classiﬁer systems (LCS). The basic idea behind all these EC variants is commonly the same from a unifying viewpoint. On availability of a population of individuals, natural selection, which is based on the principle of survival of the ﬁttest following the existence of environmental pressures, is exercised to choose individuals that could better ﬁt the environment. The operations involved are as follows. Given a ﬁtness function to be optimized, a set of candidate solutions is randomly generated. The candidate solutions are assessed by the ﬁtness function that indicates an abstract of ﬁtness measure. Based on the ﬁtness values, ﬁtter candidate solutions have a better chance to be selected to seed the next generation through the application of recombination (crossover) and/or mutation. On one hand, the recombination operator is applied to two or more parent candidate solutions and one or more new candidate solutions (oﬀspring) is/are produced. On the other hand, the mutation operator is applied to one parent candidate solution and one new candidate solution is produced. The new candidate solutions will compete among one another, based on the ﬁtness values, for a place in the next generation. Even though candidate solutions with higher ﬁtness values are favorable, in some selection schemes, the solution candidates with relatively low ﬁtness values are included in the next generation so as to maintain diversity of the population. The processes of selection, recombination, and mutation are repeated from one generation of population

1 Introduction to Evolutionary Computing in System Design

3

to another until a terminating criterion is satisﬁed, i.e., either a candidate solution with suﬃcient quality has been obtained or the algorithm has been executed for a predeﬁned number of generations. Developed by Holland [4], the GA is the most popular and widely used EC paradigm. The GA is essentially a class of population-based search strategies that utilize an iterative approach to perform a global search on the solution space of a given problem. The three most popular search operators used in the GA are selection, crossover, and mutation. The selection operator chooses a subset group of candidate solutions (from a population) that are subsequently used to generate new candidate solutions with the crossover and mutation operators. In this regard, the crossover operator is applied to combine genetic information of the existing candidate solutions whereas the mutation operator is applied to create variations of a single candidate solution. While conventional GA models utilize binary codes to keep information of a candidate solution (i.e., a chromosome), real-valued chromosomes are now widely used to solve more diﬃcult problems. Introduced by Fogel, Owens, and Walsh [5], EP was aimed to simulate intelligent behavior by means of ﬁnite-state machines. In this regard, candidate solutions to a problem are considered as a population of ﬁnite-state machines. New candidate solutions (oﬀspring) are generated from the existing candidate solutions (parents) through mutation. All candidate solutions are assessed by a ﬁtness function. The crossover operator is not used in EP. Instead, the mutation operator is employed to provide variations in the population of candidate solutions. ES [6, 7] was ﬁrst developed to optimize parameters for aero-technology devices. This method is based on the concept of the evolution of evolution. Each candidate solution in the population is formed by genetic building blocks and a set of strategy parameters that models the behavior of that candidate solution in its environment. Both genetic building blocks and strategy parameters participate in the evolution. The evolution of the genetic characteristics is governed by the strategy parameters that are also adapted from evolution. Like EP, ES uses only the mutation operator. Devised by Koza [8], GP was targeted at making the computer to solve problems without being explicitly programmed to do so. GP is an extension of conventional GA. However, the representation scheme of GP is diﬀerent from that of conventional GA; the former represents individuals as executable programs (i.e., trees) whereas the latter uses string representation. LCS [9] tackles the machine learning tasks as an evolutionary rule discovery component. The knowledge representation of LCS is a collection of production rules that are expressed in the if-then format. Each production rule, which is encoded as a binary string in Holland’s standard approach, is considered as a classiﬁer. The rules are updated according to some speciﬁc evolutionary procedure.

4

L.C. Jain et al.

1.3 Why Evolutionary Computing As explained in [10], one of the main advantages of EC is that it is conceptually simple and easy to use. An EC process normally covers random initialization of a population of candidate solutions, followed by some iterative variation and selection strategy, and by evaluation of candidate solutions in accordance with a performance index. This process is relatively simple, as compared with, for example, model-based approaches that deal with complicated mathematical formulation of the system under scrutiny. The applicability of EC is broad and comprehensive, covering most problems that can be cast as an optimization task. In contrast to some numerical approaches that require continuous values or constraint formulation, the search space of EC can be disjoint, and the representation can be intuitive to humans [10]. Besides, EC techniques can be used to adapt solutions to changing environments. Indeed, a wide range of disciplines, including computer science, mathematics, medicine, engineering, and economics, to name a few, has been explored extensively by capitalizing EC techniques. Some successful application examples of EC are presented in section 1.4. In solving complex, real-world problems, one may resort to methods that mimic the aﬃnity from the nature. In this regard, the biologically inspired EC is cohesive with the idea of designing solutions that exploit a few aspects of natural evolutionary processes. In addition, recent trends in EC have geared towards devising hybrid models with other conventional optimization (e.g. gradient-based) techniques as well as other artiﬁcial intelligence paradigms. These hybrid models normally are highly automated and can give high eﬃciency, as evidenced in the work presented in other chapters of this book. Another motivation of using EC stems from a technical standpoint. One trend from the growth of computer power has been the growing demand for problem-solving automation, which is vitally needed to cope with modern system design tasks of which their complexity is also ever increasing. EC techniques may turn out to be just as important because of the robustness and adaptation they can oﬀer when dealing with modern computer-based applications. They are versatile enough and can normally provide satisfactory solutions within acceptance time with the advent of fast, powerful computing machines.

1.4 Applications of Evolutionary Computing to System Design The applications of EC to system design are far reaching, especially from the perspective of meeting the requirements of software and hardware designs. A snapshot of successful applications of EC is presented, as follows. EC can be deployed to solve problems that require any kind of optimization. Examples of optimization problems include operations-research tasks

1 Introduction to Evolutionary Computing in System Design

5

such as scheduling (e.g., [11, 12] and design optimization (e.g., [13, 14]). EC-based methods have been proposed for identifying an optimal broadcasting strategy in urban areas [15]. Automatic test data generation problems [16] can be formulated as an optimization problem and be solved using EC. Classiﬁcation and regression tasks are one of the fundamental intelligent processes in data analysis. Of many approaches, artiﬁcial neural networks (ANN) models, which are inspired from the functionality of biological brains, are a feasible method for tackling classiﬁcation and regression tasks. Recently, various combinations between ANNs and EC have been investigated (e.g., [17]–[22]). The motivation of this synergistic machine learning lies in the retention of well-deﬁned identity of the two entities, for which ANNs provide a framework of accurate and exact computation whereas EC provides a robust and eﬃcient search, for better solution. Hybrid models combining EC and fuzzy systems (e.g., [23, 24]) as well as EC, ANNs, and fuzzy systems (e.g., [25–27]) have also been proposed, with the purpose of obtaining improved solutions. Evolutionary robotics is an active ﬁeld of research that utilizes EC to develop controllers for autonomous robots. In this regard, a population of candidate controllers is repeatedly evolved and tested with respect to a ﬁtness function. The most feasible controller is identiﬁed for achieving an eﬃcient robot learning application that could provide desired actions for the robot to perform. Successful applications of evolutionary robotics include research work by, to name a few, Paul et al. [28], Krasny and Orin [29], Reil and Husbands [30], Kodjabachian and Meyer [31], and Gallagher et al. [32]. Evolvable hardware is an emerging sub-specialty of EC that is aimed to automatically design and optimize electrical and mechanical structures, such as digital [33, 34]/analog [34–36] circuits, antennas [37], and robots [38–40] Research in this area has deployed various EC techniques including GA, GP, EP, adaptive GA [41], parallel GA [42]), compact GA [43], and Cartesian GP [44]. Recent advances in molecular biology and genomic technology have led to an explosive growth in the generation of biological data. Bioinformatics, as its name suggests, is a research area that utilizes computational methods to make discoveries from a huge biological database. The role of EC in bioinformatics, and data mining in general, is to assist in analyzing and, thus, extracting meaningful information from large data sets in a robust and computationally eﬃcient way. EC has been applied to handle a broad classiﬁcation of the various bioinformatics tasks, for instances, gene expression [45] and sequence alignment [46, 47], as well as for other data mining problems [48]. Other innovative application of EC to designing intelligent systems for solving a variety of practical problems include machine condition monitoring [49], medical diagnosis [50, 51], web-information search and retrieval [52, 53], and pedagogical educational learning system [54], as well as other problems in engineering and information systems [55]. In short, as computer technology continues to advance, the advantages of EC will become more profound, and

6

L.C. Jain et al.

EC techniques will continue to be adopted as a practical problem-solving tool in diverse ﬁelds of system design.

1.5 Chapters Included in this Book This book includes thirteen chapters. Chapter one introduces EC in system design. Chapter two by Castellano et al., is focused on evolutionary neuronfuzzy systems and their applications. A scheme using neural learning and genetic optimization to learn a fuzzy model is presented. Chapter three, by Pratihar and Hui, describes the evolution of fuzzy controllers and applications. Some applications of the reported techniques are presented. Chapter four by Ranawana and Palade, is on neuro-genetic techniques for designing multiclassiﬁer systems. An application for promoter recognition in DNA sequences is presented. Chapter ﬁve by Xu and Liu is on evolutionary grooming of traﬃc in WDM optical networks. Future challenges and research directions are also included. Chapter six by Miranda presents evolutionary particle swarm optimization as an evolutionary meta-heuristic that implements a scheme of self adaptive recombination. Chapter seven by Tan and Wu is on the design of computationally eﬃcient type-reduction strategies for type-2 fuzzy logic systems using genetic algorithm. Chapter eight by Ang et al. is on the design of neural network-based controller using an artiﬁcial immune system. Chapter nine by Srinivasan and Choy is on distributed problem solving using evolutionary learning in multi-agent systems. Chapter ten by Tiwari et al. reports their work on evolutionary computing in grid-based environment. Chapter eleven by Vasilakos and Anastasopoulos is on the application of evolutionary game theory in wireless mesh networks. Chapter twelve by Garrett el al. is on the application of hybrid multiobjective evolutionary algorithms to the sailor assignment problem. The ﬁnal chapter by Sanchez and Squillero presents the application of evolutionary computing techniques to hardware optimization problems related to the test and veriﬁcation of advanced processors.

1.6 Conclusion This chapter has presented an introduction to EC in handling system design problems. A variety of EC techniques, which include GA, EP, ES, GP, and LCS, have been described. The advantages of using EC techniques, especially for tackling optimization and search problems, as well as some successful applications of EC to diﬀerent system design problems have also been elaborated. It is envisaged that EC techniques will ultimately play a signiﬁcant role in solving practical system design tasks, especially those complex, nonlinear systems in various domains including computer and information sciences, engineering,

1 Introduction to Evolutionary Computing in System Design

7

management, and social and cognitive sciences, that have become more and more common in today’s knowledge-based society.

References 1. Dumitrescu, D., Lazzerini, B., Jain, L.C., Dumitrescu, A., (2000), Evolutionary Computation, CRC Press, Boca Raton, FL, USA. 2. Suzuki, R., Arita, T., (2007), The dynamic changes in roles of learning through the Baldwin eﬀect, Artiﬁcial Life, 13, 31–43. 3. http://en.wikipedia.org/wiki/System 4. Holland, J.H., (1962), Outline for a logical theory of adaptive systems, J. ACM, 3, 297–314. 5. Fogel, L.J., Owens, A.J., Walsh, M.J., (1966), Artiﬁcial Intelligence through Simulated Evolution, John Wiley, New York. 6. Rechenberg, I., (1994), Evolutionary strategy, Computational Intelligence: Imitating Life, Zurada, J.M., Marks II, R., Robinson, C., (Eds.), IEEE Press, 147– 159. 7. Schwefel, H.-P., (1981), Numerical Optimization of Computer Models, John Wiley, Chichester, UK. 8. Koza, J.R., (1992), Genetic Programming, MIT Press, Cambridge, MA. 9. Holland, J.H., (1986), Escaping brittleness: The possibility of general-purpose learning algorithms applied to parallel rule-based systems, Mach. Learn., 2, Michalski, R.S., Carbonell, J.G., Mitchell, T.M., (Eds.), Morgan Kaufmann, Los Altos, CA, 593–624. 10. Fogel, D.B. (1997), The advantages of evolutionary computation, BioComputing and Emergent Computation, Lundh, D., Olsson, B., Narayanan A., (Eds.), Skve, Sweden, World Scientiﬁc Press, Singapore, 1–11. 11. Lee, L.H., Lee, C.U., Tan, Y.P., (2007), A multi-objective genetic algorithm for robust ﬂight scheduling using simulation, Eur. J. Oper. Res., 177, 1948–1968. 12. Jozefowska, J., Mika, M., Rozycki, R., Waligora, G., Weglarz, J., (2002), A heuristic approach to allocating the continuous resource in discrete-continuous scheduling problems to minimize the makespan, J. Scheduling, 5, 487–499. 13. Greiner, H., (1996), Robust optical coating design with evolution strategies, Appl. Opt., 35, 5477–5483. 14. Wiesmann, D., Hammel, U., B¨ ack, T., (1998), Robust design of multiplayer optical coatings by means of evolutionary algorithms, IEEE Trans. Evol. Comput., 2, 162–167. 15. Alba, E., Dorronsoro, B., Luna, F., Nebro, A.J., Bouvry, P., Hogie, L., (2007), A cellular multi-objective genetic algorithm for optimal broadcasting strategy in metropolitan MANETs, Comput. Commun., 30, 685–697. 16. Celso, C. Ribeiro, C.C., Martins, S.L., Rosseti, I., (2007), Metaheuristics for optimization problems in computer communications, Comput. Commun., 30, 656–669. 17. Han, S.-J., Cho S.-B., (2006), Evolutionary neural networks for anomaly detection based on the behavior of a program, IEEE Trans. Syst., Man Cybernet.— Part B, 36, 559–570. 18. Fieldsend, J.E., Singh, S., (2005), Pareto Evolutionary Neural Networks, IEEE Trans. Neural Netw.,16, 338–354.

8

L.C. Jain et al.

19. Bonissone, P.P., Chen, Y.-T., Goebel, K., Khedkar, P.S, (1999), Hybrid soft computing systems: industrial and commercial applications, Proc. IEEE, 87, pp. 1641–1667. 20. Yao, X., (1999), Evolving artiﬁcial neural networks, Proc. IEEE, 87, 1423–1447. 21. Vonk, E., Jain, L.C., and Johnson, R.P., (1997), Automatic Generation of Neural Networks Architecture Using Evolutionary Computing, World Scientiﬁc Publishing Company, Singapore. 22. Van Rooij, A., Jain, L.C., and Johnson, R.P., (1996), Neural Network Training Using Genetic Algorithms, World Scientiﬁc Publishing Company, Singapore. 23. Bonissone, P.P., Subbu, R., Eklund, N., Kiehl, T.R., (2006), Evolutionary algorithms + domain knowledge = real-world evolutionary computation, IEEE Trans. Evol. Comput., 10, 256–280. 24. Hoﬀmann, F., (2001), Evolutionary algorithms for fuzzy control system design, Proc. IEEE, 89, 1318–1333. 25. Lin, C.J., Xu, Y.-J., (2006), A self-adaptive neural fuzzy network with groupbased symbiotic evolution and its prediction applications, Fuzzy Set Syst., 157, 1036–1056. 26. Oh, S.-K., Pedrycz, W., (2005), A new approach to self-organizing fuzzy polynomial neural networks guided by genetic optimization, Phys. Lett. A, 345, 88–100. 27. Jain, L.C. and Martin, N.M. (Eds), (1999), Fusion of Neural Networks, Fuzzy Logic and Evolutionary Computing and their Applications, CRC Press USA. 28. Paul, C., Valero-Cuevas, F.J., Lipson, H., 2006, Design and Control of Tensegrity Robots for Locomotion, IEEE Trans. Robot., 22, 944–957. 29. Krasny, D.P., Orin, D.E., (2004), Generating High-Speed Dynamic Running Gaits in a Quadruped Robot Using an Evolutionary Search, IEEE Trans. Syst., Man Cybernet.—Part B, 34, 1685–1696. 30. Reil, T., Husbands, P., (2002), Evolution of Central Pattern Generators for Bipedal Walking in a Real-Time Physics Environment, IEEE Trans. Evol. Comput., 6, 159–168. 31. Kodjabachian, J., Meyer, J.-A., (1998), Evolution and development of neural networks controlling locomotion, gradient following and obstacle avoidance in artiﬁcial insects, IEEE Trans. Neural Netw., 9, 796–812. 32. Gallagher, J., Beer, R., Espenschiel, K., Quinn, R., (1996), Application of evolved locomotion controllers to a hexapod robot, Robot. Autonomous Syst., 19, 95–103. 33. Hartmann, M., Haddow, P.C. (2004), Evolution of fault-tolerant and noiserobust digital designs, Proc. Inst. Elect Eng.—Comput. Digit. Tech., 151, 287–294. 34. Higuchi, T., Iwata, M., Keymeulen, D., Sakanashi, H., Murakawa, H., Iajitani, I., Takahashi, E., Toda, K.,Salami, M., Kajihara, N., Oesu, N., (1999), Real-world applications of analog and digital evolvable hardware, IEEE Trans. Evol. Comput., 220–235. 35. Lohn, J.D., (1999), Experiments on evolving software models of analog circuits, Commun. ACM, 42, 67–69. 36. Keymeulen, D., Zebulum, R.S., Jin, Y., Stoica, A., (2000), Fault-tolerant evolvable hardware using ﬁeld-programmable transistor arrays, IEEE Trans. Rel., 49, 305–316. 37. Hum, S.V., Okoniewski, M., Davies, R.J., (2005), An evolvable antenna platform based on reconﬁgurable reﬂect arrays, Proc. NASA/DoD Conf. Evolvable Hardware, Washington, DC, 139–146.

1 Introduction to Evolutionary Computing in System Design

9

38. Lohn, J.D., Hornby, G.S., (2006), Evolvable hardware: using evolutionary computation to design and optimize hardware systems, IEEE Computational Intelligence Magazine, 1, 19–27. 39. Terrile, R.J., Aghazarian, H., Ferguson, M.I., Fink, W., Huntsberger, T.L., Keymeulen, D., Klimeck, G., Kordon, M.A., Seungwon, L., Allmen, P.V., (2005), Evolutionary computation technologies for the automated design of space systems, Proc. NASA/DoD Conf. Evolvable Hardware, Washington, DC, 131–138. 40. Yao, X., Higuchi, T., (1999), Promises and challenges of evolvable hardware, IEEE Trans. Syst., Man Cybernet.—Part C, 29, 87–97. 41. Ko, M.-S., Kang, T.-W., Hwang, C.-S., (1997), Function optimization using an adaptive crossover operator based on locality, Eng. Appl. Artif. Intell., 10, 519–524. 42. Alba, E., Tomassini, M., (2002), Parallelism and evolutionary algorithms, IEEE Trans. Evol. Comput., 6, 443–462. 43. Harik, G., Lobo, F., Goldberg, D., (1999), The compact genetic algorithm, IEEE Trans. Evol. Comput., 3, 287–297. 44. Miller, J., (1999), An empirical study of the eﬃciency of learning Boolean functions using a Cartesian genetic programming approach, Proc. Genetic Evol. Comput. Conf., Orlando, FL, 1, 1135–1142. 45. Tsai, H.-K., Yang, J.-M., Tsai, Y.-F., Kao, C.-Y., (2004), An evolutionary approach for gene expression patterns, IEEE Trans. Inf. Technol. Biol., 8, 69–78. 46. Hung, C.-M., Huang, Y.-M., Chang, M.-S., (2006), Alignment using genetic programming with causal trees for identiﬁcation of protein functions, Nonlinear Analysis, 65, 1070–1093. 47. Ngom, A., (2006), Parallel evolution strategy on grids for the protein threading problem, J. Parallel Distrib. Comput., 66, 1489–1502. 48. Ghosh, A. and Jain, L.C. (Eds) (2005), Evolutionary Computation in Data Mining, Springer, Germany. 49. Guo, H., Jack, L.B., Nandi, A.K., (2005), Feature generation using genetic programming with application to fault classiﬁcation, IEEE Trans. Syst., Man Cybernet.—Part B, 35, 89–99. 50. Toro, F.D., Ros E., Mota, S., Ortega, J., (2006), Evolutionary algorithms for multiobjective and multimodal optimization of diagnostic schemes, IEEE Trans. Bio-Med. Eng., 53, 178–189. 51. Freitas, H.S., Bojarczuk, A.A., Lopes, C.C., (2000), Genetic programming for knowledge discovery in chest-pain diagnosis, IEEE Eng. Med. Biol. Mag., 19, 38–44. 52. Gordon, M., Fan W.-G., Pathak, P., (2006), Adaptive web search: evolving a program that ﬁnds information, IEEE Intell. Syst., 21, 72–77. 53. Kuo, R.J., Liao, J.L., Tu, C., (2005), Integration of ART2 neural network and genetic K-means algorithm for analyzing web browsing paths in electronic commerce, Decis. Support Syst., 40, 355–374. 54. Bode’n, M., Bode’n, M., (2007), Evolving spelling exercises to suit individual student needs, Applied Soft Computing, 7, 126–135. 55. Jain, L.C. (Ed.), (2000), Evolution of Engineering and Information Systems, CRC Press USA.

2 Evolutionary Neuro-Fuzzy Systems and Applications G. Castellano1 , C. Castiello1 , A.M. Fanelli1 , and L. Jain2 1

2

Computer Science Department, University of Bari, Via Orabona 4, 70126 Bari, Italy [castellano, castiello, fanelli]@di.uniba.it School of Electrical and Information Engineering, University of South Australia, Adelaide, Mawson Lakes Campus, South Australia SA 5095, Australia [email protected]

Summary. In recent years, the use of hybrid Soft Computing methods has shown that in various applications the synergism of several techniques is superior to a single technique. For example, the use of a neural fuzzy system and an evolutionary fuzzy system hybridises the approximate reasoning mechanism of fuzzy systems with the learning capabilities of neural networks and evolutionary algorithms. Evolutionary neural systems hybridise the neurocomputing approach with the solution-searching ability of evolutionary computing. Such hybrid methodologies retain limitations that can be overcome with full integration of the three basic Soft Computing paradigms, and this leads to evolutionary neural fuzzy systems. The objective of this chapter is to provide an account of hybrid Soft Computing systems, with special attention to the combined use of evolutionary algorithms and neural networks in order to endow fuzzy systems with learning and adaptive capabilities. After an introduction to basic Soft Computing paradigms, the various forms of hybridisation are considered, which results in evolutionary neural fuzzy systems. The chapter also describes a particular approach that jointly uses neural learning and genetic optimisation to learn a fuzzy model from the given data and to optimise it for accuracy and interpretability.

2.1 Introduction Hybridisation of intelligent systems is a promising research ﬁeld in the area of Soft Computing. It arises from an awareness that a combined approach may be necessary to eﬃciently solve real-world problems. Each of the basic Soft Computing paradigms has advantages and disadvantages. Fuzzy systems do not have the automatic learning of neural networks, hence they can not be used in the absence of expert knowledge. Fuzzy systems have an advantage over neural networks and evolutionary strategies, as they can express knowledge in a linguistic form. This is familiar and forms part of the human reasoning process. Neural networks are developed speciﬁcally for learning, and are fundamental when a set of signiﬁcant examples of the G. Castellano et al.: Evolutionary Neuro-Fuzzy Systems and Applications, Studies in Computational Intelligence (SCI) 66, 11–45 (2007) c Springer-Verlag Berlin Heidelberg 2007 www.springerlink.com

12

G. Castellano et al.

problem to be solved are available, rather than a solution algorithm for the problem. In general, neural networks can learn from examples. The way in which a knowledge base is extracted through a learning process is not easy for humans to understand. Evolutionary strategies, despite the learning speed being slower than neural networks, are able to use much more general functions than diﬀerentiable ones. This is because they do not require the computation of the gradient of the functions. Finally, as evolutionary algorithms explore in several directions the search space, they are less aﬀected than neural learning by the problem of ﬁnding local minima. A hybrid technique which makes use of a combination of the three Soft Computing paradigms is an interesting prospect. When these constituents are combined, they operate synergistically rather than competitively. Their mutual dependence may produce unexpected performance improvement. In the last decade, several Soft Computing frameworks have been developed for a wide range of domains. Many of them solve a computational task by using a combination of diﬀerent methodologies. The aim is to overcome the limitations and weakness of the several techniques. This chapter gives an overview of diﬀerent viewpoints of hybridisation among Soft Computing paradigms. The main features of principal Soft Computing paradigms are introduced. The attention is focused on all the possible ways of integrating the characteristics of two paradigms. This results in neural fuzzy systems, which is a hybrid of approximate reasoning method of fuzzy systems with the learning capabilities of neural networks, evolutionary fuzzy systems that use evolutionary algorithms to adapt fuzzy systems, and evolutionary neural systems, integrating the solution-searching ability of evolutionary computing with the neurocomputing approach. In conclusion, full hybridisation of the three paradigms is addressed. We summarise the research done on evolutionary neuro-fuzzy systems that integrate the solution-searching feature of evolutionary computing and the learning ability of neurocomputing with the explicit knowledge representation provided by fuzzy computing. An example of full integration between Soft Computing paradigms is presented that makes use of neural learning and genetic algorithms to learn fuzzy models from data. This improves the accuracy and the ease of interpretation. The chapter is organised as follows. Section 2.2 gives the fundamentals on the three main Soft Computing paradigms. That is, neural networks, fuzzy inference systems and evolutionary algorithms. Sections 2.3 to 2.5 provide a review of all possible forms of combination among the paradigms. In section 2.6 diﬀerent examples of evolutionary-neuro-fuzzy hybridisation are considered, and some applications are reported. Section 2.7 presents a hybrid approach for the optimisation of neuro-fuzzy models based on an evolutionary algorithm. An example of the application is included in this section. Section 2.8 concludes the chapter.

2 Evolutionary Neuro-Fuzzy Systems and Applications

13

2.2 Soft Computing Systems The term Soft Computing [193] indicates a number of methodologies used to ﬁnd approximate solutions for real-world problems which contain various kinds of inaccuracies and uncertainties. The guiding principle of Soft Computing is to develop a tolerance for imprecision, uncertainty, partial truth, and approximation to achieve tractability, robustness and low solution cost. The underlying paradigms of Soft Computing are Neural Computing, Fuzzy Logic Computing and Evolutionary Computing. Systems based on such paradigms are Artiﬁcial Neural Networks (ANN’s), Fuzzy Systems (FS’s), and Evolutionary Algorithms (EA’s). Rather than a collection of diﬀerent paradigms, Soft Computing is better regarded as a partnership in which each of the partners provides a methodology for addressing problems in a diﬀerent manner. From this perspective, the Soft Computing methodologies are complementary rather than competitive. This relationship enables the creation of hybrid computing schemes which use neural networks, fuzzy systems and evolutionary algorithms in combination. Figure 2.1 shows their use in a Soft Computing investigation. In the following, after a description of the three basic Soft Computing paradigms, some hybrid Soft Computing systems are overviewed. These show the rapid growth in the number and variety of combinations using ANN’s, FS’s and EA’s. 2.2.1 Neural Networks Artiﬁcial Neural Networks (ANN’s) are computational models that are loosely modelled on biological systems and exhibit in a particular way some of the

Fig. 2.1. Integration possibilities in Soft Computing

14

G. Castellano et al.

Fig. 2.2. The structure of three layers ANN

properties of the brain. They are composed by a number of simple processors (neurons) working in parallel, without any centralised control. The neurons are arranged in a particular structure which is usually organised in layers. A system of weighted connections determines the information ﬂow through the network. In ﬁgure 2.2 a neural network is depicted with two layers, plus an input layer. The behaviour of an ANN is determined by the topology of the connections and by the properties of every processing unit, which typically evaluates a mathematical function. The numeric weights of the connections are modiﬁed in order to give ANN’s a dynamic nature. ANN’s possess also adaptive properties, since they are able to modify their output depending on the actual weight. The weight value is dependent on past experience. The learning of an ANN is based on the analysis of input data (training set) and the particular learning algorithms. The knowledge is eventually embedded into the weight conﬁguration of the network. Neural networks are commonly regarded as learning machines that work on the basis of empirical data. The only means of acquiring knowledge about the world in a connectionist system comes from observational instances. There are no a priori conceptual patterns that could lead to a learning process. The neurons composing the network only perform a very simple task. This consists of evaluating a mathematical function and transmitting a signal. This is similar to the neuronal cells of biological brains. The aim is to adapt to the surrounding data from the system. This rather simpliﬁed structure enables ANN’s to perform quite complex tasks, endowing connectionist systems with the capability of approximating continuous functions to any required degree of accuracy. The task of a neural network is to realise a mapping of the type: φ : x ∈ Rn → y ∈ Rm ,

(2.1)

where x is a n-dimensional input vector and y is an m-dimensional output vector. ANN’s can be variously distinguished in terms of their conﬁgurations, topology and learning algorithms. Generally speaking, two main types of

2 Evolutionary Neuro-Fuzzy Systems and Applications

15

conﬁgurations can be identiﬁed: feedforward networks, these possess connections which propagate the information ﬂow in one direction only. The other type is recurrent networks, whose nodes may have feedback connections, where the input for a neuron may be the output from another neuron. Here the input may be from the particular neuron in question. Each connectionist system is characterised by a particular topology. That is, the organisation of its own neurons. The most typical neural architecture is represented by the multilayered network, where the processing units are fully interconnected in three or more layers and the output is interpolated among all the nodes of the network. On the other hand, in competitive networks the nodes compete to respond to inputs. A particular scenario may have some units that are active, while all others are inhibited. Finally, learning algorithms play an important role in characterising the working mechanism of an ANN. The most commonly employed learning algorithm is the error back-propagation training. This involves adjusting the connection weights under a supervised learning scheme. Other types of learning algorithms include counter propagation mechanisms, Hebbian learning, Boltzmann algorithm and self-organising maps. Further details concerning the characterisation of neural networks, may be obtained from references which include [7, 15, 71, 75, 149]. 2.2.2 Fuzzy Systems Fuzzy mechanisms suitable for tasks involving reasoning have been proposed as an extension to classical formal logic. They were ﬁrst introduced in set theory. The concept of a “fuzzy set” has been employed to extend classical sets, which are characterised by crisp boundaries. This addition permits a degree of ﬂexibility for each object belonging to a particular set. This quality is realised by the deﬁnition of membership functions that give fuzzy sets the capacity of modelling linguistic, vague expression [192]. Fuzzy sets constitute the basis for fuzzy logic, a novel way for developing reasoning models by handling imprecise information, where truth can assume a continuum of values between 0 and 1. This kind of information is often referred to as fuzziness: it should be noted that fuzziness does not come from randomness, but from the uncertain and imprecise nature of abstract thoughts and concepts. Fuzzy reasoning realises a form of approximate reasoning that, using particular mathematical inferences, derives conclusions based on a set of fuzzy IF-THEN rules, where linguistic variables could be involved. In this way, fuzzy logic is suitable for describing the behaviour of systems which are either too complex or too illdeﬁned to be amenable to precise mathematical analysis. Classical systems cannot cope with inexact or incomplete information, because they do not provide any means of representing imprecise propositions and do not possess any mechanism that can make an inference from such propositions. Fuzzy logic systems commonly contain expert IF-THEN rules and can be characterised in terms of their fundamental constituents: fuzziﬁcation, rule base, inference, defuzziﬁcation. Figure 2.3 is a schematic representation of such

16

G. Castellano et al.

Fig. 2.3. The basic components of a fuzzy system

a fuzzy system. Fuzziﬁcation is a mapping from a crisp input space to fuzzy sets in a deﬁned universe: U : xi ∈ R → X ∈ U ⊂ Rq .

(2.2)

Here xi represents a crisp value and q is the number of fuzzy classes. The fuzzy sets are characterised by membership functions which portray the degree of belonging of xi to the values in U, µF (xi ) : U → [0, 1]. The rule base is constituted by an ensemble of fuzzy rules and the knowledge is expressed in the following form: IF x1 is Ak1 AND . . . AND xn is Akn THEN y1 is bk1 AND . . . AND ym is bkm ,

(2.3)

where the index k = 1, . . . , K indicates the k-th rule among the K rules in the rule base; Aki and bkj are fuzzy sets. These are deﬁned over the input components xi , i = 1, . . . , n, and the output components yj , j = 1, . . . , m, respectively. The rule is a fuzzy implication that is usually represented by a Cartesian product of the membership functions of antecedents and consequents. The fuzzy inference process can be described by starting with the deﬁnition of the membership functions µik (·) related to the k-th fuzzy rule and evaluated for each input component of a sample vector x = (x1 , . . . , xn ). The most commonly employed membership functions are the triangular and the Gaussian functions. The values obtained by the fuzziﬁcation contribute to the AND conjunction of each rule, which is interpreted by a particular T-norm. This is most commonly the min or the algebraic product operators. After evaluating the degrees of satisfaction for the entire set of fuzzy rules, the K activation strengths are used in the OR disjunction represented by the alternative rules. These are interpreted by a particular S-norm, the max or the algebraic sum operators are most commonly used. Finally, the defuzziﬁcation process is used to reconvert the fuzzy output values, deriving from the inference mechanism, into crisp values. These can

2 Evolutionary Neuro-Fuzzy Systems and Applications

17

then be eventually employed in diﬀerent contexts. The most common strategy for defuzziﬁcation is to use the centre of area method which gives the centre of gravity of the output membership function. Further details concerning the characteristics of fuzzy systems can be found in [22, 43, 147]. 2.2.3 Evolutionary Algorithms “Evolutionary Algorithms” is a general term for indicating a number of computational strategies that are based on the principle of evolution. Populationbased generate-and-test algorithms are used. An evolutionary algorithm is an iterative probabilistic program, employed in optimisation problems. A population of individuals, P (t) = {pt1 , . . . , ptn } is involved for each iteration t [52, 79, 115]. Each individual pti represents a potential solution for the problem and is represented as some data structure S. The metaphor underlying evolutionary computation enables us to review this encoding process as a translation from a phenotype representation to chromosomal representation or genotype encoding. The population members are successively evaluated in terms of a particular function. This function is speciﬁcally designed for the problem, in order to express the ﬁtness of the individuals. In this way, at iteration t + 1 a new population can be generated on the basis of the ﬁttest elements. These are selected according to some preﬁxed rule to produce an oﬀspring by means of “genetic” operators. They are usually represented by unary transformations, mi : S → S (mutation), and higher order transformations, cj : S × . . . × S → S called crossover, or recombination. The successive generations of several populations hopefully converge to a ﬁnal oﬀspring encoding an optimal solution. In ﬁgure 2.4 the general scheme of an evolutionary algorithm is shown. The most common evolutionary computation techniques are genetic algorithms, so that the terms “Evolutionary Computation” and “GA-based methods” are used almost as synonyms. However, diﬀerent evolutionary techniques should be distinguished. They can be characterised in terms of the data structures S employed to represent the chromosomes, the genetic operators and the individual selection schemes involved. In particular, Genetic Algorithms [59, 79, 115] codify chromosomes as binary strings. Genetic operators are applied only to the genotypical representations where the role of recombination is emphasised. Evolutionary Programming (EP) [52, 53] is a stochastic optimisation strategy similar to GA’s, that gives emphasis on the behavioural linkage between parents and their oﬀspring, rather than seeking to emulate speciﬁc genetic operators as observed in nature. The only genetic operator used in EP is mutation, applied to the phenotypical representations. Similarly to EP, Evolution Strategies [155, 156] make use of real-valued vectors to represent individuals. An individual comprises not only the speciﬁc object parameter set and its ﬁtness function value, but usually also a set of endogenous (i.e. evolvable) strategy parameters that are used to control

18

G. Castellano et al.

Fig. 2.4. The general structure of an evolutionary program

certain statistical properties of the genetic operators. Another evolutionary approach consists in Genetic Programming [5, 99, 101–104], where structures of possible computer programs (trees or even neural networks) evolve to represent the best solution to a problem.

2.3 Neuro-Fuzzy Systems The neuro-fuzzy hybridisation represents by far the most fruitful and the most investigated strategy of integration in the context of Soft Computing. Both neural networks and fuzzy systems are dynamical, parallel processing systems that estimate input-output functions [118]. Fuzzy logic is capable of modelling vagueness, handling uncertainty and supporting human-type reasoning. Connectionist systems rely on numerical and distributed representation of knowledge and they are often regarded as black boxes, revealing little about their inductive engines. Also in terms of application requirements fuzzy logic and neural systems show very contrasting positions. Neural networks are capable of learning from scratch, without needing any a-priori intervention, provided that suﬃcient data are available or measurable. On the other hand, fuzzy systems make use of linguistic knowledge of the process, which can be supplied by human experts. To a large extent, the key-points and the shortcomings of connectionist and fuzzy approaches appear to be complementary. It seems to be a natural practice to build up integrated strategies combining the concepts of the two paradigms [89,94,100,111,122,130,135]. It should be observed that neuro-fuzzy systems are the most prominent representatives of hybridisations in terms of

2 Evolutionary Neuro-Fuzzy Systems and Applications

19

the number of practical implementations. A number of successful applications can be observed in engineering and industrial applications [19]. An inspection of the multitude of hybridisation strategies proposed in literature which involve neural networks and fuzzy logic would be somewhat impractical. It is however straightforward to indicate the general lines underlying this kind of integration. This appears to be twofold. If we consider the fuzzy inference system as the main subject of the hybridisation, neural networks can add learning capabilities to an inference engine which reveals to be spoilt by some self-constructing feature. This approach is referred to as NeuroFuzzy System (NFS). However, when a connectionist system is regarded as the main subject of the hybridisation, fuzzy logic may assist by incorporating fuzziness into the neural framework. This may enable a better view of the black box. This argument is commonly referred to as Fuzzy-Neural System (FNS). In the following we will brieﬂy review some of the most prominent examples of neuro-fuzzy hybridisation proposed in literature. 2.3.1 Fuzzy-Neural Systems Fuzziness could be incorporated into neural networks at diﬀerent levels. Examples are given in literature concerning the fuzziﬁcation of input data, output results, learning procedures, error functions [69, 82, 95, 117, 119, 120, 129, 175]. One of the most intriguing examples of fuzzy-neural hybridisation is investigated by a ﬁeld of research which constructs systems based on fuzzy neurons. These speciﬁc neurons are designed to realise the common operations of fuzzy set theory. That is, fuzzy union, intersection, aggregation. This is instead of the usual standard algebraic functions. Using this technique it may be possible to take advantage of the transparency and readability of the structures. This may result in improved interpretation for the overall system [76, 132, 136]. 2.3.2 Neuro-Fuzzy Systems A number of results have been presented to demonstrate the formal equivalence between fuzzy rule-based systems and neural networks [11,20,21,70]. In particular, a functional equivalence has been shown to exist between radial basis function networks and fuzzy systems [88]. On the basis of this correspondence, various neuro-fuzzy integrations have been proposed to endow fuzzy inference engines with learning capabilities using the neural components. In practice, connectionist learning can be exploited to tune the parameters of an existing fuzzy system and/or to compile the structure of the rule base [34, 36, 50, 92, 96, 128, 166, 168, 170]. Among the most popular and pioneering neuro-fuzzy systems can be cited: GARIC [12], NEFCON [123], ANFIS [87]. All of these are models for successive hybridisation strategies [26, 98, 124–126, 167, 169].

20

G. Castellano et al.

2.4 Evolutionary-Fuzzy Systems The eﬀectiveness of evolutionary techniques in performing complex search processes over spaces has been successfully applied to provide learning capabilities to fuzzy systems. This kind of hybridisation led to the development of genetic fuzzy systems (GFS’s) [6, 42, 134, 152]. A GFS is therefore represented by a fuzzy rule based system which relies on genetic components to accomplish the main processes related to the system design: parameter optimisation and rule generation. Actually, the optimisation process usually concerns the tuning procedure applied over the variables of a fuzzy system. This may be already established in the structure. This conﬁgures the adaptation process. A more proper learning approach is realised when the genetic components are involved in the generation of the rule base, without reference to any preexisting structure. It should be noted that, without loss of generality, we refer to the employment of genetic components. This is because genetic algorithms appear to be the most applied techniques inside this ﬁeld of research. The genetic search processes involved with GFS’s are with a Knowledge Base constituted of: – –

a Data Base, including the deﬁnitions of the parameters and variables of the fuzzy system; a Rule Base, comprising of a number of fuzzy rules.

As previously observed, the exploration of the diﬀerent components of the Knowledge Base is correlated with the diﬀerent objectives that can be attained. These range from parameter adaptation to pure learning activity. It should be observed that, similarly to any other search process, the application of evolutionary techniques to the Knowledge Base is subject to a trade-oﬀ between the dimension of the search space and the eﬃciency of the search. In this way, the smaller the search space will likely lead to a faster process yielding suboptimal solutions. A larger search space is more likely to contain optimal solutions, even if the search process would be less eﬃcient. For the purposes of our presentation, we are ﬁrst going to describe the common applications of genetic procedures when used to create fuzzy rule-based systems, distinguishing diﬀerent genetic adaptation and learning processes, and then some other kinds of hybridisation strategies proposed in literature. 2.4.1 Adaptation of the GFS Data Base The tuning process applied to the Data Base of a GFS aims to optimise the membership functions in the fuzzy rule base. This is commonly predeﬁned for the system. Each individual may encode within its chromosomes the parameters of the Data Base. This codiﬁes the particular shape of the membership functions or even the entire fuzzy partitions used to design the fuzzy rule base. This is especially true when dealing with descriptive fuzzy systems which involve linguistic

2 Evolutionary Neuro-Fuzzy Systems and Applications

21

variables. A number of references are addressed in literature as examples of this approach, and both real-valued and binary-valued chromosomes have been adopted [40, 58, 65, 68, 73, 159]. 2.4.2 Learning of the GFS Rule Base The employment of genetic learning for the generation of fuzzy rules has been traditionally proposed in literature. There are three main approaches and these are the Pittsburgh [163], the Michigan [80] and the Iterative Rule Learning [173]. In the Pittsburgh approach an entire fuzzy rule base is encoded as a chromosome, and thus is one of the individuals of the candidate population. The Michigan approach, on the other hand, codiﬁes each individual rule in a chromosome, so that the population represents the fuzzy rule base. The IRL approach encodes separately each rule and builds up the fuzzy rule base in an iterative fashion. This is done by adding a rule at every application of the genetic algorithm. In every case, when we consider approaches devoted to GFS Rule Base learning, we should be concerned with a particular scenario where the Data Base or the set of the involved membership functions is pre-conﬁgurated. The most common way to codify a rule is to represent it in the disjunctive normal form. A set of rules is generally encoded as a list. In this way, the chromosomes can be given by the code of a single rule or by the assembly of several rules. Bibliographical references can be classiﬁed as belonging to the Pittsburgh approach [78,141], the Michigan approach [16,84] and the IRL approach [39, 61]. 2.4.3 Learning of the GFS Knowledge Base The genetic component can be employed to determine the entire Knowledge Base pertaining to a GFS. Here both the parameter adaptation and the rule base learning are assessed by the genetic search process. The fundamental strategies introduced in the previous section can be restated to characterise the approaches in literature. They distinguish between the Pittsburgh-based [8, 23], the Michigan-based [131, 172] and the IRL-based [40, 41] learning methodologies. 2.4.4 Diﬀerent Hybridisation Strategies In order to complete the presentation, we mention a number of hybridisation strategies which integrate evolutionary techniques and fuzzy logic diﬀerently to the genetic adaptation and learning processes discussed in the previous sections. First of all, it is pointed out how diﬀerent approaches, other than genetic algorithms, have been employed to evolve fuzzy rule bases. In particular,

22

G. Castellano et al.

Genetic Programming can be utilised for this purpose, provided that the fuzzy rules are codiﬁed by coherent computer program structures, such as syntactic trees [2, 9, 35, 57, 77]. Recent sophisticated versions of genetic algorithms have been successfully adopted, and these include parallel GA’s and coevolutionary algorithms [33, 45, 138, 143, 146]. Evolutionary algorithms have been also applied to solve peculiar questions connected with the management of a fuzzy rule base. It must be mentioned the employment of genetic techniques in high-dimensional problems to perform the selection of fuzzy rules [60,83,148], and the selection of features [24,62,108]. Recent works concern the growing interest in issues of comprehensibility. They should be regarded with the same attention as issues of accuracy arising from the application of fuzzy systems [83, 85, 91]. In a diﬀerent way, GA’s have found application in conjunction with fuzzy logic to improve the capabilities of clustering algorithms such as fuzzy C-means [14]. In particular, genetic strategies can be adopted to optimise the parameters of a FCM-type algorithm [66, 171], used to deﬁne the distance norm [191] and to directly produce the clustering results [18]. We point out that the evolutionary-fuzzy hybridisation can be analysed in the reverse direction, with respect to that so far considered. This is done even if it appears to be less proﬁtable. Instead of considering the learning capabilities that genetic strategies could add to fuzzy system, fuzzy evolutionary algorithms focus on the management of GA parameters such as mutation rate or population size, using fuzzy systems [72, 137, 165]. To conclude our discussion, two bibliographical references assess the application possibility of GFS’s [17, 46].

2.5 Evolutionary-Neural Systems The combination of evolutionary algorithms and neural networks has led to the development of the so-called “Evolutionary Artiﬁcial Neural Networks” (EANN’s) [181]. Using this term it is common to indicate a particular class of connectionist systems having augmented capabilities of adaptation to the external environment. If we regard the learning capabilities of a neural network as a form of adaptation based on the basis of data investigation, EANN’s are able to show enhanced adaptability in a dynamic environment, due to the evolutionary components. This kind of ﬁtting process may take place at several levels. These range from connection weights to network topology or even to learning rules. Strategies of evolution of connection weights are in many cases oriented to replace the classical back-propagation and conjugate gradient algorithms. It is hoped to overcome the drawbacks of gradient-descent techniques. Adaptation of architectures is also useful to avoid the tedious trial and error mechanism used by the human knowledge expert, which up to now remains the most common way of designing connectionist systems. With the evolution possibilities of learning rules a kind of “learning to learn” process

2 Evolutionary Neuro-Fuzzy Systems and Applications

23

could be developed. This would enable neural networks to become ﬂexible and adjust to the problems to be solved. This could lead to an improvement of the learning performance, that is known to be dependent on the environment conditions and the task at hand. This would be instead of relying on individual training algorithms [179]. The adaptation process supported by the evolutionary components of an EANN is subjected to the usual cycle of computation. It is noted that some problems are inherently connected with the evaluation process of the individuals. The noisy ﬁtness evaluation problem [182] compromises a fair estimate of the generated individuals. That is due mainly to the unavoidable oneto-many mapping between genotype and phenotype, whenever evolution of network topology is performed without consideration of the weight connections. The noise apparent during the evaluations of the ﬁtness is related to the approximation of the genotype ﬁtness value, that is the network encoding, with the phenotype ﬁtness value. The network is evaluated by means of random initial weights during the evolutionary iterations. The permutation problem [67] could also occur when a many-to-one mapping from the genotypical to the phenotypical representation is done. This is due to the fact that a permutation of the hidden nodes of a network produces equivalent ANN’s with diﬀerent chromosome representations. Further analyses concerning the genotype/phenotype mapping can be found in [51, 52]. We will now brieﬂy describe the diﬀerent types of integration strategies of evolutionary algorithms and neural networks, as given in the literature. 2.5.1 Adaptation of the Connection Weights The idea of applying evolutionary algorithms to adapt neural network weights could appear straightforward if we consider the network learning as a search for an optimal parameter conﬁguration. Actually, EA’s can be helpful in order to overcome the limitations of gradient descent methods. They may avoid the problem of them becoming trapped in local minima of the error function. They are then able to seek globally optimal solutions, without the need of referring to continuos and diﬀerentiable functions. The employment of classical genetic algorithms is to determine a representation of connection weights via binary strings. Each individual can be reproduced by concatenating the binary code of the single weight conﬁguration. This yields a general representation and allows a straightforward application of both crossover and mutation operators [90, 164, 178]. A more compact representation could be obtained by using real numbers which directly correspond to the weight values [63, 142, 153]. This approach can produce a more compact chromosome representation. Some diﬃculties can arise with the application of traditional mutation and crossover operators. For this reason, diﬀerent strategies have been employed in this situation. These include evolutionary strategies implementing only mutation [154, 187].

24

G. Castellano et al.

The adoption of EA’s for weight adaptation may be shown to be useful in enhancing the accuracy of neural networks. However, the most promising applications appear to be the development of cooperative training processes. In fact, the ability of EA’s to ﬁnd global basin of attraction may be coupled with some techniques of local search. The employment of evolutionary techniques to ﬁnd near-optimal initial condition, followed by back-propagation local search has been successfully applied [107, 189]. 2.5.2 Adaptation of the Architectures In a similar way to the search for connection weights, the design of a network architecture can be regarded as a peculiar search process. In principle it can be done by means of evolutionary methods. The problem here can be embedded into the chromosome representation. The direct encoding approach aims at encoding all the parameters useful for realising a correspondence between the network topology and the employed data structure. This can be represented by a matrix recording the connections between the nodes [114,116]. This kind of approach is likely to increase the dimension of the coding structure and therefore some indirect encoding strategies have been proposed. These make use only of a set of the most important factors needed to deﬁne the network architecture. Other details may be either disregarded, predeﬁned or speciﬁed by means of deterministic rules [139, 174, 188]. In order to avoid the noisy ﬁtness evaluation problem mentioned before, architecture and weight connections have sometimes evolved together. A fully speciﬁed neural network may be codiﬁed in a one-to-one mapping [48, 144, 182, 183, 185]. 2.5.3 Adaptation of the Learning Rules The ultimate application of EA’s to neural network optimisation relies on the adaptation possibilities of the learning rules. The selection of an appropriate training algorithm, together with the weight initialisation and the topology design, plays a fundamental role in determining the behaviour of a connectionist system. It can be argued that the principle of applying the same learning rules for every task is impracticable. For these reasons, the help supplied by an evolutionary search in dynamically determining a proper learning strategy in diﬀerent contexts would be valuable. The process of encoding a learning rule represents a challenging task. Some proposals can be found in literature, and these range from the plain evolution of the algorithmic parameters of back-propagation methods [97], to the eﬀective evolution of learning rules, which have been assumed to be particular functions. These have been evolved in terms of their coeﬃcients [10, 32].

2 Evolutionary Neuro-Fuzzy Systems and Applications

25

2.5.4 Diﬀerent Hybridisation Strategies The hybridisation of evolutionary algorithms and neural networks can be done in ways which are quite diﬀerent to those described in the previous sections. Instead of considering the support provided by EA’s for designing a connectionist system in terms of weight conﬁgurations, topology and learning rules, a preprocessing activity can be used which is suitable for eﬃcient neural training. Evolutionary methods are employed to develop an input feature selection process [64, 106]. The integration of the paradigms can also be considered in the use of an ANN to support evolving procedures. Some have suggested the employment of neural networks for controlling the parameters involved in ﬁtness evaluation [37]. A fruitful ﬁeld of research is represented by the use of ensembles of neural networks. An ensemble of ANN’s is regarded as a population which may oﬀer better generalisation opportunities when considered as an integrated system [113, 184].

2.6 Evolutionary-Neuro-Fuzzy Systems An evolutionary neuro-fuzzy system (ENFS) is the result of adding evolutionary search procedures to systems, integrating fuzzy logic computing and neural learning. Using these three Soft Computing paradigms together can overcome some limitations of simpler forms of hybridisation. One of the main problems with neuro-fuzzy systems is that the learning algorithm is typically based on a steepest descent optimisation technique minimising an error function. That is, back-propagation training is not guaranteed to converge. The algorithm may be trapped in a local minimum. It can never ﬁnd the global solution. The tuning of the membership function parameters through neural learning is also not guaranteed. As previously stated (see section 2.5.1), experimental evidence has shown cases where evolutionary algorithms are ineﬃcient for ﬁne tuning solutions. They are better at ﬁnding global solutions. A hybrid learning scheme combining neural learning and evolutionary strategies may be able to solve this problem. A common approach involves using a genetic algorithm to rapidly locate a good region in the solution space and to initialise membership function parameters. The parameters are then ﬁne tuned by applying a gradient descent learning algorithm that performs a local search in the good region to ﬁnd a near optimal solution [194]. A problem of (neuro-)fuzzy modelling is the diﬃculty of determining the proper number of rules and the number of membership functions for each rule (see section 2.4). Evolutionary approaches can be integrated into neuro-fuzzy modelling to overcome this diﬃculty and to perform structure and parameter optimisation of the fuzzy rule base.

26

G. Castellano et al.

Several methodologies have been proposed in the literature to develop a form of evolutionary-neural-fuzzy hybridisation. We now review some of the existing approaches, grouping methodologies to reach the goal of optimisation. Some of them aim to learn parameters of a fuzzy rule base, whose structure is assumed to be ﬁxed in advance. Other methods are developed to perform both structure and parameter learning of a fuzzy rule base. 2.6.1 Parameter Learning A common approach given in the literature to assess an ENFS is by adding evolutionary learning capabilities to a neuro-fuzzy network. This network is usually a feed-forward multilayered network, which incorporates some fuzzy concepts. The result is a feed-forward multilayered network having both fuzzy and genetic characteristics. In simplest approaches, a GA is used to learn or tune all the parameters of a neuro-fuzzy network which is assumed to have a ﬁxed structure [112, 127, 145, 157, 177]. In these cases chromosomes of the GA encode all the parameters of the neuro-fuzzy model, including parameters of membership functions as deﬁned in the antecedent and consequent of each fuzzy rule. In [49] the parameter identiﬁcation of a fuzzy model is achieved in three separate learning phases. The ﬁrst phase uses the Kohonen’s self organising feature map algorithm to ﬁnd the initial parameters of the membership functions. A maximum matching-factor algorithm is applied, in the second phase, to ﬁnd the correct consequence part of each rule. After the fuzzy rules have been found and the whole network structure is established, the third learning phase ﬁne-tunes the membership function parameters using a multiresolutional dynamic genetic algorithm (MRD-GA) that dynamically adapts the fuzzy-model conﬁguration during the optimisation process. 2.6.2 Structure and Parameter Learning Genetic algorithms may also be applied for structure optimisation of a fuzzy rule base. That is, to deﬁne the proper number of rules and the number of antecedents and consequents in each rule, as in [54]. Evolutionary strategies can be conveniently applied to simultaneously learning the parameters and the structure of a fuzzy model. Several approaches are given in the literature that involve the membership function parameters together with the fuzzy rule set, including the number of rules. In some cases the antecedent and consequent part of each rule are also included. As an example, the strategy proposed in [4] optimises the whole fuzzy system, represented as a neural network, in order to derive a minimum rule number ﬁtting the given speciﬁcations, while training the network parameters. In [55] a GA-based method is used for a rough search for proper structures in the antecedent of fuzzy rules. The ﬁne tuning of the parameters of the fuzzy model

2 Evolutionary Neuro-Fuzzy Systems and Applications

27

is done successively using neural learning. A complex hybrid genetic-neurofuzzy scheme is implemented in FuGeNeSys [150] that can learn a fuzzy rule base from the data. This uses the membership function parameters, the number of rules and the structure of each rule. These are simultaneously deﬁned via a genetic algorithm that incorporates neural learning in the evolution procedure. Each individual in the population is made up of a set of rules, where each rule has an adjustable number of antecedents and consequents. Whenever a better individual is generated, it is transformed in a neuro-fuzzy network and trained with a back-propagation procedure until the error is reduced. Finally, the trained neuro-fuzzy network is retransformed in a genetic individual and used to replace the original individual. In other approaches, such as the EvoNF (Evolving Neuro Fuzzy) model proposed in [1], all the adjustable parts of a fuzzy system, such as the membership function parameters, structure of the rule base are involved. This includes items such as the number of rules, representation of antecedents and consequents. The fuzzy operators are derived through an evolutionary strategy. Coding all such informations into a chromosome leads to individuals characterised by a complex layered structure. The optimisation strategy may become computationally expensive. Neural learning can be used to further tune the membership functions located by the evolutionary algorithm. Finally, genetic algorithms can be used to evolve fuzzy neural networks. That is, neural networks that incorporate fuzzy numbers as weights, perform fuzzy operations in the nodes, and/or consider fuzzy nodes to represent membership functions. In this case, the learning process uses GA’s to obtain the weights of the neural network, to adapt the transfer functions of the nodes, and/or to adapt the topology of the network, as in [105]. 2.6.3 Applications Hybrid systems using a combination of neural, fuzzy and evolutionary computing have been successfully employed in many applications, such as control [162], manufacturing [121], ﬁnancial prediction [190], pharmacology [151], consumer products [161], telecommunications [13], modelling and decision making. Some of these ENFS’s application samples are reported in the literature. Our account does not claim to represent a complete overview. In [161], the authors report an application of evolutionary computation in combination with neural networks and fuzzy systems for intelligent consumer products. The role of the evolutionary algorithm is to adapt the number of rules and to tune the membership functions to improve the performance of particular fuzzy systems. They are involved in predicting the number of dishes to be cleaned by a dish washer, estimating the amount of rice in a rice cooker and controlling a microwave oven. The paper also mentions evolutionary computation for fuzzy rule generation when applied to process control.

28

G. Castellano et al.

In [190] an evolutionary fuzzy neural network is proposed for ﬁnancial prediction with hybrid input data sets from diﬀerent ﬁnancial domains. In [3] a genetic-neuro-fuzzy system is applied to on-line recognition of on-line Arabic cursive handwriting. In this system, a genetic algorithm is used to select the best combination of characters recognised by a fuzzy neural network. The value returned from the ﬁtness function for each gene represents the degree of match between the word represented by that gene and the real handwritten word. The calculation of the ﬁtness value is based on the fuzzy values assigned to the recognition of each character by the fuzzy neural network. In [121] the authors present a neuro-genetic-fuzzy system for computerised colour prediction, which is a challenging problem in paint production. A fuzzy knowledge base for predicting the pigment concentration of ten diﬀerent colours for a given surface spectral reﬂectance is obtained by means of a neuro-fuzzy system. The fuzzy population generator uses this knowledge to seed the ﬁrst generation of colour chromosomes. In addition, expert knowledge of the colour technician about the correct proportions of colours, the number of necessary colours and conﬂicts between complementary and similar colours is summarised in a fuzzy rule base that is introduced in the ﬁrst generation of colour chromosomes. The GA calculates one component of colour chromosome ﬁtness according to the compliance of the chromosome’s colour with the fuzzy expert rules.

2.7 A Genetic-Based Approach to Optimise Neuro-Fuzzy Models Research on NFS’s and EFS’s has for long been considered the objective of the learning process in terms of accuracy. Consequently, the error measure as minimised by neural learning, or the ﬁtness function as maximised by the genetic algorithm was stated in terms of errors or distances from the target output. This causes a lack of interpretability and transparency in the resulting fuzzy rules. This is due to generation of meaningless and overlapping fuzzy sets. Indeed, the most important motivation to use a fuzzy system is that the produced model is characterised by a clear reliance on linguistic terms that allows for easy understanding by humans. Hence, methods for training fuzzy models from data should not only ﬁnd the best approximation of the data, but also and more importantly, they should extract knowledge from the data in the form of fuzzy rules that can be readily understood and interpreted. Recently, the concepts of linguistic fuzzy modelling, interpretability, and similar ideas were considered to have qualities almost the opposite to accuracy. These have now been reconsidered [25] and today are viewed as an interesting part of the design process of a FS using neural learning and genetic algorithms. Several works have suggested both accuracy and interpretability as

2 Evolutionary Neuro-Fuzzy Systems and Applications

29

objectives in genetic-based learning systems [83, 85, 91]. This sets a situation for the learning process where several diﬀering objectives have to be simultaneously considered. Some of the measures used to determine the level of interpretability of a fuzzy system are degree of compactness [83]. This may be regarded as the number of rules in the rule base, and rule simplicity [31]. This simplicity of rules may be evaluated through the number of input variables involved in each rule. In this section we present an approach to identify accurate and interpretable fuzzy models based on a combination of neural learning and genetic algorithms [30]. The approach is aimed to optimise both the structure and the parameters of the fuzzy model extracted from data by a neural learning scheme through a multi-objective genetic algorithm. An iterative process is performed in each step using two optimisation procedures. A structure optimisation procedure that reduces the number of rules, and a multi-objective genetic algorithm that tunes membership function parameters. This is done by enforcing constraints on the fuzzy sets to ensure they result well-formed, leading to fuzzy rules that can be easily understood. Returning to the fuzzy rule expressed in (2.3): IF x1 is Ak1 AND . . . AND xn is Akn THEN y1 is bk1 AND . . . AND ym is bkm , we may suppose that the fuzzy model, generated by the neural learning algorithm, makes use of fuzzy singletons bkj (j = 1, . . . , m). These are deﬁned over k the output variables yj and the fuzzy represented by Gaussian sets Ai are

−cik ) , where cik and aik are the membership functions µik (xi ) = exp − (xi2a 2 ik centre and the width of the Gaussian function, respectively. Based on a set of K rules, the output of the fuzzy inference system for a given input x(0) is obtained as follows: 2

K yˆj (0) = k=1 K

µk (x(0))bkj

k=1

µk (x(0))

j = 1, . . . , m.

(2.4)

n Here µk (x(0)) = i=1 µik (xi (0)), where k = 1, . . . , K is the degree of fulﬁlment for the k-th rule, for k = 1, . . . , K. To better formalise the scheme of the GA-based approach, we denote by F RB(w, K) a Fuzzy Rule Base with parameter vector w and structure size K. That is, with a number of K fuzzy rules. We indicate by F RB(w0 , K0 ) the rule base of the fuzzy model initially derived by neural learning. Firstly, the structure optimisation procedure is applied to F RB(w0 , K0 ) to remove iteratively rules, until the diﬀerence between the accuracy of the reduced model and the accuracy of the initial model drops below a threshold . The result is a new fuzzy rule base F RB(wS , KS ), with 0 ≤ S ≤ (K0 − 1). Here KS ≤ K0 and ERR(wS ) − ERR(w0 ) ≤ , where ERR(·) is an error function used to evaluate the accuracy (generalisation ability) of the fuzzy model. Such a fuzzy

30

G. Castellano et al.

Fig. 2.5. Scheme of the GA-based optimisation approach

rule base is taken as starting point for the GA-based optimisation procedure. The result of this parameter optimisation is a rule base F RB(w S , KS ) that satisﬁes the following conditions: ERR(wS ) ≤ ERR(wS ),

IN T (wS ) ≥ IN T (wS ),

where IN T (·) is a qualitative measure of the interpretability of the antecedent fuzzy sets, that will be detailed in Section 2.7.2. If the parameter optimisation process is able to improve the accuracy, that is, if ERR(wS ) < ERR(wS ), then the structure-parameter optimisation cycle can be reiterated, using the fuzzy rule base F RB(wS , KS ) as starting point. Otherwise, that is, if the local parameter tuning remains unchanged the accuracy, the whole optimisation algorithm is stopped. Providing the rule base F RB(wS , KS ) is the ﬁnal rule base. The whole iterative algorithm ends when no further improvement in the accuracy and interpretability of the fuzzy rules can be observed. That is, when the best possible compromise between accuracy and interpretability of the fuzzy model is achieved. Figure 2.5 shows the ﬂow chart of the whole optimisation approach. We now describe both of the optimisation procedures.

2 Evolutionary Neuro-Fuzzy Systems and Applications

31

2.7.1 Structure Optimisation The structure optimisation procedure is referred to in [27,28] and is a systematic process that sequentially removes rules and investigates the model with fewer rules. At each step, rules to be removed are identiﬁed and the remaining rules are updated so the accuracy of the reduced model remains unchanged. This is achieved by updating only the consequent parameters, while premise parameters are left unchanged. The removal of a rule involves the elimination of the fuzzy sets in its antecedent. Thus, after the structure simpliﬁcation process, the number of fuzzy sets is also reduced, so as to produce a simple (hence easily interpretable) rule base. There is no guarantee that the remaining fuzzy sets still cover the entire input space. The use of Gaussian membership functions assures enough coverage and provides an acceptable accuracy when few rules are removed. As soon as many rules are removed, the coverage is no longer guaranteed. The remaining fuzzy sets may reveal a high degree of overlapping. That is, they may lack in distinguishability. To deal with these problems, a multi-objective genetic algorithm is used to tune the premise parameters of reduced fuzzy rules. This is done to improve interpretability in the sense of readable fuzzy sets with no loss in accuracy of the model. In the following section we describe in detail the multi-objective genetic algorithm. 2.7.2 Multi-Objective GA for Parameter Optimisation The parameters of membership functions in the antecedents of each fuzzy rule are encoded into an individual (chromosome). Since we have nK premise membership functions, each function being represented by two real values (i.e. cik and aik ), the length of each individual is 2nK. To limit the length of individuals, we adopt a coding with real-valued genes in the chromosome which directly represent the premise parameters of a fuzzy rule base. Thus, the i-th chromosome is a string of the form: (i)

(i)

(i)

(i)

(i)

(i)

(i)

(i)

si = (c11 , a11 , · · · , cn1 , an1 , · · · , c1K , a1K , · · · , cnK , anK ). premise of rule 1

premise of rule K

The ﬁrst individual of the initial population is generated as a copy of the premise parameters of the fuzzy rules produced by the structure optimisation procedure. The remaining individuals are initialised with random values constrained within certain permissible ranges, that is: cik = rand · range{xi (t)} + min {xi (t)} t=1,...,N

and aik = rand ·

2 range{xi (t)} + ai,min , 3 KS

32

G. Castellano et al.

where rand is a random number in [0, 1], range{xi } is deﬁned as:

range{xi } = max {xi (t)} − min {xi (t)} t=1,...,N

t=1,...,N

and ai,min is the minimum width among all the widths in the ﬁrst individual. By doing so, the centres of membership functions are constrained within the corresponding input range, while the widths are greater than zero and not wider than the whole input range. The reproduction probability of an individual is deﬁned from the values of the ﬁtness function according to: pR (si ) = NfIit(si ) . To speed up the j=1

f it(sj )

convergence, the best individual in the population is always selected and kept unchanged in the next generations, according to the elitist strategy [59]. The simplest form of crossover, which is the single point crossover, is adopted. Individuals are mutated by adding a random value generated within the permissible ranges, and deﬁned by: = N (α · range{xi }, (β · range{xi })2 ), where α and β assume diﬀerent values depending whether the gene is the centre or the width of a Gaussian function. For each input variable xi , genes representing the centres of extreme membership functions are ﬁxed to the extremes of range{xi } and are never changed by mutation. A readable fuzzy partition for xi requires the leftmost (or the rightmost) membership function to attain its maximum value at the extreme point deﬁned by the minimum (or the maximum) of range{xi }. Dynamic crossover and mutation probability rates are used, since they provide faster convergence [160]. The crossover probability is set high at the beginning of the evolution and decreases exponentially with the increasing number of generations. The mutation probability increases exponentially with each succeeding generation. The expression for the ﬁtness function was derived to model the two objectives of the optimisation process. These are the interpretability and the accuracy of the fuzzy model. To deﬁne the ﬁtness term with respect to interpretability, the concept of interpretability is translated into the concepts of completeness and distinguishability of the fuzzy sets, that are expressed through a fuzzy similarity measure. If the similarity of two neighbouring fuzzy sets is zero or too small, this indicates that either the fuzzy partitioning is incomplete or they do not overlap enough. If the similarity value is too large, then it means that the two fuzzy sets overlap too much and the distinguishability between them is lost. The following measure [158] has been used to deﬁne the similarity between the two fuzzy sets A and B: M (A B) , (2.5) S(A, B) = M (A) + M (B) − M (A B) −∞ where M (A) = +∞ µA (x)dx is the size of fuzzy set A. This is computed as the area of the triangle approximating the Gaussian function µA (· ). To

2 Evolutionary Neuro-Fuzzy Systems and Applications

33

keep the fuzzy sets in a proper shape, the fuzzy similarity measure of any two neighbouring membership functions must satisfy the following condition: Slow ≤ S(Al , Al+1 ) ≤ Sup ,

(2.6)

where Al and Al+1 are the two neighbouring fuzzy sets, Slow and Sup are the desired lower and upper bound of the fuzzy similarity measure, respectively. For each input variable xi , the quantity: 1

βil = {0

if Slow ≤S(Ail ,Ai,l+1 )≤Sup otherwise

(2.7)

is deﬁned. We can derive the following term which must be inserted into the ﬁtness function of an individual: f itint (s) =

β n(KS − 1)

,

(2.8)

where KS is the number of fuzzy sets deﬁned on each input variable. This is equal to the number of rule units in the network and β is deﬁned as: β=

n K S −1

i=1

βil .

(2.9)

l=1

The value f itint (s) is maximum (that is, equal to 1) when all adjacent fuzzy sets coded into the individual s satisfy the condition (2.6). Conversely, if this condition is not satisﬁed by most of adjacent fuzzy sets, the corresponding ﬁtness term will be assigned a very low value which means that the corresponding individual is unlikely to survive. To preserve and, hopefully, improve the accuracy of the fuzzy model, the modelling error ERR(wi ) should be minimised. The value wi is obtained by joining the premise parameters coded into the individual si and the vector of consequent parameters. This vector is taken from the rule base F RB(wS , KS ) resulting from the structure optimisation procedure, where wi = [si , bS ]. The corresponding ﬁtness term is deﬁned as: (2.10) f itacc (si ) = exp −λERR(wi )2 , where λ is a ﬁxed parameter with a large value. A high value of f itacc (si ) corresponds to a very low error, or high accuracy, of the fuzzy model coded into the individual si . To summarise, the analytical expression of the ﬁtness function for an individual si is given by: f it(si ) = f itacc (si ) + γf itint (si ),

(2.11)

where γ is a factor that controls the inﬂuence of the term f itint (si ) during the whole GA evolution. It is made less relevant during the ﬁrst generations, and has more and more inﬂuence as the evolution proceeds.

34

G. Castellano et al.

2.7.3 Illustrative Example In this section, we will examine the application of this approach when used to derive fuzzy models for a problem of medical diagnosis. In particular, the Heart Disease data set from the University of California, Irvine, is considered. The database contains 270 samples. Each sample is represented by 13 attributes. There are two classes: presence or absence of heart-disease. By using the neuro-fuzzy learning method proposed in [29] and performing a 10-fold cross validation strategy, we derived 10 diﬀerent fuzzy models with a number of rules varying between 10 and 16. The application of the GA-based approach gave the results summarised in Table 2.1. It can be seen that a great improvement can be achieved in terms of complexity reduction and accuracy on the test set. The ﬁnal number of rules was between 2 and 4. Almost in all trials, the structure-parameter optimisation required two iterations to improve the neuro-fuzzy model. As an illustrative example of such a behaviour, Table 2.2 gives the detailed results obtained for the model no. 6 where, starting from a 15-rule approximate model with low classiﬁcation rate Table 2.1. Results obtained after the optimisation of neuro-fuzzy models with 10fold cross validation model 1 2 3 4 5 6 7 8 9 10 average

before optimisation number of rules class. rate 16 59.25 12 70.37 13 70.37 10 74.07 11 74.07 15 70.37 12 55.56 10 74.07 13 74.07 11 59.26 12.3 68.15

after optimisation number of rules class. rate 4 70.37 3 70.37 2 81.48 2 74.07 3 70.37 3 74.07 2 55.56 4 88.89 2 85.18 4 88.89 2.9 75.92

Table 2.2. Results obtained at each stage of the GA-based optimisation approach in the case of the model no. 6 model initial model

number of rules

15 iteration no. 1 structure opt. 10 GA-based opt. 10 iteration no. 2 structure opt. 3 GA-based opt. 3

class. rate train. set test set 70.37 70.37 69.96 66.67

66.67 74.07

66.67 65.43

70.37 74.07

2 Evolutionary Neuro-Fuzzy Systems and Applications

35

1

1 x1

0.5

0.5 0

0 40

30

50

60

70

30

1

40

50

60

70

300

400

500

1 x3

0.5 0

0.5 0

200

300

400

500

200

1

1 x5

0.5 0

0.5 0

100

150

200

100

150

200

1

1 x7

0.5

0.5 0

0 100

150

(a)

200

100

150

200

(b)

Fig. 2.6. Fuzzy sets of some input variables before (a) and (b) after the application of the GA-based optimisation process

on the test set, a ﬁnal fuzzy model with only 3 rules and improved classiﬁcation rate was obtained. Also, the interpretability of the ﬁnal model is clearly improved, as shown by the well-formed ﬁnal fuzzy sets in ﬁg. 2.6.

2.8 Conclusions In this chapter we have discussed the integration possibilities among the three paradigms that normally constitute the key components of the Soft Computing ﬁeld of research. These are artiﬁcial neural networks, fuzzy logic and evolutionary algorithms. We reviewed the hybridisation mechanisms at the basis of the coupling approaches. The intent was to address the issues related to the intrinsic beneﬁts and diﬃculties connected with their implementation. We considered the more complex integration of the three paradigms, which gives rise to the development of Evolutionary Neuro-Fuzzy Systems. We proposed a technique concerning design techniques used for the realisation of EFNF’s, together with their application feasibility. In addition, we have presented a particular approach of evolutionaryneuro-fuzzy integration, devoted to data-driven learning of fuzzy inference models. Particular attention has been paid in organising a hybrid system which could prove to be eﬀective in constructing an accurate and comprehensible fuzzy rule base.

36

G. Castellano et al.

We hope that this work could trigger the increase in the research eﬀorts devoted to progress the Soft Computing ﬁeld of investigation. To further motivate such an endeavour, we oﬀered an elaborated series of bibliographical references which could serve as a directory for the inclined reader to be oriented among both classical and novel literature contributes.

References 1. Abraham A (2002) EvoNF: A framework for optimization of fuzzy inference systems using neural network learning and evolutionary computation. In: Proc. 2002 IEEE International Symposium on Intelligent Control (ISIC’02), Canada, IEEE Press, 327–332 2. Alba E, Cotta C, Troya JM (1999) Evolutionary design of fuzzy logic controllers using strongly-typed GP. Mathware Soft Comput. 6(1):109–124 3. Alimi AM (1997) An evolutionary neuro-fuzzy approach to recognize on-line Arabic handwriting. In: Proceedings of the Fourth International Conference on Document Analysis and Recognition, 382–386 4. Alpaydin G, Dandar G, Balkir S (2002) Evolution-based design of neural fuzzy networks using self-adapting genetic parameters. IEEE Trans. Fuzzy Systems 10(2):211–221 5. Angeline PJ, Kinnear KE (eds) (1996) Advances in Genetic Programming II. MIT Press, Cambridge, MA 6. Angelov PP (2002) Evolving Rule-Based Models. A Tool for Design of Flexible Adaptive Systems. Physica-Verlag, Wurzburg 7. Arbib MA (ed) (2002) The Handbook of Brain Theory and Neural Networks (2nd Edition). MIT Press, Cambridge, MA 8. Baron L, Achiche S, Balazinski M (2001) Fuzzy decision support system knowledge base generation using a genetic algorithm. Int. J. Approximate Reasoning 28(1):125–148 9. Bastian A (2000) Identifying fuzzy models utilizing genetic programming. Fuzzy Sets and Systems 113:333–350 10. Bengio S, Bengio Y, Cloutier J, Gecsei J (1992) On the optimization of a synaptic learning rule. In: Preprints Conf. Optimality in Artiﬁcial and Biological Neural Networks 11. Benitez JM, Castro JL, Requena I (1997) Are artiﬁcial neural networks black boxes?. IEEE Trans. Neural Networks 8:1156–1164 12. Berenji HR, Khedkar P (1992) Learning and tuning fuzzy logic controllers through reinforcements. IEEE Trans. Neural Networks 3:724–740 13. Beritelli F, Casale S, Russo M (1995) Robust phase reversal tone detection using soft computing. In: Proc. of ISUMA - NAFIPS ’95 The Third International Symposium on Uncertainty Modeling and Analysis and Annual Conference of the North American Fuzzy Information Processing Society, 589–594 14. Bezdek JC (1981) Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum Press, New York 15. Bishop CM (1995) Neural Networks for Pattern Recognition. Oxford University Press 16. Bonarini A (1996) Evolutionary learning of fuzzy rules: competition and cooperation. In: [133]

2 Evolutionary Neuro-Fuzzy Systems and Applications

37

17. Bonissone PP, Khedkar PS, Chen Y (1996) Genetic algorithms for automated tuning of fuzzy controllers: a transportation application. In: Proc. Fifth IEEE Int. Conf. on Fuzzy Systems (FUZZ-IEEE96), pp. 674–680 18. Buckles BP, Petry FE, Prabhu D, George R, Srikanth R (1994) Fuzzy clustering with genetic search. In: Proc. 1st IEEE Conf. on Evolutionary Computation (ICEC94), pp. 46–50 19. Buckley JJ, Feuring T (1999) Fuzzy and Neural: Interactions and Applications. Studies in Fuzziness and Soft Computing. Physica-Verlag, Heidelberg, Germany 20. Buckley JJ, Hayashi Y (1993) Numerical relationship between neural networks, continuous functions and fuzzy systems. Fuzzy Sets Syst. 60(1):1–8 21. Buckley JJ, Hayashi Y, Czogala E (1993) On the equivalence of neural nets and fuzzy expert systems. Fuzzy Sets Syst. 53(2):129–134 22. (2003) Advanced Fuzzy Systems Design and Applications. Studies in Fuzziness and Soft Computing (112), Physica-Verlag 23. Carse B, Fogarty TC, Munro A (1996) Evolving fuzzy rule based controllers using genetic algorithms. Fuzzy Sets and Systems 80:273–294 24. Casillas J, Cordon O, del Jesus MJ, Herrera F (2001) Genetic feature selection in a fuzzy rule-based classiﬁcation system learning process for high dimensional problems. Inform. Sci. 136:169–191 25. Casillas J, Cordon O, Herrera F, Magdalena L (2003) Interpretability Issues in Fuzzy Modeling, Springer 26. Castellano G, Castiello C, Fanelli AM, Mencar C (2005) Knowledge Discovery by a Neuro-Fuzzy Modeling Framework. Fuzzy Sets and Systems 149:187–207 27. Castellano G, Fanelli AM (1996) Simplifying a neuro-fuzzy model. Neural Proc. Letters 4:75–81 28. Castellano G, Fanelli AM (1997) An approach to structure identiﬁcation of fuzzy models. In: Proc. of the Sixth IEEE International Conference on Fuzzy Systems (FUZZ-IEEE 1997) 531–536 29. Castellano G, Fanelli AM (2000) Fuzzy inference and rule extraction using a neural network. Neural Network World Journal 3:361–371 30. Castellano G, Fanelli AM, Gentile E, Roselli T (2002) A GA-based approach to optimize fuzzy models learned from data. Proc. of Workshop on Approximation and Learning in Evolutionary Computing, part of Genetic and Evolutionary Computation Conference (GECCO 2002), New York, USA, 5–8 31. Castillo L, Gonzalez A, Perez R (2001) Including a simplicity criterion in the selection of the best rule in a fuzzy genetic learning algorithm. Fuzzy Sets and Systems 120(2):309–321 32. Chalmers DJ (1990) The evolution of learning: An experiment in genetic connectionism. In: Proc. 1990 Connectionist Models Summer School, pp. 81–90 33. Cheong F, Lai R (2000) Constraining the optimization of a fuzzy logic controller using an enhanced genetic algorithm. IEEE Trans. Systems Man Cybernet. B 30:31–46 34. Chen J, Xi Y (1998) Nonlinear system modeling by competitive learning and adaptive fuzzy inference system. IEEE Trans. Syst., Man, Cybern. 28:231–238 35. Chien B-C, Lin JY, Hong T-P (2002) Learning discriminant functions with fuzzy attributes for classiﬁcation using genetic programming. Expert Systems Appl. 23(1):31–37

38

G. Castellano et al.

36. Cho KB, Wang BH (1996) Radial basis function based adaptive fuzzy systems and their applications to system identiﬁcation and prediction. Fuzzy Sets Syst. 83:325339 37. Chou L-D, Wu J-LC (1995) Parameter adjustment using neural-network-based genetic algorithms for guaranteed QOS in atm networks. IEICE Trans. Commun. E78-B(4):572–579 38. Chung I-F, Lin CJ, Lin CT (2000) A GA-based fuzzy adaptive learning control network. Fuzzy Sets and Systems 112(1):65–84 39. Cordon O, del Jesus MJ, Herrera F, Lozano M (1999) MOGUL: a methodology to obtain genetic fuzzy rule-based systems under the iterative rule learning approach. Int. J. Intelligent Systems 14(11):1123–1153 40. Cordon O, Herrera F (1997) A three-stage evolutionary process for learning descriptive and approximate fuzzy logic controller knowledge bases from examples. Int. J. Approximate Reasoning 17(4):369–407 41. Cordon O, Herrera F (2001) Hybridizing genetic algorithms with sharing scheme and evolution strategies for designing approximate fuzzy rule-based systems, Fuzzy Sets and Systems 118(2):235–255 42. Cordon O, Herrera F, Hoﬀmann F, Magdalena L (2001) Genetic Fuzzy Systems - Evolutionary Tuning and Learning of Fuzzy Knowledge Bases. World Scientiﬁc, Singapore 43. Czogala E, Leski J (2000) Fuzzy and Neuro-Fuzzy Intelligent Systems. Studies in Fuzziness and Soft Computing (47), Physica-Verlag 44. de Oliveira JV (1999) Towards Neuro-Linguistic Modeling: Constraints for Optimization of Membership Functions. Fuzzy Sets and Systems 106:357–380 45. Delgado MR, Von Zuben F, Gomide F (2004) Coevolutionary genetic fuzzy systems: a hierarchical collaborative approach. Fuzzy Sets and Systems 141:89– 106 46. Dote Y, Ovaska SJ (2001) Industrial applications of soft computing: a review. Proc. IEEE 89(9):1243–1265 47. Eiben AE, Back T, Schoenauer M, Schwefel HP (eds) (1998) Parallel Problem Solving from Nature (PPSN) V. Lecture Notes in Computer Science, vol. 1498 Springer-Verlag, Berlin, Germany 48. Fang J, Xi Y (1997) Neural network design based on evolutionary programming. Artiﬁcial Intell. Eng. 11(2):155–161 49. Farag WA, Quintana VH, Lambert-Torres G (1998) A Genetic-Based NeuroFuzzy Approach for Modeling and Control of Dynamical Systems. IEEE Trans. on Neural Networks 9(5):756–767 50. Feng JC, Teng LC (1998) An Online Self Constructing Neural Fuzzy Inference Network and its Applications. IEEE Trans. on Fuzzy Systems 6(1):12–32 51. Fogel D (1995) Phenotypes, genotypes, and operators in evolutionary computation. In Proc. 1995 IEEE Int. Conf. Evolutionary Computation (ICEC95), pp. 193–198 52. Fogel D (1999) Evolutionary Computation: Towards a New Philosophy of Machine Intelligence (2nd edition). IEEE Press 53. Fogel LJ, Owens AJ, Walsh MJ (1966) Artiﬁcial Intelligence Through Simulated Evolution. John Wiley & Sons, Chichester, UK 54. Fukuda T, Ishigami H, Shibata T, Arai F (1993) Structure Optimization of Fuzzy Neural Network by Genetic Algorithm, Fifth JFSA World Congress, 964–967

2 Evolutionary Neuro-Fuzzy Systems and Applications

39

55. Furuhashi T, Matsushita S, Tsutsui H (1997) Evolutionary Fuzzy Modeling Using Fuzzy Neural Networks and Genetic Algorithm, IEEE International Conference on Evolutionary Computation, 623–627 56. Furuhashi T (2001) Fusion of fuzzy/neuro/evolutionary computing for knowledge acquisition, Proceedings of the IEEE 89(9):1266–1274 57. Geyer-Schulz A (1995) Fuzzy Rule-Based Expert Systems and Genetic Machine Learning. Physica-Verlag, Heidelberg 58. Glorennec PY (1997) Coordination between autonomous robots. Int. J. Approximate Reasoning 17(4):433–446 59. Goldberg DE (1989) Genetic Algorithms in Search, Optimization and Machine Learning. Addison Wesley, Reading, MA 60. Gomez-Skarmeta AF, Jimenez F (1999) Fuzzy modeling with hybrid systems, Fuzzy Sets and Systems 104:199–208 61. Gonzalez A, Perez R (1999) SLAVE: a genetic learning system based on an iterative approach, IEEE Trans. Fuzzy Systems 7(2):176–191 62. Gonzalez A, Perez R (2001) An experimental study about the search mechanism in SLAVE learning algorithm: hill-climbing methods versus genetic algorithm. Inform. Sci. 136(14):159–174 63. Greenwood GW (1997) Training partially recurrent neural networks using evolutionary strategies. IEEE Trans. Speech Audio Processing 5:192–194 64. Guo Z, Uhrig RE (1992) Using genetic algorithms to select inputs for neural networks. In: Proc. Int. Workshop Combinations of Genetic Algorithms and Neural Networks (COGANN-92), pp. 223–234 65. Gurocak HB (1999) A genetic-algorithm-based method for tuning fuzzy logic controllers. Fuzzy Sets and Systems 108(1):39–47 66. Hall LO, Bezdek JC, Boggavarapu S, Bensaid A (1994) Genetic fuzzy clustering. In: Proc. NAFIPS94, pp. 411–415 67. Hancock PJB (1992) Genetic algorithms and permutation problems: A comparison of recombination operators for neural net structure speciﬁcation. In Proc. Int. Workshop Combinations of Genetic Algorithms and Neural Networks (COGANN-92), pp. 108–122 68. Hanebeck UD, Schmidt GK (1996) Genetic optimization of fuzzy networks. Fuzzy Sets and Systems 79 (1):59–68 69. Hayashi Y (1994) Neural expert system using fuzzy teaching input and its application to medical diagnosis. Inform. Sci. Applicat. 1:47–58 70. Hayashi Y, Buckley JJ (1994) Approximations between fuzzy expert systems and neural networks. Int. J. Approx. Reas. 10:63–73 71. Haykin S (1999) Neural Networks - A Comprehensive Foundation (2nd edition). Prentice Hall 72. Herrera F, Lozano M (1996) Adaptation of genetic algorithm parameters based on fuzzy logic controllers. In: [74] 73. Herrera F, Lozano M, Verdegay JL (1995) Tuning fuzzy controllers by genetic algorithms. Int. J. Approx. Reasoning 12:299–315 74. Herrera F, Verdegay JL (eds) (1996) Genetic Algorithms and Soft Computing. Physica-Verlag, Wurzburg 75. Hertz J, Krogh A, Palmer RG (1991) Introduction to the theory of neural computation. Addison Wesley, Reading, MA 76. Hirota K, Pedrycz W (1994) OR/AND neuron in modeling fuzzy set connectives. IEEE Trans. on Fuzzy Systems 2:151–161

40

G. Castellano et al.

77. Hoﬀmann F, Nelles O (2001) Genetic programming for model selection of TSKfuzzy systems. Inform. Sci. 136(14):7–28 78. Hoﬀmann F, Pﬁster G (1997) Evolutionary design of a fuzzy knowledge base for a mobile robot. Int. J. Approximate Reasoning 17(4):447–469 79. Holland J (1975) Adaptation in Natural and Artiﬁcial Systems. University of Michigan Press, Ann Arbor 80. Holland J, Reitman JS (1978) Cognitive systems based on adaptive algorithms. In: [176] 81. Homaifar A, McCormick E (1995) Simultaneous Design of Membership Functions and Rule Sets for Fuzzy Controllers using Genetic Algorithms. IEEE Trans. on Fuzzy Systems 3(2):129–138 82. Hudson DL, Cohen ME, Anderson MF (1991) Use of neural network techniques in a medical expert system. Int. J. Intell. Syst. 6:213–223 83. Ishibuchi H, Murata T, Turksen IB (1997) Single-objective and two-objective genetic algorithms for selecting linguistic rules for pattern classiﬁcation problems. Fuzzy Sets and Systems 89:135–150 84. Ishibuchi H, Nakashima T, Murata T (1999) Performance evaluation of fuzzy classiﬁer systems for multidimensional pattern classiﬁcation problems. IEEE Trans. System Man Cybernet. 29:601–618 85. Ishibuchi H, Nakashima T, Murata T (2001) Three-objective genetics-based machine learning for linguistic rule extraction. Inform. Sci. 136(14):109–133 86. Ishibuchi H, Nozaki K, Yamamoto N, Tanaka H (1995) Selecting Fuzzy If-Then Rules for Classiﬁcation Problems Using Genetic Algorithms. IEEE Trans. on Fuzzy Systems 3(3):260–270 87. Jang JR (1993) ANFIS: Adaptive-network-based fuzzy inference system. IEEE Trans. Syst., Man, Cybern. 23(3):665–685 88. Jang JSR, Sun CT (1993) Functional equivalence between radial basis function networks and fuzzy inference systems. IEEE Trans. Neural Networks 4:156–159 89. Jang JSR, Sun CT, Mizutani E (1997) NeuroFuzzy and Soft Computing. Prentice-Hall, Englewood Cliﬀs, NJ 90. Janson DJ, Frenzel JF (1993) Training product unit neural networks with genetic algorithms. IEEE Expert, 8:26–33 91. Jin Y (2000) Fuzzy modeling of high-dimensional systems: complexity reduction and interpretability improvement. IEEE Trans. Fuzzy Systems 8(2):212– 220 92. Juang C, Lin C (1998) An on-line self-constructing neural fuzzy inference network and its applications. IEEE Trans. Fuzzy Syst. 6:12–32 93. Kang SJ, Woo CH, Hwang HS, Woo KB (2000) Evolutionary Design of Fuzzy Rule Base for Nonlinear System Modeling and Control. IEEE Trans. on Fuzzy Systems 8(1):37–45 94. Kasabov N (1996) Foundations of Neural Networks, Fuzzy Systems and Knowledge Engineering. MIT Press, Cambridge, MA 95. Keller JK, Hunt DJ (1985) Incorporating fuzzy membership functions into the perceptron algorithm. IEEE Trans. Pattern Anal. Mach. Intell. 7:693–699 96. Keller JM, Yager RR, Tahani H (1992) Neural network implementation of fuzzy logic. Fuzzy Sets Syst. 45:1–12 97. Kim HB, Jung SH, Kim TG, Park KH (1996) Fast learning method for backpropagation neural network by evolutionary adaptation of learning rates. Neurocomput. 11(1):101–106

2 Evolutionary Neuro-Fuzzy Systems and Applications

41

98. Kim J, Kasabov N (1999) HyFIS: Adaptive Neuro-Fuzzy Inference Systems and Their Application to Nonlinear Dynamical Systems. Neural Networks 12:1301– 1319 99. Kinnear KE (ed) (1994) Advances in Genetic Programming. MIT Press, Cambridge, MA 100. Kosko B (1991) Neural Networks and Fuzzy Systems. Prentice-Hall, Englewood Cliﬀs, NJ 101. Koza JR (1992) Genetic Programming: On the Programming of Computers by Means of Natural Selection. MIT Press, Cambridge, MA 102. Koza JR (1994) Genetic Programming II: Automatic Discovery of Reusable Programs. MIT Press, Cambridge, MA 103. Koza JR, Bennett FH III, Andre D, Keane MA (1999) Genetic Programming III: Darwinian Invention and Problem Solving. Morgan Kaufmann 104. Koza JR, Keane MA, Streeter MJ, Mydlowec W, Yu J, Lanza G (2003) Genetic Programming IV: Routine Human-Competitive Machine Intelligence. Kluwer Academic Publishers 105. Krishnamraju PV, Buckley JJ, Reilly KD, Hayashi Y (1994) Genetic learning algorithms for fuzzy neural nets. In Proc. Third IEEE Int. Conf. on Fuzzy Systems (FUZZ-IEEE94), Orlando, FL, USA 1969–1974 106. Kupinski MA, Maryellen ML (1997) Feature selection and classiﬁers for the computerized detection of mass lesions in digital mammography. In: Proc. 1997 IEEE Int. Conf. Neural Networks., pp. 2460–2463 107. Lee S-W (1996) Oﬀ-line recognition of totally unconstrained handwritten numerals using multilayer cluster neural network. IEEE Trans. Pattern Anal. Machine Intell. 18:648–652 108. Lee H-M, Chen C-M, Chen J-M, Jou Y-L (2001) An eﬃcient fuzzy classiﬁer with feature selection based on fuzzy entropy. IEEE Trans. Systems Man Cybernet. B 31(3):426–432 109. Leitch DD (1995) A New Genetic Algorithm for the Evolution of Fuzzy Systems. PhD thesis, Department of Engineering, Oxford University 110. Lim MH, Rahardja S, Gwee BH (1996) A GA Paradigm for Learning Fuzzy Rules. Fuzzy Sets and Systems 82:177–186 111. Lin CT, George Lee CS (1996) Neural Fuzzy Systems A NeuroFuzzy Synergism to Intelligent Systems. Prentice-Hall, Englewood Cliﬀs, NJ 112. Linkens DA, Nyongesa HO (1995) Evolutionary learning in fuzzy neural control systems. In Proc. Third European Congress on Fuzzy and Intelligent Technologies (EUFIT95), Aachen, Germany, 990–995 113. Liu Y, Yao X (1998) Toward designing neural network ensembles by evolution. In: [47] 114. Marin FJ, Sandoval F (1993) Genetic synthesis of discrete-time recurrent neural network. In: Proc. Int. Workshop Artiﬁcial Neural Networks (IWANN93), pp. 179–184 115. Michalewicz Z (1992) Genetic Algorithms + Data Structure = Evolution Programs. Springer-Verlag 116. Miller GF, Todd PM, Hegde SU (1989) Designing neural networks using genetic algorithms. In: Proc. 3rd Int. Conf. Genetic Algorithms and Their Applications, pp. 379–384 117. Mitra S (1994) Fuzzy MLP based expert system for medical diagnosis. Fuzzy Sets Syst. 65:285–296

42

G. Castellano et al.

118. Mitra S, Hayashi Y (2000) Neuro-Fuzzy Rule Generation: Survey in Soft Computing Framework. IEEE Trans. Neural Networks 11(3):748–768 119. Mitra S, Pal SK (1995) Fuzzy multilayer perceptron, inferencing and rule generation. IEEE Trans. Neural Networks 6:51–63 120. Mitra S, Pal SK (1996) Fuzzy self organization, inferencing and rule generation. IEEE Trans. Syst., Man, Cybern. 26:608–620 121. Mizutani E, Takagi H, Auslander DM, Jang J-SR (2000) Evolving colour recipes. IEEE Trans. Systems Man Cybernet. 30(4):537–550 122. Nauck D, Klawonn F, Kruse R (1997) Foundations of NeuroFuzzy Systems. Wiley, Chichester, U.K. 123. Nauck D, Kruse R (1992) A Neuro-Fuzzy Controller Learning by Fuzzy Error Propagation. In: Proc. of Conf. of the North Amer. Fuzzy Inf. Proc. Soc. (NAFIPS92), pp. 388–397 124. Nauck D, Kruse R (1997) New Learning Strategies for NEFCLASS. In: Proc. of Seventh Intern. Fuzzy Systems Ass. World Congress (IFSA97), pp. 50–55 125. Nauck D, Kruse R (1999) Neuro-Fuzzy Systems for function approximation. Fuzzy Sets and Systems 101:261–271 126. Nauck D, Nauck U, Kruse R (1996) Generating Classiﬁcation Rules with Neuro-Fuzzy System NEFCLASS. In: Proc. of the Biennal Conf. of the North Amer. Fuzzy Inf. Proc. Soc. (NAFIPS96) 127. Paetz J (1995) Evolutionary Optimization of Weights of a Neuro-Fuzzy Classiﬁer and the Eﬀects on Benchmark Data and Complex Chemical Data. In Proc. of NAFIPS 2005 - Annual Meeting of the North American Fuzzy Information Processing Society, 615–620 128. Pal SK, Ghosh A (1996) Neuro-fuzzy computing for image processing and pattern recognition. Int. J. Syst. Sci. 27:1179–1193 129. Pal SK, Mitra S (1992) Multilayer perceptron, fuzzy sets and classiﬁcation. IEEE Trans. Neural Networks 3:683–697 130. Pal SK, Mitra S (1999) Neuro-fuzzy Pattern Recognition: Methods in Soft Computing. Wiley, New York 131. Parodi A, Bonelli P (1993) A new approach to fuzzy classiﬁer systems. In: Proc. of Fifth International Conference on Genetic Algorithms (ICGA93), Morgan Kaufmann, pp. 223–230 132. Pedrycz W (1993) Fuzzy neural networks and neurocomputations. Fuzzy Sets and Systems 56:1–28 133. Pedrycz W (ed) (1996) Fuzzy Modelling: Paradigms and Practice. Kluwer Academic Press, Norwell, MA 134. Pedrycz W (ed) (1997) Fuzzy Evolutionary Computation. Kluwer Academic Publishers, Dordrecht 135. Pedrycz W (199) Computational Intelligence: An Introduction. CRC, Boca Raton, FL 136. Pedrycz W, Rocha A (1993) Knowledge-based neural networks, IEEE Trans. on Fuzzy Systems 1:254–266 137. Pedrycz W, Reformat M (1996) Genetic Optimization with Fuzzy Coding. In: [74] 138. Pena-Reyes CA, Sipper M (2001) Fuzzy CoCo: a cooperative coevolutionary approach to fuzzy modeling. IEEE Trans. Fuzzy Systems 9(5):727–737 139. Perez CA, Holzmann CA (1997) Improvements on handwritten digit recognition by genetic selection of neural network topology and by augmented training. In: Proc. 1997 IEEE Int. Conf. Systems, Man, and Cybernetics, pp. 1487–1491

2 Evolutionary Neuro-Fuzzy Systems and Applications

43

140. Perneel C, Themlin JM, Renders JM, Acheroy M (1995) Optimization of Fuzzy Expert Systems Using Genetic Algorithms and Neural Networks. IEEE Trans. on Fuzzy Systems, 3(3):300–312 141. Pham DT, Karaboga D (1991) Optimum design of fuzzy logic controllers using genetic algorithms. J. System Eng. 1:114–118 142. Porto VW, Fogel DB, Fogel LJ (1995) Alternative neural network training methods. IEEE Expert 10:16–22 143. Potter MA, De Jong KA (2000) Cooperative coevolution: an architecture for evolving coadapted subcomponents. Evolutionary Comput. 8(1):1–29 144. Pujol JCF, Poli R (1998) Evolving the topology and the weights of neural networks using a dual representation. Appl. Intell. 8(1):73–84 145. Rahmoun A, Berrani S (2001) A Genetic-Based Neuro-Fuzzy Generator: NEFGEN, ACS/IEEE International Conference on Computer Systems and Applications, 18–23 146. Rojas I, Gonzalez J, Pomares H, Rojas FJ, Fernandez FJ, Prieto A (2001) Multidimensional and multideme genetic algorithms for the construction of fuzzy systems. Int. J. Approx. Reasoning 26(3):179–210 147. Ross T (1997) Fuzzy Logic with Engineering Applications. McGraw-Hill 148. Roubos H, Setnes M (2001) Compact and transparent fuzzy models through iterative complexity reduction. IEEE Trans. Fuzzy Systems 9(4):515–524 149. Rumelhart DE, McClelland JL (1986) Parallel Distributing Processing. MIT Press, Cambridge, MA 150. Russo M (1998) FuGeNeSys - A Fuzzy Genetic Neural System for Fuzzy Modeling. IEEE Trans. on Fuzzy Systems 6(3):373–388 151. Russo M, Santagati NA, Lo Pinto E (1998) Medicinal chemistry and fuzzy logic. Inform. Sci. 105(14):299–314 152. Sanchez E, Shibata T, Zadeh L (eds) (1997) Genetic Algorithms and Fuzzy Logic Systems. Soft Computing Perspectives. World Scientiﬁc, Singapore 153. Saravanan N, Fogel DB (1995) Evolving neural control systems. IEEE Expert 10:23–27 154. Sarkar M, Yegnanarayana B (1997) Evolutionary programming-based probabilistic neural networks construction technique. In: Proc. 1997 IEEE Int. Conf. Neural Networks, pp. 456–461 155. Schwefel HP (1994) On the Evolution of Evolutionary Computation. In: [195]: 116–124 156. Schwefel HP (1995) Evolution and Optimum Seeking. John Wiley & Sons, Chichester, UK 157. Seng TL, Khalid MB, Yusof R (1999) Tuning of a Neuro-Fuzzy Controller by Genetic Algorithm. IEEE Transactions on Systems, Man and Cybernetics-Part B 29(2):226-236 158. Setnes M, Babuska R, Kaymak U, van Nauta Lemke HR (1998) Similarity Measures in Fuzzy Rule Base Simpliﬁcation. IEEE Trans. on Systems, Man and Cybernetics - Part B 28(3):376–386 159. Setnes M, Roubos H (2000) GA-fuzzy modeling and classiﬁcation: complexity and performance, IEEE Trans. Fuzzy Systems 8(5):509–522 160. Sheble GB, Maifeld TT (1995) Reﬁned Genetic Algorithm - Economic Dispatch Example. IEEE Trans. Power Syst. 10:117–124 161. Shim M, Seong S, Ko B, So M (1999) Application of evolutionary computation at LG electronics. In Proc. of IEEE Int. Conf. on Fuzzy Systems FUZZ-IEEE99, Seoul, South Korea, 1802–1806

44

G. Castellano et al.

162. Silva N, Macedo H, Rosa A (1998) Evolutionary fuzzy neural networks automatic design of rule based controllers of nonlinear delayed systems. In Proc. of IEEE World Congress on Computational Intelligence - Fuzzy Systems Proceedings 2:1271–1276 163. Smith SF (1980) A learning system based on genetic adaptive algorithms. PhD Thesis, University of Pittsburgh 164. Srinivas M, Patnaik LM (1991) Learning neural network weights using genetic algorithms Improving performance by search-space reduction. In: Proc. 1991 IEEE Int. Joint Conf. Neural Networks (IJCNN91), vol. 3, pp. 2331–2336 165. Subbu R, Anderson A, Bonissone PP (1998) Fuzzy logic controlled genetic algorithms versus tuned genetic algorithms: an agile manufacturing application. In: Proc. IEEE Int. Symp. on Intelligent Control (NIST) 166. Sulzberger SM, Tschicholg-Gurman NN, Vestli SJ (1998) FUN: Optimization of Fuzzy Rule Based Systems Using Neural Networks. In Proc. of IEEE Conf. on Neural Networks, pp. 312–316 167. Sun CT (1994) Rule-Base Structure Identiﬁcation in an Adaptive-NetworkBased Fuzzy Inference System. IEEE Trans. on Fuzzy Systems 2(1): 64–73 168. Takagi H, Suzuki N, Koda T, Kojima Y (1992) Neural networks designed on approximate reasoning architecture and their applications. IEEE Trans. Neural Networks 3:752–760 169. Tano S, Oyama T, Arnauld T (1996) Deep Combination of Fuzzy Inference and Neural Network in Fuzzy Inference. Fuzzy Sets and Systems 82(2):151–160 170. Tschichold-Gurmann N (1995) Generation and Improvement of Fuzzy Classiﬁers with Incremental Learning Using Fuzzy Rulenet. In: Proc. of ACM Sympos. on Appl. Comp., pp. 466–470 171. Van Le T (1995) Evolutionary fuzzy clustering. In: Proc. 2nd IEEE Conf. on Evolutionary Computation (ICEC95), Vol. 2, pp. 753–758 172. Velasco JR (1998) Genetic-based on-line learning for fuzzy process control. Int. J. Intelligent System 13(1011):891–903 173. Venturini G (1993) SIA: a supervised inductive algorithm with genetic search for learning attribute based concepts. In Proc. European Conf. on Machine Learning, Vienna, pp. 280–296 174. Vonk E, Jain LC, Johnson R (1995) Using genetic algorithms with grammar encoding to generate neural networks. In: Proc. IEEE Int. Conf. Neural Networks, pp. 1928–1931 175. Wang D, Keller JM, Carson CA, McAdoo-Edwards KK, Bailey CW (1998) Use of fuzzy-logic-inspired features to improve bacterial recognition through classiﬁer fusion. IEEE Trans. Syst., Man, Cybern. 28:583–591 176. Waterman DA, Hayes-Roth F (eds) (1978) Pattern-Directed Inference Systems. Academic Press, New York 177. Watts MJ, Kasabov K (1998) Genetic Algorithms for the design of fuzzy neural networks. In: Proc. of the Fifth Int. Conf. on Neural Information Processing (ICONIP’R98), pp. 793–796 178. Whitley D, Starkweather T, Bogart C (1990) Genetic algorithms and neural networks: Optimizing connections and connectivity. Parallel Comput. 14(3) pp. 347–361 179. Wolpert DH, Macready WG (1997) No-Free Lunch Theorems for Optimization. IEEE Trans. Evolutionary Computation 1:67–82 180. Wong CC, Feng SM (1995) Switching-Type Fuzzy Controller Design by Genetic Algorithms. Fuzzy Sets and Systems, 74(2):175–185

2 Evolutionary Neuro-Fuzzy Systems and Applications

45

181. Yao X (1999) Evolving Artiﬁcial Neural Networks. Proc. IEEE 87(9):1423–1447 182. Yao X, Liu Y (1997) A new evolutionary system for evolving artiﬁcial neural networks. IEEE Trans. Neural Networks 8:694–713 183. Yao X, Liu Y (1997) EPNet for chaotic time-series prediction. In: [186] 184. Yao X, Liu Y (1998) Making use of population information in evolutionary artiﬁcial neural networks. IEEE Trans. Syst., Man, Cyber. B 28:417–425 185. Yao X, Liu Y (1998) Toward designing artiﬁcial neural networks by evolution. Appl. Math. Computation 91(1):83–90 186. Yao X, Kim J-H, Furuhashi T (eds) (1997) Select. Papers 1st Asia-Paciﬁc Conf. Simulated Evolution and Learning (SEAL96). Lecture Notes in Artiﬁcial Intelligence, vol. 1285, Springer-Verlag, Berlin, Germany 187. Yao X, Liu Y (1997) Fast Evolution Strategies. Control Cybern. 26(3) pp. 467–496 188. Yao X, Shi Y (1995) A preliminary study on designing artiﬁcial neural networks using co-evolution. In: Proc. IEEE Singapore Int. Conf. Intelligent Control and Instrumentation, pp. 149–154 189. Yan W, Zhu Z, Hu R (1997) Hybrid genetic/BP algorithm and its application for radar target classiﬁcation. In: Proc. 1997 IEEE National Aerospace and Electronics Conf. (NAECON), pp. 981–984 190. Yu L, Zhang Y-Q(2005) Evolutionary fuzzy neural networks for hybrid ﬁnancial prediction. IEEE Transactions on Systems, Man and Cybernetics - Part C 35(2):244249 191. Yuan B, Klir GJ, Swan-Stone JF (1995) Evolutionary fuzzy c-means clustering algorithm. In: Proc. 4th IEEE Int. Conf. on Fuzzy Systems (FUZZ-IEEE95), pp. 2221–2226 192. Zadeh LA (1965) Fuzzy Sets. Information and Control 8:338–353 193. Zadeh LA (1994) Fuzzy logic, neural networks, and soft computing, Commun. ACM 37:77–84 194. Zhang Y.-Q., Kandel A. (1998) Compensatory Genetic Fuzzy Neural Networks and Their Applications. Ser. Machine Perception Artiﬁcial Intelligence. Singapore: World Scientiﬁc, vol. 30 195. Zurada J, Marks R, Robinson C (eds) (1994) Computational Intelligence: Imitating Life. IEEE Press

3 Evolution of Fuzzy Controllers and Applications Dilip Kumar Pratihar1∗ and Nirmal Baran Hui2 1

2

Associate Professor [email protected] Research Scholar [email protected] Soft Computing Lab. Department of Mechanical Engineering Indian Institute of Technology, Kharagpur Kharagpur-721 302 India

Summary. The present chapter deals with the issues related to the evolution of optimal fuzzy logic controllers (FLC) by proper tuning of its knowledge base (KB), using diﬀerent tools, such as least-square techniques, genetic algorithms, backpropagation (steepest descent) algorithm, ant-colony optimization, reinforcement learning, Tabu search, Taguchi method and simulated annealing. The selection of a particular tool for the evolution of the FLC, generally depends on the application. Some of the applications have also been included in this chapter.

Keywords: Fuzzy logic controller, Evolution, Least-square technique, Genetic-fuzzy system, Neural-fuzzy system, Ant-colony optimization, Reinforcement learning, Tabu search, Taguchi method, Simulated annealing

3.1 Introduction Real-world problems are generally associated with diﬀerent types of uncertainties. In the past, considerable eﬀort has been made to model these uncertainties. Prior to 1965, it was considered that probability theory working based on Aristotelian two-valued logic was the sole agent available to deal with uncertainties. This particular logic uses the concept of the classical crisp set. That is a set with a ﬁxed boundary. Prof. Zadeh developed the concept ∗

Corresponding author: Associate Professor, Department of Mechanical Engineering, Indian Institute of Technology, Kharagpur-721 302, India; http:// www.facweb.iitkgp.ernet.in/∼dkpra

D.K. Pratihar and N.B. Hui: Evolution of Fuzzy Controllers and Applications, Studies in Computational Intelligence (SCI) 66, 47–69 (2007) c Springer-Verlag Berlin Heidelberg 2007 www.springerlink.com

48

D.K. Pratihar and N.B. Hui

of fuzzy sets, in the year 1965 [1]. Those are the sets having the vague boundaries. He argued that probability theory can handle only one out of several diﬀerent types of possible uncertainties. Thus, there are uncertainties, which cannot be dealt with by using the probability theory. Taking an example, in which Mr. X requests Mr. Y, to bring some red apples for him from the market. There are two uncertainties at least, which relate to the following: (i) the availability of the apples, and (ii) a guarantee that the apple is red. Depending on the season, there is a probability of obtaining the apples, which varies between 0 and 1. But, the colour – red cannot be deﬁned by the classical set. It is not between red (1) and not-red (0). In the fuzzy set, the colour – red can be deﬁned as follows (Fig. 3.1) using the concept of membership of an element to a class. That is the function value (µ): If the colour is perfectly red PR, then it may be said red with a membership value of 1.0; if it is R, then it is considered to be red with a membership value of 0.65; if it is slightly red SR, then it is red with a membership value of 0.39. If it is not red (NR), then also it is red with a membership value of 0.0. In this way, the uncertainty related to the colour of the apples can be handled. Thus, a fuzzy set may be considered to be a more general concept than the classical set. The concept of fuzzy set theory has been used in a number of applications, such as the Fuzzy Logic Controller (FLC), fuzzy clustering, fuzzy mathematical programming, fuzzy graph theory and other examples. Out of all such applications, FLC is the most popular application for the following reasons – (i) ease of understanding and implementations, (ii) ability to handle uncertainty etc. An exact mathematical formulation of the problem is not required for the development of an FLC. This feature makes it a natural choice for solving complex real-world problems. These are either diﬃcult to model mathematically or the mathematical model becomes highly non-linear. It is to be noted that a fuzzy logic controller was ﬁrst developed by Mamdani and Assilian, in the year 1975 [2]. The concept of fuzzy set was published in the

1.0 0.65 0.39

NR

SR

R

µ

0.0 PR

Colour Red Fig. 3.1. A schematic diagram explaining the concept of membership function distribution.

3 Evolution of Fuzzy Controllers and Applications

49

year 1965. Human beings have the natural ability of determining the inputoutput relationships of a process. The behavior of a human being is modeled artiﬁcially, when designing a suitable FLC. The performance of an FLC depends on its knowledge base (KB), which in turn consists of both Data Base (DB) and a Rule Base (RB). The DB consists of data related to membership function distributions of the variables of the process to be controlled. Designing a proper KB of an FLC is a diﬃcult task, which may be implemented in one of the following ways: – – – –

Optimization Optimization Optimization Optimization

of of of of

the the the the

data base only, rule base only, data base and rule base in stages, data base and rule base simultaneously.

The membership function distributions are assumed to be either Linear such as, triangular, trapezoidal or Non-Linear. The Non-Linear can be Gaussian, bell-shaped, sigmoidal in nature. To design and develop a suitable FLC for controlling a process, its variables need to be expressed in the form of some linguistic terms (such as VN: Very Near, VF: Very Far, A: Ahead for example). The relationships between the input (antecedent) and output (consequent) variables are expressed in the form of rules. For example, a rule can be expressed as indicated in Fig. 3.2: IF I1 is N AND I2 is A THEN O is AR, The number of such rules will be present in the rule base. The number of linguistic terms used to represent the variables increases in order to improve the accuracy of the prediction. The computational complexity of the controller will increase with a larger number of rules. For easy implementation in either

VN

N

F

VF LT AL

1

2

3

A

AR

0

45

RT

4

I (m) 1 LT AL

A

AR

RT

– 90

– 45

90

O (degrees)

– 90

– 45

0

45

90

I (degrees) 2

Fig. 3.2. A diagram showing some membership function distributions of input and output variables of the Fuzzy Logic Controller.

50

D.K. Pratihar and N.B. Hui

the software or the hardware, the number of rules present in the Rule Base should be as small as possible. Consequently, some investigators have tried to design and develop a hierarchical FLC, in which the number of rules will be kept to the minimum [3, 4]. It has been observed that the performance of an FLC largely depends on the rule base and optimizing the data base is a ﬁne tuning process [5]. A fuzzy logic controller does not have an internal optimization module. An external optimizer is used to develop an optimal Knowledge Base through a proper tuning and this helps to improve the performance. In this chapter, the focus is on the issues related to design and development of an optimal fuzzy logic controller using diﬀerent optimization tools. Some of the applications of FLC are cited. The remainder of the text is organized as follows. Two major forms of FLC are discussed in Section 2. Various methods of designing optimal FLCs are given in Section 3. A summary of this work is presented in Section 4.

3.2 Two Major Forms of Fuzzy Logic Controller System modeling done by using the fuzzy set concept can be classiﬁed into two groups. That is linguistic fuzzy modeling and precise fuzzy modeling. Linguistic fuzzy modeling, such as Mamdani Approach is characterized by its high interpretability and low accuracy. The aim of precise fuzzy modeling such as Takagi and Sugeno’s Approach, is to obtain high accuracy at the cost of interpretability. Interpretability of a fuzzy modeling is deﬁned as a capability to express the behavior of a system in an understandable form. This is expressed in terms of compactness, completeness, consistency and transparency. The accuracy of a fuzzy model indicates how closely it can represent the system modeled. The working principles of both these approaches are brieﬂy explained below. 3.2.1 Mamdani Approach [2] An FLC consists of four modules namely, a fuzzy rule base, a fuzzy inference engine, fuzziﬁcation and de-fuzziﬁcation. Fig. 3.3 shows a schematic diagram explaining the working of an FLC. (a) The condition known as the antecedent and the action called the consequent variables needed to control a process are identiﬁed and measurements are taken of all the condition variables. (b) The measurements taken in the previous step are converted into appropriate fuzzy sets to express measurement uncertainties. This process is known as fuzziﬁcation. (c) The fuzziﬁed measurements are then used by the inference engine to evaluate the control rules stored in the fuzzy rule base and a fuzziﬁed output is determined.

3 Evolution of Fuzzy Controllers and Applications

Defuzzification module

Fuzzy rule base

51

Actions

Fuzzy inference engine

Process to be controlled

Fuzzification module

Conditions

FLC

Fig. 3.3. The working cycle of an FLC. µA1

µB1

µC1

Rule 1

s1 µA2

f

s2 µB2

µC2

Rule 2

s1*

s1

s2*

f

s2 µC

Uf

f

Fig. 3.4. A schematic diagram showing the working principle of an FLC.

(d) The fuzziﬁed output is then converted into a single crisp value. This conversion is called de-fuzziﬁcation. The de-fuzziﬁed values represent actions which need to be taken by the FLC in controlling the process. The fuzzy reasoning process is illustrated in Figure 3.4. Let us assume for simplicity that only two fuzzy control rules (out of many rules present in the

52

D.K. Pratihar and N.B. Hui

rule base) are being ‘FIRED’ as shown below. This is for a set of inputs – (s1∗ , s2∗ ). RULE 1: IF s1 is A1 and s2 is B1 THEN f is C1 RULE 2: IF s1 is A2 and s2 is B2 THEN f is C2. If s1∗ and s2∗ are the inputs for fuzzy variables s1 and s2. If µA1 and µB1 are the membership function values for A and B, respectively, then the grade of membership of s1∗ in A1 and the grade of membership of s2∗ in B1 are represented by µA1 (s1∗ ) and µB1 (s2∗ ), for rule 1. Similarly, for rule 2, where µA2 (s1∗ ) and µB2 (s2∗ ), are used to represent the membership function values. The ﬁring strengths of the ﬁrst and second rules are calculated as follows: α1 = min (µA1 (s1∗ ), µB1 (s2∗ )) , α2 = min (µA2 (s1∗ ), µB2 (s2∗ )) .

(3.1) (3.2)

The membership function of the combined control action C is given by µC (f ) = max (µ∗C1 (f ), µ∗C2 (f )) .

(3.3)

There are several methods of defuzziﬁcation (shown in Fig. 3.5). These are explained below. 1. Center of Sums Method: According to this method of defuzziﬁcation (refer to Fig. 3.5(a)), the crisp output can be determined by the following.

0.5

0.5 3

µ

µ

0.3

0.3 2 1

0.0

4 5 6

0.0 1 2 3 4 5 6 7 8 9 10 (a)

Output

1 2 3 4 5 6 7 8 9 10 (b)

0.5 µ

0.3

0.0 1 2 3 4 5 6 7 8 9 10 (c)

Output

Fig. 3.5. Diﬀerent methods of defuzziﬁcation.

Output

3 Evolution of Fuzzy Controllers and Applications

53

p Uf =

j=1 A(αj ) × fj p j=1 A(αj )

,

(3.4)

Where Uf is the output of the controller. A(αj ) represents the ﬁring area of the j-th rule. p is the total number of the ﬁred rules. fj represents the centroid of a membership function. 2. Centroid Method: The total area of the membership function distribution used to represent the combined control action is divided into a number of standard sub-areas. Their area and the center of area can be determined easily (refer to Fig. 3.5(b)). The crisp output of the controller can be calculated by using the expression given below. N Uf = i=1 N

Ai fi

i=1

Ai

,

(3.5)

Where N indicates the number of small areas or regions, Ai and fi represent the area and the center of area of i-th small region. 3. Mean of Maxima Method: From the membership function distribution of the combined control action, the range of the output variable is located. This is where the maximum value of the membership function is reached. The mid-value of this range is considered to be the crisp output of the controller (refer to Fig. 3.5(c)). 3.2.2 Takagi and Sugeno’s Approach [6] Here, a rule consists of the fuzzy antecedent and the functional consequent parts. Thus, a rule can be represented as follows: If x1 is Ai1 and x2 is Ai2 ..... and xn is Ain then y i = ai0 + ai1 x1 + . . . + ain xn where a0 , a1 , . . . , an are the coeﬃcients. In this way, nonlinear system is considered as a combination of several linear systems. Control action of i-th rule can be determined for a set of inputs (x1 , x2 , . . . , xn ) as follows. wi = µiA1 (x1 )µiA2 (x2 ) . . . µiAn (xn ),

(3.6)

Where A1 , A2 , . . . , An indicate the membership function distributions of the linguistic terms used to represent the input variables. The membership function value is given by µ. Thus, the combined control action can be determined as k wi y i y = i=1 , (3.7) k i i=1 w where k is the total number of rules.

54

D.K. Pratihar and N.B. Hui

3.3 Methods of Designing Optimal Fuzzy Logic Controllers In order to establish the input-output relationships of a process, a designer tries to design the KB of an FLC manually, based on a knowledge of the process. In most of the cases, it is diﬃcult to gather prior information of a process. The manually-designed KB of the FLC may not be optimal. As an FLC does not have a in-built optimizer, an optimization tool is used, while tuning a KB. Several methods have been developed and some of these are discussed below. 3.3.1 Least-square Method Attempts were made to determine an appropriate shape of the membership function distributions by using least-square methods. In this connection, see Pham and Valliappan [7], Bustince et al. [8]. The membership function distribution of a fuzzy set was assumed to follow a power function such as µA (xi ) = axbi . Here x indicates a variable represented by a fuzzy set A, i = 1, 2, . . . , n, n is the number of training cases, µA is the membership function value of the fuzzy set A lying between 0 and 1, a (greater than zero) and b are the constants to be determined by the least-square method. Two equations were solved for this [8]: n n

lnxi b = lnµA (xi ) (3.8) nlna + n

i=1

lnxi

lna +

n

i=1

i=1

ln2 xi

i=1

b=

n

lnxi lnµA (xi )

(3.9)

i=1

where axbi ≤ 1. 3.3.2 Genetic-Fuzzy System Genetic algorithm (GA) [9] is a population-based search and optimization technique based on the principle of natural selection and mechanics of natural genetics, was used by several researchers, for a genetic-fuzzy system. The performance of a Fuzzy Logic Controller (FLC) is dependent on its KB. Fig. 3.6 shows the schematic diagram of the genetic-fuzzy system. Here, a GA is used to determine optimal KB of the FLC. Thus, the GA improves the performance of the FLC. During optimization of the FLC, the feedback which is a deviation in prediction is calculated. This is based on a set of training cases and it is utilized as the ﬁtness of the GA. A GA is computationally expensive and the tuning is done oﬀ-line. Once optimized, the FLC will be able to predict the outputs for a set of inputs, within a reasonable accuracy

3 Evolution of Fuzzy Controllers and Applications

55

GA–based tuning Off–line Knowledge Base On–line

Inputs

Fuzzy Logic Controller

Outputs

Fig. 3.6. A schematic diagram showing a genetic-fuzzy system

limit. This concept has been used to solve a number of physical problems. See Karr [10], Thrift [11], Pham and Karaboga [14]. A detailed review on this scheme is done by Cordon et al. [12]. There are three basic approaches of this scheme, the Pittsburgh [13, 14], Michigan [15] and iterative rule learning [16, 17] approaches. In Pittsburgh approach, the entire rule base of the FLC is represented by a GA-string. Thus, the GA-population indicates the population of candidate rule sets. The genetic operators are used to modify the rule sets and obtain the optimal rule base. In the Michigan approach, members of the population are individual rules. Thus, a rule set is represented by the entire population. The main drawback of these two approaches lies in the fact that for the large number of fuzzy rules, the GA requires a huge amount of computer memory. To overcome the problem, using an iterative rule learning approach, chromosomes code individual rule, a new rule is added to the rule set, in an iterative fashion, for every run of GA. It requires a proper encoding scheme for extracting the rules from a chromosome. In this approach, the evolved RB of the FLC may contain some redundant rules, due to the iterative nature of the GA. A considerable amount of work has been carried out in this ﬁeld of research. Some of these attempts are mentioned below. Furuhashi et al. [18] developed a variable length decoding method, known as the Nagoya Approach. Using this approach, as the lengths of the chromosomes are not ﬁxed, it is diﬃcult to implement the necessary crossover operation in GA. Again the simultaneous design of the data base and rule base requires a proper optimization procedure. This can tackle both continuous as well as integer variables. Wang and Yen [19] proposed a method, in which a GA was used to extract the rule base, and the data base of an FLC was optimized using a Kalman ﬁltering technique. Farag et al. [20] developed a new multi-resolutional dynamic GA for this purpose. In this, the initial parameters of the data base of an FLC were determined by

56

D.K. Pratihar and N.B. Hui

using Kohonen’s self-organizing feature map algorithm and optimization was done by using a GA. Fuzzy rule generation and tuning using a GA was also tried by Ishibuchi et al. [21]. Recently, Abdessemed et al. [22] proposed a GAbased procedure for designing an FLC, to control the end eﬀector’s motion of a planar manipulator. Yupu et al. [23] used a GA to search for appropriate fuzzy rules. The membership function distributions were optimized by using a neural network. The FLC is becoming more popular nowadays, developing a suitable knowledge base for it, is not easy. The designer requires much time, to initially design the knowledge base (KB). It is further improved by using GA-based tuning. Thus, the designer must have a knowledge of the process to be controlled by the FLC. To overcome this requirement, a few investigators [24, 25] tried to automatically design the FLC by using a GA. Using search, the GA will develop the optimized data base and rule base for the FLC. A GA is basically a ﬁtness function-driven search method, therefore, it is blind for any other aspect that is not explicitly considered on ﬁtness function. Hence, a GA might evolve some redundant rules, that have limited inﬂuence on the process to be controlled. Redundant rules are to be removed to make the rule base compact. This makes the implementation of the controller easier, particularly when it is done by hard-ware. Thus, there is a need to determine the contribution of each rule. In this context, the work of Nawa et al. [26], Ishibuchi and Nakashima [27], Ghosh and Nath [28], Hui and Pratihar [35] are important. Nawa et al. [26] measured the quality of a rule by determining its accumulated truth value. The accumulated truth value was considered to be the sum of probability of occurrences of a rule in the training data. A rule is said to be good, if its accumulated truth value is high. Ishibuchi and Nakashima [27] made an attempt to assign an importance factor to each rule. They calculated the importance factor of a rule, by considering the way it interacts with the neighbors. An evolutionary technique was utilized to ﬁnd the interaction eﬀect. Ghosh and Nath [28] investigated the eﬀectiveness of a rule by measuring three parameters, namely support count, comprehensibility and interestingness. Support count of an item set is deﬁned by the number of records in the data base that contains all the items of that set. Comprehensibility is used to justify the understandability of a rule. A rule is said to be more comprehensive, if the number of attributes associated with the antecedent part of the rule is less and interestingness is represented by the probability of generating a rule during the learning process. It was a theoretical approach of ﬁnding interesting rules in the rule base and is unable to predict the importance of a rule for a ﬁxed number of attributes in both antecedent as well as in the consequent parts. The above methods considered the probability of occurrence of a rule only, for the determination of a good rule base. No attention was paid to calculate the contribution eﬀect of a rule with respect to a speciﬁc objective. Hui and Pratihar [35] proposed a method of determining importance factor for each rule contained in the RB of an FLC, to check the redundancy, if any. The importance factor of a rule is calculated

3 Evolution of Fuzzy Controllers and Applications

57

by considering its probability of occurrence and worth (goodness). A rule is said to be redundant and thus may be eliminated, if its importance factor comes out to be smaller than a pre-speciﬁed value and the removal of which does not lead to any non-ﬁring situation. The genetic-fuzzy system has been developed by the authors also, following the two diﬀerent approaches discussed below. Approach 1: GA-based tuning of the manually-constructed KB of the FLC. The KB of the FLC is designed manually and is based on the designer’s experience of the problem to be solved. But, it may not be optimal in any sense. GA-based tuning is adopted, to further optimize the KB, to improve the performance. As a GA is found to be computationally expensive, the GA-based tuning is carried out, oﬀ-line. During optimization, the GA-string will carry information for both the data base as well as the rule base. The GA-search will ﬁnd the optimal KB of the FLC. Once optimized, the FLC is able to determine its outputs in the optimal sense. Approach 2: Automatic design of KB using a GA. In Approach 1, much time is spent on manual design of the KB of an FLC. It might be diﬃcult beforehand to foresee the characteristics of the process to be controlled. Thus, designing a proper KB might be a diﬃcult task. To overcome this, a method for automatic design of the KB is developed by using a GA. Here the task of designing a suitable KB is given to the GA. The GA through its exhaustive search will try to determine the optimal KB of the FLC. The above concept has been used by the authors, to solve a number of physical problems. One of them is explained below. Optimal Path and Gait Planning of a Six-legged Robot A six-legged robot will have to plan its time-optimal, collision-free path as well as the optimal gait, setting simultaneously the minimum number of ground-legs having the maximum average kinematic margin. This is while moving on a ﬂat terrain with occasional hurdles, such as ditches and some moving obstacles. Its stability margin should always be positive to ensure static stability. This is a complicated task because the path planning and gait planning must be done simultaneously [29]. Fig. 3.7 shows the optimal path and gait for a six-legged robot. It has planned its optimal path and gait, after starting from an initial position S to reach the ﬁnal position G. It faces three moving obstacles and a ditch on the way to-wards its goal. The total movement of the robot has been achieved through a number of segments called motion segments. The robot plans its optimal path and gait on-line, for each motion segment. The robot shown in Fig. 3.7, is found to reach its goal in the time-optimal sense at 79-th motion segment, after avoiding collision with the moving obstacles and generating its optimal gaits.

58

D.K. Pratihar and N.B. Hui G G Ditch

Obs 2, 0.12 m/s

Ditch

Obs 2, 0.12 m/s

Obs 3, 0.15 m/s

Obs 3, 0.15 m/s

Obs 1, 0.1 m/s

Obs 1, 0.1 m/s

S

S Note: Positions at 10–th motion segment

Note: Positions at 20–th motion segment G

G Ditch

Obs 2, 0.12 m/s

Ditch Obs 2, 0.12 m/s

Obs 3, 0.15 m/s

Obs 1, 0.1 m/s

Obs 3, 0.15 m/s

Obs 1, 0.1 m/s

S

S

Note: Positions at 35–th motion segment

Note: Positions at 50–th motion segment

G Ditch G Ditch Obs 2, 0.12 m/s Obs 3, 0.15 m/s Obs 2, 0.12 m/s Obs 3, 0.15 m/s

Obs 1, 0.1 m/s Obs 1, 0.1 m/s

S

S Note: Positions at 65–th motion segment

Note: Positions at 79–th motion segment

Fig. 3.7. Optimal path and gaits of a six-legged robot obtained using the geneticfuzzy system [29]

3 Evolution of Fuzzy Controllers and Applications

59

3.3.3 Neural-Fuzzy System The purpose of developing a neural-fuzzy system is to improve the performance of an FLC by using neural network-based learning. It had been utilized by a number of researchers to solve a variety of problems. Some of these are mentioned below. Marichal et al. [30] proposed a neuro-fuzzy approach to generate the motion of a car-like robot navigating among static obstacles. In their approach, a least mean squared algorithm was used for the learning purposes and Kohonen’s self organizing feature map algorithm was considered to obtain the initial number of fuzzy rules and fuzzy membership function centers. They did not optimize the traveling time nor the approach was tested in a dynamic environment. Song and Sheen [31] suggested a pattern recognition approach based on a fuzzy-neuro network for the reactive navigation of a car-like robot. Li et al. [32] developed a neuro-fuzzy architecture for behavior-based control of a car-like robot, that navigates among static obstacles. The present chapter includes two schemes of neural-fuzzy system developed by the authors. These are discussed below [33]. Scheme 1: Neural-fuzzy system based on Mamdani Approach. In the developed neural-fuzzy system, a fuzzy logic controller using Mamdani Approach is expressed by utilizing the structure of a Neural Network (NN) and a back-propagation algorithm is utilized to optimize the KB of the FLC. The back-propagation algorithm is a steepest descent algorithm. Fig. 3.8 shows the schematic diagram of the ﬁve layer neural-fuzzy system– Layer 1 is the input layer, fuzziﬁcation is done in Layer 2, Layer 3 indicates the AND operation. The OR operation is carried out in Layer 4, and Layer 5 is the output layer. The training cases are passed through the network and the total error is calculated. The average error is propagated in the backward direction, to determine the updated weights. The network will try to ﬁnd an optimal set of weights, corresponding to which the error is minimum. Layer 1

Layer 2

[V]

1

Layer 4

FR

1 2

2

Layer 5

[W]

1

NR I1

Layer 3

1

LT

3 4

I2 2

VF

3

5

LT

1

6

AH RT

2

AH

O 1

7 2 3

8

3

RT

9

Fig. 3.8. A schematic diagram of the neural network-structured FLC

60

D.K. Pratihar and N.B. Hui

Three diﬀerent approaches to Scheme 1 are developed. These are discussed in brief below. Approach 1: NN-tuned FLC. The initial weights of the neural network representing the FLC are generated, at random. A batch mode of training is adopted. Training cases are passed through the NN (i.e., forward propagation) and average error is determined. As this error depends on the weights, it can be minimized by updating the weight values. A back-propagation algorithm is used to minimize the error. Approach 2: Genetic-Neural-Fuzzy system. In Approach 1, the error is minimized using a steepest descent method. This may have the local minima problems. To overcome this problem, the backpropagation algorithm is replaced by a GA. As GA is a populationbased search and optimization method, the chance of its solutions for getting trapped into the local minima is less. Thus, Approach 2 may be expected to perform better than Approach 1. Approach 3: Automatic design of neural-fuzzy system. To increase the search space of the GA, a method for automatic design of neural-fuzzy system is proposed. In this approach, the outputs of different rules are evolved solely by the GA itself. The GA through its exhaustive search, determines a good rule base for the FLC. There might be some redundant rules present in the GA-designed rule base. It may happen due to the iterative nature of the GA. To identify the redundant rules, a method is proposed, in which importance of a rule is decided by considering its frequency of occurrence and its worth with respect to the objective function of the optimization problem. Based on the value of this importance factor, a decision is taken whether a particular rule will be declared as redundant. Scheme 2: Neural-fuzzy system based on Takagi and Sugeno Approach. A neural-fuzzy system has been developed based on the Takagi and Sugeno Approach. This is known as the ANFIS (i.e., Adaptive Neuro-Fuzzy Inference Systems) [34]. An ANFIS is a multi-layered feed forward network, in which each layer performs a particular task. The layers are characterized by the fuzzy operations they perform. Fig. 3.9 shows the schematic diagram of the ANFIS structure, which consists of six layers – Layer 1 (input layer), Layer 2 (condition layer), Layer 3 (rule base layer), Layer 4 (normalization layer), Layer 5 (consequence layer) and Layer 6 (output layer). Let us assume that there are two inputs – I1 and I2 and one output O of the network. The ﬁrst two layers perform similar tasks to those done by Layers 1 and 2 of the neuro-fuzzy system developed in Scheme 1. The functions of other layers are explained below. Layer 3: This layer deﬁnes the rules of the fuzzy inference system. As three linguistic terms are used to represent each of the two inputs, there is a maximum of 3 × 3 = 9 rules present in the rule base. Each neuron in this layer represents a fuzzy rule and is termed as a rule node. The output of each neuron lying in this layer is the multiplication of their

3 Evolution of Fuzzy Controllers and Applications Layer 1

Layer 2 [V]

I1

I2

1

2

NR FR

1 2

VF

3

LT

1

AH RT

2 3

Layer 3

Layer 4

Layer 5

1

1

1

2

2

2

3

3

3

4

4

4

5

5

5

6

6

6

7

7

7

8

8

8

9

9

9

61

Layer 6

1

O

Fig. 3.9. A schematic diagram of the ANFIS architecture

respective two membership values. It is to be noted that each node output represents the ﬁring strength of a rule. Layer 4: This layer has the same number of nodes as the previous layer. It calculates the normalized ﬁring strength of each node. Layer 5: The output of a particular neuron (say, the q-th) lying on this layer is determined by O5q = (aq I1 + bq I2 + cq )

(3.10)

where (aq , bq , cq ) represents one set of coeﬃcients associated with the q-th node. Layer 6: The output of the node lying on Layer 6, can be determined by summing up all incoming signals. O61 =

R

O5q ,

(3.11)

q=1

where R indicates the total number of rules. A maximum of four rules (out of nine) will be ﬁred, for one set of input variables. The performance of an ANFIS depends on the selection of consequence parameters and premise parameters. That is the half base-widths of the input membership function distributions. For the selection of optimal parameters, a GA might be used together with the ANFIS. The developed neural-fuzzy systems have been used to plan collision-free, time-optimal paths of a car-like robot. This is explained below. 3.3.3.1 Collision-free, Time-optimal Path Planning for a Car-like Robot [33, 35] A car-like mobile robot needs to ﬁnd its time-optimal and collision-free path while navigating among some moving obstacles, and satisfy its kinematic (nonholonomic) constraints and dynamic constraints (such as sliding constraint,

62

D.K. Pratihar and N.B. Hui

motor torque constraint, curvature constraint). A detailed discussion on these constraints is beyond the scope of this chapter. Interested readers may refer to [33], for the same. The total path of the robot is divided into a number of distance steps having varying lengths, each of which is traveled during a time step. To calculate total traveling time of a robot to reach its destination, the time steps are summed and the time required to align its main axis towards the goal is added. There can be a saving in traveling time, particularly if the robot does not change its direction in two successive distance steps. It is subtracted from the total traveling time. The aim is to minimize the traveling time after ensuring a collision-free movement of the robot. A high positive value penalty is added to the total traveling time, if the robot collides with any one of the obstacles. Fig. 3.10 shows the near-optimal, collision-free paths of a robot in the presence of 16 moving obstacles. This is as obtained by using the three approaches of Scheme 1, Scheme 2 (explained above) and a

Fig. 3.10. Navigation of a robot among 16 moving obstacles

3 Evolution of Fuzzy Controllers and Applications

63

traditional motion planning scheme (potential ﬁeld method) [36]. The initial position, size, velocity and direction of movement of the obstacles are created at random. The planning robot starts from the point S and reaches the goal G, by avoiding collisions with the obstacles. Soft computing-based approaches have proved their supremacy over the potential ﬁeld method. It could be due to the reason that there is a chance that the solutions of the potential ﬁeld method will get trapped at the local minima. On the other hand, the chance of the solutions of GA-tuned fuzzy logic controller for getting trapped into the local minima is less and it could be due to an exhaustive search carried out by the GA. Moreover, the GA is able to inject adaptability to the FLC, which has been observed from the performances of Approaches 2 and 3 of Scheme 1. Approach 3 of Scheme 1 is found to be the best of all approaches. It could be due to the fact that using this approach, a good KB of the FLC is evolved by the GA, after carrying the search in a wider space. 3.3.4 Optimization of FLC Using Ant Colony Optimization [37] In Ant Colony Optimization (ACO) algorithm, an optimization problem is represented in the form of a graph – G = (C, L). Here, C is the set of components of the problem and L indicates the possible connection or transition among the elements C. The solutions are expressed in terms of feasible paths on the graph G, after satisfying a set of constraints. Thus, the Fuzzy Rule Learning Problem (FRLP) using the ACO, is formulated as a combinatorial optimization problem. Its operational mode is composed of two stages: in the ﬁrst stage, the number and antecedents of the linguistic rules are deﬁned, and a set of consequent candidates is assigned to each rule. In the second stage, a combinatorial search is carried out to ﬁnd the best consequent of each rule, according to a global error measure over the training set. The ﬁtness of a solution consists of two parts, namely the functional ﬁtness and the objective ﬁtness. The functional ﬁtness deals with the functionality of the solutions. That is how good is the solution. The objective ﬁtness is the measure of the quality of the solution, in terms of optimization objectives, such as area, delay, gate count, power consumption, and others. To apply ACO algorithm to a speciﬁc problem, the following steps need to be followed: – Represent the problem in the form of a graph or a similar easily covered structure, – Deﬁne the way of assigning a heuristic preference to each choice that needs to be taken in each step in order to generate the solution, – Establish an appropriate way of initializing the pheromone, – Deﬁne the ﬁtness function to be optimized, – Select an ACO algorithm to determine the optimal solutions. The Fuzzy Rule Learning Problem (FRLP) aims to obtain the rules combining the labels of the antecedents and to assign a speciﬁc consequent to

64

D.K. Pratihar and N.B. Hui

each antecedent combination. This problem is interpreted as a way of assigning consequents to the rules with respect to an optimality criterion. An ant iteratively goes over each rule and chooses a consequent with a probability that depends on the pheromone trail τij and the heuristic information ηij . 3.3.5 Tuning of FLC Using Reinforcement Learning Fuzzy rules for control can be eﬀectively tuned by means of reinforcement learning. In this approach, the rules with their associated antecedent and consequent fuzzy sets are represented with the help of a fuzzy-neural network. For this an action selection network (ASN) is used. This network provides continuous action value and records the state of the environment and also determines the next action required. Thereafter, the actions are evaluated by means of a critic element (CE), which is a two-layer feed forward action evaluation network (AEN). It predicts the reinforcements associated with diﬀerent input states and whether or not a failure has occurred. If a failure occurs, it identiﬁes the steps leading to the failure and modiﬁes the fuzzy sets associated with the rules. A gradient descent technique in conjunction with an average reward is used to train both the action selection network (ASN) and the action evaluation network (AEN) over a set of trials. During training, a reward is provided until a failure occurs and then a high value penalty is given. This approach had been used by Berenji and Khedkar [38], to solve the problem of a cart-pole balancing system. 3.3.6 Optimization of FLC Using Tabu Search Denna et al. [39] presented an approach for automatic deﬁnition of the fuzzy rules based on the Tabu Search (TS) algorithm. To determine the most appropriate rule base for solving the problem, they employed the reactive form of TS algorithm. To apply the Reactive Tabu Search (RTS) algorithm in determining the rules of a fuzzy controller, the consequent of each rule is expressed with a binary string. The learning procedure is shown in Figure 3.11. The learning begins with an initial rule base, chosen randomly at each iteration. Initial states can also be selected by following a uniform distribution over the entire state space. In such conditions, regions of interest are assigned a higher probability during the learning procedure. Performance of the rule base is then evaluated by using an error function E(•), over a set of typical control rules. It is important to mention that during the evaluation of E(•), some rules, with a smaller contribution to the system are not used. This procedure continues until a termination criterion is reached. The termination condition for each execution may be based on the following parameters: – The number of iterations carried out, – The Current State of the error function, – The properties of the solution found.

3 Evolution of Fuzzy Controllers and Applications

rules represented as a binary string 0010....110..01...1

65

Fuzzy Controller Rules

typical control cases Reactive Tabu Search

Evaluation

Model

E( )

Fig. 3.11. A schematic diagram showing the learning of fuzzy rules using Tabu search [39]

Bagis [40] described a method for the determination of optimum fuzzy membership function distribution used for controlling a reservoir system of dams during ﬂoods. 3.3.7 Design of a Fuzzy Controller using the Taguchi Method The Taguchi Method determines the parameter settings, which maximize the signal to noise (S/N ) ratio in each problem by systematically performing the designed experiment. The designed experiment is composed of an inner array and an outer array. The inner array is a designed experiment using the control factors and the outer array consists of the noise factors. To design an FLC using the Taguchi method, control factors are considered as the membership parameters and diﬀerent system conditions are assumed to be the noise factors. If the inner array is made up of m rows and the outer array contains n rows, then each of the m rows can obtain n performance characteristics. These n data are used to calculate the S/N ratio, for each row of the inner array. The optimal parameter settings are determined by analyzing the S/N ratio data. To check the adequacy of the model, Analysis of Mean (ANOM) and Analysis of Variance (ANOVA) are carried out. Later, a veriﬁcation experiment is conducted to test the performance of the model. Kim and Rhee [41] utilized the Taguchi method, to design a suitable fuzzy logic controller, in which the following steps were used: – – – –

Identify the performance characteristic to be observed, Identify important noise factors and their ranges, Identify the control factors and their levels, Construct the inner array and the outer array,

66

D.K. Pratihar and N.B. Hui

– Conduct the designed experiment, – Analyze the data and determine optimal levels for the control factors, – Conduct the veriﬁcation experiment. 3.3.8 Fuzzy Logic Controller Tuned by Simulated Annealing Simulated Annealing (SA) is one of the most popular non-traditional methods of optimization, in which the cooling process of a molten metal has been artiﬁcially modeled. Alfaro and Garcia [42] described a method for development of a fuzzy logic controller applied to path planning and navigation of mobile robots, by using a simulated annealing. Most of the researchers tried to optimize the membership function distributions of the FLC by utilizing the SA. In this approach,the cost function was deﬁned as follows: F =

N 1 2 (yk − yˆk ) , N

(3.12)

k

where k = 1, 2, . . . , N . N is the number of learning samples, (xk , yk ) is the k th learning sample and yˆk is the output of the fuzzy system corresponding to the input vector xk . The optimization algorithm tunes the parameters (spread and shape) of membership function distributions. This is in order to minimize the cost function. Consider the membership functions of the input variables to be Gaussian in nature, as shown below. Gaussian(x; σ; c) = e−(

x−c 2 σ )

(3.13)

where c and σ indicate the Center and Width, respectively, of the membership function distribution. In SA, the following steps are to be considered in order to optimize the Gaussian membership function distribution: 1. Set an Initial Temperature T to a high value and generate initial parameters cij and σji , randomly and compute the cost function (Fold ). 2. Generate a set of new parameters cij and σji and compute the new cost function (Fnew ). Obtain the change in the cost function δ = Fnew − Fold . If δ < 0, memorize the new set of membership functions and proceed until the termination criterion is reached. Otherwise, go to Step 3. 3. If δ > 0 and probability of accepting the new set of membership functions P (δ) = exp(−δ/T ) ≤ random[0, 1]), the center and width values are not changed. Now, go to Step 2 by reducing the temperature T to the half of its previous value. 4. Repeat Steps 2 and 3 until an acceptable solution has been found or until a speciﬁed number of iterations has been reached.

3 Evolution of Fuzzy Controllers and Applications

67

3.4 Summary Fuzzy logic controllers have proved their worth and are popular nowadays to solve real-world complex problems. As the performance of an FLC depends on its KB, several attempts had been made to design a suitable KB. Several methods had been tried by various investigators, to solve the problem. There is a chance of further improvement and much further work is necessary. Both the Linguistic as well as Precise Fuzzy Modeling have been used separately, to solve a variety of problems and some satisfactory results have been obtained. Linguistic fuzzy modeling ensures better interpretability, but precise fuzzy modeling aims to achieve higher accuracy. It is obvious that as interpretability of the fuzzy model increases, its accuracy will decrease and vice-versa. Thus, depending on the physical problem, a particular type of fuzzy modeling is chosen. It is challenging to obtain a proper balance between interpretability and accuracy of a fuzzy model. These two properties are inversely related and it is important to investigate as to whether a pareto-optimal front exists.

References 1. Zadeh, L.A.: Fuzzy sets. Information and Control 8 (1965) 338–353 2. Mamdani, E.H., Assilian, S.: An experiment in linguistic synthesis with a fuzzy logic controller. International Journal on Man-Machine Studies 7 (1975) 1–13 3. Wang, L.X.: Analysis and design of hierarchical fuzzy systems. IEEE Trans. on Fuzzy Systems 7 5 (1999) 617–624 4. Lee, M.L., Chung, H.Y., Yu, F.M.: Modeling of hierarchical fuzzy systems. Fuzzy Sets and Systems 138 (2003) 343–361 5. Pratihar, D.K., Deb, K., Ghosh, A.: A genetic-fuzzy approach for mobile robot navigation among moving obstacles. International Journal of Approximate Reasoning 20 (1999) 145–172 6. Takagi, T., Sugeno, M.: Fuzzy identiﬁcation of systems and its application to modeling and control. IEEE Trans. on Systems, Man and Cybernetics SMC-15 (1985) 116–132 7. Pham, T.D., Valliappan, S.: A least square model for fuzzy rules of inference. Fuzzy Sets and Systems 64 (1994) 207–212 8. Bustince, H., Calderon, M., Mohedano, V.: Some considerations about a least square model for fuzzy rules of inference. Fuzzy Sets and Systems 97 (1998) 315–336 9. Goldberg, D.E.: Genetic algorithms in search, optimization, machine learning. Addison-Wesley, Reading, Mass, USA (1989) 10. Karr, C.: Genetic algorithms for fuzzy controllers. AI Expert (1991) 38–43 11. Thrift, P.: Fuzzy logic synthesis with genetic algorithms, Proc. of Fourth International Conference on Genetic Algorithms. (1991) 509–513 12. Cordon, O.: Gomide, F., Herrera, F., Hoﬀmann, F., Magdalena, L.: Ten years of genetic-fuzzy systems: current framework and new trends. Fuzzy Sets and Systems 141 (2004) 5–31

68

D.K. Pratihar and N.B. Hui

13. Hoﬀman F., Pﬁster, G.: Evolutionary design of a fuzzy knowledge base for a mobile robot. Intl. Jl. of Approximate Reasoning 17 4 (1997) 447–469 14. Pham, D.T., Karaboga, D.: Optimum design of fuzzy logic controllers using genetic algorithms. Journal of Syst. Engg. 1 (1991) 114–118 15. Ishibuchi, H., Nakashima, T., Murata, T.: Performance evaluation of fuzzy classiﬁer systems for multidimensional pattern classiﬁcation problems. IEEE Trans. on Systems Man and Cybernetics 29 (1999) 601–618 16. Cordon, O., DeeJesus, M.J., Herrera, F., Lozano, M.: MOGUL: a methodology to obtain genetic fuzzy rule-based systems under the iterative rule learning approach. Intl. Jl. of Intelligent Systems 14 11 (1999) 1123–1153 17. Gonzalez, A., Perez, R.: SLAVE: a genetic learning system based on an iterative approach. IEEE Trans. on Fuzzy Systems 7 2 (2001) 176–191 18. Furuhashi, T., Miyata, Y., Nakaoka, K., Uchikawa, Y.: A new approach to genetic based machine learning and an eﬃcient ﬁnding of fuzzy rules-proposal of Nagoya approach. Lecture notes on Artiﬁcial Intelligence 101 (1995) 178–189 19. Wang, L., Yen, J.: Extracting fuzzy rules for system modeling using a hybrid of genetic algorithms and Kalman ﬁltering. Fuzzy Sets and Systems 101 (1999) 353–362 20. Farag, W.A., Quintana, V.H., Lambert-Torres, G.: A genetic-based neuro-fuzzy approach for modeling and control of dynamical systems. IEEE Trans. on Neural Networks 9 (1998) 576–767 21. Ishibuchi, H., Nil, M., Murata, T.: Linguistic rule extraction from neural networks and genetic algorithm-based rule selection, Proc. of IEEE Intl. Conf. on Neural Networks. Houston, TX (1997) 2390–2395 22. Abdessemed, F., Benmahammed, K., Monacelli, E.: A fuzzy-based reactive controller for a non-holonomic mobile robot. Robotics and Autonomous Systems 47 (2004) 1–22 23. Yupu, Y., Xiaoming, X., Wengyuan, Z.: Real-time stable self learning FNN controller using genetic algorithm. Fuzzy Sets and Systems 100 (1998) 173–178 24. Angelov, P.P., Buswell, R.A.: Automatic generation of fuzzy rule-based models from data by genetic algorithms. Information Science 50 (2003) 17–31 25. Nandi, A.K.: Pratihar, D.K.: Automatic design of fuzzy logic controller using a genetic algorithm-to predict power requirement and surface ﬁnish in grinding. Journal of Materials Processing Technology 148 (2004) 288–300 26. Nawa, N.E., Hashiyama, T., Furuhashi, T., Uchikawa, Y.: A Study on fuzzy rules discovery using pseudo-bacterial genetic algorithm with adaptive operator, Proc. IEEE Int. Conf. Evolutionary Computation. Indianapolis, USA, (1997) 13–16 27. Ishibuchi H., Nakashima, T.: Eﬀect of rule weights in fuzzy rule-based classiﬁcation systems. IEEE Trans. on Fuzzy Systems 9 4 (2001) 506–515 28. Ghosh A., Nath, B.: Multi-objective rule mining using genetic algorithms. Information Sciences 163 (2004) 123–133 29. Pratihar, D.K., Deb, K., Ghosh, A.: Optimal path and gait generations simultaneously of a six-legged robot using a GA-Fuzzy approach. Robotics and Autonomous Systems 41 1 (2002) 1–20 30. Marichal, G.N., Acosta, L., Moreno, L., Mendez, J.A., Rodrigo, J.J., Sigut, M.: Obstacle avoidance for a mobile robot: a neuro-fuzzy approach. Fuzzy Sets and Systems 124 (2001) 171–170 31. Song, K.T., Sheen, L.H.: Heuristic fuzzy-neuro network and its application to reactive navigation of a mobile robot. Fuzzy Sets and Systems 110 (2000) 331–340

3 Evolution of Fuzzy Controllers and Applications

69

32. Li, W., Ma, C., Wahl, F.M.: A Neuro-fuzzy system architecture for behavior based control of a mobile robot in unknown environments. Fuzzy Sets and Systems 87 (1997) 133–140 33. Hui, N.B., Mahendar, V., Pratihar, D.K.: Time-optimal, collision-free navigation of a car-like mobile robot using a neuro-fuzzy approach. Fuzzy Sets and Systems 157 16 (2006) 2171–2204 34. Jang, J.S.R., Sun, C.T., Mizutani, E.: Neuro-Fuzzy and Soft Computing. Prentice-Hall of India Pvt. Ltd., New Delhi (2002) 35. Hui, N.B., Pratihar, D.K.: Automatic design of fuzzy logic controller using a genetic algorithm for collision-free, time-optimal navigation of a car-like robot. International Journal of Hybrid Intelligent Systems 5 3 (2005) 161–187 36. Latombe, J.C.: Robot motion planning. Kluwer Academic Publishers (1991) 37. Casillas, J., Cordon, O., Herrera, F.: Learning fuzzy rules using ant colony optimization algorithms, Proc. of 2nd Intl. Workshop on Ant Algorithms. Brussels, Belgium (2000) 13–21 38. Berenji, H., Khedkar, P.: Learning and tuning fuzzy controllers through reinforcements. IEEE Transactions on Neural Networks 3 5 (1992) 724–740 39. Denna, M., Mauri, G., Zanaboni, A.M.: Learning fuzzy rules with Tabu search– an application to control. IEEE Transactions on Fuzzy Systems 7 2 (1999) 295–318 40. Bagis, A.: Determining fuzzy membership functions with tabu search – an application to control. Fuzzy Sets and Systems 139 (2003) 209–225 41. Kim, D., Rhee, S.: Design of a robust fuzzy controller for the arc stability of CO2 welding process using the Taguchi method. IEEE Transactions on Systems, Man and Cybernetics–Part B 32 2 (2002) 157–162 42. Alfaro, H.M., Garcia, S.G.: Mobile robot path planning and tracking using simulated annealing and fuzzy logic control. Expert Systems with Applications 15 (1998) 421–429

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design: An Application to Promoter Recognition in DNA Sequences Romesh Ranawana and Vasile Palade University of Oxford Computing Laboratory, Oxford, United Kingdom January 16, 2006 [email protected] [email protected] Summary. This chapter presents a novel methodology for the customization of neural network based multi-classiﬁers used for the recognition of promoter regions in genomic DNA. We present a framework that utilizes genetic algorithms (GA’s) for the determination of optimal neural network parameters for better promoter recognition. The framework also presents a GA based method for the combination of the designed neural networks into the multi-classiﬁer system.

4.1 Introduction All sequenced genomes1 are extremely data-rich. The information contained within these sequences includes the answer for a lot of the problems faced by humans, but, due to their complexity, it is near impossible to analyze them using traditional methodologies. Thus, to fully realize the value of any sequenced genome and to fully understand it, eﬃcient and intelligent computational methods are needed in order to identify the biological relevant features in the sequences and to provide insight into their structure and function. Due to the diﬃculty involved in deﬁning these features in a manner understandable to a computer algorithm, designers are being forced into using machine learning techniques for their implementation. Machine learning algorithms allow us to design eﬃcient programs for pattern recognition tasks which do not have well deﬁned data sets. Some of the recently used methodologies include the usage of neural networks [13, 30, 31], decision trees [27], genetic programming [10], hidden markov models [2, 9, 11, 12] and fuzzy logic [17, 32].

1

The genome of an organism is its set of chromosomes, containing all of its genes and associated DNA. Please see Section 4.2 for further details.

R. Ranawana and V. Palade: A Neuro-Genetic Framework for Multi-Classiﬁer Design: An Application to Promoter Recognition in DNA Sequences, Studies in Computational Intelligence (SCI) 66, 71–94 (2007) c Springer-Verlag Berlin Heidelberg 2007 www.springerlink.com

72

R. Ranawana and V. Palade

One observation that has been made with the use of these types of methodologies for gene recognition is the diﬀerence in results obtained through the use of each type of algorithm. Although each type of system provides similar results in terms of the accuracy displayed, slight diﬀerences in the correct and incorrect genes recognized have been observed. Thus, it can be concluded that each methodology focuses the system to specialize in speciﬁc types of genes. It is due to this reason that hybrid multi-classiﬁers have now become increasingly popular in the implementation of pattern recognition systems. These hybrid multi-classiﬁer systems are implemented by building a series of classiﬁers to solve the same task, but by using a diﬀerent implementation algorithm for each of them. The results of each of these individual classiﬁers are then combined to obtain the ﬁnal result of the classifying system. Due to their robustness and their remarkable ability in deriving meaning from complicated or imprecise data, neural networks can be used to extract patterns and detect trends that are too complex to be noticed by either humans or other computer based techniques. Thus, they have been used extensively in the solving of problems related to bioinformatics. Nevertheless, one of the central dogmas involved in the design of a neural network for the solving of any problem lies in the decision on the characteristics or parameters that deﬁne the network. In this chapter, we present a framework which can be utilized for the automatic design and implementation of a neural network based multi-classiﬁer system for the identiﬁcation of protein coding regions (i.e. genes) in strings of DNA.

4.2 Biological Background 4.2.1 Genes and Promoters Gene expression involves the usage of the information contained within a gene to produce a protein, which in turn leads to the production of the phenotype determined by that particular gene. The production of these proteins is a multi-step process, which begins with transcription and translation and is followed by folding, post-translation modiﬁcation and targeting. During transcription, the DNA is copied into RNA2 by a polymerase called RNA polymerase (RNAP). This transcription yields mRNA, which is the ﬁrst step of the protein production process. Promoters within genomes sequences are sequences of DNA located upstream of protein coding regions. They are usually short sequences and are bounded by speciﬁc transcription factors3 . RNAP has the ability to recognize 2

3

Is a nucleic acid, and is used as intermediatory during the production of proteins from genomic information. A transcription factor (TF) is a protein which binds onto speciﬁc DNA sequences on a genome sequence.

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design Promoter

upstream

5’

73

-35

-10

+1

T

3’

mRNA

Fig. 4.1. A simple Prokaryotic Gene model

promoters of protein coding genes. Thus, RNAP uses the successful identiﬁcation of promoters for the identiﬁcation of protein coding regions. This paper uses the same method for the identiﬁcation of genes in genomic DNA. The proposed system works on the premise that the successful identiﬁcation of a promoter region leads to the successful identiﬁcation of the start position of a gene. Figure 4.1 displays a simple model of a Prokaryotic gene. The ‘−35’ positions and ‘−10’ positions located within the promoter. The ‘+1’ position marks the beginning of the gene and the ‘T’ signiﬁes the ‘terminator’ which marks the end of the gene. It is the DNA between these two positions which is transcribed in to mRNA, which is turn used for the production of the protein. ’In prokaryotes, the sequence of a promoter is recognized by the Sigma(σ) factor of the RNA polymerase. These sites are represented by the ‘−35’ and ‘−10’ symbols in Figure 4.1. E.Coli promoters are composed of two of these sites4 . These are the locations to which E.Coli polymerase, the RNAP used by E.Coli, binds onto in order to begin the transcription of protein. These two binding sites are always located at the points known as the −35 hexamer box and the −10 hexamer box. Consensus sequences have been identiﬁed for each each of these two locations. A consensus sequence is the most probably sequence to occur at a certain position. The spacer between the −10 hexamer box and the transcriptional has a variable length, the most probable length being 7. The spacer between the −10 site and the −35 site is also of variable length and can vary between 15 and 21 bases. An “ideal” E.Coli promoter would look something like the sequence displayed in Figure 4.2. It displays the most probable sequences that could occur at the two hexamer boxes. It is this variation that can make the recognition of these promoters diﬃcult with traditional methodologies. Many promoter sequences have the pyrimidine5 at position +1 (one nucleotide upstream of the transcription start site or the gene), and purine6 at the transcriptional start site. In addition to the more obvious features described, a few more non-obvious features have been observed by [5] and [16]. [13] used the software available at http://www-lecb.ncifcrf.gov/toms/ delia.html to display the sequence logos of 438 E.Coli promoters that were aligned according to their transcriptional start sites. Their results are displayed 4 5 6

These sites are also known as binding sites. C or T A or G

74

R. Ranawana and V. Palade

5’

TTGACA -35 site

17 bp

TATAAT

7 bp

+1

3’

-10 site

Fig. 4.2. A simple Prokaryotic Gene model

Fig. 4.3. A ﬂow-chart of the framework utilized for the development of the system

in Figure 4.3. This software used the Shannon Entropy to measure the nonrandomness of the individual bases at each position independently. We use these observations later in the paper to display the ability of our neural networks to learn these distributions automatically. 4.2.2 The Importance of Studying the E.Coli Gene E.Coli K-12 and its close relatives have been extensively studied in genetic, molecular and biochemical studies from the dawn of modern microbiology and molecular biology. The development of a ‘virtual cell’ maybe far into the future, but it is a well known concept that the place to start the consideration of what will work and how to set about it lies in the development of a simple but adaptable system. In this regard, E.Coli K-12 provides an opportunity to test out approaches to modelling cellular and organismal systems. The Bioinformatics Initiative7 lists a set of reasons giving out reasons as to why E.Coli is currently the organism of choice for such a study, both as a reference organism for many studies on prokaryotic systems and as a source of information on proteins and metabolic pathways that are shared by eukaryotics as well. 4.2.3 Methods Used for the Identiﬁcation of E.Coli Promoters Most of the methods that have been used for the detection of E.Coli promoters involve the alignment of all the given sequences at the transcriptional start site 7

http://www.ecoli.princeton.edu/E coli Bioinformatics Document.pdf

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

75

(i.e. the ‘+1’ position) and then scanning them for the consensus sequences. They look for the primary consensus sequences at positions ‘−35’ and ‘−10’. Some of the more successful methods scan for some of the weaker motifs that have been identiﬁed at positions ‘+1’, ‘−22’ and the ‘−44’ position. Some examples of the use of these methods are listed out in [20] and [26]. Each of these systems programmed known knowledge about the E.Coli promoter into the system in order correctly classify unknown sequences. The main problem associated with these systems and other systems which use this technique is the uncertainty involved in the placement of the consensus sequence within the promoter. This can be caused due to the consensus sequence varying from the typical or by the consensus sequence being placed at a slightly diﬀerent position. The spacer between consensus sequences is also variable. Another method involves the use of a scoring matrix to calculate the variation of the diﬀerent bases within the sequence. An example of such an implementation can be found at [4]. Neural networks have also been successfully utilized for the recognition of these E.Coli promoters [13, 14, 30]. Statistical methods like Hidden Markov Models and the EM algorithm have also been successfully utilized for this purpose. The system presented in this paper also attempts to solve this problem, but utilizes a hybridization of a few of the methods mentioned above to achieve the objective. We built nine neural networks, each of which was individually trained and customized for better performance. These neural networks were then combined using a genetic algorithm. The results show us that the customized neural networks were able to identify the consensus sequences and identify important subsections of the DNA sequences shown to them. Manual entry of the consensus sequence information was not required. 4.2.4 The Data Set The data set used consisted of 473 E.Coli sequences and 600 non-promoter sequences. The positive data sequence was obtained from the PromEC8 website. Each of the obtained E.Coli sequences were 101 nucleotides in length and were aligned at the transcriptional start site. Each of the transcriptional start sites were positioned such that they each appeared at position 76 of the string. Thus, each of the positive sequences consisted of nucleotides from 101 nucleotides starting from the −75 position, 75 nucleotides upstream of the transcriptional start site, and ending at the +25 position, 25 nucleotides downstream of the transcriptional start site. The negative data set was obtained from the database compiled by the E.Coli GenomeProject9 , at the University of Wisconsin-Madison. These negative data consisted of E.Coli genes with the preceding promoter region deleted. Using the above mentioned data set, we constructed three data sets: a training sequence, a positive test sequence and a negative test sequence. 8 9

http://bioinfo.md.huji.ac.il/marg/promec/ http://www.genome.wisc.edu/sequencing/k12.htm#seq

76

R. Ranawana and V. Palade

The training set consisted of 314 positive sequences and 429 negative sequences. The remaining positive and negative data were divided up into a positive test set of 159 sequences and a negative test set of 171 sequences. For the purpose of initial training and testing the originating neural networks, all three data sets were encoded using −2, −1, 1 and 2 to represent the bases A, T, C and G, respectively.

4.3 Multi-Classiﬁer Systems 4.3.1 Classiﬁers Classiﬁcation, in computer science terminology, is the act of placing a given data item within a set of categories based on the properties of that particular data item. A classiﬁer is a computer based agent who would perform this classiﬁcation. For any pattern classiﬁcation application research area currently under investigation, there are many possible classiﬁcation algorithmic solutions which could be utilized. Any learning problem involves the usage of a set of training examples of the form {(x1 , y1 ), ..., (xm , ym )} for the approximation of a function f (x). The x values are usually of the form < xi,1 , xi,2 , ..., xi,n >, and are composed of either real or discrete values. The y values are the expected outputs for the corresponding x vales, and are usually drawn from discrete sets of classes. Thus, when presented with the training set, the learning paradigm attempts to approximate the function f (x) and produces a classiﬁer. The produced function is an approximation of the function f (x) as the training set, more often than not, contains random noise. Thus, with the use of multiple learning paradigms, one obtains a set of classiﬁers h1 , h2 , ..., hL , each of which attempts to approximate the function f (x). Due to diﬀerences between the learning paradigms, the way that the approximation is evolved also diﬀers. Thus, the results obtained through each classiﬁer can also be diﬀerent. 4.3.2 Why use Multiple Classiﬁers? An ensemble of classiﬁers or a multi classiﬁer, is a collection of classiﬁers whose individual decisions are combined in some way to obtain a combined decision. [8] claims that such an ensemble can only be more successful than the individual classiﬁers if the individual classiﬁers disagree with each other. As mentioned in [33], there are two main groups of methodologies for the combinations of classiﬁers, namely, feature-vector-based methods10 and syntactic-and-structural methods11 . In addition to this distribution, each group is composed of a large variety of identiﬁable subgroups. For many of the 10 11

E.g. Neural networks E.g. Fuzzy rule bases

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

77

classiﬁcation problems encountered, it has been experimentally proven that diﬀerent classiﬁers exhibit diﬀerent degrees of success. No classiﬁer has yet has proven to be totally perfect in the classiﬁcation of any problem domain. Often, it has also been observed that diﬀerent classiﬁers can disagree with each other on classiﬁcation decisions when presented with the same training data and testing data. Thus, Individual classiﬁer models are recently being challenged by combined pattern recognition systems, which often show better performance. This is due to the reason that, when a system that incorporates many methodologies is used, the inadequacies of one methodology are usually nulliﬁed by the characteristics of another. That is, a multiple classiﬁer can nullify the predictive inaccuracy obtained through the user of singular classiﬁcation systems. Thus, genes not recognized by one classiﬁer will be recognized by another and genes incorrectly recognized by one classiﬁer will be rejected by another (Figure 4.4). The challenge in building such a system is the selection of a combination method for the results provided by each classiﬁer in order to come up with the optimal and most popular result. Here, the performance of the entire system can be proven to be never much worse than that of the best expert [8]. [3] also states three causes as to why multiple classiﬁers exhibit better performance when compared to singular systems, and then goes on to state that the use of multiple classiﬁcation solutions and their combinations leads to the nulliﬁcation of many of the problems encountered through the use of singular classiﬁers. In a situation where a number of classiﬁers are available, the simplest approach would be to select the best performing classiﬁer and use it for the classiﬁcation task [18]. This approach, although simple and easy to implement,

Classifier 2 Classifier 1

Classifier 3

Input Space

True Positives True Positives identified by the classifier

Fig. 4.4. A Classiﬁer problem with diﬀerent classiﬁers being used for the same data set. The shaded region represents the true positives and the ellipses represent the space classiﬁed by each classiﬁer as being true positives

78

R. Ranawana and V. Palade

does not guarantee good performance [22]. It is highly probable that a combination of classiﬁers would outperform a singular classiﬁer [25]. On the other hand, diﬀerent and worse performing classiﬁers might only add to the complexity of the problem and provide even worse result than the worst classiﬁer. Thus, it is a well known fact that if a multi-classiﬁer system is to be successful, the diﬀerent classiﬁers should have good individual performances and be suﬃciently diﬀerent from each other [28]. But, neither individual performances [21, 35] nor the diversity [24, 29] provides an ideal measure of how successful the combination of the classiﬁers will be. As explained in [25], the core element of classiﬁer selection is the selection criteria. The most natural choice being the combined performance, which will also the criterion for selection of the combiner. The only drawback of this methodology is the exponential complexity of testing out all possible combinations of a given set of classiﬁers. It has been proven that the usage of an ensemble of neural networks for certain classiﬁcation problems can improve classiﬁcation performance, when compared to the use of singular neural networks. [19] and [23] provide results tested on protein secondary structure prediction. [1] also lists out an overview of applications in molecular biology. Of the diﬀerent methodologies used for the combination of multiple neural net classiﬁers, majority voting, neural networks, Bayesian inference and the Dempster-Shafer theories have proven the most popular [15, 33, 34]. The Dempster-Schafer method has proven to be successful, but has a considerable dependency on the function used for the alignment of probability. For the type of output produced by neural networks, posterior class-conditional probabilities can be calculated. The calculation of these probabilities becomes relatively simple, specially when the number of output classes is small. In this Chapter, we test two methods (LAP and LOP) that can be used for the combination of results obtained through multiple neural network classiﬁcations and are compared with the use of a variation of the two methods, the LOP2 method. These two methods, the LAP and LOP methods, were introduced in (Hansen and Krogh, 1999). 4.3.3 The LAP and LOP Methods for Combining Classiﬁers Introduction to the LAP and LOP Methods [7] presents a general composition method that considers the average of results of several neural networks to obtain a combined result that would improve the overall classiﬁcation performance. The method that we used for the combination of the classiﬁers designed in this research was a variation of the LOP (Logarithmic Opinion Pool) [7] method, LOP2. We compare this method with the LAP (Linear Average Predictor) method, which is more commonly used. The LOP method is a general ensemble method which shows how the error or an ensemble can be written

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

79

as the average error between members (called the ensemble ambiguity). [7] claims that this proves that ensembles always improve average performance. Description of the Ensemble Composition Method used in this Research An ensemble consists of M predictors fi , which are combined into the combined predictor F. Let each predictor produce an output in the form of a probability vector (fi1 ,...,fiN ), where is the estimated probability that input belongs to class cj . There is also a coeﬃcient M associated with each ensemble M of the form (α1 ,...,αM ) where i=1 αi = 1. The LAP Method The combined predictor for the LAP of the ensemble is then deﬁned in Equation (4.1). j = FLAP

M

αi fij

(4.1)

i=1

The LOP Method The combined predictor for the LOP of the ensemble is then deﬁned in Equation (4.2), where Z is a normalization factor given by Equation (4.3). M

1 j j α1 logfi (4.2) FLOP = exp Z i=1 M N

j Z= exp α1 logfi (4.3) j=1

i=1

The LOP2 Method The LOP algorithm was changed by removing the condition (α1 ,...,αM ) M where i=1 αi = 1 , with α also being given the capability to take on negative values. We found that this slight alteration allowed the system to nullify small errors created by the more successful classiﬁers. This research attempted to utilize this methodology for the combination of the classiﬁers designed, and utilized genetic algorithms for the optimization of the unknown coeﬃcients (α1 ,...,αM ).

80

R. Ranawana and V. Palade

4.4 Design of the System 4.4.1 Justiﬁcation for the Development of the Framework The development of this system was promoted through the development of the MultiNNProm System [http://web.comlab.ox.ac.uk/oucl/work/romesh. ranawana/RP2004a.pdf] and the problems encountered during its implementation. MultiNNProm provided us with very promising results in terms of the accuracy and precision displayed. The main problem encountered during the development was the large amount of time taken to properly conﬁgure the neural networks and to determine the optimal encoding methods for the given data set. All these design decisions had to be made on an ad-hoc basis, with the testing of many diﬀerent conﬁgurations and encoding methods being required in order to obtain an optimal set of parameters. As each test run required the training and testing of a neural network, this process was extremely time consuming. In order to negate these problems, we successfully designed and implemented a framework which would automatically determine the optimal parameters required for the design of the multi-classiﬁer system. Within this framework, we utilized a genetic algorithm based method for the determination of the optimal number of layers and the optimal number of neurons on each layer. We also implemented a component of this framework to automatically determine an optimal set of encodings for the presented DNA data set. Thus, when presented with a DNA data set, the framework successfully designed and trained a set of neural networks for inclusion in a multi-classiﬁer system. The following sub-sections describe the implementation of the framework and also lists out the results obtained. 4.4.2 Determination of Optimal Neural Networks Conﬁgurations for the given DNA Data Set Determination of the Optimal Number of Hidden Layers In order to obtain an optimal neural network conﬁguration for the DNA data considered in this paper, we ran a series of simulations, each testing out the accuracy of the system for diﬀerent numbers of hidden layers. We began the simulation run with a neural network that contained one hidden layer and 10 neurons on that hidden layer. The network was trained until an error rate of 1e − 4 was reached. The resulting network was then tested on a separate testing set. A similar set of training/testing runs were then conduced on separate networks, each containing one neuron more than the amount present on the previously trained/tested network. Figure 4.5 graphically illustrates the normalized error of each conﬁguration ﬁtted onto a third order polynomial curve. This curve identiﬁes the fact that as the number of neurons on the ﬁrst layer increases, the generalization capability of the system in turn decreases. Each of these networks demonstrated an accuracy in excess of 96% when presented with the training set.

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

81

Fig. 4.5. The ﬁrst order polynomial curve ﬁt of the error of the system with respect to the number of neurons on the hidden layer of a one hidden layer neural network

Fig. 4.6. The polynomially ﬁt curve of the errors for the one hidden layer and two hidden layer neural networks

We then performed a similar set of simulations on networks with two hidden layers. We maintained a constant number of neurons on the ﬁrst hidden layer and varied the number of units on the second hidden layer. Figure 4.6 shows the accuracy of the networks that contained 50 neurons on the ﬁrst hidden layer and a varying number of neurons on the second hidden layer. It was also observed that the performance of the system with either one or two hidden layers was similar in terms of the accuracy with respect to the number of neurons on the layer before the output layer.

82

R. Ranawana and V. Palade

Fig. 4.7. A comparison of the errors for the systems with 1, 2, 3, 4 and 5 hidden layers Table 4.1. Neural network conﬁgurations for the types of neural network trained and tested. Here, ∗ indicates the variable value Type One hidden Layer Two hidden layer Three hidden layer Four hidden layer Five hidden layer

Conﬁguration ∗

50:∗ 100:50:∗ 100:80:50:∗ 125:100:80:50:∗

A marked improvement in the generalization of the system was observed once the number of hidden layers exceeded 2. Figure 4.7 compares the accuracy of the systems trained and tested using 1, 2, 3, 4 and 5 with the number of neurons on the ﬁnal layer being varied. Table 4.1 lists the conﬁgurations used as the base for the conﬁguration of each of these networks. It was observed that the performance did not vary drastically once the number of layers exceeded 3. While the performance of the systems with 3 hidden layers diminished by small increments, the performances of the remaining two systems remained almost constant. We accredited this increase in performance along with the increase in the number of hidden layers to the networks being able to identify the complexity of the given promoters more successfully. We concluded that neural networks with a higher number of hidden layers were able to recognize relationships of a higher order better than networks with one or two hidden layers. That is, they were able to correctly learn and classify second and third order relationships among the bases of the DNA string, which in turn leads to more genes being classiﬁed.

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

83

Table 4.2. The best performing neural network conﬁgurations for varying numbers of hidden layers with the error percentage observed Number of hidden layers 1 2 3 4 5

Optimal Conﬁguration 18 50:6 100:50:5 100:80:50:18 125:100:80:50:24

Error(%) 12.84 12.73 11.84 11.55 11.43

Listed in Table 4.2 are the minimum errors obtained for each type of neural network along with the number of neurons present on the ﬁnal hidden of that optimal network. Thus, it was concluded that the optimal network with regard to accuracy, generalization and complexity was the network that contained three hidden layers, with 10 neurons on the ﬁnal hidden layer. Although the network with 5 hidden layers did perform moderately better, the network with 3 hidden layers was deemed to be better due to the its reduced complexity and the lesser time required for training. Determination of the Number of Units on Each Layer Once the optimal number of neural network hidden layers for the dataset was determined, the next task was the determination of the number of units required on each layer in order to enhance generalization. For this purpose, we initialized a random set of vectors to represent the number of neurons on each layer. Each vector consisted of 12 integer values, where the ﬁrst four values varied between 1 and 50, the second four between 1 and 20 and the remainder between 1 and 10. The sum of the ﬁrst, the second and ﬁnal four values were used to represent the number of neurons on the ﬁrst, second and third hidden layers respectively. We divided the number on neurons in each layer within each chromosome in order to maximize the eﬀect of mutation and crossover when used within the neural network. If each unit count was used by itself, individual chromosomes would only have included three values, and would have consequently nulliﬁed most of the advantageous eﬀects of the genetic algorithm. Four initial populations, each consisting of 60 vectors were initialized. Each of these populations was then exercised through a genetic algorithm in order to obtain an optimal set of vectors representing the conﬁguration of the network. The 15 best performing vectors of each population was then extracted to create a new population, and was exercised through a genetic algorithm in order to obtain an optimal set of conﬁgurations for the neural network. This stage was required for the nulliﬁcation of any local minimums that the genetic algorithm would have encountered. The ﬁtness function utilized for the selection of vectors was the classiﬁcation error exhibited by the neural network in terms of the percentage of

84

R. Ranawana and V. Palade

Table 4.3. The best performing neural network conﬁgurations with respect to the number of units on each hidden layer and the error displayed by each system. Label a b c

Hidden Layer 1 103 75 56

Hidden Layer 2 63 19 43

Hidden Layer 3 22 17 54

Error(%) 9.82 8.36 8.36

false positives and false negatives with respect to the total amount of testing data presented12 . Thus, the genetic algorithm attempted to minimize the total error displayed by the system with respect to the test data. Each neural network was trained until the error reached a value of 1E − 3 or the number of training epochs reached 100. The genetic algorithm was implemented using a precision of 0.9, which meant that the 6 best performing vectors were included in their entirety within the next generation. Also, a crossover probability of 0.7 and a mutation probability of 0.1 was used along with roulette wheel selection [6] for its implementation. Each evolutionary function was iterated 60 times before the selection of the winning vectors was made. Table 4.3 lists out the 3 best performing conﬁgurations along with the error displayed by each of them. The best performing system during the determination of the number of layers displayed an error rate of 11.43%, whereas with the introduction of the genetic algorithm and better unit number selections, the recognition rate was increased to produce a minimum error rate of 8.36%. Thus, it can be inferred that the correct conﬁguration of the layers leads to a better rate of recognition. When the true positives and true negatives of conﬁgurations b and c were compared, it was found that the sequences correctly classiﬁed by each network diﬀered by around 3%. This lead us to the conclusion that diﬀerent network conﬁgurations can lead diﬀerent networks specializing in speciﬁc classes of promoters and non-promoters. 4.4.3 Method used for the Determination of Optimal Encoding Methods To test the hypothesis that the encoding method can aﬀect the performance of the system, we ran a series of genetic algorithms which tested diﬀerent encoding methods. We initialized ten random populations of 60 vectors, where each vector was represented by four random real values which ranged from −10 to +10. Here, the four values on the vector corresponded to the four DNA bases. Each population was then exercised through a genetic algorithm which used the accuracy of the system of the system with respect to the testing set as the ﬁtness function. Each population was iterated 60 times through the evolutionary algorithm with a mutation probability of 0.8, a precision of 0.8. The algorithm used roulette wheel selection for cross-overs selections. The 12

The inverse of the Precision

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

85

Error

10 best performing vectors from each population were then extracted to form a new population and was then exercised through a genetic algorithm which identical attributes. This step was performed in order to obtain an optimal set of encoding schemes by using the best chromosomes identiﬁed by the individual populations and also assists with the negating of negative eﬀects caused through local minimums. The training graph for the best performing vectors of each generation of the ﬁnal optimized population is displayed in Figure 4.8 by open circles. Here, the average error of each generation is represented by a cross. These results show us that the accuracy of the system was optimized from an error rate of around 8% to a value close to 4%. Thus, the total accuracy of the system was increased from 92% to 96% through the utilization of the genetic algorithm. Table 4.8 lists out the three pest performing encoding methods along with the accuracy displayed by the usage of each. Displayed in Figure 4.9(a) and (b) are value comparisons of the 3 best performing vectors. Figure 4.9(a) shows displays the distribution of each vector where it can be seen that the distribution of the values for the diﬀerent bases displays a very distinct pattern. The numerical values of A and C remain close to each other and vary between 0 and 1, whereas, the values of T and C remain numerically distanced from each others values. It was also observed that the values of T and C always remain

Population Number

Fig. 4.8. Training graph for the encoding optimization genetic algorithm Table 4.4. The three best performing encoding methods along with the error rate displayed by each Label e1 e2 e3

A 0.4788 0.1309 0.2702

T 1.3868 1.9883 0.0693

C 0.1692 0.6046 0.4255

G −0.3659 −0.0513 1.5572

Error(%) 0.0436 0.0489 0.0409

86

R. Ranawana and V. Palade 2.5 2 1.5 Value

A T

1

C 0.5

G

0 2

1

3

-0.5

2.5 2

Values

1.5 Vector 1 Vector 2

1

Vector 3 0.5 0 AT

CG

-0.5 Bases

Fig. 4.9. A comparison of the best performing encoding vectors

either larger or smaller, conversely, to the values of A and C. Thus, the curve always remains either convex or concave. The line never assumes a zigzagged value. These conclusions are substantiated by Figure 4.9(b), which shows the distribution of the diﬀerent bases for the three wining vectors, demonstrating the more or less constant values maintained by both A and C. The best performing network with the optimized number of neurons on each layer displayed an error rate of 8.36%, whereas, with the introduction of an optimized set of encoding methods, the error rate was reduced to 4.36%. As with the results obtained through the optimization of the network conﬁguration, the results obtained through the use of each encoding method produced slightly varying results with relation to the promoters classiﬁed. Thus, the underlying conclusion being that each encoding method helped the network specialize on diﬀerent types of promoters present within the data set. 4.4.4 An Overview of the System As shown in Figure 4.10, the system is a neural-network based multi-classiﬁer system. Through the optimization process of the framework, we obtained three optimal neural network conﬁgurations; a, b and c, and three optimal sequence encoding methodologies; e1, e2 and e3. By combining each encoding method with each conﬁguration, we developed and trained 9 neural networks, NNe1a, NNe1b, NNe1c, NNe2a, NNe2b, NNe2c, NNe3a, NNe3b and NNe3c. Each

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

87

A T C

DNA string

G T ... ... ... A G C C T

Fig. 4.10. NNGenProm Architecture

neural network was initially trained to produce an output of -1 if it predicted the sequence to be a non-promoter and an output of 1 if it predicted a promoter. The outputs of the individual neural networks were then passed onto a probability builder function which assigned probabilities as to whether the presented sequence was an E.Coli promoter or not. Finally, the outputs of the probability functions were passed onto a result combiner which combined the results and presented a ﬁnal result as to whether the given sequence was an E.Coli promoter or not. The ﬁnal output was of the ‘yes’(1) or ‘no’(0) form. The probability function used in this implementation was identical to that used during the implementation of the MultiNNPRom system. 4.4.5 The Result Combiner The results were combined using the LOP2 method described in Section 4.3.3.2. The system was tested on three combination methods, namely, classical majority voting, LAP, and LOP2. The resulting observations showed us that the results obtained through the LOP2 method provided us with a much better recognition rate in terms of the test data, both positive and negative. The comparison of results between these three combinatorial methods is listed out in Section 4.5. The LOP2 method was implemented as follows. We let the outputs of the nine neural networks be symbolized by Oi , where 1 ≤ i ≤ 9. We also deﬁned 9 coeﬃcients αi , where 1 ≤ i ≤ 9. Then, the combined predictor was deﬁned by Equations (4.4) and (4.5) where (4.6). 9

1 αi (log(Oi .P ositive)) (4.4) O.P ositive = exp Z i=1

88

R. Ranawana and V. Palade

9

1 O.N egative = exp αi (log(Oi .N egative)) Z i=1 Z = exp

9

αi (log(Oi .P ositive))

i=1

+ exp

9

(4.5)

αi (log(Oi .N egative))

i=1

(4.6) It is one again obvious that O.P ositive + O.negative = 1. The ﬁnal conclusion on whether the given sequence was an E.Coli promoter or not was reached using equation (4.7). C(O.P ositive, O.N egative) =

Y es ; O.P ositive > O.N egative N o ; Else

(4.7)

The methodology used to determine values for the coeﬃcients is described in the next section. The coeﬃcient values were determined using a method identical to that used during the design and implementation of MultiNNProm.

4.5 Results 4.5.1 Evaluation of the Performance of the Individual Classiﬁers Each neural network performed perfectly on the training set and displayed a recognition rate of approximately 100% when presented with them after training. The test sequences were exercised through the 9 neural networks and the results obtained are listed out in Table 4.11. These performances are graphically in Figure 4.11. It can be observed that while some networks perform better with respect to the recognition of true negatives, others specialize in the recognition of true positives. For example, nne2b has a better Table 4.5. The Speciﬁcity, Sensitivity and Precision of the nine trained neural networks Network nne1a nne2a nne3c nne1b nne2b nne3b nne1c nne2c nne3c

True Negatives 97.66 97.66 98.83 98.23 98.25 98.83 94.15 98.83 94.74

True Positives 91.20 93.71 92.45 88.68 91.82 91.20 92.45 92.45 93.71

Precision(%) 94.43 95.69 95.64 93.46 95.03 95.01 93.30 95.64 94.22

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

89

Accuracy

100.00 98.00

nne1a

96.00

nne2a nne3a

94.00

nne1b

92.00

nne2b

90.00

nne3b

88.00

nne1c

86.00

nne2c

84.00

nne3c

82.00 True Negatives

True Positives

Precision

Fig. 4.11. Comparison of the true positives and true negatives recognized by each neural network Table 4.6. The results obtained when the test data sets were exercised through the combined system Attribute Speciﬁcity Sensitivity Precision

Value 0.9882 0.9748 0.9815

recognition rate with respect to true negatives, whereas nne2c has a better recognition rate with respect to true positives. nne3a, nne3b and nne2c provide the best recognition of true negatives, whereas the best recognition rate of true positives was displayed by nne2a. 4.5.2 Performance evaluation of the Combined System The speciﬁcity, sensitivity and precision of the neural network results combined using the LOP2 method are listed out in Table 4.6. These results are compared with the results obtained for the individual neural networks in Figure 4.12. These results indicate that the combination of these nine neural networks provides us with a far better recognition rate in terms of the precision, speciﬁcity and sensitivity. Thus, the performance of the system has been improved by a considerable margin. It was observed that the combination of the nine neural networks using the LOP2 method did not diminish the performance by either the true positive accuracy or the true negative accuracy of the overall precision. Whereas the true negative recognition rate remains equal to the recognition rate of the best negative classiﬁer after the application of the combinatorial function, the true positive recognition rate has been increased by a considerable fraction. Thus, as claimed by [3], the performance of the combination was not worse than the performance of the best classiﬁer.

90

R. Ranawana and V. Palade 100.00

nne 1a

98.00

nne 2a

Accuracy

96.00

nne 3a

94.00

nne 1b

92.00

nne 2b

90.00

nne 3b

88.00

nne 1c nne 2c

86.00

nne 3c

84.00

Combination

82.00 True Negatives

True Positives

Precision

Fig. 4.12. Comparison of the speciﬁcity, sensitivity and precision of the four neural networks and the combined system Table 4.7. Comparison of the speciﬁcity, sensitivity and precision of the research obtained using diﬀerent combinatorial methods Combinatorial method Majority Voting LOP2

Sp 0.9882 0.9882

Sn 0.9559 0.9748

P 0.9720 0.9815

Table 4.8. Comparison of the speciﬁcity, sensitivity and precision of MultiGenProm with previous research on the same dataset System NNGENProm Ma et al.,2001 [13] Mahadevan and Ghosh, 1994 [14]

Sp 0.9882 0.9176 0.9020

Sn 0.9748 0.9920 0.9800

P 0.9815 0.9194 0.9040

We attribute this increase in accuracy to the diﬀerent specializations exhibited by each network. The combination of the networks lead to a committee of classiﬁers that performed better than the individual classiﬁers. Thus, the inaccurate predictions of one network were negated by the correct predictions of another. The results of this combination were also compared with the results obtained through the usage of majority voting. Table 4.8 lists out this comparison. These results point our the fact that, although the combination of the 9 neural networks using either method provides us with better results than the individual neural networks, the results obtained through the LOP2 method are better. Table 21 compares the results obtained through this rerearch with previous work done on the same problem. Thus, our system is seen as a considerable improvement compared with recent research on the recognition of E.Coli promoters.

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

91

Table 4.9. The performance of the system after each step of the framework was completed Step Lowest Error (%) Determination of the number of hidden layers 11.43 Determination of the number of units on each hidden 8.36 layer Determination of the optimal encoding methods for the 4.36 DNA data set Design and implementation of the networks using opti4.36 mal conﬁgurations and encodings Combination of the neural networks 1.85 0.6

0.5

Error

0.4

0.3

0.2

0.1

Hidden layer and unit count Encoding Determination Combination

0

Fig. 4.13. The performance of the system through the diﬀerent steps of the framework

4.5.3 Summary of the Results Obtained through the use of the Presented Framework As tabulated in Table 4.9, the diﬀerent steps presented within the framework gradually help decrease the error in the recognition rate displayed by the system. These observations are graphically illustrated in Figure 4.13. The spikes on the graph are due to random chromosomes being used for each initial population. This framework presents a methodology that assumes the determination of all the unknown parameters to be an optimization problem, and uses genetic algorithms in order to determine them.

4.6 Discussion and Future Work In this chapter, we presented a novel approach for the recognition of E.Coli promoters in strings of DNA. The proposed system showed a substantial improvement in the recognition rate of these promoters, speciﬁcally in the

92

R. Ranawana and V. Palade

recognition of true positives, i.e. rejection of non-promoters. This caused our system to display a higher speciﬁcity than all other systems developed thus far. We conclude the reasons for this improvement to be threefold. Firstly, it was the use of larger neural networks than those that have been used thus far. This led to a far better rate of recognition and generalization. We also noted that an increase in the number of hidden layers on the neural network also led to a better recognition rate. Secondly, the use of multiple neural networks, each accepting the same set of inputs in diﬀerent forms. We observed the fact that the false positives and false negatives of one network could be wiped out by true positives and true negatives in other networks. Thus, the combination of the opinions of more classiﬁers led us to a system that performed far better than the individual components. Our results substantiates Dietterich’s (Dietterich, 1997) conclusion. Finally, the use of diﬀerent encoding methods, which were customized to suite the particular data set, increased the recognition capabilities of the neural networks in terms of the data presented. The above three reasons, coupled with the fact that each optimized parameter was used in conjunction with permutations of other optimized parameters lead to a system, when combined, provided us with far better accuracy when compared to singular systems or other multi-classiﬁer systems using similar parameters. As was pointed out earlier, one of the major obstacles in the development of the presented system was the time taken for the determination of optimized parameters for the development of the system. This problem was negated by the introduction of the neuro-genetic framework, which provides the designer with each of the optimized parameters necessary for the development of a system specially customized towards the given data set. The neuro-genetic framework provided us with the optimal number of layers to include within the neural networks, the number of neurons to include with each layer and also a set of optimized encoding methods specially customized for the given data set.

References 1. P. Baldi and S. Brunak. Bioinformatics - The Machine Learning Approach, volume 1. MIT Press, 1998. 2. E. Birney. Hidden markov models in biological sequence analysis. IBM Journal of Research and Development, 45(3/4), 2001. 3. T.G. Dietterich. Machine-learning research: Four current directions. The AI Magazine, 18(4):97–136, 1998. 4. D Frishman, A Mironov, HW Mewes, and M Gelfand. Combining diverse evidence for gene recognition in completely sequenced bacterial genomes [published erratum appears in Nucleic Acids Res 1998 Aug 15;26(16):following 3870]. Nucl. Acids Res., 26(12):2941–2947, 1998.

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

93

5. D.J. Galas, M. Eggert, and M.S. Waterman. Rigorous pattern-recognition methods for dna sequences: Analysis of promoter sequences from E.Coli. Journal of Molecular Biology, 186(1):117–128, 1985. 6. D.E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley Longman Publishing Co., Inc., 1989. 7. J.V. Hansen and A. Krogh. A general method for combining in predictors tested on protein secondary structure prediction. In Proceedings of Artiﬁcial Neural Networks in Medicine and Biology, pages 259–264. Springer-Verlag, May 2000. 8. L. K. Hansen and P. Salamon. Neural network ensembles. IEEE Trans. Pattern Anal. Mach. Intell., 12(10):993–1001, 1990. 9. J Henderson, S Salzberg, and K Fasman. Finding genes in dna with a hidden markov model. Journal of Computational Biology, 4(2):127–141, 1997. 10. John R. Koza and David Andre. Automatic discovery of protein motifs using genetic programming. In Xin Yao, editor, Evolutionary Computation: Theory and Applications. World Scientiﬁc, Singapore, 1996. 11. A Krogh. Two methods for improving performance of a hmm and their application for gene ﬁnding. In Proceedings of the 5th International Conference on Intelligent Systems for Molecular Biology, pages 179–186. AAAI Press, 1997. 12. D. Kulp, D. Haussler, M. G. Reese, and F. H. Eeckman. A generalized hidden markov model for the recognition of human genes in DNA. ISMB-96, pages 134–141, 1996. 13. Q. Ma, J.T.L. Wang, D. Shasha, and C.H. Wu. Dna sequence classiﬁcation via an expectation maximization algorithm and neural networks: A case study. IEEE Transactions on Systems, Man, and Cybernetics, part C: Applications and Reviews, Special Issue on Knowledge Management, 31(4):468–475, November 2001. 14. I Mahadevan and I Ghosh. Analysis of E.Coli promoter structures using neural networks. Nucl. Acids Res., 22(11):2158–2165, 1994. 15. E.J. Mandler and J. Schurmann. Combining the classiﬁcation results of independent classiﬁers based on the dempster/schafer theory of evidence. Pattern Recognition and Artiﬁcial Intelligence, X:381–393, 1988. 16. G. Mengeritsky and T.F. Smith. Recognition of characteristic patterns in sets of functionally equivalent dna sequences. Comput. Appl. Biosci., 3(3):223–227, 1987. 17. Lucila Ohno-Machado, Staal A. Vinterbo, and Griﬃn Weber. Classiﬁcation of gene expression data using fuzzy logic. Journal of Intelligent and Fuzzy Systems, 12(1):19–24, 2002. 18. D. Partridge and W.B. Yates. Engineering multiversion neural-net systems. Neural Comput., 8(4):869–893, 1996. 19. S.K. Riis and A. Krogh. Improving prediction of protein secondary structure using neural networks and multiple sequence alignments. Journal of Computational Biology, 3:163–183, 1996. 20. K. Robison. A comprehensive library of dna-binding site matrices for 55 proteins applied to the complete escherichia coli k-12 genome, 1998. 21. G. Rogova. Combining the results of several neural network classiﬁers. Neural Netw., 7(5):777–781, 1994. 22. F. Roli and G. Giacinto. Hybrid Methods in Pattern Recognition, chapter Design of multiple classiﬁer systems, pages 199–226. Worldwide Scientiﬁc Publishing, 2002.

94

R. Ranawana and V. Palade

23. B. Rost and C. Sander. Prediction of protein secondary structure at better than 70% accuracy. Journal of Molecular Biology, 232(2):584–599, July 1993. 24. D. Ruta and B. Gabrys. Analysis of the correlation between majority voting error and the diversity measures in multiple classiﬁer systems. In Proceedings of the 4th International Symposium on Soft Computing, paper 1824–025, 2001. 25. D. Ruta and B. Gabrys. Classiﬁer selection for majority voting. Special issue of the journal of INFORMATION FUSION on Diversity in Multiple Classiﬁer Systems, 2004. 26. H Salgado, A Santos, U Garza-Ramos, J van Helden, E Diaz, and J ColladoVides. RegulonDB (version 2.0): a database on transcriptional regulation in Escherichia coli. Nucl. Acids Res., 27(1):59–60, 1999. 27. Steven Salzberg, Arthur L. Delcher, Kenneth H. Fasman, and John Henderson. A decision tree system for ﬁnding genes in dna. Journal of Computational Biology, 5(4):667–680, 1998. 28. A.J.C. Sharkey and N.E. Sharkey. Combining diverse neural nets. Knowl. Eng. Rev., 12(3):231–247, 1997. 29. C.A. Shipp and L.I. Kuncheva. An investigation into how adaboost aﬀects classiﬁer diversity. In Proc. IPMU 2002, pages 203–208, 2002. 30. E.E. Snyder. Identiﬁcation of protein coding regions in genomic dna. Journal of Molecular Biology, 248:1–18, 1995. 31. EC Uberbacher and RJ Mural. Locating Protein-Coding Regions in Human DNA Sequences by a Multiple Sensor-Neural Network Approach. PNAS, 88(24):11261–11265, 1991. 32. PJ. Woolf and Y Wang. A fuzzy logic approach to analyzing gene expression data. Physiol. Genomics, 3(1):9–15, 2000. 33. L. Xu, A. Krzyzak, and C.Y. Suen. Several methods for combining multiple classiﬁers and their applications in handwritten character recognition. IEEE Trans. on System, Man and Cybernetics, SMC-22(3):418–435, 1992. 34. L. Xu, A. Krzyzak, and C.Y. Suen. Associative switch for combining multiple classiﬁers. Journal of Artiﬁcial Neural Networks, 1(1):77–100, 1994. 35. G. Zenobi and P. Cunningham. Using diversity in preparing ensembles of classiﬁers based on diﬀerent feature subsets to minimize generalization error. Lecture Notes in Computer Science, 2167:576–587, 2001.

5 Evolutionary Grooming of Traﬃc in WDM Optical Networks Yong Xu1 and Kunhong Liu2 1

2

School of Medicine, University of Birmingham, Edgbaston, Birmingham, B15 2TT, UK. Email: [email protected] Intelligent Computing Lab, Hefei Institute of Intelligent Machines, Chinese Academy of Sciences, P.O. Box 1130, Hefei Anhui 230031, China. Email: [email protected]

Summary. The widespread deployment of WDM optical networks posts lots of new challenges to network designers. Traﬃc grooming is one of the most common and interesting problems. Eﬃcient grooming of traﬃc can eﬀectively reduce the overall cost of the network. But unfortunately, it has been shown to be NP-hard. Therefore, new heuristics must be devised to tackle them. Among those approaches, metaheuristics are probably the most promising ones. In this chapter, we present a thorough and comprehensive discussion on various metaheuristic approaches to the grooming of traﬃc in both static and dynamic patterns in WDM optical networks. Some future challenges and research directions are also discussed in this chapter.

Keywords: WDM Optical network, Traﬃc grooming, Genetic algorithm, Tabu search, Simulated annealing.

5.1 Introduction Wavelength-division multiplexing (WDM) technology will undoubtedly dominate next generation’s backbone transport networks. In WDM optical networks, hundreds of wavelength channels can be established in a single ﬁber to make the best use of its huge bandwidth. In such a network, a set of lightpaths on the optical layer, which is used for transmitting optical signal between its end nodes, deﬁne the virtual topology and the interconnections of them are implemented via optical cross-connects (OXCs). Each OXC can switch the optical signal coming in on a wavelength of an input ﬁber to the same wavelength in an output ﬁber. Such network architecture makes it ﬂexible to various services and protection requirements. The deployment and design of such WDM optical networks are currently the subjects of intensive research and development eﬀorts. Y. Xu and K. Liu: Evolutionary Grooming of Traﬃc in WDM Optical Networks, Studies in Computational Intelligence (SCI) 66, 95–137 (2007) c Springer-Verlag Berlin Heidelberg 2007 www.springerlink.com

96

Y. Xu and K. Liu

Currently, the typical transmission rate on a wavelength is 10 Gbit/s (OC-192) and higher rate (40 Gbit/s or OC-768) is also commercially available. On the contrary, the requirement of end-users, such as Internet service providers (ISPs), universities, and industries is generally several to a few hundreds megabits, much lower than that rate. The huge gap between these two requires that a number of such low rate traﬃc streams be multiplexed onto high rate wavelengths to make full use of their capacity and at the same time to minimize the network’s cost. However, multiplexing/demultiplexing a traﬃc stream to/from a wavelength requires the wavelength to drop at the two end nodes of it at which an electronic equipment must be used to process such operation. In SONET/WDM networks, for example, a SONET add/drop multiplexer (ADM) is required to electronically combine such lowrate traﬃc streams onto a high-rate circuit (wavelength). Furthermore, such electronic equipment is often very expensive for high-rate traﬃc transformations. Therefore it is beneﬁcial if each wavelength drops only at those nodes onto which the low-rate streams are congregated in order to reduce the cost. This is realized by the addition of a wavelength ADM (WADM) at each node. With the WADM, each wavelength is allowed to either drop at a node or optically bypass it with no electronic equipment requirement at that node for that wavelength. A question that arises in such a system is how to groom low-rate traﬃc streams onto each wavelength so that the amount of electronic equipment needed in the network is minimized while as few wavelengths as possible are used. This problem, often referred to as the traﬃc-grooming (TG) problem, has been proved to be NP-complete [1] and drawn increasing interest in the literature recently [1–6]. The topic of TG has always been one of the popular problems in optical communication ﬁeld since it was ﬁrst proposed in 1998. It has been the main focus in the recent special issue on WDM-Based Network Architectures in IEEE Journal on Selected Areas in Communications [7]. The First and Second Workshop on TG in WDM Networks have been held recently [8]. Comprehensive review and speciﬁc review have also been done in this subject [4,9,10]. Due to the NP-complete property of this problem, research showed that evolutionary approaches are a very good option for traﬃc-grooming problems. In this chapter of the book we will provide a detailed discussion on how evolutionary approaches can be used to various grooming problems. Some future challenges and research directions are also discussed in this chapter. One purpose of this chapter is to provide the researchers and practitioners in evolutionary computation community a new application area and those in traﬃc grooming community a new tool to improve the grooming performance. The rest of this chapter is organized as follows. Section 5.2 gives the basic concept of traﬃc grooming in WDM optical networks. Sections 5.3 and 5.4 give a thorough discussion on evolutionary grooming of both static and dynamic traﬃc patterns. An application of evolutionary approaches based traﬃc grooming technique will be discussed in Section 5.5. Some future

5 Evolutionary Grooming of Traﬃc in WDM Optical Networks

97

challenges and research directions about traﬃc grooming are discussed in Section 5.6. Finally, conclusions are drawn in Section 5.7.

5.2 Traﬃc Grooming in WDM Optical Networks TG is one of the most important traﬃc engineering techniques which is deﬁned as the allocation of sub-wavelength traﬃc tributaries onto full wavelength channels in order to achieve eﬃcient utilization of network resources, e.g., minimizing cost or blocking probability, or maximizing revenue. This is a complicated combinatorial optimization problem and has been shown to be NP-complete even in a single-hub unidirectional ring network with nonuniform traﬃc requests between each pair of nodes [1]. It was ﬁrst introduced in 1998 [11–13] and has aroused much interest from the researchers in both academia and industrial societies due to its high academic and commercial values. It is also one of the most attractive issues in journals related to optical networks. In WDM optical networks, each ﬁber transmits a number of wavelengths simultaneously in a WDM manner and at the same time every wavelength can carry several traﬃc streams in a Time Division Multiplexing (TDM) manner. By using the ADM, low rate traﬃc can be multiplexed onto high rate circuits. 4 OC-3 traﬃc streams, for instance, can be multiplexed onto an OC-12 circuit and 16 OC-3’s onto an OC-48, etc. The number of low rate traﬃc streams a wavelength can accommodate is referred to as traﬃc granularity. Since high speed ADM is very expensive, the use of ADM’s usually dominates the total cost of today’s backbone networks. Hence TG is often used to avoid each wavelength dropping at too many nodes on the network to reduce such cost. The beneﬁt of TG in WDM optical networks can be shown by the simple example given below. Consider a 4-node unidirectional ring. The traﬃc request between each pair of nodes is 8 OC-3’s and the wavelength capacity is OC-48. Thus each wavelength can carry two node pairs’ amount of traﬃc. Since there are 6 pairs of traﬃc on the ring, 3 wavelengths are needed. Fig. 5.1(a) shows the traﬃc assignment with no grooming, where wavelength λ1 accommodates traﬃc requests between node pairs 1 ↔ 2 and 3 ↔ 4, λ2 2 ↔ 3 and 1 ↔ 4, λ3 1 ↔ 3 and 2 ↔ 4. In this conﬁguration, every wavelength drops at every node so that 3 ADM’s are needed at every node. 12 ADM’s are needed on the ring. Figure 5.1(b) gives another conﬁguration with proper grooming. In this scenario, nodes 1, 2, and 4 equip with a WADM, which allows only wavelengths that carry traﬃc to or from that node to be added/dropped at it and bypasses all the other wavelengths. The assignment of traﬃc is as follows: λ1 : 1 ↔ 2 and 1 ↔ 3; λ2 : 2 ↔ 3 and 2 ↔ 4; λ3 : 1 ↔ 4 and 3 ↔ 4. There are also 3 wavelengths on the ring but since each wavelength drops at only three nodes, 9 ADM’s are needed this time, 3 ADM’s are saved compared to the scenario in Fig. 5.1(a).

98

Y. Xu and K. Liu OC-3's SONET ADM SONET ADM

λ1

SONET ADM OC-3's

SONET ADM

λ2 SONET ADM

λ2

SONET ADM

λ3

WDM

λ3

λ1 SONET

1

ADM

λ1, λ2, λ3 4

WDM

2

λ1

WDM

λ3

3

λ1 SONET ADM

WDM

λ2

λ2

SONET ADM

OC-3's

SONET ADM

λ3 SONET ADM

SONET ADM OC-3's

(a) A possible grooming scenario with 12 ADM’s OC-3's SONET ADM

λ1

SONET ADM

OC-3's

SONET ADM

λ2

SONET ADM

λ3

WADM

λ3

λ1

1

λ1, λ2, λ3 4

WADM

2

WADM

SONET ADM

λ2

SONET ADM

OC-3's

3

λ3 SONET ADM

WDM

λ2

λ1 SONET ADM

SONET ADM

OC-3's

(b) Another possible grooming scenario with only 9 ADM’s

Fig. 5.1. A grooming example on a 4-node unidirectional ring with two diﬀerent scenarios

In these two grooming scenarios, the traﬃc is carried by the same number of wavelengths and every wavelength works at their full capacity, but the number of ADM’s required is diﬀerent. This shows that after appropriate grooming, not only can the burden of the electric devices be alleviated but the amount of them can be reduced. In large-scale networks, the reduction of

5 Evolutionary Grooming of Traﬃc in WDM Optical Networks

99

the amount of AMDs due to good grooming will bring about a remarkable saving to the cost of constructing them. Therefore, traﬃc grooming is one of the problems that must be taken into serious account in the design of today’s high-speed telecommunication networks. According to whether traﬃc changes or not, there are basically two grooming categories: static grooming and dynamic grooming. The former is to groom a known and unchanged traﬃc pattern in order to use the minimum number of ADM’s to support it. The latter is to groom a set of changing traﬃc patterns, either known or unknown, in order to use the minimum number of ADM’s to support this changing traﬃc. It was shown that traﬃc grooming is NP-hard even for a unidirectional ring with only one egress node and nonuniform traﬃc. The grooming of arbitrary traﬃc on general network topology is even diﬃcult. Therefore, evolutionary approach is a good candidate for it. The following sections will discuss the evolutionary approaches to the grooming of static and dynamic traﬃc.

5.3 Evolutionary Grooming of Static Traﬃc This section presents some important algorithms for grooming of static traﬃc in WDM networks. Grooming of static traﬃc has been one of the most interested topics in the literature. Although it is more practical to groom dynamic traﬃc, the grooming of static traﬃc can not only serve as an approximate to the grooming of dynamic traﬃc but also provide a basic methodology to it. All the major evolutionary techniques, i.e., genetic algorithm (GA), tabu search (TS), and simulated annealing (SA), have been applied to this kind of grooming. Up to now, the main eﬀorts are devoted to ring and mesh topologies, and we will discuss each of these approaches to this kind of grooming in the following subsections. 5.3.1 Grooming of Static Traﬃc in WDM Ring Networks It is reasonable to groom traﬃc in ring networks due to its widespread use in today’s infrastructural networks. Although many researchers have already worked in this ﬁeld right from the beginning of the TG problem being proposed, and great achievements have been reached [1–3, 5, 6, 10–23], there are still some regions remained to be explored. Today, many researchers still try to develop new heuristics for solving TG problem in WDM ring networks eﬃciently [24, 25]. In this part, we explain the applications of evolutionarybased approaches for the static traﬃc grooming in WDM ring networks. Although there are many diﬀerent approaches, we mainly focuse on two typical approaches: circle construction based two-step model and directly traﬃc matrix based operation.

100

Y. Xu and K. Liu

5.3.1.1 Circle Construction Based Two-Step Model The circle construction based model was ﬁrst proposed in [2]. In this model, all the traﬃc connections are ﬁrst divided into the basic traﬃc rate, and then as few circles as possible are constructed to include all the requested connections, where each circle consists of multiple nonoverlapping (i.e., link disjoint) connections. The circle construction heuristic is shown in Fig. 5.2. In this algorithm, it searches the traﬃc matrix from the beginning to the end, and then selects the traﬃc requests in the order of the longest to the shortest stride to construct circles. The connection will be added into the circle under the condition that at least one end node of it already existing in the circle, otherwise, it will be put into a so-called GapMaker list which is initially empty. When this process ends, a greedy algorithm is employed to check the GapMaker list to construct as few circles as possible with minimum number of end nodes. For example, given a 5-node ring network with a traﬃc matrix shown in Fig. 5.3(a). Let i and j denote the source and the destination node of a traﬃc request and rij represent the traﬃc request from node i to node j, a possible circle may be constructed by the traﬃc stream 0-4 and 4-0. But as rij is not necessarily equal to rji , only ﬁve circles can be constructed for r04 and r40 , and two basic traﬃc requests belonging to r40 will be put into the GapMaker list. After the circles have been constructed, another heuristic will be used to groom circles onto wavelengths with overlapping as many end nodes belonging to diﬀerent circles as possible so as to reduce the number of ADM’s. If the granularity of a wavelength is 8, eight circles can be groomed onto a wavelength. This model was widely applied to the grooming of static traﬃc in ring networks and two SA based and one GA based approaches were proposed based on it in [3, 15, 20] respectively. For the purpose of comparison, only the GA approach is shown below. The GA approach proposed in [15] aimed to minimize the total number of ADM’s. The operation of the GA can be illustrated as follows. A binary coding method is used and each bit xi w represents an ADM needed on wavelength w at node i if it equals “1” or indicates no ADM necessary if it is “0”. The ﬁtness function can be presented as minimizing: W N w w=1 i=1 xi (where W and N are the total number of wavelengths and nodes in the ring respectively). The traﬃc request is ﬁrst divided into basic elements (the lowest speed rate), then the circle construction heuristics is applied to construct C basic circles based on the traﬃc request. Finally a GA is employed to grooming circles onto wavelengths. A seed obtained from the greedy circle construction grooming algorithm shown in Fig. 5.2 is inserted into the initial population and the rest chromosomes are generated by random permutation of the C circles. In decoding process, two methods, next-ﬁt (NF) and best-ﬁt (BF), are applied to assign circles to wavelengths in accordance with the order of them in the chromosome.

5 Evolutionary Grooming of Traﬃc in WDM Optical Networks

101

Fig. 5.2. Circle construction heuristics (source: reference [2])

NF algorithm distributes circles evenly among all wavelengths but BF tries to fully pack every wavelength except possibly the last one. The corresponding algorithms are shown in Fig. 5.4. An exponential ranking selection scheme is adopted with the probability pi = (r −1)∗ rM −i /(rM −1) for the ith individual to be selected in M total individuals, where r is set to be 0.99. Two crossovers, partially mapped crossover (PMX) and order crossover (OX), are used to produce oﬀspring. In mutation, the elements of an individual are right-shifted by a random number of positions in a cyclic manner.

102

Y. Xu and K. Liu n 0 1 2 3 4 0 1 2 3 4

0 4 1 2 7

5 0 3 3 2

3 1 0 6 4

(a)

1 3 5 0 3

5 4 3 5 0

n 0 1 2 3 4

n 0 1 2 3 4

0 1 2 3 4

0 1 2 3 4

0 3 5 6 3

4 0 4 2 7

6 3 0 5 1

(b)

6 3 3 0 2

2 4 4 3 0

0 4 5 6 7

5 0 4 3 7

6 3 0 6 4

6 3 5 0 3

5 4 4 5 0

(c)

Fig. 5.3. Three sample traﬃc matrices in a WDM ring network with 5 nodes (source: reference [19])

Fig. 5.4. NF and BF algorithms (source: reference [15])

5.3.1.2 Direct Traﬃc Matrix Based Grooming Method The GA based circle construction model may not be very eﬃcient for large scale networks. Therefore, another GA-based approach directly operating on

5 Evolutionary Grooming of Traﬃc in WDM Optical Networks

103

traﬃc matrix and grooming the traﬃc in a comprehensive way was proposed in [16]. The framework of the GA is described below. First, the traﬃc matrix is converted into a vector V, and a random permutation of the traﬃc items is used as a chromosome, V = (x0 , . . ., xk , . . ., xn(n−1) − 1) if there are n nodes in a ring. The position of each item is used to determine the decoding order. The connection between the component’s subscript k of the vector V and those of the traﬃc matrix can be formulated as: ⎧ k = (2n − i − 1) × i/2 + j − i, 0 ≤ i < n − 1, i + 1 ≤ j < n ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ or 0 ≤ k < n(n − 1)/2 (5.1) n (n − 1) + i (i − 1) ⎪ k= + j, 1 ≤ i < n, 0 ≤ j < i ⎪ ⎪ ⎪ 2 ⎪ ⎩ or n(n − 1)/2 ≤ k < n(n − 1) The algorithm in Fig. 5.5 can be used to calculate i, j from k: The basic ideas behind the decoding algorithm proposed in this approach can be described as follows. If a wavelength drops at π nodes, it can at most accommodate π(π − 1) traﬃc requests. If all the π(π − 1) traﬃc requests are assigned to the wavelength, the numbers of ADM’s and wavelengths used on the ring must be minimized. Based on this observation, a ﬁrst-ﬁt heuristic incorporated with a greedy local improvement was proposed to decode each chromosome, as is shown in Fig. 5.6. The algorithm assigns the ﬁrst encountered traﬃc items onto each wavelength until one link load on the wavelength exceeds its capacity. Then the local improvement algorithm is called to assign such a traﬃc item to it with no additional node added. The objective of this algorithm is to minimize the number of ADM’s. Therefore, the fewer ADM’s a chromosome uses, the ﬁtter it is. If two chromosomes require the same number of ADM’s, the one with smaller number of wavelengths is assigned higher ﬁtness value. The order-mapped crossover (OMX) proposed in [26] is used to produce oﬀspring and the inversion of a section of some randomly selected genes is used as mutation in this algorithm. if(k

Studies in Computational Intelligence, Volume 66 Editor-in-chief Prof. Janusz Kacprzyk Systems Research Institute Polish Academy of Sciences ul. Newelska 6 01-447 Warsaw Poland E-mail: [email protected] Further volumes of this series can be found on our homepage: springer.com

Vol. 55. Xianyi Zeng, Yi Li, Da Ruan and Ludovic Koehl (Eds.) Computational Textile, 2007 ISBN 978-3-540-70656-4

Vol. 44. Nadia Nedjah, Luiza de Macedo Mourelle, Mario Neto Borges, Nival Nunes de Almeida (Eds.) Intelligent Educational Machines, 2007 ISBN 978-3-540-44920-1

Vol. 56. Akira Namatame, Satoshi Kurihara and Hideyuki Nakashima (Eds.) Emergent Intelligence of Networked Agents, 2007 ISBN 978-3-540-71073-8

Vol. 45. Vladimir G. Ivancevic, Tijana T. Ivancevic Neuro-Fuzzy Associative Machinery for Comprehensive Brain and Cognition Modeling, 2007 ISBN 978-3-540-47463-0

Vol. 57. Nadia Nedjah, Ajith Abraham and Luiza de Macedo Mourella (Eds.) Computational Intelligence in Information Assurance and Security, 2007 ISBN 978-3-540-71077-6

Vol. 46. Valentina Zharkova, Lakhmi C. Jain Artificial Intelligence in Recognition and Classification of Astrophysical and Medical Images, 2007 ISBN 978-3-540-47511-8

Vol. 58. Jeng-Shyang Pan, Hsiang-Cheh Huang, Lakhmi C. Jain and Wai-Chi Fang (Eds.) Intelligent Multimedia Data Hiding, 2007 ISBN 978-3-540-71168-1

Vol. 47. S. Sumathi, S. Esakkirajan Fundamentals of Relational Database Management Systems, 2007 ISBN 978-3-540-48397-7

Vol. 59. Andrzej P. Wierzbicki and Yoshiteru Nakamori (Eds.) Creative Environments, 2007 ISBN 978-3-540-71466-8

Vol. 48. H. Yoshida (Ed.) Advanced Computational Intelligence Paradigms in Healthcare-1, 2007 ISBN 978-3-540-47523-1

Vol. 60. Vladimir G. Ivancevic and Tijana T. Ivacevic Computational Mind: A Complex Dynamics Perspective, 2007 ISBN 978-3-540-71465-1

Vol. 49. Keshav P. Dahal, Kay Chen Tan, Peter I. Cowling (Eds.) Evolutionary Scheduling, 2007 ISBN 978-3-540-48582-7

Vol. 61. Jacques Teller, John R. Lee and Catherine Roussey (Eds.) Ontologies for Urban Development, 2007 ISBN 978-3-540-71975-5

Vol. 50. Nadia Nedjah, Leandro dos Santos Coelho, Luiza de Macedo Mourelle (Eds.) Mobile Robots: The Evolutionary Approach, 2007 ISBN 978-3-540-49719-6

Vol. 62. Lakhmi C. Jain, Raymond A. Tedman and Debra K. Tedman (Eds.) Evolution of Teaching and Learning Paradigms in Intelligent Environment, 2007 ISBN 978-3-540-71973-1

Vol. 51. Shengxiang Yang, Yew Soon Ong, Yaochu Jin Honda (Eds.) Evolutionary Computation in Dynamic and Uncertain Environment, 2007 ISBN 978-3-540-49772-1 Vol. 52. Abraham Kandel, Horst Bunke, Mark Last (Eds.) Applied Graph Theory in Computer Vision and Pattern Recognition, 2007 ISBN 978-3-540-68019-2 Vol. 53. Huajin Tang, Kay Chen Tan, Zhang Yi Neural Networks: Computational Models and Applications, 2007 ISBN 978-3-540-69225-6 Vol. 54. Fernando G. Lobo, Cl´audio F. Lima and Zbigniew Michalewicz (Eds.) Parameter Setting in Evolutionary Algorithms, 2007 ISBN 978-3-540-69431-1

Vol. 63. Wlodzislaw Duch and Jacek Ma´ndziuk (Eds.) Challenges for Computational Intelligence, 2007 ISBN 978-3-540-71983-0 Vol. 64. Lorenzo Magnani and Ping Li (Eds.) Model-Based Reasoning in Science, Technology, and Medicine, 2007 ISBN 978-3-540-71985-4 Vol. 65. S. Vaidya, L.C. Jain and Hiro Yoshida (Eds.) Advanced Computational Intelligence Paradigms in Healthcare-2, 2007 ISBN 978-3-540-72374-5 Vol. 66. Lakhmi C. Jain, Vasile Palade and Dipti Srinivasan (Eds.) Advances in Evolutionary Computing for System Design, 2007 ISBN 978-3-540-72376-9

Lakhmi C. Jain Vasile Palade Dipti Srinivasan (Eds.)

Advances in Evolutionary Computing for System Design With 134 Figures and 55 Tables

123

Prof. Lakhmi C. Jain

Dr. Dipti Srinivasan

KES Centre University of South Australia Adelaide, Mawson Lakes Campus South Australia SA 5095 Australia E-mail:- [email protected]

Department of Electrical and Computer Engineering National University of Singapore Block E4-06-08 4 Engineering Drive 3 Singapore 117576 E-mail:- [email protected]

Dr. Vasile Palade Oxford University Computing Laboratory Wolfson Building Parks Road Oxford, OX1 3QD England E-mail:- [email protected]

Library of Congress Control Number: 2007926313 ISSN print edition: 1860-949X ISSN electronic edition: 1860-9503 ISBN 978-3-540-72376-9 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, reuse of illustrations, recitation, broadcasting, reproduction on microfilm 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-Verlag. Violations are liable to prosecution under the German Copyright Law. Springer is a part of Springer Science+Business Media springer.com c Springer-Verlag Berlin Heidelberg 2007 ° The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Cover design: deblik, Berlin Typesetting by the SPi using a Springer LATEX macro package Printed on acid-free paper SPIN: 11551430 89/SPi

543210

Preface

In the last two decades, there has been a huge trend towards building intelligent systems that are inspired by biological systems from nature. Many nature-inspired algorithms and approaches have been developed and have become popular for designing intelligent systems with application to various real-world problems. Evolutionary computing paradigms, which are powerful search techniques inspired from natural genetics, proved to be at the forefront of this endeavour. Undoubtedly, many authored and edited books have already revealed the ﬁeld of evolutionary computing from many facets, which include deep theoretical analyses as well as more application-oriented studies. Edited volumes containing conference papers on the theory and application of evolutionary computing have also been very popular. In this book, the editors tried to put together a collection of extended and consistent chapters presenting original and latest research work from leading researchers on evolutionary computing. The focus of the book was to present challenging real-world applications of evolutionary computing paradigms as well as to provide an up-to-date reference manual for someone interested to use evolutionary techniques to system design. This edited book consists of thirteen chapters covering a wide area of topics on evolutionary computing and applications. The application areas are diverse and include system control, bioinformatics, hardware optimization, traﬃc grooming, games theory, grid computing, and so on. Challenging and captivating theoretical issues are also introduced and discussed. These refer especially to the use of evolutionary computing techniques in conjunction with other machine learning approaches, and include: evolutionary neurofuzzy systems, fuzzy systems and genetic algorithms, evolutionary algorithms and immune learning for neural network design, evolutionary techniques for multiple classiﬁer design, hybrid multi-objective evolutionary algorithms, evolutionary particle swarms. Our hope is that this book will serve as a reference to researchers in evolutionary computing and to system designers and practitioners working

VI

Preface

in various application domains who are interested in using evolutionary computing paradigms in system design and implementation. This book can also be used by students and lecturers as an advanced reading material for courses on evolutionary computing. The editors are grateful to the authors for their excellent contributions to the book. Thanks are due to the reviewers for providing precious feedback. The editorial support provided by Springer-Verlag is acknowledged. We hope that this book on evolutionary computing in system design will prove valuable to its readers. Lakhmi C. Jain University of South Australia Australia Vasile Palade Oxford University UK Dipti Srinivasan National University of Singapore Singapore

Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V 1 Introduction to Evolutionary Computing in System Design Lakhmi C. Jain, Shing Chiang Tan, Chee Peng Lim . . . . . . . . . . . . . . . . . .

1

2 Evolutionary Neuro-Fuzzy Systems and Applications G. Castellano, C. Castiello, A.M. Fanelli, L. Jain . . . . . . . . . . . . . . . . . . . . 11 3 Evolution of Fuzzy Controllers and Applications Dilip Kumar Pratihar, Nirmal Baran Hui . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 4 A Neuro-Genetic Framework for Multi-Classiﬁer Design: An Application to Promoter Recognition in DNA Sequences Romesh Ranawana and Vasile Palade . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 5 Evolutionary Grooming of Traﬃc in WDM Optical Networks Yong Xu and Kunhong Liu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 6 EPSO: Evolutionary Particle Swarms V. Miranda, Hrvoje Keko, Alvaro Jaramillo . . . . . . . . . . . . . . . . . . . . . . . . . 139 7 Design of Type-Reduction Strategies for Type-2 Fuzzy Logic Systems using Genetic Algorithms Woei-Wan Tan, Dongrui Wu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 8 Designing a Recurrent Neural Network-based Controller for Gyro-Mirror Line-of-Sight Stabilization System using an Artiﬁcial Immune Algorithm Ji Hua Ang, Chi Keong Goh, Eu Jin Teoh, and Kay Chen Tan . . . . . . . . 189 9 Distributed Problem Solving using Evolutionary Learning in Multi-Agent Systems Dipti Srinivasan, Min Chee Choy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

VIII

Contents

10 Evolutionary Computing within Grid Environment Ashutosh Tiwari, Gokop Goteng, and Rajkumar Roy . . . . . . . . . . . . . . . . . 229 11 Application of Evolutionary Game Theory to Wireless Mesh Networks Athanasios Vasilakos, Markos Anastasopoulos . . . . . . . . . . . . . . . . . . . . . . . 249 12 Applying Hybrid Multiobjective Evolutionary Algorithms to the Sailor Assignment Problem Deon Garrett, Dipankar Dasgupta, Joseph Vannucci, James Simien . . . . 269 13 Evolutionary Techniques Applied to Hardware Optimization Problems: Test and Veriﬁcation of Advanced Processors Ernesto Sanchez and Giovanni Squillero . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303

1 Introduction to Evolutionary Computing in System Design Lakhmi C. Jain1 , Shing Chiang Tan2 , Chee Peng Lim3 1

School of Electrical and Information Engineering University of South Australia, Australia 2 Faculty of Information Science and Technology Multimedia University, Malaysia 3 School of Electrical and Electronic Engineering University of Science Malaysia, Malaysia Summary. In this chapter, an introduction on the use of evolutionary computing techniques, which are considered as global optimization and search techniques inspired from biological evolutions, in the domain of system design is presented. A variety of evolutionary computing techniques are ﬁrst explained, and the motivations of using evolutionary computing techniques in tackling system design tasks are then discussed. In addition, a number of successful applications of evolutionary computing to system design tasks are described.

1.1 Introduction Evolutionary Computing (EC) [1] techniques have gained a lot of interests from researchers and practitioners in various ﬁelds as they provide robust and powerful adaptive search mechanisms that are useful for system design. In addition, the underlying biological concepts on which EC techniques are based on also contribute to their attractiveness in tackling system design problems. Although many diﬀerent EC techniques have been proposed over the years, they, in general, share a common idea, i.e., evolving a population of potential solutions through the process of selection, recombination, and mutation. The search process of EC can be interpreted as a collective learning process in a population. Such interpretation is in agreement with the claims of some evolutionary models, such as the Baldwin evolution [2], that put emphasis on the role of the evolutionary learning process in a population. On the other hand, the emergence of systems plays an important role in our daily life. According to [3], a system is deﬁned as a set of entities, real or abstract, comprising a whole where each component interacts with or is related to at least one other component and they all serve a common objective. Researches in various types of systems are widespread and interdisciplinary, ranging from information and computer science, social and cognitive sciences, L.C. Jain et al.: Introduction to Evolutionary Computing in System Design, Studies in Computational Intelligence (SCI) 66, 1–9 (2007) c Springer-Verlag Berlin Heidelberg 2007 www.springerlink.com

2

L.C. Jain et al.

to engineering and management research. For example, as stated in [3], in information and computer science, a system could also be a method or an algorithm. From the perspective of engineering, system engineering deals with the application of engineering techniques to form a structured development process that covers various stages of the process, from concept to production to operation and disposal. In social and cognitive sciences, systems comprise human brain functions and human mental processes as well as social/cultural behavioral patterns; while in management research, operations research, and organizational development, human organizations are viewed as conceptual systems of interacting components. In this chapter, an introduction to EC in system designs is presented. In sections 1.2 and 1.3, what is EC, and why EC is useful in system design tasks are explained. In section 1.4, some successful applications of EC to a variety of system design problems are described. A description on each chapter in this book is given in Section 1.5, and a summary of this chapter is included in Section 1.6.

1.2 What is Evolutionary Computing Evolutionary computing is referred to as computer methods that are inspired by biological evolutions [1]. From the literature, there are generally ﬁve basic types of EC, viz., genetic algorithms (GA), evolutionary programming (EP), evolution strategies (ES), genetic programming (GP), and learning classiﬁer systems (LCS). The basic idea behind all these EC variants is commonly the same from a unifying viewpoint. On availability of a population of individuals, natural selection, which is based on the principle of survival of the ﬁttest following the existence of environmental pressures, is exercised to choose individuals that could better ﬁt the environment. The operations involved are as follows. Given a ﬁtness function to be optimized, a set of candidate solutions is randomly generated. The candidate solutions are assessed by the ﬁtness function that indicates an abstract of ﬁtness measure. Based on the ﬁtness values, ﬁtter candidate solutions have a better chance to be selected to seed the next generation through the application of recombination (crossover) and/or mutation. On one hand, the recombination operator is applied to two or more parent candidate solutions and one or more new candidate solutions (oﬀspring) is/are produced. On the other hand, the mutation operator is applied to one parent candidate solution and one new candidate solution is produced. The new candidate solutions will compete among one another, based on the ﬁtness values, for a place in the next generation. Even though candidate solutions with higher ﬁtness values are favorable, in some selection schemes, the solution candidates with relatively low ﬁtness values are included in the next generation so as to maintain diversity of the population. The processes of selection, recombination, and mutation are repeated from one generation of population

1 Introduction to Evolutionary Computing in System Design

3

to another until a terminating criterion is satisﬁed, i.e., either a candidate solution with suﬃcient quality has been obtained or the algorithm has been executed for a predeﬁned number of generations. Developed by Holland [4], the GA is the most popular and widely used EC paradigm. The GA is essentially a class of population-based search strategies that utilize an iterative approach to perform a global search on the solution space of a given problem. The three most popular search operators used in the GA are selection, crossover, and mutation. The selection operator chooses a subset group of candidate solutions (from a population) that are subsequently used to generate new candidate solutions with the crossover and mutation operators. In this regard, the crossover operator is applied to combine genetic information of the existing candidate solutions whereas the mutation operator is applied to create variations of a single candidate solution. While conventional GA models utilize binary codes to keep information of a candidate solution (i.e., a chromosome), real-valued chromosomes are now widely used to solve more diﬃcult problems. Introduced by Fogel, Owens, and Walsh [5], EP was aimed to simulate intelligent behavior by means of ﬁnite-state machines. In this regard, candidate solutions to a problem are considered as a population of ﬁnite-state machines. New candidate solutions (oﬀspring) are generated from the existing candidate solutions (parents) through mutation. All candidate solutions are assessed by a ﬁtness function. The crossover operator is not used in EP. Instead, the mutation operator is employed to provide variations in the population of candidate solutions. ES [6, 7] was ﬁrst developed to optimize parameters for aero-technology devices. This method is based on the concept of the evolution of evolution. Each candidate solution in the population is formed by genetic building blocks and a set of strategy parameters that models the behavior of that candidate solution in its environment. Both genetic building blocks and strategy parameters participate in the evolution. The evolution of the genetic characteristics is governed by the strategy parameters that are also adapted from evolution. Like EP, ES uses only the mutation operator. Devised by Koza [8], GP was targeted at making the computer to solve problems without being explicitly programmed to do so. GP is an extension of conventional GA. However, the representation scheme of GP is diﬀerent from that of conventional GA; the former represents individuals as executable programs (i.e., trees) whereas the latter uses string representation. LCS [9] tackles the machine learning tasks as an evolutionary rule discovery component. The knowledge representation of LCS is a collection of production rules that are expressed in the if-then format. Each production rule, which is encoded as a binary string in Holland’s standard approach, is considered as a classiﬁer. The rules are updated according to some speciﬁc evolutionary procedure.

4

L.C. Jain et al.

1.3 Why Evolutionary Computing As explained in [10], one of the main advantages of EC is that it is conceptually simple and easy to use. An EC process normally covers random initialization of a population of candidate solutions, followed by some iterative variation and selection strategy, and by evaluation of candidate solutions in accordance with a performance index. This process is relatively simple, as compared with, for example, model-based approaches that deal with complicated mathematical formulation of the system under scrutiny. The applicability of EC is broad and comprehensive, covering most problems that can be cast as an optimization task. In contrast to some numerical approaches that require continuous values or constraint formulation, the search space of EC can be disjoint, and the representation can be intuitive to humans [10]. Besides, EC techniques can be used to adapt solutions to changing environments. Indeed, a wide range of disciplines, including computer science, mathematics, medicine, engineering, and economics, to name a few, has been explored extensively by capitalizing EC techniques. Some successful application examples of EC are presented in section 1.4. In solving complex, real-world problems, one may resort to methods that mimic the aﬃnity from the nature. In this regard, the biologically inspired EC is cohesive with the idea of designing solutions that exploit a few aspects of natural evolutionary processes. In addition, recent trends in EC have geared towards devising hybrid models with other conventional optimization (e.g. gradient-based) techniques as well as other artiﬁcial intelligence paradigms. These hybrid models normally are highly automated and can give high eﬃciency, as evidenced in the work presented in other chapters of this book. Another motivation of using EC stems from a technical standpoint. One trend from the growth of computer power has been the growing demand for problem-solving automation, which is vitally needed to cope with modern system design tasks of which their complexity is also ever increasing. EC techniques may turn out to be just as important because of the robustness and adaptation they can oﬀer when dealing with modern computer-based applications. They are versatile enough and can normally provide satisfactory solutions within acceptance time with the advent of fast, powerful computing machines.

1.4 Applications of Evolutionary Computing to System Design The applications of EC to system design are far reaching, especially from the perspective of meeting the requirements of software and hardware designs. A snapshot of successful applications of EC is presented, as follows. EC can be deployed to solve problems that require any kind of optimization. Examples of optimization problems include operations-research tasks

1 Introduction to Evolutionary Computing in System Design

5

such as scheduling (e.g., [11, 12] and design optimization (e.g., [13, 14]). EC-based methods have been proposed for identifying an optimal broadcasting strategy in urban areas [15]. Automatic test data generation problems [16] can be formulated as an optimization problem and be solved using EC. Classiﬁcation and regression tasks are one of the fundamental intelligent processes in data analysis. Of many approaches, artiﬁcial neural networks (ANN) models, which are inspired from the functionality of biological brains, are a feasible method for tackling classiﬁcation and regression tasks. Recently, various combinations between ANNs and EC have been investigated (e.g., [17]–[22]). The motivation of this synergistic machine learning lies in the retention of well-deﬁned identity of the two entities, for which ANNs provide a framework of accurate and exact computation whereas EC provides a robust and eﬃcient search, for better solution. Hybrid models combining EC and fuzzy systems (e.g., [23, 24]) as well as EC, ANNs, and fuzzy systems (e.g., [25–27]) have also been proposed, with the purpose of obtaining improved solutions. Evolutionary robotics is an active ﬁeld of research that utilizes EC to develop controllers for autonomous robots. In this regard, a population of candidate controllers is repeatedly evolved and tested with respect to a ﬁtness function. The most feasible controller is identiﬁed for achieving an eﬃcient robot learning application that could provide desired actions for the robot to perform. Successful applications of evolutionary robotics include research work by, to name a few, Paul et al. [28], Krasny and Orin [29], Reil and Husbands [30], Kodjabachian and Meyer [31], and Gallagher et al. [32]. Evolvable hardware is an emerging sub-specialty of EC that is aimed to automatically design and optimize electrical and mechanical structures, such as digital [33, 34]/analog [34–36] circuits, antennas [37], and robots [38–40] Research in this area has deployed various EC techniques including GA, GP, EP, adaptive GA [41], parallel GA [42]), compact GA [43], and Cartesian GP [44]. Recent advances in molecular biology and genomic technology have led to an explosive growth in the generation of biological data. Bioinformatics, as its name suggests, is a research area that utilizes computational methods to make discoveries from a huge biological database. The role of EC in bioinformatics, and data mining in general, is to assist in analyzing and, thus, extracting meaningful information from large data sets in a robust and computationally eﬃcient way. EC has been applied to handle a broad classiﬁcation of the various bioinformatics tasks, for instances, gene expression [45] and sequence alignment [46, 47], as well as for other data mining problems [48]. Other innovative application of EC to designing intelligent systems for solving a variety of practical problems include machine condition monitoring [49], medical diagnosis [50, 51], web-information search and retrieval [52, 53], and pedagogical educational learning system [54], as well as other problems in engineering and information systems [55]. In short, as computer technology continues to advance, the advantages of EC will become more profound, and

6

L.C. Jain et al.

EC techniques will continue to be adopted as a practical problem-solving tool in diverse ﬁelds of system design.

1.5 Chapters Included in this Book This book includes thirteen chapters. Chapter one introduces EC in system design. Chapter two by Castellano et al., is focused on evolutionary neuronfuzzy systems and their applications. A scheme using neural learning and genetic optimization to learn a fuzzy model is presented. Chapter three, by Pratihar and Hui, describes the evolution of fuzzy controllers and applications. Some applications of the reported techniques are presented. Chapter four by Ranawana and Palade, is on neuro-genetic techniques for designing multiclassiﬁer systems. An application for promoter recognition in DNA sequences is presented. Chapter ﬁve by Xu and Liu is on evolutionary grooming of traﬃc in WDM optical networks. Future challenges and research directions are also included. Chapter six by Miranda presents evolutionary particle swarm optimization as an evolutionary meta-heuristic that implements a scheme of self adaptive recombination. Chapter seven by Tan and Wu is on the design of computationally eﬃcient type-reduction strategies for type-2 fuzzy logic systems using genetic algorithm. Chapter eight by Ang et al. is on the design of neural network-based controller using an artiﬁcial immune system. Chapter nine by Srinivasan and Choy is on distributed problem solving using evolutionary learning in multi-agent systems. Chapter ten by Tiwari et al. reports their work on evolutionary computing in grid-based environment. Chapter eleven by Vasilakos and Anastasopoulos is on the application of evolutionary game theory in wireless mesh networks. Chapter twelve by Garrett el al. is on the application of hybrid multiobjective evolutionary algorithms to the sailor assignment problem. The ﬁnal chapter by Sanchez and Squillero presents the application of evolutionary computing techniques to hardware optimization problems related to the test and veriﬁcation of advanced processors.

1.6 Conclusion This chapter has presented an introduction to EC in handling system design problems. A variety of EC techniques, which include GA, EP, ES, GP, and LCS, have been described. The advantages of using EC techniques, especially for tackling optimization and search problems, as well as some successful applications of EC to diﬀerent system design problems have also been elaborated. It is envisaged that EC techniques will ultimately play a signiﬁcant role in solving practical system design tasks, especially those complex, nonlinear systems in various domains including computer and information sciences, engineering,

1 Introduction to Evolutionary Computing in System Design

7

management, and social and cognitive sciences, that have become more and more common in today’s knowledge-based society.

References 1. Dumitrescu, D., Lazzerini, B., Jain, L.C., Dumitrescu, A., (2000), Evolutionary Computation, CRC Press, Boca Raton, FL, USA. 2. Suzuki, R., Arita, T., (2007), The dynamic changes in roles of learning through the Baldwin eﬀect, Artiﬁcial Life, 13, 31–43. 3. http://en.wikipedia.org/wiki/System 4. Holland, J.H., (1962), Outline for a logical theory of adaptive systems, J. ACM, 3, 297–314. 5. Fogel, L.J., Owens, A.J., Walsh, M.J., (1966), Artiﬁcial Intelligence through Simulated Evolution, John Wiley, New York. 6. Rechenberg, I., (1994), Evolutionary strategy, Computational Intelligence: Imitating Life, Zurada, J.M., Marks II, R., Robinson, C., (Eds.), IEEE Press, 147– 159. 7. Schwefel, H.-P., (1981), Numerical Optimization of Computer Models, John Wiley, Chichester, UK. 8. Koza, J.R., (1992), Genetic Programming, MIT Press, Cambridge, MA. 9. Holland, J.H., (1986), Escaping brittleness: The possibility of general-purpose learning algorithms applied to parallel rule-based systems, Mach. Learn., 2, Michalski, R.S., Carbonell, J.G., Mitchell, T.M., (Eds.), Morgan Kaufmann, Los Altos, CA, 593–624. 10. Fogel, D.B. (1997), The advantages of evolutionary computation, BioComputing and Emergent Computation, Lundh, D., Olsson, B., Narayanan A., (Eds.), Skve, Sweden, World Scientiﬁc Press, Singapore, 1–11. 11. Lee, L.H., Lee, C.U., Tan, Y.P., (2007), A multi-objective genetic algorithm for robust ﬂight scheduling using simulation, Eur. J. Oper. Res., 177, 1948–1968. 12. Jozefowska, J., Mika, M., Rozycki, R., Waligora, G., Weglarz, J., (2002), A heuristic approach to allocating the continuous resource in discrete-continuous scheduling problems to minimize the makespan, J. Scheduling, 5, 487–499. 13. Greiner, H., (1996), Robust optical coating design with evolution strategies, Appl. Opt., 35, 5477–5483. 14. Wiesmann, D., Hammel, U., B¨ ack, T., (1998), Robust design of multiplayer optical coatings by means of evolutionary algorithms, IEEE Trans. Evol. Comput., 2, 162–167. 15. Alba, E., Dorronsoro, B., Luna, F., Nebro, A.J., Bouvry, P., Hogie, L., (2007), A cellular multi-objective genetic algorithm for optimal broadcasting strategy in metropolitan MANETs, Comput. Commun., 30, 685–697. 16. Celso, C. Ribeiro, C.C., Martins, S.L., Rosseti, I., (2007), Metaheuristics for optimization problems in computer communications, Comput. Commun., 30, 656–669. 17. Han, S.-J., Cho S.-B., (2006), Evolutionary neural networks for anomaly detection based on the behavior of a program, IEEE Trans. Syst., Man Cybernet.— Part B, 36, 559–570. 18. Fieldsend, J.E., Singh, S., (2005), Pareto Evolutionary Neural Networks, IEEE Trans. Neural Netw.,16, 338–354.

8

L.C. Jain et al.

19. Bonissone, P.P., Chen, Y.-T., Goebel, K., Khedkar, P.S, (1999), Hybrid soft computing systems: industrial and commercial applications, Proc. IEEE, 87, pp. 1641–1667. 20. Yao, X., (1999), Evolving artiﬁcial neural networks, Proc. IEEE, 87, 1423–1447. 21. Vonk, E., Jain, L.C., and Johnson, R.P., (1997), Automatic Generation of Neural Networks Architecture Using Evolutionary Computing, World Scientiﬁc Publishing Company, Singapore. 22. Van Rooij, A., Jain, L.C., and Johnson, R.P., (1996), Neural Network Training Using Genetic Algorithms, World Scientiﬁc Publishing Company, Singapore. 23. Bonissone, P.P., Subbu, R., Eklund, N., Kiehl, T.R., (2006), Evolutionary algorithms + domain knowledge = real-world evolutionary computation, IEEE Trans. Evol. Comput., 10, 256–280. 24. Hoﬀmann, F., (2001), Evolutionary algorithms for fuzzy control system design, Proc. IEEE, 89, 1318–1333. 25. Lin, C.J., Xu, Y.-J., (2006), A self-adaptive neural fuzzy network with groupbased symbiotic evolution and its prediction applications, Fuzzy Set Syst., 157, 1036–1056. 26. Oh, S.-K., Pedrycz, W., (2005), A new approach to self-organizing fuzzy polynomial neural networks guided by genetic optimization, Phys. Lett. A, 345, 88–100. 27. Jain, L.C. and Martin, N.M. (Eds), (1999), Fusion of Neural Networks, Fuzzy Logic and Evolutionary Computing and their Applications, CRC Press USA. 28. Paul, C., Valero-Cuevas, F.J., Lipson, H., 2006, Design and Control of Tensegrity Robots for Locomotion, IEEE Trans. Robot., 22, 944–957. 29. Krasny, D.P., Orin, D.E., (2004), Generating High-Speed Dynamic Running Gaits in a Quadruped Robot Using an Evolutionary Search, IEEE Trans. Syst., Man Cybernet.—Part B, 34, 1685–1696. 30. Reil, T., Husbands, P., (2002), Evolution of Central Pattern Generators for Bipedal Walking in a Real-Time Physics Environment, IEEE Trans. Evol. Comput., 6, 159–168. 31. Kodjabachian, J., Meyer, J.-A., (1998), Evolution and development of neural networks controlling locomotion, gradient following and obstacle avoidance in artiﬁcial insects, IEEE Trans. Neural Netw., 9, 796–812. 32. Gallagher, J., Beer, R., Espenschiel, K., Quinn, R., (1996), Application of evolved locomotion controllers to a hexapod robot, Robot. Autonomous Syst., 19, 95–103. 33. Hartmann, M., Haddow, P.C. (2004), Evolution of fault-tolerant and noiserobust digital designs, Proc. Inst. Elect Eng.—Comput. Digit. Tech., 151, 287–294. 34. Higuchi, T., Iwata, M., Keymeulen, D., Sakanashi, H., Murakawa, H., Iajitani, I., Takahashi, E., Toda, K.,Salami, M., Kajihara, N., Oesu, N., (1999), Real-world applications of analog and digital evolvable hardware, IEEE Trans. Evol. Comput., 220–235. 35. Lohn, J.D., (1999), Experiments on evolving software models of analog circuits, Commun. ACM, 42, 67–69. 36. Keymeulen, D., Zebulum, R.S., Jin, Y., Stoica, A., (2000), Fault-tolerant evolvable hardware using ﬁeld-programmable transistor arrays, IEEE Trans. Rel., 49, 305–316. 37. Hum, S.V., Okoniewski, M., Davies, R.J., (2005), An evolvable antenna platform based on reconﬁgurable reﬂect arrays, Proc. NASA/DoD Conf. Evolvable Hardware, Washington, DC, 139–146.

1 Introduction to Evolutionary Computing in System Design

9

38. Lohn, J.D., Hornby, G.S., (2006), Evolvable hardware: using evolutionary computation to design and optimize hardware systems, IEEE Computational Intelligence Magazine, 1, 19–27. 39. Terrile, R.J., Aghazarian, H., Ferguson, M.I., Fink, W., Huntsberger, T.L., Keymeulen, D., Klimeck, G., Kordon, M.A., Seungwon, L., Allmen, P.V., (2005), Evolutionary computation technologies for the automated design of space systems, Proc. NASA/DoD Conf. Evolvable Hardware, Washington, DC, 131–138. 40. Yao, X., Higuchi, T., (1999), Promises and challenges of evolvable hardware, IEEE Trans. Syst., Man Cybernet.—Part C, 29, 87–97. 41. Ko, M.-S., Kang, T.-W., Hwang, C.-S., (1997), Function optimization using an adaptive crossover operator based on locality, Eng. Appl. Artif. Intell., 10, 519–524. 42. Alba, E., Tomassini, M., (2002), Parallelism and evolutionary algorithms, IEEE Trans. Evol. Comput., 6, 443–462. 43. Harik, G., Lobo, F., Goldberg, D., (1999), The compact genetic algorithm, IEEE Trans. Evol. Comput., 3, 287–297. 44. Miller, J., (1999), An empirical study of the eﬃciency of learning Boolean functions using a Cartesian genetic programming approach, Proc. Genetic Evol. Comput. Conf., Orlando, FL, 1, 1135–1142. 45. Tsai, H.-K., Yang, J.-M., Tsai, Y.-F., Kao, C.-Y., (2004), An evolutionary approach for gene expression patterns, IEEE Trans. Inf. Technol. Biol., 8, 69–78. 46. Hung, C.-M., Huang, Y.-M., Chang, M.-S., (2006), Alignment using genetic programming with causal trees for identiﬁcation of protein functions, Nonlinear Analysis, 65, 1070–1093. 47. Ngom, A., (2006), Parallel evolution strategy on grids for the protein threading problem, J. Parallel Distrib. Comput., 66, 1489–1502. 48. Ghosh, A. and Jain, L.C. (Eds) (2005), Evolutionary Computation in Data Mining, Springer, Germany. 49. Guo, H., Jack, L.B., Nandi, A.K., (2005), Feature generation using genetic programming with application to fault classiﬁcation, IEEE Trans. Syst., Man Cybernet.—Part B, 35, 89–99. 50. Toro, F.D., Ros E., Mota, S., Ortega, J., (2006), Evolutionary algorithms for multiobjective and multimodal optimization of diagnostic schemes, IEEE Trans. Bio-Med. Eng., 53, 178–189. 51. Freitas, H.S., Bojarczuk, A.A., Lopes, C.C., (2000), Genetic programming for knowledge discovery in chest-pain diagnosis, IEEE Eng. Med. Biol. Mag., 19, 38–44. 52. Gordon, M., Fan W.-G., Pathak, P., (2006), Adaptive web search: evolving a program that ﬁnds information, IEEE Intell. Syst., 21, 72–77. 53. Kuo, R.J., Liao, J.L., Tu, C., (2005), Integration of ART2 neural network and genetic K-means algorithm for analyzing web browsing paths in electronic commerce, Decis. Support Syst., 40, 355–374. 54. Bode’n, M., Bode’n, M., (2007), Evolving spelling exercises to suit individual student needs, Applied Soft Computing, 7, 126–135. 55. Jain, L.C. (Ed.), (2000), Evolution of Engineering and Information Systems, CRC Press USA.

2 Evolutionary Neuro-Fuzzy Systems and Applications G. Castellano1 , C. Castiello1 , A.M. Fanelli1 , and L. Jain2 1

2

Computer Science Department, University of Bari, Via Orabona 4, 70126 Bari, Italy [castellano, castiello, fanelli]@di.uniba.it School of Electrical and Information Engineering, University of South Australia, Adelaide, Mawson Lakes Campus, South Australia SA 5095, Australia [email protected]

Summary. In recent years, the use of hybrid Soft Computing methods has shown that in various applications the synergism of several techniques is superior to a single technique. For example, the use of a neural fuzzy system and an evolutionary fuzzy system hybridises the approximate reasoning mechanism of fuzzy systems with the learning capabilities of neural networks and evolutionary algorithms. Evolutionary neural systems hybridise the neurocomputing approach with the solution-searching ability of evolutionary computing. Such hybrid methodologies retain limitations that can be overcome with full integration of the three basic Soft Computing paradigms, and this leads to evolutionary neural fuzzy systems. The objective of this chapter is to provide an account of hybrid Soft Computing systems, with special attention to the combined use of evolutionary algorithms and neural networks in order to endow fuzzy systems with learning and adaptive capabilities. After an introduction to basic Soft Computing paradigms, the various forms of hybridisation are considered, which results in evolutionary neural fuzzy systems. The chapter also describes a particular approach that jointly uses neural learning and genetic optimisation to learn a fuzzy model from the given data and to optimise it for accuracy and interpretability.

2.1 Introduction Hybridisation of intelligent systems is a promising research ﬁeld in the area of Soft Computing. It arises from an awareness that a combined approach may be necessary to eﬃciently solve real-world problems. Each of the basic Soft Computing paradigms has advantages and disadvantages. Fuzzy systems do not have the automatic learning of neural networks, hence they can not be used in the absence of expert knowledge. Fuzzy systems have an advantage over neural networks and evolutionary strategies, as they can express knowledge in a linguistic form. This is familiar and forms part of the human reasoning process. Neural networks are developed speciﬁcally for learning, and are fundamental when a set of signiﬁcant examples of the G. Castellano et al.: Evolutionary Neuro-Fuzzy Systems and Applications, Studies in Computational Intelligence (SCI) 66, 11–45 (2007) c Springer-Verlag Berlin Heidelberg 2007 www.springerlink.com

12

G. Castellano et al.

problem to be solved are available, rather than a solution algorithm for the problem. In general, neural networks can learn from examples. The way in which a knowledge base is extracted through a learning process is not easy for humans to understand. Evolutionary strategies, despite the learning speed being slower than neural networks, are able to use much more general functions than diﬀerentiable ones. This is because they do not require the computation of the gradient of the functions. Finally, as evolutionary algorithms explore in several directions the search space, they are less aﬀected than neural learning by the problem of ﬁnding local minima. A hybrid technique which makes use of a combination of the three Soft Computing paradigms is an interesting prospect. When these constituents are combined, they operate synergistically rather than competitively. Their mutual dependence may produce unexpected performance improvement. In the last decade, several Soft Computing frameworks have been developed for a wide range of domains. Many of them solve a computational task by using a combination of diﬀerent methodologies. The aim is to overcome the limitations and weakness of the several techniques. This chapter gives an overview of diﬀerent viewpoints of hybridisation among Soft Computing paradigms. The main features of principal Soft Computing paradigms are introduced. The attention is focused on all the possible ways of integrating the characteristics of two paradigms. This results in neural fuzzy systems, which is a hybrid of approximate reasoning method of fuzzy systems with the learning capabilities of neural networks, evolutionary fuzzy systems that use evolutionary algorithms to adapt fuzzy systems, and evolutionary neural systems, integrating the solution-searching ability of evolutionary computing with the neurocomputing approach. In conclusion, full hybridisation of the three paradigms is addressed. We summarise the research done on evolutionary neuro-fuzzy systems that integrate the solution-searching feature of evolutionary computing and the learning ability of neurocomputing with the explicit knowledge representation provided by fuzzy computing. An example of full integration between Soft Computing paradigms is presented that makes use of neural learning and genetic algorithms to learn fuzzy models from data. This improves the accuracy and the ease of interpretation. The chapter is organised as follows. Section 2.2 gives the fundamentals on the three main Soft Computing paradigms. That is, neural networks, fuzzy inference systems and evolutionary algorithms. Sections 2.3 to 2.5 provide a review of all possible forms of combination among the paradigms. In section 2.6 diﬀerent examples of evolutionary-neuro-fuzzy hybridisation are considered, and some applications are reported. Section 2.7 presents a hybrid approach for the optimisation of neuro-fuzzy models based on an evolutionary algorithm. An example of the application is included in this section. Section 2.8 concludes the chapter.

2 Evolutionary Neuro-Fuzzy Systems and Applications

13

2.2 Soft Computing Systems The term Soft Computing [193] indicates a number of methodologies used to ﬁnd approximate solutions for real-world problems which contain various kinds of inaccuracies and uncertainties. The guiding principle of Soft Computing is to develop a tolerance for imprecision, uncertainty, partial truth, and approximation to achieve tractability, robustness and low solution cost. The underlying paradigms of Soft Computing are Neural Computing, Fuzzy Logic Computing and Evolutionary Computing. Systems based on such paradigms are Artiﬁcial Neural Networks (ANN’s), Fuzzy Systems (FS’s), and Evolutionary Algorithms (EA’s). Rather than a collection of diﬀerent paradigms, Soft Computing is better regarded as a partnership in which each of the partners provides a methodology for addressing problems in a diﬀerent manner. From this perspective, the Soft Computing methodologies are complementary rather than competitive. This relationship enables the creation of hybrid computing schemes which use neural networks, fuzzy systems and evolutionary algorithms in combination. Figure 2.1 shows their use in a Soft Computing investigation. In the following, after a description of the three basic Soft Computing paradigms, some hybrid Soft Computing systems are overviewed. These show the rapid growth in the number and variety of combinations using ANN’s, FS’s and EA’s. 2.2.1 Neural Networks Artiﬁcial Neural Networks (ANN’s) are computational models that are loosely modelled on biological systems and exhibit in a particular way some of the

Fig. 2.1. Integration possibilities in Soft Computing

14

G. Castellano et al.

Fig. 2.2. The structure of three layers ANN

properties of the brain. They are composed by a number of simple processors (neurons) working in parallel, without any centralised control. The neurons are arranged in a particular structure which is usually organised in layers. A system of weighted connections determines the information ﬂow through the network. In ﬁgure 2.2 a neural network is depicted with two layers, plus an input layer. The behaviour of an ANN is determined by the topology of the connections and by the properties of every processing unit, which typically evaluates a mathematical function. The numeric weights of the connections are modiﬁed in order to give ANN’s a dynamic nature. ANN’s possess also adaptive properties, since they are able to modify their output depending on the actual weight. The weight value is dependent on past experience. The learning of an ANN is based on the analysis of input data (training set) and the particular learning algorithms. The knowledge is eventually embedded into the weight conﬁguration of the network. Neural networks are commonly regarded as learning machines that work on the basis of empirical data. The only means of acquiring knowledge about the world in a connectionist system comes from observational instances. There are no a priori conceptual patterns that could lead to a learning process. The neurons composing the network only perform a very simple task. This consists of evaluating a mathematical function and transmitting a signal. This is similar to the neuronal cells of biological brains. The aim is to adapt to the surrounding data from the system. This rather simpliﬁed structure enables ANN’s to perform quite complex tasks, endowing connectionist systems with the capability of approximating continuous functions to any required degree of accuracy. The task of a neural network is to realise a mapping of the type: φ : x ∈ Rn → y ∈ Rm ,

(2.1)

where x is a n-dimensional input vector and y is an m-dimensional output vector. ANN’s can be variously distinguished in terms of their conﬁgurations, topology and learning algorithms. Generally speaking, two main types of

2 Evolutionary Neuro-Fuzzy Systems and Applications

15

conﬁgurations can be identiﬁed: feedforward networks, these possess connections which propagate the information ﬂow in one direction only. The other type is recurrent networks, whose nodes may have feedback connections, where the input for a neuron may be the output from another neuron. Here the input may be from the particular neuron in question. Each connectionist system is characterised by a particular topology. That is, the organisation of its own neurons. The most typical neural architecture is represented by the multilayered network, where the processing units are fully interconnected in three or more layers and the output is interpolated among all the nodes of the network. On the other hand, in competitive networks the nodes compete to respond to inputs. A particular scenario may have some units that are active, while all others are inhibited. Finally, learning algorithms play an important role in characterising the working mechanism of an ANN. The most commonly employed learning algorithm is the error back-propagation training. This involves adjusting the connection weights under a supervised learning scheme. Other types of learning algorithms include counter propagation mechanisms, Hebbian learning, Boltzmann algorithm and self-organising maps. Further details concerning the characterisation of neural networks, may be obtained from references which include [7, 15, 71, 75, 149]. 2.2.2 Fuzzy Systems Fuzzy mechanisms suitable for tasks involving reasoning have been proposed as an extension to classical formal logic. They were ﬁrst introduced in set theory. The concept of a “fuzzy set” has been employed to extend classical sets, which are characterised by crisp boundaries. This addition permits a degree of ﬂexibility for each object belonging to a particular set. This quality is realised by the deﬁnition of membership functions that give fuzzy sets the capacity of modelling linguistic, vague expression [192]. Fuzzy sets constitute the basis for fuzzy logic, a novel way for developing reasoning models by handling imprecise information, where truth can assume a continuum of values between 0 and 1. This kind of information is often referred to as fuzziness: it should be noted that fuzziness does not come from randomness, but from the uncertain and imprecise nature of abstract thoughts and concepts. Fuzzy reasoning realises a form of approximate reasoning that, using particular mathematical inferences, derives conclusions based on a set of fuzzy IF-THEN rules, where linguistic variables could be involved. In this way, fuzzy logic is suitable for describing the behaviour of systems which are either too complex or too illdeﬁned to be amenable to precise mathematical analysis. Classical systems cannot cope with inexact or incomplete information, because they do not provide any means of representing imprecise propositions and do not possess any mechanism that can make an inference from such propositions. Fuzzy logic systems commonly contain expert IF-THEN rules and can be characterised in terms of their fundamental constituents: fuzziﬁcation, rule base, inference, defuzziﬁcation. Figure 2.3 is a schematic representation of such

16

G. Castellano et al.

Fig. 2.3. The basic components of a fuzzy system

a fuzzy system. Fuzziﬁcation is a mapping from a crisp input space to fuzzy sets in a deﬁned universe: U : xi ∈ R → X ∈ U ⊂ Rq .

(2.2)

Here xi represents a crisp value and q is the number of fuzzy classes. The fuzzy sets are characterised by membership functions which portray the degree of belonging of xi to the values in U, µF (xi ) : U → [0, 1]. The rule base is constituted by an ensemble of fuzzy rules and the knowledge is expressed in the following form: IF x1 is Ak1 AND . . . AND xn is Akn THEN y1 is bk1 AND . . . AND ym is bkm ,

(2.3)

where the index k = 1, . . . , K indicates the k-th rule among the K rules in the rule base; Aki and bkj are fuzzy sets. These are deﬁned over the input components xi , i = 1, . . . , n, and the output components yj , j = 1, . . . , m, respectively. The rule is a fuzzy implication that is usually represented by a Cartesian product of the membership functions of antecedents and consequents. The fuzzy inference process can be described by starting with the deﬁnition of the membership functions µik (·) related to the k-th fuzzy rule and evaluated for each input component of a sample vector x = (x1 , . . . , xn ). The most commonly employed membership functions are the triangular and the Gaussian functions. The values obtained by the fuzziﬁcation contribute to the AND conjunction of each rule, which is interpreted by a particular T-norm. This is most commonly the min or the algebraic product operators. After evaluating the degrees of satisfaction for the entire set of fuzzy rules, the K activation strengths are used in the OR disjunction represented by the alternative rules. These are interpreted by a particular S-norm, the max or the algebraic sum operators are most commonly used. Finally, the defuzziﬁcation process is used to reconvert the fuzzy output values, deriving from the inference mechanism, into crisp values. These can

2 Evolutionary Neuro-Fuzzy Systems and Applications

17

then be eventually employed in diﬀerent contexts. The most common strategy for defuzziﬁcation is to use the centre of area method which gives the centre of gravity of the output membership function. Further details concerning the characteristics of fuzzy systems can be found in [22, 43, 147]. 2.2.3 Evolutionary Algorithms “Evolutionary Algorithms” is a general term for indicating a number of computational strategies that are based on the principle of evolution. Populationbased generate-and-test algorithms are used. An evolutionary algorithm is an iterative probabilistic program, employed in optimisation problems. A population of individuals, P (t) = {pt1 , . . . , ptn } is involved for each iteration t [52, 79, 115]. Each individual pti represents a potential solution for the problem and is represented as some data structure S. The metaphor underlying evolutionary computation enables us to review this encoding process as a translation from a phenotype representation to chromosomal representation or genotype encoding. The population members are successively evaluated in terms of a particular function. This function is speciﬁcally designed for the problem, in order to express the ﬁtness of the individuals. In this way, at iteration t + 1 a new population can be generated on the basis of the ﬁttest elements. These are selected according to some preﬁxed rule to produce an oﬀspring by means of “genetic” operators. They are usually represented by unary transformations, mi : S → S (mutation), and higher order transformations, cj : S × . . . × S → S called crossover, or recombination. The successive generations of several populations hopefully converge to a ﬁnal oﬀspring encoding an optimal solution. In ﬁgure 2.4 the general scheme of an evolutionary algorithm is shown. The most common evolutionary computation techniques are genetic algorithms, so that the terms “Evolutionary Computation” and “GA-based methods” are used almost as synonyms. However, diﬀerent evolutionary techniques should be distinguished. They can be characterised in terms of the data structures S employed to represent the chromosomes, the genetic operators and the individual selection schemes involved. In particular, Genetic Algorithms [59, 79, 115] codify chromosomes as binary strings. Genetic operators are applied only to the genotypical representations where the role of recombination is emphasised. Evolutionary Programming (EP) [52, 53] is a stochastic optimisation strategy similar to GA’s, that gives emphasis on the behavioural linkage between parents and their oﬀspring, rather than seeking to emulate speciﬁc genetic operators as observed in nature. The only genetic operator used in EP is mutation, applied to the phenotypical representations. Similarly to EP, Evolution Strategies [155, 156] make use of real-valued vectors to represent individuals. An individual comprises not only the speciﬁc object parameter set and its ﬁtness function value, but usually also a set of endogenous (i.e. evolvable) strategy parameters that are used to control

18

G. Castellano et al.

Fig. 2.4. The general structure of an evolutionary program

certain statistical properties of the genetic operators. Another evolutionary approach consists in Genetic Programming [5, 99, 101–104], where structures of possible computer programs (trees or even neural networks) evolve to represent the best solution to a problem.

2.3 Neuro-Fuzzy Systems The neuro-fuzzy hybridisation represents by far the most fruitful and the most investigated strategy of integration in the context of Soft Computing. Both neural networks and fuzzy systems are dynamical, parallel processing systems that estimate input-output functions [118]. Fuzzy logic is capable of modelling vagueness, handling uncertainty and supporting human-type reasoning. Connectionist systems rely on numerical and distributed representation of knowledge and they are often regarded as black boxes, revealing little about their inductive engines. Also in terms of application requirements fuzzy logic and neural systems show very contrasting positions. Neural networks are capable of learning from scratch, without needing any a-priori intervention, provided that suﬃcient data are available or measurable. On the other hand, fuzzy systems make use of linguistic knowledge of the process, which can be supplied by human experts. To a large extent, the key-points and the shortcomings of connectionist and fuzzy approaches appear to be complementary. It seems to be a natural practice to build up integrated strategies combining the concepts of the two paradigms [89,94,100,111,122,130,135]. It should be observed that neuro-fuzzy systems are the most prominent representatives of hybridisations in terms of

2 Evolutionary Neuro-Fuzzy Systems and Applications

19

the number of practical implementations. A number of successful applications can be observed in engineering and industrial applications [19]. An inspection of the multitude of hybridisation strategies proposed in literature which involve neural networks and fuzzy logic would be somewhat impractical. It is however straightforward to indicate the general lines underlying this kind of integration. This appears to be twofold. If we consider the fuzzy inference system as the main subject of the hybridisation, neural networks can add learning capabilities to an inference engine which reveals to be spoilt by some self-constructing feature. This approach is referred to as NeuroFuzzy System (NFS). However, when a connectionist system is regarded as the main subject of the hybridisation, fuzzy logic may assist by incorporating fuzziness into the neural framework. This may enable a better view of the black box. This argument is commonly referred to as Fuzzy-Neural System (FNS). In the following we will brieﬂy review some of the most prominent examples of neuro-fuzzy hybridisation proposed in literature. 2.3.1 Fuzzy-Neural Systems Fuzziness could be incorporated into neural networks at diﬀerent levels. Examples are given in literature concerning the fuzziﬁcation of input data, output results, learning procedures, error functions [69, 82, 95, 117, 119, 120, 129, 175]. One of the most intriguing examples of fuzzy-neural hybridisation is investigated by a ﬁeld of research which constructs systems based on fuzzy neurons. These speciﬁc neurons are designed to realise the common operations of fuzzy set theory. That is, fuzzy union, intersection, aggregation. This is instead of the usual standard algebraic functions. Using this technique it may be possible to take advantage of the transparency and readability of the structures. This may result in improved interpretation for the overall system [76, 132, 136]. 2.3.2 Neuro-Fuzzy Systems A number of results have been presented to demonstrate the formal equivalence between fuzzy rule-based systems and neural networks [11,20,21,70]. In particular, a functional equivalence has been shown to exist between radial basis function networks and fuzzy systems [88]. On the basis of this correspondence, various neuro-fuzzy integrations have been proposed to endow fuzzy inference engines with learning capabilities using the neural components. In practice, connectionist learning can be exploited to tune the parameters of an existing fuzzy system and/or to compile the structure of the rule base [34, 36, 50, 92, 96, 128, 166, 168, 170]. Among the most popular and pioneering neuro-fuzzy systems can be cited: GARIC [12], NEFCON [123], ANFIS [87]. All of these are models for successive hybridisation strategies [26, 98, 124–126, 167, 169].

20

G. Castellano et al.

2.4 Evolutionary-Fuzzy Systems The eﬀectiveness of evolutionary techniques in performing complex search processes over spaces has been successfully applied to provide learning capabilities to fuzzy systems. This kind of hybridisation led to the development of genetic fuzzy systems (GFS’s) [6, 42, 134, 152]. A GFS is therefore represented by a fuzzy rule based system which relies on genetic components to accomplish the main processes related to the system design: parameter optimisation and rule generation. Actually, the optimisation process usually concerns the tuning procedure applied over the variables of a fuzzy system. This may be already established in the structure. This conﬁgures the adaptation process. A more proper learning approach is realised when the genetic components are involved in the generation of the rule base, without reference to any preexisting structure. It should be noted that, without loss of generality, we refer to the employment of genetic components. This is because genetic algorithms appear to be the most applied techniques inside this ﬁeld of research. The genetic search processes involved with GFS’s are with a Knowledge Base constituted of: – –

a Data Base, including the deﬁnitions of the parameters and variables of the fuzzy system; a Rule Base, comprising of a number of fuzzy rules.

As previously observed, the exploration of the diﬀerent components of the Knowledge Base is correlated with the diﬀerent objectives that can be attained. These range from parameter adaptation to pure learning activity. It should be observed that, similarly to any other search process, the application of evolutionary techniques to the Knowledge Base is subject to a trade-oﬀ between the dimension of the search space and the eﬃciency of the search. In this way, the smaller the search space will likely lead to a faster process yielding suboptimal solutions. A larger search space is more likely to contain optimal solutions, even if the search process would be less eﬃcient. For the purposes of our presentation, we are ﬁrst going to describe the common applications of genetic procedures when used to create fuzzy rule-based systems, distinguishing diﬀerent genetic adaptation and learning processes, and then some other kinds of hybridisation strategies proposed in literature. 2.4.1 Adaptation of the GFS Data Base The tuning process applied to the Data Base of a GFS aims to optimise the membership functions in the fuzzy rule base. This is commonly predeﬁned for the system. Each individual may encode within its chromosomes the parameters of the Data Base. This codiﬁes the particular shape of the membership functions or even the entire fuzzy partitions used to design the fuzzy rule base. This is especially true when dealing with descriptive fuzzy systems which involve linguistic

2 Evolutionary Neuro-Fuzzy Systems and Applications

21

variables. A number of references are addressed in literature as examples of this approach, and both real-valued and binary-valued chromosomes have been adopted [40, 58, 65, 68, 73, 159]. 2.4.2 Learning of the GFS Rule Base The employment of genetic learning for the generation of fuzzy rules has been traditionally proposed in literature. There are three main approaches and these are the Pittsburgh [163], the Michigan [80] and the Iterative Rule Learning [173]. In the Pittsburgh approach an entire fuzzy rule base is encoded as a chromosome, and thus is one of the individuals of the candidate population. The Michigan approach, on the other hand, codiﬁes each individual rule in a chromosome, so that the population represents the fuzzy rule base. The IRL approach encodes separately each rule and builds up the fuzzy rule base in an iterative fashion. This is done by adding a rule at every application of the genetic algorithm. In every case, when we consider approaches devoted to GFS Rule Base learning, we should be concerned with a particular scenario where the Data Base or the set of the involved membership functions is pre-conﬁgurated. The most common way to codify a rule is to represent it in the disjunctive normal form. A set of rules is generally encoded as a list. In this way, the chromosomes can be given by the code of a single rule or by the assembly of several rules. Bibliographical references can be classiﬁed as belonging to the Pittsburgh approach [78,141], the Michigan approach [16,84] and the IRL approach [39, 61]. 2.4.3 Learning of the GFS Knowledge Base The genetic component can be employed to determine the entire Knowledge Base pertaining to a GFS. Here both the parameter adaptation and the rule base learning are assessed by the genetic search process. The fundamental strategies introduced in the previous section can be restated to characterise the approaches in literature. They distinguish between the Pittsburgh-based [8, 23], the Michigan-based [131, 172] and the IRL-based [40, 41] learning methodologies. 2.4.4 Diﬀerent Hybridisation Strategies In order to complete the presentation, we mention a number of hybridisation strategies which integrate evolutionary techniques and fuzzy logic diﬀerently to the genetic adaptation and learning processes discussed in the previous sections. First of all, it is pointed out how diﬀerent approaches, other than genetic algorithms, have been employed to evolve fuzzy rule bases. In particular,

22

G. Castellano et al.

Genetic Programming can be utilised for this purpose, provided that the fuzzy rules are codiﬁed by coherent computer program structures, such as syntactic trees [2, 9, 35, 57, 77]. Recent sophisticated versions of genetic algorithms have been successfully adopted, and these include parallel GA’s and coevolutionary algorithms [33, 45, 138, 143, 146]. Evolutionary algorithms have been also applied to solve peculiar questions connected with the management of a fuzzy rule base. It must be mentioned the employment of genetic techniques in high-dimensional problems to perform the selection of fuzzy rules [60,83,148], and the selection of features [24,62,108]. Recent works concern the growing interest in issues of comprehensibility. They should be regarded with the same attention as issues of accuracy arising from the application of fuzzy systems [83, 85, 91]. In a diﬀerent way, GA’s have found application in conjunction with fuzzy logic to improve the capabilities of clustering algorithms such as fuzzy C-means [14]. In particular, genetic strategies can be adopted to optimise the parameters of a FCM-type algorithm [66, 171], used to deﬁne the distance norm [191] and to directly produce the clustering results [18]. We point out that the evolutionary-fuzzy hybridisation can be analysed in the reverse direction, with respect to that so far considered. This is done even if it appears to be less proﬁtable. Instead of considering the learning capabilities that genetic strategies could add to fuzzy system, fuzzy evolutionary algorithms focus on the management of GA parameters such as mutation rate or population size, using fuzzy systems [72, 137, 165]. To conclude our discussion, two bibliographical references assess the application possibility of GFS’s [17, 46].

2.5 Evolutionary-Neural Systems The combination of evolutionary algorithms and neural networks has led to the development of the so-called “Evolutionary Artiﬁcial Neural Networks” (EANN’s) [181]. Using this term it is common to indicate a particular class of connectionist systems having augmented capabilities of adaptation to the external environment. If we regard the learning capabilities of a neural network as a form of adaptation based on the basis of data investigation, EANN’s are able to show enhanced adaptability in a dynamic environment, due to the evolutionary components. This kind of ﬁtting process may take place at several levels. These range from connection weights to network topology or even to learning rules. Strategies of evolution of connection weights are in many cases oriented to replace the classical back-propagation and conjugate gradient algorithms. It is hoped to overcome the drawbacks of gradient-descent techniques. Adaptation of architectures is also useful to avoid the tedious trial and error mechanism used by the human knowledge expert, which up to now remains the most common way of designing connectionist systems. With the evolution possibilities of learning rules a kind of “learning to learn” process

2 Evolutionary Neuro-Fuzzy Systems and Applications

23

could be developed. This would enable neural networks to become ﬂexible and adjust to the problems to be solved. This could lead to an improvement of the learning performance, that is known to be dependent on the environment conditions and the task at hand. This would be instead of relying on individual training algorithms [179]. The adaptation process supported by the evolutionary components of an EANN is subjected to the usual cycle of computation. It is noted that some problems are inherently connected with the evaluation process of the individuals. The noisy ﬁtness evaluation problem [182] compromises a fair estimate of the generated individuals. That is due mainly to the unavoidable oneto-many mapping between genotype and phenotype, whenever evolution of network topology is performed without consideration of the weight connections. The noise apparent during the evaluations of the ﬁtness is related to the approximation of the genotype ﬁtness value, that is the network encoding, with the phenotype ﬁtness value. The network is evaluated by means of random initial weights during the evolutionary iterations. The permutation problem [67] could also occur when a many-to-one mapping from the genotypical to the phenotypical representation is done. This is due to the fact that a permutation of the hidden nodes of a network produces equivalent ANN’s with diﬀerent chromosome representations. Further analyses concerning the genotype/phenotype mapping can be found in [51, 52]. We will now brieﬂy describe the diﬀerent types of integration strategies of evolutionary algorithms and neural networks, as given in the literature. 2.5.1 Adaptation of the Connection Weights The idea of applying evolutionary algorithms to adapt neural network weights could appear straightforward if we consider the network learning as a search for an optimal parameter conﬁguration. Actually, EA’s can be helpful in order to overcome the limitations of gradient descent methods. They may avoid the problem of them becoming trapped in local minima of the error function. They are then able to seek globally optimal solutions, without the need of referring to continuos and diﬀerentiable functions. The employment of classical genetic algorithms is to determine a representation of connection weights via binary strings. Each individual can be reproduced by concatenating the binary code of the single weight conﬁguration. This yields a general representation and allows a straightforward application of both crossover and mutation operators [90, 164, 178]. A more compact representation could be obtained by using real numbers which directly correspond to the weight values [63, 142, 153]. This approach can produce a more compact chromosome representation. Some diﬃculties can arise with the application of traditional mutation and crossover operators. For this reason, diﬀerent strategies have been employed in this situation. These include evolutionary strategies implementing only mutation [154, 187].

24

G. Castellano et al.

The adoption of EA’s for weight adaptation may be shown to be useful in enhancing the accuracy of neural networks. However, the most promising applications appear to be the development of cooperative training processes. In fact, the ability of EA’s to ﬁnd global basin of attraction may be coupled with some techniques of local search. The employment of evolutionary techniques to ﬁnd near-optimal initial condition, followed by back-propagation local search has been successfully applied [107, 189]. 2.5.2 Adaptation of the Architectures In a similar way to the search for connection weights, the design of a network architecture can be regarded as a peculiar search process. In principle it can be done by means of evolutionary methods. The problem here can be embedded into the chromosome representation. The direct encoding approach aims at encoding all the parameters useful for realising a correspondence between the network topology and the employed data structure. This can be represented by a matrix recording the connections between the nodes [114,116]. This kind of approach is likely to increase the dimension of the coding structure and therefore some indirect encoding strategies have been proposed. These make use only of a set of the most important factors needed to deﬁne the network architecture. Other details may be either disregarded, predeﬁned or speciﬁed by means of deterministic rules [139, 174, 188]. In order to avoid the noisy ﬁtness evaluation problem mentioned before, architecture and weight connections have sometimes evolved together. A fully speciﬁed neural network may be codiﬁed in a one-to-one mapping [48, 144, 182, 183, 185]. 2.5.3 Adaptation of the Learning Rules The ultimate application of EA’s to neural network optimisation relies on the adaptation possibilities of the learning rules. The selection of an appropriate training algorithm, together with the weight initialisation and the topology design, plays a fundamental role in determining the behaviour of a connectionist system. It can be argued that the principle of applying the same learning rules for every task is impracticable. For these reasons, the help supplied by an evolutionary search in dynamically determining a proper learning strategy in diﬀerent contexts would be valuable. The process of encoding a learning rule represents a challenging task. Some proposals can be found in literature, and these range from the plain evolution of the algorithmic parameters of back-propagation methods [97], to the eﬀective evolution of learning rules, which have been assumed to be particular functions. These have been evolved in terms of their coeﬃcients [10, 32].

2 Evolutionary Neuro-Fuzzy Systems and Applications

25

2.5.4 Diﬀerent Hybridisation Strategies The hybridisation of evolutionary algorithms and neural networks can be done in ways which are quite diﬀerent to those described in the previous sections. Instead of considering the support provided by EA’s for designing a connectionist system in terms of weight conﬁgurations, topology and learning rules, a preprocessing activity can be used which is suitable for eﬃcient neural training. Evolutionary methods are employed to develop an input feature selection process [64, 106]. The integration of the paradigms can also be considered in the use of an ANN to support evolving procedures. Some have suggested the employment of neural networks for controlling the parameters involved in ﬁtness evaluation [37]. A fruitful ﬁeld of research is represented by the use of ensembles of neural networks. An ensemble of ANN’s is regarded as a population which may oﬀer better generalisation opportunities when considered as an integrated system [113, 184].

2.6 Evolutionary-Neuro-Fuzzy Systems An evolutionary neuro-fuzzy system (ENFS) is the result of adding evolutionary search procedures to systems, integrating fuzzy logic computing and neural learning. Using these three Soft Computing paradigms together can overcome some limitations of simpler forms of hybridisation. One of the main problems with neuro-fuzzy systems is that the learning algorithm is typically based on a steepest descent optimisation technique minimising an error function. That is, back-propagation training is not guaranteed to converge. The algorithm may be trapped in a local minimum. It can never ﬁnd the global solution. The tuning of the membership function parameters through neural learning is also not guaranteed. As previously stated (see section 2.5.1), experimental evidence has shown cases where evolutionary algorithms are ineﬃcient for ﬁne tuning solutions. They are better at ﬁnding global solutions. A hybrid learning scheme combining neural learning and evolutionary strategies may be able to solve this problem. A common approach involves using a genetic algorithm to rapidly locate a good region in the solution space and to initialise membership function parameters. The parameters are then ﬁne tuned by applying a gradient descent learning algorithm that performs a local search in the good region to ﬁnd a near optimal solution [194]. A problem of (neuro-)fuzzy modelling is the diﬃculty of determining the proper number of rules and the number of membership functions for each rule (see section 2.4). Evolutionary approaches can be integrated into neuro-fuzzy modelling to overcome this diﬃculty and to perform structure and parameter optimisation of the fuzzy rule base.

26

G. Castellano et al.

Several methodologies have been proposed in the literature to develop a form of evolutionary-neural-fuzzy hybridisation. We now review some of the existing approaches, grouping methodologies to reach the goal of optimisation. Some of them aim to learn parameters of a fuzzy rule base, whose structure is assumed to be ﬁxed in advance. Other methods are developed to perform both structure and parameter learning of a fuzzy rule base. 2.6.1 Parameter Learning A common approach given in the literature to assess an ENFS is by adding evolutionary learning capabilities to a neuro-fuzzy network. This network is usually a feed-forward multilayered network, which incorporates some fuzzy concepts. The result is a feed-forward multilayered network having both fuzzy and genetic characteristics. In simplest approaches, a GA is used to learn or tune all the parameters of a neuro-fuzzy network which is assumed to have a ﬁxed structure [112, 127, 145, 157, 177]. In these cases chromosomes of the GA encode all the parameters of the neuro-fuzzy model, including parameters of membership functions as deﬁned in the antecedent and consequent of each fuzzy rule. In [49] the parameter identiﬁcation of a fuzzy model is achieved in three separate learning phases. The ﬁrst phase uses the Kohonen’s self organising feature map algorithm to ﬁnd the initial parameters of the membership functions. A maximum matching-factor algorithm is applied, in the second phase, to ﬁnd the correct consequence part of each rule. After the fuzzy rules have been found and the whole network structure is established, the third learning phase ﬁne-tunes the membership function parameters using a multiresolutional dynamic genetic algorithm (MRD-GA) that dynamically adapts the fuzzy-model conﬁguration during the optimisation process. 2.6.2 Structure and Parameter Learning Genetic algorithms may also be applied for structure optimisation of a fuzzy rule base. That is, to deﬁne the proper number of rules and the number of antecedents and consequents in each rule, as in [54]. Evolutionary strategies can be conveniently applied to simultaneously learning the parameters and the structure of a fuzzy model. Several approaches are given in the literature that involve the membership function parameters together with the fuzzy rule set, including the number of rules. In some cases the antecedent and consequent part of each rule are also included. As an example, the strategy proposed in [4] optimises the whole fuzzy system, represented as a neural network, in order to derive a minimum rule number ﬁtting the given speciﬁcations, while training the network parameters. In [55] a GA-based method is used for a rough search for proper structures in the antecedent of fuzzy rules. The ﬁne tuning of the parameters of the fuzzy model

2 Evolutionary Neuro-Fuzzy Systems and Applications

27

is done successively using neural learning. A complex hybrid genetic-neurofuzzy scheme is implemented in FuGeNeSys [150] that can learn a fuzzy rule base from the data. This uses the membership function parameters, the number of rules and the structure of each rule. These are simultaneously deﬁned via a genetic algorithm that incorporates neural learning in the evolution procedure. Each individual in the population is made up of a set of rules, where each rule has an adjustable number of antecedents and consequents. Whenever a better individual is generated, it is transformed in a neuro-fuzzy network and trained with a back-propagation procedure until the error is reduced. Finally, the trained neuro-fuzzy network is retransformed in a genetic individual and used to replace the original individual. In other approaches, such as the EvoNF (Evolving Neuro Fuzzy) model proposed in [1], all the adjustable parts of a fuzzy system, such as the membership function parameters, structure of the rule base are involved. This includes items such as the number of rules, representation of antecedents and consequents. The fuzzy operators are derived through an evolutionary strategy. Coding all such informations into a chromosome leads to individuals characterised by a complex layered structure. The optimisation strategy may become computationally expensive. Neural learning can be used to further tune the membership functions located by the evolutionary algorithm. Finally, genetic algorithms can be used to evolve fuzzy neural networks. That is, neural networks that incorporate fuzzy numbers as weights, perform fuzzy operations in the nodes, and/or consider fuzzy nodes to represent membership functions. In this case, the learning process uses GA’s to obtain the weights of the neural network, to adapt the transfer functions of the nodes, and/or to adapt the topology of the network, as in [105]. 2.6.3 Applications Hybrid systems using a combination of neural, fuzzy and evolutionary computing have been successfully employed in many applications, such as control [162], manufacturing [121], ﬁnancial prediction [190], pharmacology [151], consumer products [161], telecommunications [13], modelling and decision making. Some of these ENFS’s application samples are reported in the literature. Our account does not claim to represent a complete overview. In [161], the authors report an application of evolutionary computation in combination with neural networks and fuzzy systems for intelligent consumer products. The role of the evolutionary algorithm is to adapt the number of rules and to tune the membership functions to improve the performance of particular fuzzy systems. They are involved in predicting the number of dishes to be cleaned by a dish washer, estimating the amount of rice in a rice cooker and controlling a microwave oven. The paper also mentions evolutionary computation for fuzzy rule generation when applied to process control.

28

G. Castellano et al.

In [190] an evolutionary fuzzy neural network is proposed for ﬁnancial prediction with hybrid input data sets from diﬀerent ﬁnancial domains. In [3] a genetic-neuro-fuzzy system is applied to on-line recognition of on-line Arabic cursive handwriting. In this system, a genetic algorithm is used to select the best combination of characters recognised by a fuzzy neural network. The value returned from the ﬁtness function for each gene represents the degree of match between the word represented by that gene and the real handwritten word. The calculation of the ﬁtness value is based on the fuzzy values assigned to the recognition of each character by the fuzzy neural network. In [121] the authors present a neuro-genetic-fuzzy system for computerised colour prediction, which is a challenging problem in paint production. A fuzzy knowledge base for predicting the pigment concentration of ten diﬀerent colours for a given surface spectral reﬂectance is obtained by means of a neuro-fuzzy system. The fuzzy population generator uses this knowledge to seed the ﬁrst generation of colour chromosomes. In addition, expert knowledge of the colour technician about the correct proportions of colours, the number of necessary colours and conﬂicts between complementary and similar colours is summarised in a fuzzy rule base that is introduced in the ﬁrst generation of colour chromosomes. The GA calculates one component of colour chromosome ﬁtness according to the compliance of the chromosome’s colour with the fuzzy expert rules.

2.7 A Genetic-Based Approach to Optimise Neuro-Fuzzy Models Research on NFS’s and EFS’s has for long been considered the objective of the learning process in terms of accuracy. Consequently, the error measure as minimised by neural learning, or the ﬁtness function as maximised by the genetic algorithm was stated in terms of errors or distances from the target output. This causes a lack of interpretability and transparency in the resulting fuzzy rules. This is due to generation of meaningless and overlapping fuzzy sets. Indeed, the most important motivation to use a fuzzy system is that the produced model is characterised by a clear reliance on linguistic terms that allows for easy understanding by humans. Hence, methods for training fuzzy models from data should not only ﬁnd the best approximation of the data, but also and more importantly, they should extract knowledge from the data in the form of fuzzy rules that can be readily understood and interpreted. Recently, the concepts of linguistic fuzzy modelling, interpretability, and similar ideas were considered to have qualities almost the opposite to accuracy. These have now been reconsidered [25] and today are viewed as an interesting part of the design process of a FS using neural learning and genetic algorithms. Several works have suggested both accuracy and interpretability as

2 Evolutionary Neuro-Fuzzy Systems and Applications

29

objectives in genetic-based learning systems [83, 85, 91]. This sets a situation for the learning process where several diﬀering objectives have to be simultaneously considered. Some of the measures used to determine the level of interpretability of a fuzzy system are degree of compactness [83]. This may be regarded as the number of rules in the rule base, and rule simplicity [31]. This simplicity of rules may be evaluated through the number of input variables involved in each rule. In this section we present an approach to identify accurate and interpretable fuzzy models based on a combination of neural learning and genetic algorithms [30]. The approach is aimed to optimise both the structure and the parameters of the fuzzy model extracted from data by a neural learning scheme through a multi-objective genetic algorithm. An iterative process is performed in each step using two optimisation procedures. A structure optimisation procedure that reduces the number of rules, and a multi-objective genetic algorithm that tunes membership function parameters. This is done by enforcing constraints on the fuzzy sets to ensure they result well-formed, leading to fuzzy rules that can be easily understood. Returning to the fuzzy rule expressed in (2.3): IF x1 is Ak1 AND . . . AND xn is Akn THEN y1 is bk1 AND . . . AND ym is bkm , we may suppose that the fuzzy model, generated by the neural learning algorithm, makes use of fuzzy singletons bkj (j = 1, . . . , m). These are deﬁned over k the output variables yj and the fuzzy represented by Gaussian sets Ai are

−cik ) , where cik and aik are the membership functions µik (xi ) = exp − (xi2a 2 ik centre and the width of the Gaussian function, respectively. Based on a set of K rules, the output of the fuzzy inference system for a given input x(0) is obtained as follows: 2

K yˆj (0) = k=1 K

µk (x(0))bkj

k=1

µk (x(0))

j = 1, . . . , m.

(2.4)

n Here µk (x(0)) = i=1 µik (xi (0)), where k = 1, . . . , K is the degree of fulﬁlment for the k-th rule, for k = 1, . . . , K. To better formalise the scheme of the GA-based approach, we denote by F RB(w, K) a Fuzzy Rule Base with parameter vector w and structure size K. That is, with a number of K fuzzy rules. We indicate by F RB(w0 , K0 ) the rule base of the fuzzy model initially derived by neural learning. Firstly, the structure optimisation procedure is applied to F RB(w0 , K0 ) to remove iteratively rules, until the diﬀerence between the accuracy of the reduced model and the accuracy of the initial model drops below a threshold . The result is a new fuzzy rule base F RB(wS , KS ), with 0 ≤ S ≤ (K0 − 1). Here KS ≤ K0 and ERR(wS ) − ERR(w0 ) ≤ , where ERR(·) is an error function used to evaluate the accuracy (generalisation ability) of the fuzzy model. Such a fuzzy

30

G. Castellano et al.

Fig. 2.5. Scheme of the GA-based optimisation approach

rule base is taken as starting point for the GA-based optimisation procedure. The result of this parameter optimisation is a rule base F RB(w S , KS ) that satisﬁes the following conditions: ERR(wS ) ≤ ERR(wS ),

IN T (wS ) ≥ IN T (wS ),

where IN T (·) is a qualitative measure of the interpretability of the antecedent fuzzy sets, that will be detailed in Section 2.7.2. If the parameter optimisation process is able to improve the accuracy, that is, if ERR(wS ) < ERR(wS ), then the structure-parameter optimisation cycle can be reiterated, using the fuzzy rule base F RB(wS , KS ) as starting point. Otherwise, that is, if the local parameter tuning remains unchanged the accuracy, the whole optimisation algorithm is stopped. Providing the rule base F RB(wS , KS ) is the ﬁnal rule base. The whole iterative algorithm ends when no further improvement in the accuracy and interpretability of the fuzzy rules can be observed. That is, when the best possible compromise between accuracy and interpretability of the fuzzy model is achieved. Figure 2.5 shows the ﬂow chart of the whole optimisation approach. We now describe both of the optimisation procedures.

2 Evolutionary Neuro-Fuzzy Systems and Applications

31

2.7.1 Structure Optimisation The structure optimisation procedure is referred to in [27,28] and is a systematic process that sequentially removes rules and investigates the model with fewer rules. At each step, rules to be removed are identiﬁed and the remaining rules are updated so the accuracy of the reduced model remains unchanged. This is achieved by updating only the consequent parameters, while premise parameters are left unchanged. The removal of a rule involves the elimination of the fuzzy sets in its antecedent. Thus, after the structure simpliﬁcation process, the number of fuzzy sets is also reduced, so as to produce a simple (hence easily interpretable) rule base. There is no guarantee that the remaining fuzzy sets still cover the entire input space. The use of Gaussian membership functions assures enough coverage and provides an acceptable accuracy when few rules are removed. As soon as many rules are removed, the coverage is no longer guaranteed. The remaining fuzzy sets may reveal a high degree of overlapping. That is, they may lack in distinguishability. To deal with these problems, a multi-objective genetic algorithm is used to tune the premise parameters of reduced fuzzy rules. This is done to improve interpretability in the sense of readable fuzzy sets with no loss in accuracy of the model. In the following section we describe in detail the multi-objective genetic algorithm. 2.7.2 Multi-Objective GA for Parameter Optimisation The parameters of membership functions in the antecedents of each fuzzy rule are encoded into an individual (chromosome). Since we have nK premise membership functions, each function being represented by two real values (i.e. cik and aik ), the length of each individual is 2nK. To limit the length of individuals, we adopt a coding with real-valued genes in the chromosome which directly represent the premise parameters of a fuzzy rule base. Thus, the i-th chromosome is a string of the form: (i)

(i)

(i)

(i)

(i)

(i)

(i)

(i)

si = (c11 , a11 , · · · , cn1 , an1 , · · · , c1K , a1K , · · · , cnK , anK ). premise of rule 1

premise of rule K

The ﬁrst individual of the initial population is generated as a copy of the premise parameters of the fuzzy rules produced by the structure optimisation procedure. The remaining individuals are initialised with random values constrained within certain permissible ranges, that is: cik = rand · range{xi (t)} + min {xi (t)} t=1,...,N

and aik = rand ·

2 range{xi (t)} + ai,min , 3 KS

32

G. Castellano et al.

where rand is a random number in [0, 1], range{xi } is deﬁned as:

range{xi } = max {xi (t)} − min {xi (t)} t=1,...,N

t=1,...,N

and ai,min is the minimum width among all the widths in the ﬁrst individual. By doing so, the centres of membership functions are constrained within the corresponding input range, while the widths are greater than zero and not wider than the whole input range. The reproduction probability of an individual is deﬁned from the values of the ﬁtness function according to: pR (si ) = NfIit(si ) . To speed up the j=1

f it(sj )

convergence, the best individual in the population is always selected and kept unchanged in the next generations, according to the elitist strategy [59]. The simplest form of crossover, which is the single point crossover, is adopted. Individuals are mutated by adding a random value generated within the permissible ranges, and deﬁned by: = N (α · range{xi }, (β · range{xi })2 ), where α and β assume diﬀerent values depending whether the gene is the centre or the width of a Gaussian function. For each input variable xi , genes representing the centres of extreme membership functions are ﬁxed to the extremes of range{xi } and are never changed by mutation. A readable fuzzy partition for xi requires the leftmost (or the rightmost) membership function to attain its maximum value at the extreme point deﬁned by the minimum (or the maximum) of range{xi }. Dynamic crossover and mutation probability rates are used, since they provide faster convergence [160]. The crossover probability is set high at the beginning of the evolution and decreases exponentially with the increasing number of generations. The mutation probability increases exponentially with each succeeding generation. The expression for the ﬁtness function was derived to model the two objectives of the optimisation process. These are the interpretability and the accuracy of the fuzzy model. To deﬁne the ﬁtness term with respect to interpretability, the concept of interpretability is translated into the concepts of completeness and distinguishability of the fuzzy sets, that are expressed through a fuzzy similarity measure. If the similarity of two neighbouring fuzzy sets is zero or too small, this indicates that either the fuzzy partitioning is incomplete or they do not overlap enough. If the similarity value is too large, then it means that the two fuzzy sets overlap too much and the distinguishability between them is lost. The following measure [158] has been used to deﬁne the similarity between the two fuzzy sets A and B: M (A B) , (2.5) S(A, B) = M (A) + M (B) − M (A B) −∞ where M (A) = +∞ µA (x)dx is the size of fuzzy set A. This is computed as the area of the triangle approximating the Gaussian function µA (· ). To

2 Evolutionary Neuro-Fuzzy Systems and Applications

33

keep the fuzzy sets in a proper shape, the fuzzy similarity measure of any two neighbouring membership functions must satisfy the following condition: Slow ≤ S(Al , Al+1 ) ≤ Sup ,

(2.6)

where Al and Al+1 are the two neighbouring fuzzy sets, Slow and Sup are the desired lower and upper bound of the fuzzy similarity measure, respectively. For each input variable xi , the quantity: 1

βil = {0

if Slow ≤S(Ail ,Ai,l+1 )≤Sup otherwise

(2.7)

is deﬁned. We can derive the following term which must be inserted into the ﬁtness function of an individual: f itint (s) =

β n(KS − 1)

,

(2.8)

where KS is the number of fuzzy sets deﬁned on each input variable. This is equal to the number of rule units in the network and β is deﬁned as: β=

n K S −1

i=1

βil .

(2.9)

l=1

The value f itint (s) is maximum (that is, equal to 1) when all adjacent fuzzy sets coded into the individual s satisfy the condition (2.6). Conversely, if this condition is not satisﬁed by most of adjacent fuzzy sets, the corresponding ﬁtness term will be assigned a very low value which means that the corresponding individual is unlikely to survive. To preserve and, hopefully, improve the accuracy of the fuzzy model, the modelling error ERR(wi ) should be minimised. The value wi is obtained by joining the premise parameters coded into the individual si and the vector of consequent parameters. This vector is taken from the rule base F RB(wS , KS ) resulting from the structure optimisation procedure, where wi = [si , bS ]. The corresponding ﬁtness term is deﬁned as: (2.10) f itacc (si ) = exp −λERR(wi )2 , where λ is a ﬁxed parameter with a large value. A high value of f itacc (si ) corresponds to a very low error, or high accuracy, of the fuzzy model coded into the individual si . To summarise, the analytical expression of the ﬁtness function for an individual si is given by: f it(si ) = f itacc (si ) + γf itint (si ),

(2.11)

where γ is a factor that controls the inﬂuence of the term f itint (si ) during the whole GA evolution. It is made less relevant during the ﬁrst generations, and has more and more inﬂuence as the evolution proceeds.

34

G. Castellano et al.

2.7.3 Illustrative Example In this section, we will examine the application of this approach when used to derive fuzzy models for a problem of medical diagnosis. In particular, the Heart Disease data set from the University of California, Irvine, is considered. The database contains 270 samples. Each sample is represented by 13 attributes. There are two classes: presence or absence of heart-disease. By using the neuro-fuzzy learning method proposed in [29] and performing a 10-fold cross validation strategy, we derived 10 diﬀerent fuzzy models with a number of rules varying between 10 and 16. The application of the GA-based approach gave the results summarised in Table 2.1. It can be seen that a great improvement can be achieved in terms of complexity reduction and accuracy on the test set. The ﬁnal number of rules was between 2 and 4. Almost in all trials, the structure-parameter optimisation required two iterations to improve the neuro-fuzzy model. As an illustrative example of such a behaviour, Table 2.2 gives the detailed results obtained for the model no. 6 where, starting from a 15-rule approximate model with low classiﬁcation rate Table 2.1. Results obtained after the optimisation of neuro-fuzzy models with 10fold cross validation model 1 2 3 4 5 6 7 8 9 10 average

before optimisation number of rules class. rate 16 59.25 12 70.37 13 70.37 10 74.07 11 74.07 15 70.37 12 55.56 10 74.07 13 74.07 11 59.26 12.3 68.15

after optimisation number of rules class. rate 4 70.37 3 70.37 2 81.48 2 74.07 3 70.37 3 74.07 2 55.56 4 88.89 2 85.18 4 88.89 2.9 75.92

Table 2.2. Results obtained at each stage of the GA-based optimisation approach in the case of the model no. 6 model initial model

number of rules

15 iteration no. 1 structure opt. 10 GA-based opt. 10 iteration no. 2 structure opt. 3 GA-based opt. 3

class. rate train. set test set 70.37 70.37 69.96 66.67

66.67 74.07

66.67 65.43

70.37 74.07

2 Evolutionary Neuro-Fuzzy Systems and Applications

35

1

1 x1

0.5

0.5 0

0 40

30

50

60

70

30

1

40

50

60

70

300

400

500

1 x3

0.5 0

0.5 0

200

300

400

500

200

1

1 x5

0.5 0

0.5 0

100

150

200

100

150

200

1

1 x7

0.5

0.5 0

0 100

150

(a)

200

100

150

200

(b)

Fig. 2.6. Fuzzy sets of some input variables before (a) and (b) after the application of the GA-based optimisation process

on the test set, a ﬁnal fuzzy model with only 3 rules and improved classiﬁcation rate was obtained. Also, the interpretability of the ﬁnal model is clearly improved, as shown by the well-formed ﬁnal fuzzy sets in ﬁg. 2.6.

2.8 Conclusions In this chapter we have discussed the integration possibilities among the three paradigms that normally constitute the key components of the Soft Computing ﬁeld of research. These are artiﬁcial neural networks, fuzzy logic and evolutionary algorithms. We reviewed the hybridisation mechanisms at the basis of the coupling approaches. The intent was to address the issues related to the intrinsic beneﬁts and diﬃculties connected with their implementation. We considered the more complex integration of the three paradigms, which gives rise to the development of Evolutionary Neuro-Fuzzy Systems. We proposed a technique concerning design techniques used for the realisation of EFNF’s, together with their application feasibility. In addition, we have presented a particular approach of evolutionaryneuro-fuzzy integration, devoted to data-driven learning of fuzzy inference models. Particular attention has been paid in organising a hybrid system which could prove to be eﬀective in constructing an accurate and comprehensible fuzzy rule base.

36

G. Castellano et al.

We hope that this work could trigger the increase in the research eﬀorts devoted to progress the Soft Computing ﬁeld of investigation. To further motivate such an endeavour, we oﬀered an elaborated series of bibliographical references which could serve as a directory for the inclined reader to be oriented among both classical and novel literature contributes.

References 1. Abraham A (2002) EvoNF: A framework for optimization of fuzzy inference systems using neural network learning and evolutionary computation. In: Proc. 2002 IEEE International Symposium on Intelligent Control (ISIC’02), Canada, IEEE Press, 327–332 2. Alba E, Cotta C, Troya JM (1999) Evolutionary design of fuzzy logic controllers using strongly-typed GP. Mathware Soft Comput. 6(1):109–124 3. Alimi AM (1997) An evolutionary neuro-fuzzy approach to recognize on-line Arabic handwriting. In: Proceedings of the Fourth International Conference on Document Analysis and Recognition, 382–386 4. Alpaydin G, Dandar G, Balkir S (2002) Evolution-based design of neural fuzzy networks using self-adapting genetic parameters. IEEE Trans. Fuzzy Systems 10(2):211–221 5. Angeline PJ, Kinnear KE (eds) (1996) Advances in Genetic Programming II. MIT Press, Cambridge, MA 6. Angelov PP (2002) Evolving Rule-Based Models. A Tool for Design of Flexible Adaptive Systems. Physica-Verlag, Wurzburg 7. Arbib MA (ed) (2002) The Handbook of Brain Theory and Neural Networks (2nd Edition). MIT Press, Cambridge, MA 8. Baron L, Achiche S, Balazinski M (2001) Fuzzy decision support system knowledge base generation using a genetic algorithm. Int. J. Approximate Reasoning 28(1):125–148 9. Bastian A (2000) Identifying fuzzy models utilizing genetic programming. Fuzzy Sets and Systems 113:333–350 10. Bengio S, Bengio Y, Cloutier J, Gecsei J (1992) On the optimization of a synaptic learning rule. In: Preprints Conf. Optimality in Artiﬁcial and Biological Neural Networks 11. Benitez JM, Castro JL, Requena I (1997) Are artiﬁcial neural networks black boxes?. IEEE Trans. Neural Networks 8:1156–1164 12. Berenji HR, Khedkar P (1992) Learning and tuning fuzzy logic controllers through reinforcements. IEEE Trans. Neural Networks 3:724–740 13. Beritelli F, Casale S, Russo M (1995) Robust phase reversal tone detection using soft computing. In: Proc. of ISUMA - NAFIPS ’95 The Third International Symposium on Uncertainty Modeling and Analysis and Annual Conference of the North American Fuzzy Information Processing Society, 589–594 14. Bezdek JC (1981) Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum Press, New York 15. Bishop CM (1995) Neural Networks for Pattern Recognition. Oxford University Press 16. Bonarini A (1996) Evolutionary learning of fuzzy rules: competition and cooperation. In: [133]

2 Evolutionary Neuro-Fuzzy Systems and Applications

37

17. Bonissone PP, Khedkar PS, Chen Y (1996) Genetic algorithms for automated tuning of fuzzy controllers: a transportation application. In: Proc. Fifth IEEE Int. Conf. on Fuzzy Systems (FUZZ-IEEE96), pp. 674–680 18. Buckles BP, Petry FE, Prabhu D, George R, Srikanth R (1994) Fuzzy clustering with genetic search. In: Proc. 1st IEEE Conf. on Evolutionary Computation (ICEC94), pp. 46–50 19. Buckley JJ, Feuring T (1999) Fuzzy and Neural: Interactions and Applications. Studies in Fuzziness and Soft Computing. Physica-Verlag, Heidelberg, Germany 20. Buckley JJ, Hayashi Y (1993) Numerical relationship between neural networks, continuous functions and fuzzy systems. Fuzzy Sets Syst. 60(1):1–8 21. Buckley JJ, Hayashi Y, Czogala E (1993) On the equivalence of neural nets and fuzzy expert systems. Fuzzy Sets Syst. 53(2):129–134 22. (2003) Advanced Fuzzy Systems Design and Applications. Studies in Fuzziness and Soft Computing (112), Physica-Verlag 23. Carse B, Fogarty TC, Munro A (1996) Evolving fuzzy rule based controllers using genetic algorithms. Fuzzy Sets and Systems 80:273–294 24. Casillas J, Cordon O, del Jesus MJ, Herrera F (2001) Genetic feature selection in a fuzzy rule-based classiﬁcation system learning process for high dimensional problems. Inform. Sci. 136:169–191 25. Casillas J, Cordon O, Herrera F, Magdalena L (2003) Interpretability Issues in Fuzzy Modeling, Springer 26. Castellano G, Castiello C, Fanelli AM, Mencar C (2005) Knowledge Discovery by a Neuro-Fuzzy Modeling Framework. Fuzzy Sets and Systems 149:187–207 27. Castellano G, Fanelli AM (1996) Simplifying a neuro-fuzzy model. Neural Proc. Letters 4:75–81 28. Castellano G, Fanelli AM (1997) An approach to structure identiﬁcation of fuzzy models. In: Proc. of the Sixth IEEE International Conference on Fuzzy Systems (FUZZ-IEEE 1997) 531–536 29. Castellano G, Fanelli AM (2000) Fuzzy inference and rule extraction using a neural network. Neural Network World Journal 3:361–371 30. Castellano G, Fanelli AM, Gentile E, Roselli T (2002) A GA-based approach to optimize fuzzy models learned from data. Proc. of Workshop on Approximation and Learning in Evolutionary Computing, part of Genetic and Evolutionary Computation Conference (GECCO 2002), New York, USA, 5–8 31. Castillo L, Gonzalez A, Perez R (2001) Including a simplicity criterion in the selection of the best rule in a fuzzy genetic learning algorithm. Fuzzy Sets and Systems 120(2):309–321 32. Chalmers DJ (1990) The evolution of learning: An experiment in genetic connectionism. In: Proc. 1990 Connectionist Models Summer School, pp. 81–90 33. Cheong F, Lai R (2000) Constraining the optimization of a fuzzy logic controller using an enhanced genetic algorithm. IEEE Trans. Systems Man Cybernet. B 30:31–46 34. Chen J, Xi Y (1998) Nonlinear system modeling by competitive learning and adaptive fuzzy inference system. IEEE Trans. Syst., Man, Cybern. 28:231–238 35. Chien B-C, Lin JY, Hong T-P (2002) Learning discriminant functions with fuzzy attributes for classiﬁcation using genetic programming. Expert Systems Appl. 23(1):31–37

38

G. Castellano et al.

36. Cho KB, Wang BH (1996) Radial basis function based adaptive fuzzy systems and their applications to system identiﬁcation and prediction. Fuzzy Sets Syst. 83:325339 37. Chou L-D, Wu J-LC (1995) Parameter adjustment using neural-network-based genetic algorithms for guaranteed QOS in atm networks. IEICE Trans. Commun. E78-B(4):572–579 38. Chung I-F, Lin CJ, Lin CT (2000) A GA-based fuzzy adaptive learning control network. Fuzzy Sets and Systems 112(1):65–84 39. Cordon O, del Jesus MJ, Herrera F, Lozano M (1999) MOGUL: a methodology to obtain genetic fuzzy rule-based systems under the iterative rule learning approach. Int. J. Intelligent Systems 14(11):1123–1153 40. Cordon O, Herrera F (1997) A three-stage evolutionary process for learning descriptive and approximate fuzzy logic controller knowledge bases from examples. Int. J. Approximate Reasoning 17(4):369–407 41. Cordon O, Herrera F (2001) Hybridizing genetic algorithms with sharing scheme and evolution strategies for designing approximate fuzzy rule-based systems, Fuzzy Sets and Systems 118(2):235–255 42. Cordon O, Herrera F, Hoﬀmann F, Magdalena L (2001) Genetic Fuzzy Systems - Evolutionary Tuning and Learning of Fuzzy Knowledge Bases. World Scientiﬁc, Singapore 43. Czogala E, Leski J (2000) Fuzzy and Neuro-Fuzzy Intelligent Systems. Studies in Fuzziness and Soft Computing (47), Physica-Verlag 44. de Oliveira JV (1999) Towards Neuro-Linguistic Modeling: Constraints for Optimization of Membership Functions. Fuzzy Sets and Systems 106:357–380 45. Delgado MR, Von Zuben F, Gomide F (2004) Coevolutionary genetic fuzzy systems: a hierarchical collaborative approach. Fuzzy Sets and Systems 141:89– 106 46. Dote Y, Ovaska SJ (2001) Industrial applications of soft computing: a review. Proc. IEEE 89(9):1243–1265 47. Eiben AE, Back T, Schoenauer M, Schwefel HP (eds) (1998) Parallel Problem Solving from Nature (PPSN) V. Lecture Notes in Computer Science, vol. 1498 Springer-Verlag, Berlin, Germany 48. Fang J, Xi Y (1997) Neural network design based on evolutionary programming. Artiﬁcial Intell. Eng. 11(2):155–161 49. Farag WA, Quintana VH, Lambert-Torres G (1998) A Genetic-Based NeuroFuzzy Approach for Modeling and Control of Dynamical Systems. IEEE Trans. on Neural Networks 9(5):756–767 50. Feng JC, Teng LC (1998) An Online Self Constructing Neural Fuzzy Inference Network and its Applications. IEEE Trans. on Fuzzy Systems 6(1):12–32 51. Fogel D (1995) Phenotypes, genotypes, and operators in evolutionary computation. In Proc. 1995 IEEE Int. Conf. Evolutionary Computation (ICEC95), pp. 193–198 52. Fogel D (1999) Evolutionary Computation: Towards a New Philosophy of Machine Intelligence (2nd edition). IEEE Press 53. Fogel LJ, Owens AJ, Walsh MJ (1966) Artiﬁcial Intelligence Through Simulated Evolution. John Wiley & Sons, Chichester, UK 54. Fukuda T, Ishigami H, Shibata T, Arai F (1993) Structure Optimization of Fuzzy Neural Network by Genetic Algorithm, Fifth JFSA World Congress, 964–967

2 Evolutionary Neuro-Fuzzy Systems and Applications

39

55. Furuhashi T, Matsushita S, Tsutsui H (1997) Evolutionary Fuzzy Modeling Using Fuzzy Neural Networks and Genetic Algorithm, IEEE International Conference on Evolutionary Computation, 623–627 56. Furuhashi T (2001) Fusion of fuzzy/neuro/evolutionary computing for knowledge acquisition, Proceedings of the IEEE 89(9):1266–1274 57. Geyer-Schulz A (1995) Fuzzy Rule-Based Expert Systems and Genetic Machine Learning. Physica-Verlag, Heidelberg 58. Glorennec PY (1997) Coordination between autonomous robots. Int. J. Approximate Reasoning 17(4):433–446 59. Goldberg DE (1989) Genetic Algorithms in Search, Optimization and Machine Learning. Addison Wesley, Reading, MA 60. Gomez-Skarmeta AF, Jimenez F (1999) Fuzzy modeling with hybrid systems, Fuzzy Sets and Systems 104:199–208 61. Gonzalez A, Perez R (1999) SLAVE: a genetic learning system based on an iterative approach, IEEE Trans. Fuzzy Systems 7(2):176–191 62. Gonzalez A, Perez R (2001) An experimental study about the search mechanism in SLAVE learning algorithm: hill-climbing methods versus genetic algorithm. Inform. Sci. 136(14):159–174 63. Greenwood GW (1997) Training partially recurrent neural networks using evolutionary strategies. IEEE Trans. Speech Audio Processing 5:192–194 64. Guo Z, Uhrig RE (1992) Using genetic algorithms to select inputs for neural networks. In: Proc. Int. Workshop Combinations of Genetic Algorithms and Neural Networks (COGANN-92), pp. 223–234 65. Gurocak HB (1999) A genetic-algorithm-based method for tuning fuzzy logic controllers. Fuzzy Sets and Systems 108(1):39–47 66. Hall LO, Bezdek JC, Boggavarapu S, Bensaid A (1994) Genetic fuzzy clustering. In: Proc. NAFIPS94, pp. 411–415 67. Hancock PJB (1992) Genetic algorithms and permutation problems: A comparison of recombination operators for neural net structure speciﬁcation. In Proc. Int. Workshop Combinations of Genetic Algorithms and Neural Networks (COGANN-92), pp. 108–122 68. Hanebeck UD, Schmidt GK (1996) Genetic optimization of fuzzy networks. Fuzzy Sets and Systems 79 (1):59–68 69. Hayashi Y (1994) Neural expert system using fuzzy teaching input and its application to medical diagnosis. Inform. Sci. Applicat. 1:47–58 70. Hayashi Y, Buckley JJ (1994) Approximations between fuzzy expert systems and neural networks. Int. J. Approx. Reas. 10:63–73 71. Haykin S (1999) Neural Networks - A Comprehensive Foundation (2nd edition). Prentice Hall 72. Herrera F, Lozano M (1996) Adaptation of genetic algorithm parameters based on fuzzy logic controllers. In: [74] 73. Herrera F, Lozano M, Verdegay JL (1995) Tuning fuzzy controllers by genetic algorithms. Int. J. Approx. Reasoning 12:299–315 74. Herrera F, Verdegay JL (eds) (1996) Genetic Algorithms and Soft Computing. Physica-Verlag, Wurzburg 75. Hertz J, Krogh A, Palmer RG (1991) Introduction to the theory of neural computation. Addison Wesley, Reading, MA 76. Hirota K, Pedrycz W (1994) OR/AND neuron in modeling fuzzy set connectives. IEEE Trans. on Fuzzy Systems 2:151–161

40

G. Castellano et al.

77. Hoﬀmann F, Nelles O (2001) Genetic programming for model selection of TSKfuzzy systems. Inform. Sci. 136(14):7–28 78. Hoﬀmann F, Pﬁster G (1997) Evolutionary design of a fuzzy knowledge base for a mobile robot. Int. J. Approximate Reasoning 17(4):447–469 79. Holland J (1975) Adaptation in Natural and Artiﬁcial Systems. University of Michigan Press, Ann Arbor 80. Holland J, Reitman JS (1978) Cognitive systems based on adaptive algorithms. In: [176] 81. Homaifar A, McCormick E (1995) Simultaneous Design of Membership Functions and Rule Sets for Fuzzy Controllers using Genetic Algorithms. IEEE Trans. on Fuzzy Systems 3(2):129–138 82. Hudson DL, Cohen ME, Anderson MF (1991) Use of neural network techniques in a medical expert system. Int. J. Intell. Syst. 6:213–223 83. Ishibuchi H, Murata T, Turksen IB (1997) Single-objective and two-objective genetic algorithms for selecting linguistic rules for pattern classiﬁcation problems. Fuzzy Sets and Systems 89:135–150 84. Ishibuchi H, Nakashima T, Murata T (1999) Performance evaluation of fuzzy classiﬁer systems for multidimensional pattern classiﬁcation problems. IEEE Trans. System Man Cybernet. 29:601–618 85. Ishibuchi H, Nakashima T, Murata T (2001) Three-objective genetics-based machine learning for linguistic rule extraction. Inform. Sci. 136(14):109–133 86. Ishibuchi H, Nozaki K, Yamamoto N, Tanaka H (1995) Selecting Fuzzy If-Then Rules for Classiﬁcation Problems Using Genetic Algorithms. IEEE Trans. on Fuzzy Systems 3(3):260–270 87. Jang JR (1993) ANFIS: Adaptive-network-based fuzzy inference system. IEEE Trans. Syst., Man, Cybern. 23(3):665–685 88. Jang JSR, Sun CT (1993) Functional equivalence between radial basis function networks and fuzzy inference systems. IEEE Trans. Neural Networks 4:156–159 89. Jang JSR, Sun CT, Mizutani E (1997) NeuroFuzzy and Soft Computing. Prentice-Hall, Englewood Cliﬀs, NJ 90. Janson DJ, Frenzel JF (1993) Training product unit neural networks with genetic algorithms. IEEE Expert, 8:26–33 91. Jin Y (2000) Fuzzy modeling of high-dimensional systems: complexity reduction and interpretability improvement. IEEE Trans. Fuzzy Systems 8(2):212– 220 92. Juang C, Lin C (1998) An on-line self-constructing neural fuzzy inference network and its applications. IEEE Trans. Fuzzy Syst. 6:12–32 93. Kang SJ, Woo CH, Hwang HS, Woo KB (2000) Evolutionary Design of Fuzzy Rule Base for Nonlinear System Modeling and Control. IEEE Trans. on Fuzzy Systems 8(1):37–45 94. Kasabov N (1996) Foundations of Neural Networks, Fuzzy Systems and Knowledge Engineering. MIT Press, Cambridge, MA 95. Keller JK, Hunt DJ (1985) Incorporating fuzzy membership functions into the perceptron algorithm. IEEE Trans. Pattern Anal. Mach. Intell. 7:693–699 96. Keller JM, Yager RR, Tahani H (1992) Neural network implementation of fuzzy logic. Fuzzy Sets Syst. 45:1–12 97. Kim HB, Jung SH, Kim TG, Park KH (1996) Fast learning method for backpropagation neural network by evolutionary adaptation of learning rates. Neurocomput. 11(1):101–106

2 Evolutionary Neuro-Fuzzy Systems and Applications

41

98. Kim J, Kasabov N (1999) HyFIS: Adaptive Neuro-Fuzzy Inference Systems and Their Application to Nonlinear Dynamical Systems. Neural Networks 12:1301– 1319 99. Kinnear KE (ed) (1994) Advances in Genetic Programming. MIT Press, Cambridge, MA 100. Kosko B (1991) Neural Networks and Fuzzy Systems. Prentice-Hall, Englewood Cliﬀs, NJ 101. Koza JR (1992) Genetic Programming: On the Programming of Computers by Means of Natural Selection. MIT Press, Cambridge, MA 102. Koza JR (1994) Genetic Programming II: Automatic Discovery of Reusable Programs. MIT Press, Cambridge, MA 103. Koza JR, Bennett FH III, Andre D, Keane MA (1999) Genetic Programming III: Darwinian Invention and Problem Solving. Morgan Kaufmann 104. Koza JR, Keane MA, Streeter MJ, Mydlowec W, Yu J, Lanza G (2003) Genetic Programming IV: Routine Human-Competitive Machine Intelligence. Kluwer Academic Publishers 105. Krishnamraju PV, Buckley JJ, Reilly KD, Hayashi Y (1994) Genetic learning algorithms for fuzzy neural nets. In Proc. Third IEEE Int. Conf. on Fuzzy Systems (FUZZ-IEEE94), Orlando, FL, USA 1969–1974 106. Kupinski MA, Maryellen ML (1997) Feature selection and classiﬁers for the computerized detection of mass lesions in digital mammography. In: Proc. 1997 IEEE Int. Conf. Neural Networks., pp. 2460–2463 107. Lee S-W (1996) Oﬀ-line recognition of totally unconstrained handwritten numerals using multilayer cluster neural network. IEEE Trans. Pattern Anal. Machine Intell. 18:648–652 108. Lee H-M, Chen C-M, Chen J-M, Jou Y-L (2001) An eﬃcient fuzzy classiﬁer with feature selection based on fuzzy entropy. IEEE Trans. Systems Man Cybernet. B 31(3):426–432 109. Leitch DD (1995) A New Genetic Algorithm for the Evolution of Fuzzy Systems. PhD thesis, Department of Engineering, Oxford University 110. Lim MH, Rahardja S, Gwee BH (1996) A GA Paradigm for Learning Fuzzy Rules. Fuzzy Sets and Systems 82:177–186 111. Lin CT, George Lee CS (1996) Neural Fuzzy Systems A NeuroFuzzy Synergism to Intelligent Systems. Prentice-Hall, Englewood Cliﬀs, NJ 112. Linkens DA, Nyongesa HO (1995) Evolutionary learning in fuzzy neural control systems. In Proc. Third European Congress on Fuzzy and Intelligent Technologies (EUFIT95), Aachen, Germany, 990–995 113. Liu Y, Yao X (1998) Toward designing neural network ensembles by evolution. In: [47] 114. Marin FJ, Sandoval F (1993) Genetic synthesis of discrete-time recurrent neural network. In: Proc. Int. Workshop Artiﬁcial Neural Networks (IWANN93), pp. 179–184 115. Michalewicz Z (1992) Genetic Algorithms + Data Structure = Evolution Programs. Springer-Verlag 116. Miller GF, Todd PM, Hegde SU (1989) Designing neural networks using genetic algorithms. In: Proc. 3rd Int. Conf. Genetic Algorithms and Their Applications, pp. 379–384 117. Mitra S (1994) Fuzzy MLP based expert system for medical diagnosis. Fuzzy Sets Syst. 65:285–296

42

G. Castellano et al.

118. Mitra S, Hayashi Y (2000) Neuro-Fuzzy Rule Generation: Survey in Soft Computing Framework. IEEE Trans. Neural Networks 11(3):748–768 119. Mitra S, Pal SK (1995) Fuzzy multilayer perceptron, inferencing and rule generation. IEEE Trans. Neural Networks 6:51–63 120. Mitra S, Pal SK (1996) Fuzzy self organization, inferencing and rule generation. IEEE Trans. Syst., Man, Cybern. 26:608–620 121. Mizutani E, Takagi H, Auslander DM, Jang J-SR (2000) Evolving colour recipes. IEEE Trans. Systems Man Cybernet. 30(4):537–550 122. Nauck D, Klawonn F, Kruse R (1997) Foundations of NeuroFuzzy Systems. Wiley, Chichester, U.K. 123. Nauck D, Kruse R (1992) A Neuro-Fuzzy Controller Learning by Fuzzy Error Propagation. In: Proc. of Conf. of the North Amer. Fuzzy Inf. Proc. Soc. (NAFIPS92), pp. 388–397 124. Nauck D, Kruse R (1997) New Learning Strategies for NEFCLASS. In: Proc. of Seventh Intern. Fuzzy Systems Ass. World Congress (IFSA97), pp. 50–55 125. Nauck D, Kruse R (1999) Neuro-Fuzzy Systems for function approximation. Fuzzy Sets and Systems 101:261–271 126. Nauck D, Nauck U, Kruse R (1996) Generating Classiﬁcation Rules with Neuro-Fuzzy System NEFCLASS. In: Proc. of the Biennal Conf. of the North Amer. Fuzzy Inf. Proc. Soc. (NAFIPS96) 127. Paetz J (1995) Evolutionary Optimization of Weights of a Neuro-Fuzzy Classiﬁer and the Eﬀects on Benchmark Data and Complex Chemical Data. In Proc. of NAFIPS 2005 - Annual Meeting of the North American Fuzzy Information Processing Society, 615–620 128. Pal SK, Ghosh A (1996) Neuro-fuzzy computing for image processing and pattern recognition. Int. J. Syst. Sci. 27:1179–1193 129. Pal SK, Mitra S (1992) Multilayer perceptron, fuzzy sets and classiﬁcation. IEEE Trans. Neural Networks 3:683–697 130. Pal SK, Mitra S (1999) Neuro-fuzzy Pattern Recognition: Methods in Soft Computing. Wiley, New York 131. Parodi A, Bonelli P (1993) A new approach to fuzzy classiﬁer systems. In: Proc. of Fifth International Conference on Genetic Algorithms (ICGA93), Morgan Kaufmann, pp. 223–230 132. Pedrycz W (1993) Fuzzy neural networks and neurocomputations. Fuzzy Sets and Systems 56:1–28 133. Pedrycz W (ed) (1996) Fuzzy Modelling: Paradigms and Practice. Kluwer Academic Press, Norwell, MA 134. Pedrycz W (ed) (1997) Fuzzy Evolutionary Computation. Kluwer Academic Publishers, Dordrecht 135. Pedrycz W (199) Computational Intelligence: An Introduction. CRC, Boca Raton, FL 136. Pedrycz W, Rocha A (1993) Knowledge-based neural networks, IEEE Trans. on Fuzzy Systems 1:254–266 137. Pedrycz W, Reformat M (1996) Genetic Optimization with Fuzzy Coding. In: [74] 138. Pena-Reyes CA, Sipper M (2001) Fuzzy CoCo: a cooperative coevolutionary approach to fuzzy modeling. IEEE Trans. Fuzzy Systems 9(5):727–737 139. Perez CA, Holzmann CA (1997) Improvements on handwritten digit recognition by genetic selection of neural network topology and by augmented training. In: Proc. 1997 IEEE Int. Conf. Systems, Man, and Cybernetics, pp. 1487–1491

2 Evolutionary Neuro-Fuzzy Systems and Applications

43

140. Perneel C, Themlin JM, Renders JM, Acheroy M (1995) Optimization of Fuzzy Expert Systems Using Genetic Algorithms and Neural Networks. IEEE Trans. on Fuzzy Systems, 3(3):300–312 141. Pham DT, Karaboga D (1991) Optimum design of fuzzy logic controllers using genetic algorithms. J. System Eng. 1:114–118 142. Porto VW, Fogel DB, Fogel LJ (1995) Alternative neural network training methods. IEEE Expert 10:16–22 143. Potter MA, De Jong KA (2000) Cooperative coevolution: an architecture for evolving coadapted subcomponents. Evolutionary Comput. 8(1):1–29 144. Pujol JCF, Poli R (1998) Evolving the topology and the weights of neural networks using a dual representation. Appl. Intell. 8(1):73–84 145. Rahmoun A, Berrani S (2001) A Genetic-Based Neuro-Fuzzy Generator: NEFGEN, ACS/IEEE International Conference on Computer Systems and Applications, 18–23 146. Rojas I, Gonzalez J, Pomares H, Rojas FJ, Fernandez FJ, Prieto A (2001) Multidimensional and multideme genetic algorithms for the construction of fuzzy systems. Int. J. Approx. Reasoning 26(3):179–210 147. Ross T (1997) Fuzzy Logic with Engineering Applications. McGraw-Hill 148. Roubos H, Setnes M (2001) Compact and transparent fuzzy models through iterative complexity reduction. IEEE Trans. Fuzzy Systems 9(4):515–524 149. Rumelhart DE, McClelland JL (1986) Parallel Distributing Processing. MIT Press, Cambridge, MA 150. Russo M (1998) FuGeNeSys - A Fuzzy Genetic Neural System for Fuzzy Modeling. IEEE Trans. on Fuzzy Systems 6(3):373–388 151. Russo M, Santagati NA, Lo Pinto E (1998) Medicinal chemistry and fuzzy logic. Inform. Sci. 105(14):299–314 152. Sanchez E, Shibata T, Zadeh L (eds) (1997) Genetic Algorithms and Fuzzy Logic Systems. Soft Computing Perspectives. World Scientiﬁc, Singapore 153. Saravanan N, Fogel DB (1995) Evolving neural control systems. IEEE Expert 10:23–27 154. Sarkar M, Yegnanarayana B (1997) Evolutionary programming-based probabilistic neural networks construction technique. In: Proc. 1997 IEEE Int. Conf. Neural Networks, pp. 456–461 155. Schwefel HP (1994) On the Evolution of Evolutionary Computation. In: [195]: 116–124 156. Schwefel HP (1995) Evolution and Optimum Seeking. John Wiley & Sons, Chichester, UK 157. Seng TL, Khalid MB, Yusof R (1999) Tuning of a Neuro-Fuzzy Controller by Genetic Algorithm. IEEE Transactions on Systems, Man and Cybernetics-Part B 29(2):226-236 158. Setnes M, Babuska R, Kaymak U, van Nauta Lemke HR (1998) Similarity Measures in Fuzzy Rule Base Simpliﬁcation. IEEE Trans. on Systems, Man and Cybernetics - Part B 28(3):376–386 159. Setnes M, Roubos H (2000) GA-fuzzy modeling and classiﬁcation: complexity and performance, IEEE Trans. Fuzzy Systems 8(5):509–522 160. Sheble GB, Maifeld TT (1995) Reﬁned Genetic Algorithm - Economic Dispatch Example. IEEE Trans. Power Syst. 10:117–124 161. Shim M, Seong S, Ko B, So M (1999) Application of evolutionary computation at LG electronics. In Proc. of IEEE Int. Conf. on Fuzzy Systems FUZZ-IEEE99, Seoul, South Korea, 1802–1806

44

G. Castellano et al.

162. Silva N, Macedo H, Rosa A (1998) Evolutionary fuzzy neural networks automatic design of rule based controllers of nonlinear delayed systems. In Proc. of IEEE World Congress on Computational Intelligence - Fuzzy Systems Proceedings 2:1271–1276 163. Smith SF (1980) A learning system based on genetic adaptive algorithms. PhD Thesis, University of Pittsburgh 164. Srinivas M, Patnaik LM (1991) Learning neural network weights using genetic algorithms Improving performance by search-space reduction. In: Proc. 1991 IEEE Int. Joint Conf. Neural Networks (IJCNN91), vol. 3, pp. 2331–2336 165. Subbu R, Anderson A, Bonissone PP (1998) Fuzzy logic controlled genetic algorithms versus tuned genetic algorithms: an agile manufacturing application. In: Proc. IEEE Int. Symp. on Intelligent Control (NIST) 166. Sulzberger SM, Tschicholg-Gurman NN, Vestli SJ (1998) FUN: Optimization of Fuzzy Rule Based Systems Using Neural Networks. In Proc. of IEEE Conf. on Neural Networks, pp. 312–316 167. Sun CT (1994) Rule-Base Structure Identiﬁcation in an Adaptive-NetworkBased Fuzzy Inference System. IEEE Trans. on Fuzzy Systems 2(1): 64–73 168. Takagi H, Suzuki N, Koda T, Kojima Y (1992) Neural networks designed on approximate reasoning architecture and their applications. IEEE Trans. Neural Networks 3:752–760 169. Tano S, Oyama T, Arnauld T (1996) Deep Combination of Fuzzy Inference and Neural Network in Fuzzy Inference. Fuzzy Sets and Systems 82(2):151–160 170. Tschichold-Gurmann N (1995) Generation and Improvement of Fuzzy Classiﬁers with Incremental Learning Using Fuzzy Rulenet. In: Proc. of ACM Sympos. on Appl. Comp., pp. 466–470 171. Van Le T (1995) Evolutionary fuzzy clustering. In: Proc. 2nd IEEE Conf. on Evolutionary Computation (ICEC95), Vol. 2, pp. 753–758 172. Velasco JR (1998) Genetic-based on-line learning for fuzzy process control. Int. J. Intelligent System 13(1011):891–903 173. Venturini G (1993) SIA: a supervised inductive algorithm with genetic search for learning attribute based concepts. In Proc. European Conf. on Machine Learning, Vienna, pp. 280–296 174. Vonk E, Jain LC, Johnson R (1995) Using genetic algorithms with grammar encoding to generate neural networks. In: Proc. IEEE Int. Conf. Neural Networks, pp. 1928–1931 175. Wang D, Keller JM, Carson CA, McAdoo-Edwards KK, Bailey CW (1998) Use of fuzzy-logic-inspired features to improve bacterial recognition through classiﬁer fusion. IEEE Trans. Syst., Man, Cybern. 28:583–591 176. Waterman DA, Hayes-Roth F (eds) (1978) Pattern-Directed Inference Systems. Academic Press, New York 177. Watts MJ, Kasabov K (1998) Genetic Algorithms for the design of fuzzy neural networks. In: Proc. of the Fifth Int. Conf. on Neural Information Processing (ICONIP’R98), pp. 793–796 178. Whitley D, Starkweather T, Bogart C (1990) Genetic algorithms and neural networks: Optimizing connections and connectivity. Parallel Comput. 14(3) pp. 347–361 179. Wolpert DH, Macready WG (1997) No-Free Lunch Theorems for Optimization. IEEE Trans. Evolutionary Computation 1:67–82 180. Wong CC, Feng SM (1995) Switching-Type Fuzzy Controller Design by Genetic Algorithms. Fuzzy Sets and Systems, 74(2):175–185

2 Evolutionary Neuro-Fuzzy Systems and Applications

45

181. Yao X (1999) Evolving Artiﬁcial Neural Networks. Proc. IEEE 87(9):1423–1447 182. Yao X, Liu Y (1997) A new evolutionary system for evolving artiﬁcial neural networks. IEEE Trans. Neural Networks 8:694–713 183. Yao X, Liu Y (1997) EPNet for chaotic time-series prediction. In: [186] 184. Yao X, Liu Y (1998) Making use of population information in evolutionary artiﬁcial neural networks. IEEE Trans. Syst., Man, Cyber. B 28:417–425 185. Yao X, Liu Y (1998) Toward designing artiﬁcial neural networks by evolution. Appl. Math. Computation 91(1):83–90 186. Yao X, Kim J-H, Furuhashi T (eds) (1997) Select. Papers 1st Asia-Paciﬁc Conf. Simulated Evolution and Learning (SEAL96). Lecture Notes in Artiﬁcial Intelligence, vol. 1285, Springer-Verlag, Berlin, Germany 187. Yao X, Liu Y (1997) Fast Evolution Strategies. Control Cybern. 26(3) pp. 467–496 188. Yao X, Shi Y (1995) A preliminary study on designing artiﬁcial neural networks using co-evolution. In: Proc. IEEE Singapore Int. Conf. Intelligent Control and Instrumentation, pp. 149–154 189. Yan W, Zhu Z, Hu R (1997) Hybrid genetic/BP algorithm and its application for radar target classiﬁcation. In: Proc. 1997 IEEE National Aerospace and Electronics Conf. (NAECON), pp. 981–984 190. Yu L, Zhang Y-Q(2005) Evolutionary fuzzy neural networks for hybrid ﬁnancial prediction. IEEE Transactions on Systems, Man and Cybernetics - Part C 35(2):244249 191. Yuan B, Klir GJ, Swan-Stone JF (1995) Evolutionary fuzzy c-means clustering algorithm. In: Proc. 4th IEEE Int. Conf. on Fuzzy Systems (FUZZ-IEEE95), pp. 2221–2226 192. Zadeh LA (1965) Fuzzy Sets. Information and Control 8:338–353 193. Zadeh LA (1994) Fuzzy logic, neural networks, and soft computing, Commun. ACM 37:77–84 194. Zhang Y.-Q., Kandel A. (1998) Compensatory Genetic Fuzzy Neural Networks and Their Applications. Ser. Machine Perception Artiﬁcial Intelligence. Singapore: World Scientiﬁc, vol. 30 195. Zurada J, Marks R, Robinson C (eds) (1994) Computational Intelligence: Imitating Life. IEEE Press

3 Evolution of Fuzzy Controllers and Applications Dilip Kumar Pratihar1∗ and Nirmal Baran Hui2 1

2

Associate Professor [email protected] Research Scholar [email protected] Soft Computing Lab. Department of Mechanical Engineering Indian Institute of Technology, Kharagpur Kharagpur-721 302 India

Summary. The present chapter deals with the issues related to the evolution of optimal fuzzy logic controllers (FLC) by proper tuning of its knowledge base (KB), using diﬀerent tools, such as least-square techniques, genetic algorithms, backpropagation (steepest descent) algorithm, ant-colony optimization, reinforcement learning, Tabu search, Taguchi method and simulated annealing. The selection of a particular tool for the evolution of the FLC, generally depends on the application. Some of the applications have also been included in this chapter.

Keywords: Fuzzy logic controller, Evolution, Least-square technique, Genetic-fuzzy system, Neural-fuzzy system, Ant-colony optimization, Reinforcement learning, Tabu search, Taguchi method, Simulated annealing

3.1 Introduction Real-world problems are generally associated with diﬀerent types of uncertainties. In the past, considerable eﬀort has been made to model these uncertainties. Prior to 1965, it was considered that probability theory working based on Aristotelian two-valued logic was the sole agent available to deal with uncertainties. This particular logic uses the concept of the classical crisp set. That is a set with a ﬁxed boundary. Prof. Zadeh developed the concept ∗

Corresponding author: Associate Professor, Department of Mechanical Engineering, Indian Institute of Technology, Kharagpur-721 302, India; http:// www.facweb.iitkgp.ernet.in/∼dkpra

D.K. Pratihar and N.B. Hui: Evolution of Fuzzy Controllers and Applications, Studies in Computational Intelligence (SCI) 66, 47–69 (2007) c Springer-Verlag Berlin Heidelberg 2007 www.springerlink.com

48

D.K. Pratihar and N.B. Hui

of fuzzy sets, in the year 1965 [1]. Those are the sets having the vague boundaries. He argued that probability theory can handle only one out of several diﬀerent types of possible uncertainties. Thus, there are uncertainties, which cannot be dealt with by using the probability theory. Taking an example, in which Mr. X requests Mr. Y, to bring some red apples for him from the market. There are two uncertainties at least, which relate to the following: (i) the availability of the apples, and (ii) a guarantee that the apple is red. Depending on the season, there is a probability of obtaining the apples, which varies between 0 and 1. But, the colour – red cannot be deﬁned by the classical set. It is not between red (1) and not-red (0). In the fuzzy set, the colour – red can be deﬁned as follows (Fig. 3.1) using the concept of membership of an element to a class. That is the function value (µ): If the colour is perfectly red PR, then it may be said red with a membership value of 1.0; if it is R, then it is considered to be red with a membership value of 0.65; if it is slightly red SR, then it is red with a membership value of 0.39. If it is not red (NR), then also it is red with a membership value of 0.0. In this way, the uncertainty related to the colour of the apples can be handled. Thus, a fuzzy set may be considered to be a more general concept than the classical set. The concept of fuzzy set theory has been used in a number of applications, such as the Fuzzy Logic Controller (FLC), fuzzy clustering, fuzzy mathematical programming, fuzzy graph theory and other examples. Out of all such applications, FLC is the most popular application for the following reasons – (i) ease of understanding and implementations, (ii) ability to handle uncertainty etc. An exact mathematical formulation of the problem is not required for the development of an FLC. This feature makes it a natural choice for solving complex real-world problems. These are either diﬃcult to model mathematically or the mathematical model becomes highly non-linear. It is to be noted that a fuzzy logic controller was ﬁrst developed by Mamdani and Assilian, in the year 1975 [2]. The concept of fuzzy set was published in the

1.0 0.65 0.39

NR

SR

R

µ

0.0 PR

Colour Red Fig. 3.1. A schematic diagram explaining the concept of membership function distribution.

3 Evolution of Fuzzy Controllers and Applications

49

year 1965. Human beings have the natural ability of determining the inputoutput relationships of a process. The behavior of a human being is modeled artiﬁcially, when designing a suitable FLC. The performance of an FLC depends on its knowledge base (KB), which in turn consists of both Data Base (DB) and a Rule Base (RB). The DB consists of data related to membership function distributions of the variables of the process to be controlled. Designing a proper KB of an FLC is a diﬃcult task, which may be implemented in one of the following ways: – – – –

Optimization Optimization Optimization Optimization

of of of of

the the the the

data base only, rule base only, data base and rule base in stages, data base and rule base simultaneously.

The membership function distributions are assumed to be either Linear such as, triangular, trapezoidal or Non-Linear. The Non-Linear can be Gaussian, bell-shaped, sigmoidal in nature. To design and develop a suitable FLC for controlling a process, its variables need to be expressed in the form of some linguistic terms (such as VN: Very Near, VF: Very Far, A: Ahead for example). The relationships between the input (antecedent) and output (consequent) variables are expressed in the form of rules. For example, a rule can be expressed as indicated in Fig. 3.2: IF I1 is N AND I2 is A THEN O is AR, The number of such rules will be present in the rule base. The number of linguistic terms used to represent the variables increases in order to improve the accuracy of the prediction. The computational complexity of the controller will increase with a larger number of rules. For easy implementation in either

VN

N

F

VF LT AL

1

2

3

A

AR

0

45

RT

4

I (m) 1 LT AL

A

AR

RT

– 90

– 45

90

O (degrees)

– 90

– 45

0

45

90

I (degrees) 2

Fig. 3.2. A diagram showing some membership function distributions of input and output variables of the Fuzzy Logic Controller.

50

D.K. Pratihar and N.B. Hui

the software or the hardware, the number of rules present in the Rule Base should be as small as possible. Consequently, some investigators have tried to design and develop a hierarchical FLC, in which the number of rules will be kept to the minimum [3, 4]. It has been observed that the performance of an FLC largely depends on the rule base and optimizing the data base is a ﬁne tuning process [5]. A fuzzy logic controller does not have an internal optimization module. An external optimizer is used to develop an optimal Knowledge Base through a proper tuning and this helps to improve the performance. In this chapter, the focus is on the issues related to design and development of an optimal fuzzy logic controller using diﬀerent optimization tools. Some of the applications of FLC are cited. The remainder of the text is organized as follows. Two major forms of FLC are discussed in Section 2. Various methods of designing optimal FLCs are given in Section 3. A summary of this work is presented in Section 4.

3.2 Two Major Forms of Fuzzy Logic Controller System modeling done by using the fuzzy set concept can be classiﬁed into two groups. That is linguistic fuzzy modeling and precise fuzzy modeling. Linguistic fuzzy modeling, such as Mamdani Approach is characterized by its high interpretability and low accuracy. The aim of precise fuzzy modeling such as Takagi and Sugeno’s Approach, is to obtain high accuracy at the cost of interpretability. Interpretability of a fuzzy modeling is deﬁned as a capability to express the behavior of a system in an understandable form. This is expressed in terms of compactness, completeness, consistency and transparency. The accuracy of a fuzzy model indicates how closely it can represent the system modeled. The working principles of both these approaches are brieﬂy explained below. 3.2.1 Mamdani Approach [2] An FLC consists of four modules namely, a fuzzy rule base, a fuzzy inference engine, fuzziﬁcation and de-fuzziﬁcation. Fig. 3.3 shows a schematic diagram explaining the working of an FLC. (a) The condition known as the antecedent and the action called the consequent variables needed to control a process are identiﬁed and measurements are taken of all the condition variables. (b) The measurements taken in the previous step are converted into appropriate fuzzy sets to express measurement uncertainties. This process is known as fuzziﬁcation. (c) The fuzziﬁed measurements are then used by the inference engine to evaluate the control rules stored in the fuzzy rule base and a fuzziﬁed output is determined.

3 Evolution of Fuzzy Controllers and Applications

Defuzzification module

Fuzzy rule base

51

Actions

Fuzzy inference engine

Process to be controlled

Fuzzification module

Conditions

FLC

Fig. 3.3. The working cycle of an FLC. µA1

µB1

µC1

Rule 1

s1 µA2

f

s2 µB2

µC2

Rule 2

s1*

s1

s2*

f

s2 µC

Uf

f

Fig. 3.4. A schematic diagram showing the working principle of an FLC.

(d) The fuzziﬁed output is then converted into a single crisp value. This conversion is called de-fuzziﬁcation. The de-fuzziﬁed values represent actions which need to be taken by the FLC in controlling the process. The fuzzy reasoning process is illustrated in Figure 3.4. Let us assume for simplicity that only two fuzzy control rules (out of many rules present in the

52

D.K. Pratihar and N.B. Hui

rule base) are being ‘FIRED’ as shown below. This is for a set of inputs – (s1∗ , s2∗ ). RULE 1: IF s1 is A1 and s2 is B1 THEN f is C1 RULE 2: IF s1 is A2 and s2 is B2 THEN f is C2. If s1∗ and s2∗ are the inputs for fuzzy variables s1 and s2. If µA1 and µB1 are the membership function values for A and B, respectively, then the grade of membership of s1∗ in A1 and the grade of membership of s2∗ in B1 are represented by µA1 (s1∗ ) and µB1 (s2∗ ), for rule 1. Similarly, for rule 2, where µA2 (s1∗ ) and µB2 (s2∗ ), are used to represent the membership function values. The ﬁring strengths of the ﬁrst and second rules are calculated as follows: α1 = min (µA1 (s1∗ ), µB1 (s2∗ )) , α2 = min (µA2 (s1∗ ), µB2 (s2∗ )) .

(3.1) (3.2)

The membership function of the combined control action C is given by µC (f ) = max (µ∗C1 (f ), µ∗C2 (f )) .

(3.3)

There are several methods of defuzziﬁcation (shown in Fig. 3.5). These are explained below. 1. Center of Sums Method: According to this method of defuzziﬁcation (refer to Fig. 3.5(a)), the crisp output can be determined by the following.

0.5

0.5 3

µ

µ

0.3

0.3 2 1

0.0

4 5 6

0.0 1 2 3 4 5 6 7 8 9 10 (a)

Output

1 2 3 4 5 6 7 8 9 10 (b)

0.5 µ

0.3

0.0 1 2 3 4 5 6 7 8 9 10 (c)

Output

Fig. 3.5. Diﬀerent methods of defuzziﬁcation.

Output

3 Evolution of Fuzzy Controllers and Applications

53

p Uf =

j=1 A(αj ) × fj p j=1 A(αj )

,

(3.4)

Where Uf is the output of the controller. A(αj ) represents the ﬁring area of the j-th rule. p is the total number of the ﬁred rules. fj represents the centroid of a membership function. 2. Centroid Method: The total area of the membership function distribution used to represent the combined control action is divided into a number of standard sub-areas. Their area and the center of area can be determined easily (refer to Fig. 3.5(b)). The crisp output of the controller can be calculated by using the expression given below. N Uf = i=1 N

Ai fi

i=1

Ai

,

(3.5)

Where N indicates the number of small areas or regions, Ai and fi represent the area and the center of area of i-th small region. 3. Mean of Maxima Method: From the membership function distribution of the combined control action, the range of the output variable is located. This is where the maximum value of the membership function is reached. The mid-value of this range is considered to be the crisp output of the controller (refer to Fig. 3.5(c)). 3.2.2 Takagi and Sugeno’s Approach [6] Here, a rule consists of the fuzzy antecedent and the functional consequent parts. Thus, a rule can be represented as follows: If x1 is Ai1 and x2 is Ai2 ..... and xn is Ain then y i = ai0 + ai1 x1 + . . . + ain xn where a0 , a1 , . . . , an are the coeﬃcients. In this way, nonlinear system is considered as a combination of several linear systems. Control action of i-th rule can be determined for a set of inputs (x1 , x2 , . . . , xn ) as follows. wi = µiA1 (x1 )µiA2 (x2 ) . . . µiAn (xn ),

(3.6)

Where A1 , A2 , . . . , An indicate the membership function distributions of the linguistic terms used to represent the input variables. The membership function value is given by µ. Thus, the combined control action can be determined as k wi y i y = i=1 , (3.7) k i i=1 w where k is the total number of rules.

54

D.K. Pratihar and N.B. Hui

3.3 Methods of Designing Optimal Fuzzy Logic Controllers In order to establish the input-output relationships of a process, a designer tries to design the KB of an FLC manually, based on a knowledge of the process. In most of the cases, it is diﬃcult to gather prior information of a process. The manually-designed KB of the FLC may not be optimal. As an FLC does not have a in-built optimizer, an optimization tool is used, while tuning a KB. Several methods have been developed and some of these are discussed below. 3.3.1 Least-square Method Attempts were made to determine an appropriate shape of the membership function distributions by using least-square methods. In this connection, see Pham and Valliappan [7], Bustince et al. [8]. The membership function distribution of a fuzzy set was assumed to follow a power function such as µA (xi ) = axbi . Here x indicates a variable represented by a fuzzy set A, i = 1, 2, . . . , n, n is the number of training cases, µA is the membership function value of the fuzzy set A lying between 0 and 1, a (greater than zero) and b are the constants to be determined by the least-square method. Two equations were solved for this [8]: n n

lnxi b = lnµA (xi ) (3.8) nlna + n

i=1

lnxi

lna +

n

i=1

i=1

ln2 xi

i=1

b=

n

lnxi lnµA (xi )

(3.9)

i=1

where axbi ≤ 1. 3.3.2 Genetic-Fuzzy System Genetic algorithm (GA) [9] is a population-based search and optimization technique based on the principle of natural selection and mechanics of natural genetics, was used by several researchers, for a genetic-fuzzy system. The performance of a Fuzzy Logic Controller (FLC) is dependent on its KB. Fig. 3.6 shows the schematic diagram of the genetic-fuzzy system. Here, a GA is used to determine optimal KB of the FLC. Thus, the GA improves the performance of the FLC. During optimization of the FLC, the feedback which is a deviation in prediction is calculated. This is based on a set of training cases and it is utilized as the ﬁtness of the GA. A GA is computationally expensive and the tuning is done oﬀ-line. Once optimized, the FLC will be able to predict the outputs for a set of inputs, within a reasonable accuracy

3 Evolution of Fuzzy Controllers and Applications

55

GA–based tuning Off–line Knowledge Base On–line

Inputs

Fuzzy Logic Controller

Outputs

Fig. 3.6. A schematic diagram showing a genetic-fuzzy system

limit. This concept has been used to solve a number of physical problems. See Karr [10], Thrift [11], Pham and Karaboga [14]. A detailed review on this scheme is done by Cordon et al. [12]. There are three basic approaches of this scheme, the Pittsburgh [13, 14], Michigan [15] and iterative rule learning [16, 17] approaches. In Pittsburgh approach, the entire rule base of the FLC is represented by a GA-string. Thus, the GA-population indicates the population of candidate rule sets. The genetic operators are used to modify the rule sets and obtain the optimal rule base. In the Michigan approach, members of the population are individual rules. Thus, a rule set is represented by the entire population. The main drawback of these two approaches lies in the fact that for the large number of fuzzy rules, the GA requires a huge amount of computer memory. To overcome the problem, using an iterative rule learning approach, chromosomes code individual rule, a new rule is added to the rule set, in an iterative fashion, for every run of GA. It requires a proper encoding scheme for extracting the rules from a chromosome. In this approach, the evolved RB of the FLC may contain some redundant rules, due to the iterative nature of the GA. A considerable amount of work has been carried out in this ﬁeld of research. Some of these attempts are mentioned below. Furuhashi et al. [18] developed a variable length decoding method, known as the Nagoya Approach. Using this approach, as the lengths of the chromosomes are not ﬁxed, it is diﬃcult to implement the necessary crossover operation in GA. Again the simultaneous design of the data base and rule base requires a proper optimization procedure. This can tackle both continuous as well as integer variables. Wang and Yen [19] proposed a method, in which a GA was used to extract the rule base, and the data base of an FLC was optimized using a Kalman ﬁltering technique. Farag et al. [20] developed a new multi-resolutional dynamic GA for this purpose. In this, the initial parameters of the data base of an FLC were determined by

56

D.K. Pratihar and N.B. Hui

using Kohonen’s self-organizing feature map algorithm and optimization was done by using a GA. Fuzzy rule generation and tuning using a GA was also tried by Ishibuchi et al. [21]. Recently, Abdessemed et al. [22] proposed a GAbased procedure for designing an FLC, to control the end eﬀector’s motion of a planar manipulator. Yupu et al. [23] used a GA to search for appropriate fuzzy rules. The membership function distributions were optimized by using a neural network. The FLC is becoming more popular nowadays, developing a suitable knowledge base for it, is not easy. The designer requires much time, to initially design the knowledge base (KB). It is further improved by using GA-based tuning. Thus, the designer must have a knowledge of the process to be controlled by the FLC. To overcome this requirement, a few investigators [24, 25] tried to automatically design the FLC by using a GA. Using search, the GA will develop the optimized data base and rule base for the FLC. A GA is basically a ﬁtness function-driven search method, therefore, it is blind for any other aspect that is not explicitly considered on ﬁtness function. Hence, a GA might evolve some redundant rules, that have limited inﬂuence on the process to be controlled. Redundant rules are to be removed to make the rule base compact. This makes the implementation of the controller easier, particularly when it is done by hard-ware. Thus, there is a need to determine the contribution of each rule. In this context, the work of Nawa et al. [26], Ishibuchi and Nakashima [27], Ghosh and Nath [28], Hui and Pratihar [35] are important. Nawa et al. [26] measured the quality of a rule by determining its accumulated truth value. The accumulated truth value was considered to be the sum of probability of occurrences of a rule in the training data. A rule is said to be good, if its accumulated truth value is high. Ishibuchi and Nakashima [27] made an attempt to assign an importance factor to each rule. They calculated the importance factor of a rule, by considering the way it interacts with the neighbors. An evolutionary technique was utilized to ﬁnd the interaction eﬀect. Ghosh and Nath [28] investigated the eﬀectiveness of a rule by measuring three parameters, namely support count, comprehensibility and interestingness. Support count of an item set is deﬁned by the number of records in the data base that contains all the items of that set. Comprehensibility is used to justify the understandability of a rule. A rule is said to be more comprehensive, if the number of attributes associated with the antecedent part of the rule is less and interestingness is represented by the probability of generating a rule during the learning process. It was a theoretical approach of ﬁnding interesting rules in the rule base and is unable to predict the importance of a rule for a ﬁxed number of attributes in both antecedent as well as in the consequent parts. The above methods considered the probability of occurrence of a rule only, for the determination of a good rule base. No attention was paid to calculate the contribution eﬀect of a rule with respect to a speciﬁc objective. Hui and Pratihar [35] proposed a method of determining importance factor for each rule contained in the RB of an FLC, to check the redundancy, if any. The importance factor of a rule is calculated

3 Evolution of Fuzzy Controllers and Applications

57

by considering its probability of occurrence and worth (goodness). A rule is said to be redundant and thus may be eliminated, if its importance factor comes out to be smaller than a pre-speciﬁed value and the removal of which does not lead to any non-ﬁring situation. The genetic-fuzzy system has been developed by the authors also, following the two diﬀerent approaches discussed below. Approach 1: GA-based tuning of the manually-constructed KB of the FLC. The KB of the FLC is designed manually and is based on the designer’s experience of the problem to be solved. But, it may not be optimal in any sense. GA-based tuning is adopted, to further optimize the KB, to improve the performance. As a GA is found to be computationally expensive, the GA-based tuning is carried out, oﬀ-line. During optimization, the GA-string will carry information for both the data base as well as the rule base. The GA-search will ﬁnd the optimal KB of the FLC. Once optimized, the FLC is able to determine its outputs in the optimal sense. Approach 2: Automatic design of KB using a GA. In Approach 1, much time is spent on manual design of the KB of an FLC. It might be diﬃcult beforehand to foresee the characteristics of the process to be controlled. Thus, designing a proper KB might be a diﬃcult task. To overcome this, a method for automatic design of the KB is developed by using a GA. Here the task of designing a suitable KB is given to the GA. The GA through its exhaustive search will try to determine the optimal KB of the FLC. The above concept has been used by the authors, to solve a number of physical problems. One of them is explained below. Optimal Path and Gait Planning of a Six-legged Robot A six-legged robot will have to plan its time-optimal, collision-free path as well as the optimal gait, setting simultaneously the minimum number of ground-legs having the maximum average kinematic margin. This is while moving on a ﬂat terrain with occasional hurdles, such as ditches and some moving obstacles. Its stability margin should always be positive to ensure static stability. This is a complicated task because the path planning and gait planning must be done simultaneously [29]. Fig. 3.7 shows the optimal path and gait for a six-legged robot. It has planned its optimal path and gait, after starting from an initial position S to reach the ﬁnal position G. It faces three moving obstacles and a ditch on the way to-wards its goal. The total movement of the robot has been achieved through a number of segments called motion segments. The robot plans its optimal path and gait on-line, for each motion segment. The robot shown in Fig. 3.7, is found to reach its goal in the time-optimal sense at 79-th motion segment, after avoiding collision with the moving obstacles and generating its optimal gaits.

58

D.K. Pratihar and N.B. Hui G G Ditch

Obs 2, 0.12 m/s

Ditch

Obs 2, 0.12 m/s

Obs 3, 0.15 m/s

Obs 3, 0.15 m/s

Obs 1, 0.1 m/s

Obs 1, 0.1 m/s

S

S Note: Positions at 10–th motion segment

Note: Positions at 20–th motion segment G

G Ditch

Obs 2, 0.12 m/s

Ditch Obs 2, 0.12 m/s

Obs 3, 0.15 m/s

Obs 1, 0.1 m/s

Obs 3, 0.15 m/s

Obs 1, 0.1 m/s

S

S

Note: Positions at 35–th motion segment

Note: Positions at 50–th motion segment

G Ditch G Ditch Obs 2, 0.12 m/s Obs 3, 0.15 m/s Obs 2, 0.12 m/s Obs 3, 0.15 m/s

Obs 1, 0.1 m/s Obs 1, 0.1 m/s

S

S Note: Positions at 65–th motion segment

Note: Positions at 79–th motion segment

Fig. 3.7. Optimal path and gaits of a six-legged robot obtained using the geneticfuzzy system [29]

3 Evolution of Fuzzy Controllers and Applications

59

3.3.3 Neural-Fuzzy System The purpose of developing a neural-fuzzy system is to improve the performance of an FLC by using neural network-based learning. It had been utilized by a number of researchers to solve a variety of problems. Some of these are mentioned below. Marichal et al. [30] proposed a neuro-fuzzy approach to generate the motion of a car-like robot navigating among static obstacles. In their approach, a least mean squared algorithm was used for the learning purposes and Kohonen’s self organizing feature map algorithm was considered to obtain the initial number of fuzzy rules and fuzzy membership function centers. They did not optimize the traveling time nor the approach was tested in a dynamic environment. Song and Sheen [31] suggested a pattern recognition approach based on a fuzzy-neuro network for the reactive navigation of a car-like robot. Li et al. [32] developed a neuro-fuzzy architecture for behavior-based control of a car-like robot, that navigates among static obstacles. The present chapter includes two schemes of neural-fuzzy system developed by the authors. These are discussed below [33]. Scheme 1: Neural-fuzzy system based on Mamdani Approach. In the developed neural-fuzzy system, a fuzzy logic controller using Mamdani Approach is expressed by utilizing the structure of a Neural Network (NN) and a back-propagation algorithm is utilized to optimize the KB of the FLC. The back-propagation algorithm is a steepest descent algorithm. Fig. 3.8 shows the schematic diagram of the ﬁve layer neural-fuzzy system– Layer 1 is the input layer, fuzziﬁcation is done in Layer 2, Layer 3 indicates the AND operation. The OR operation is carried out in Layer 4, and Layer 5 is the output layer. The training cases are passed through the network and the total error is calculated. The average error is propagated in the backward direction, to determine the updated weights. The network will try to ﬁnd an optimal set of weights, corresponding to which the error is minimum. Layer 1

Layer 2

[V]

1

Layer 4

FR

1 2

2

Layer 5

[W]

1

NR I1

Layer 3

1

LT

3 4

I2 2

VF

3

5

LT

1

6

AH RT

2

AH

O 1

7 2 3

8

3

RT

9

Fig. 3.8. A schematic diagram of the neural network-structured FLC

60

D.K. Pratihar and N.B. Hui

Three diﬀerent approaches to Scheme 1 are developed. These are discussed in brief below. Approach 1: NN-tuned FLC. The initial weights of the neural network representing the FLC are generated, at random. A batch mode of training is adopted. Training cases are passed through the NN (i.e., forward propagation) and average error is determined. As this error depends on the weights, it can be minimized by updating the weight values. A back-propagation algorithm is used to minimize the error. Approach 2: Genetic-Neural-Fuzzy system. In Approach 1, the error is minimized using a steepest descent method. This may have the local minima problems. To overcome this problem, the backpropagation algorithm is replaced by a GA. As GA is a populationbased search and optimization method, the chance of its solutions for getting trapped into the local minima is less. Thus, Approach 2 may be expected to perform better than Approach 1. Approach 3: Automatic design of neural-fuzzy system. To increase the search space of the GA, a method for automatic design of neural-fuzzy system is proposed. In this approach, the outputs of different rules are evolved solely by the GA itself. The GA through its exhaustive search, determines a good rule base for the FLC. There might be some redundant rules present in the GA-designed rule base. It may happen due to the iterative nature of the GA. To identify the redundant rules, a method is proposed, in which importance of a rule is decided by considering its frequency of occurrence and its worth with respect to the objective function of the optimization problem. Based on the value of this importance factor, a decision is taken whether a particular rule will be declared as redundant. Scheme 2: Neural-fuzzy system based on Takagi and Sugeno Approach. A neural-fuzzy system has been developed based on the Takagi and Sugeno Approach. This is known as the ANFIS (i.e., Adaptive Neuro-Fuzzy Inference Systems) [34]. An ANFIS is a multi-layered feed forward network, in which each layer performs a particular task. The layers are characterized by the fuzzy operations they perform. Fig. 3.9 shows the schematic diagram of the ANFIS structure, which consists of six layers – Layer 1 (input layer), Layer 2 (condition layer), Layer 3 (rule base layer), Layer 4 (normalization layer), Layer 5 (consequence layer) and Layer 6 (output layer). Let us assume that there are two inputs – I1 and I2 and one output O of the network. The ﬁrst two layers perform similar tasks to those done by Layers 1 and 2 of the neuro-fuzzy system developed in Scheme 1. The functions of other layers are explained below. Layer 3: This layer deﬁnes the rules of the fuzzy inference system. As three linguistic terms are used to represent each of the two inputs, there is a maximum of 3 × 3 = 9 rules present in the rule base. Each neuron in this layer represents a fuzzy rule and is termed as a rule node. The output of each neuron lying in this layer is the multiplication of their

3 Evolution of Fuzzy Controllers and Applications Layer 1

Layer 2 [V]

I1

I2

1

2

NR FR

1 2

VF

3

LT

1

AH RT

2 3

Layer 3

Layer 4

Layer 5

1

1

1

2

2

2

3

3

3

4

4

4

5

5

5

6

6

6

7

7

7

8

8

8

9

9

9

61

Layer 6

1

O

Fig. 3.9. A schematic diagram of the ANFIS architecture

respective two membership values. It is to be noted that each node output represents the ﬁring strength of a rule. Layer 4: This layer has the same number of nodes as the previous layer. It calculates the normalized ﬁring strength of each node. Layer 5: The output of a particular neuron (say, the q-th) lying on this layer is determined by O5q = (aq I1 + bq I2 + cq )

(3.10)

where (aq , bq , cq ) represents one set of coeﬃcients associated with the q-th node. Layer 6: The output of the node lying on Layer 6, can be determined by summing up all incoming signals. O61 =

R

O5q ,

(3.11)

q=1

where R indicates the total number of rules. A maximum of four rules (out of nine) will be ﬁred, for one set of input variables. The performance of an ANFIS depends on the selection of consequence parameters and premise parameters. That is the half base-widths of the input membership function distributions. For the selection of optimal parameters, a GA might be used together with the ANFIS. The developed neural-fuzzy systems have been used to plan collision-free, time-optimal paths of a car-like robot. This is explained below. 3.3.3.1 Collision-free, Time-optimal Path Planning for a Car-like Robot [33, 35] A car-like mobile robot needs to ﬁnd its time-optimal and collision-free path while navigating among some moving obstacles, and satisfy its kinematic (nonholonomic) constraints and dynamic constraints (such as sliding constraint,

62

D.K. Pratihar and N.B. Hui

motor torque constraint, curvature constraint). A detailed discussion on these constraints is beyond the scope of this chapter. Interested readers may refer to [33], for the same. The total path of the robot is divided into a number of distance steps having varying lengths, each of which is traveled during a time step. To calculate total traveling time of a robot to reach its destination, the time steps are summed and the time required to align its main axis towards the goal is added. There can be a saving in traveling time, particularly if the robot does not change its direction in two successive distance steps. It is subtracted from the total traveling time. The aim is to minimize the traveling time after ensuring a collision-free movement of the robot. A high positive value penalty is added to the total traveling time, if the robot collides with any one of the obstacles. Fig. 3.10 shows the near-optimal, collision-free paths of a robot in the presence of 16 moving obstacles. This is as obtained by using the three approaches of Scheme 1, Scheme 2 (explained above) and a

Fig. 3.10. Navigation of a robot among 16 moving obstacles

3 Evolution of Fuzzy Controllers and Applications

63

traditional motion planning scheme (potential ﬁeld method) [36]. The initial position, size, velocity and direction of movement of the obstacles are created at random. The planning robot starts from the point S and reaches the goal G, by avoiding collisions with the obstacles. Soft computing-based approaches have proved their supremacy over the potential ﬁeld method. It could be due to the reason that there is a chance that the solutions of the potential ﬁeld method will get trapped at the local minima. On the other hand, the chance of the solutions of GA-tuned fuzzy logic controller for getting trapped into the local minima is less and it could be due to an exhaustive search carried out by the GA. Moreover, the GA is able to inject adaptability to the FLC, which has been observed from the performances of Approaches 2 and 3 of Scheme 1. Approach 3 of Scheme 1 is found to be the best of all approaches. It could be due to the fact that using this approach, a good KB of the FLC is evolved by the GA, after carrying the search in a wider space. 3.3.4 Optimization of FLC Using Ant Colony Optimization [37] In Ant Colony Optimization (ACO) algorithm, an optimization problem is represented in the form of a graph – G = (C, L). Here, C is the set of components of the problem and L indicates the possible connection or transition among the elements C. The solutions are expressed in terms of feasible paths on the graph G, after satisfying a set of constraints. Thus, the Fuzzy Rule Learning Problem (FRLP) using the ACO, is formulated as a combinatorial optimization problem. Its operational mode is composed of two stages: in the ﬁrst stage, the number and antecedents of the linguistic rules are deﬁned, and a set of consequent candidates is assigned to each rule. In the second stage, a combinatorial search is carried out to ﬁnd the best consequent of each rule, according to a global error measure over the training set. The ﬁtness of a solution consists of two parts, namely the functional ﬁtness and the objective ﬁtness. The functional ﬁtness deals with the functionality of the solutions. That is how good is the solution. The objective ﬁtness is the measure of the quality of the solution, in terms of optimization objectives, such as area, delay, gate count, power consumption, and others. To apply ACO algorithm to a speciﬁc problem, the following steps need to be followed: – Represent the problem in the form of a graph or a similar easily covered structure, – Deﬁne the way of assigning a heuristic preference to each choice that needs to be taken in each step in order to generate the solution, – Establish an appropriate way of initializing the pheromone, – Deﬁne the ﬁtness function to be optimized, – Select an ACO algorithm to determine the optimal solutions. The Fuzzy Rule Learning Problem (FRLP) aims to obtain the rules combining the labels of the antecedents and to assign a speciﬁc consequent to

64

D.K. Pratihar and N.B. Hui

each antecedent combination. This problem is interpreted as a way of assigning consequents to the rules with respect to an optimality criterion. An ant iteratively goes over each rule and chooses a consequent with a probability that depends on the pheromone trail τij and the heuristic information ηij . 3.3.5 Tuning of FLC Using Reinforcement Learning Fuzzy rules for control can be eﬀectively tuned by means of reinforcement learning. In this approach, the rules with their associated antecedent and consequent fuzzy sets are represented with the help of a fuzzy-neural network. For this an action selection network (ASN) is used. This network provides continuous action value and records the state of the environment and also determines the next action required. Thereafter, the actions are evaluated by means of a critic element (CE), which is a two-layer feed forward action evaluation network (AEN). It predicts the reinforcements associated with diﬀerent input states and whether or not a failure has occurred. If a failure occurs, it identiﬁes the steps leading to the failure and modiﬁes the fuzzy sets associated with the rules. A gradient descent technique in conjunction with an average reward is used to train both the action selection network (ASN) and the action evaluation network (AEN) over a set of trials. During training, a reward is provided until a failure occurs and then a high value penalty is given. This approach had been used by Berenji and Khedkar [38], to solve the problem of a cart-pole balancing system. 3.3.6 Optimization of FLC Using Tabu Search Denna et al. [39] presented an approach for automatic deﬁnition of the fuzzy rules based on the Tabu Search (TS) algorithm. To determine the most appropriate rule base for solving the problem, they employed the reactive form of TS algorithm. To apply the Reactive Tabu Search (RTS) algorithm in determining the rules of a fuzzy controller, the consequent of each rule is expressed with a binary string. The learning procedure is shown in Figure 3.11. The learning begins with an initial rule base, chosen randomly at each iteration. Initial states can also be selected by following a uniform distribution over the entire state space. In such conditions, regions of interest are assigned a higher probability during the learning procedure. Performance of the rule base is then evaluated by using an error function E(•), over a set of typical control rules. It is important to mention that during the evaluation of E(•), some rules, with a smaller contribution to the system are not used. This procedure continues until a termination criterion is reached. The termination condition for each execution may be based on the following parameters: – The number of iterations carried out, – The Current State of the error function, – The properties of the solution found.

3 Evolution of Fuzzy Controllers and Applications

rules represented as a binary string 0010....110..01...1

65

Fuzzy Controller Rules

typical control cases Reactive Tabu Search

Evaluation

Model

E( )

Fig. 3.11. A schematic diagram showing the learning of fuzzy rules using Tabu search [39]

Bagis [40] described a method for the determination of optimum fuzzy membership function distribution used for controlling a reservoir system of dams during ﬂoods. 3.3.7 Design of a Fuzzy Controller using the Taguchi Method The Taguchi Method determines the parameter settings, which maximize the signal to noise (S/N ) ratio in each problem by systematically performing the designed experiment. The designed experiment is composed of an inner array and an outer array. The inner array is a designed experiment using the control factors and the outer array consists of the noise factors. To design an FLC using the Taguchi method, control factors are considered as the membership parameters and diﬀerent system conditions are assumed to be the noise factors. If the inner array is made up of m rows and the outer array contains n rows, then each of the m rows can obtain n performance characteristics. These n data are used to calculate the S/N ratio, for each row of the inner array. The optimal parameter settings are determined by analyzing the S/N ratio data. To check the adequacy of the model, Analysis of Mean (ANOM) and Analysis of Variance (ANOVA) are carried out. Later, a veriﬁcation experiment is conducted to test the performance of the model. Kim and Rhee [41] utilized the Taguchi method, to design a suitable fuzzy logic controller, in which the following steps were used: – – – –

Identify the performance characteristic to be observed, Identify important noise factors and their ranges, Identify the control factors and their levels, Construct the inner array and the outer array,

66

D.K. Pratihar and N.B. Hui

– Conduct the designed experiment, – Analyze the data and determine optimal levels for the control factors, – Conduct the veriﬁcation experiment. 3.3.8 Fuzzy Logic Controller Tuned by Simulated Annealing Simulated Annealing (SA) is one of the most popular non-traditional methods of optimization, in which the cooling process of a molten metal has been artiﬁcially modeled. Alfaro and Garcia [42] described a method for development of a fuzzy logic controller applied to path planning and navigation of mobile robots, by using a simulated annealing. Most of the researchers tried to optimize the membership function distributions of the FLC by utilizing the SA. In this approach,the cost function was deﬁned as follows: F =

N 1 2 (yk − yˆk ) , N

(3.12)

k

where k = 1, 2, . . . , N . N is the number of learning samples, (xk , yk ) is the k th learning sample and yˆk is the output of the fuzzy system corresponding to the input vector xk . The optimization algorithm tunes the parameters (spread and shape) of membership function distributions. This is in order to minimize the cost function. Consider the membership functions of the input variables to be Gaussian in nature, as shown below. Gaussian(x; σ; c) = e−(

x−c 2 σ )

(3.13)

where c and σ indicate the Center and Width, respectively, of the membership function distribution. In SA, the following steps are to be considered in order to optimize the Gaussian membership function distribution: 1. Set an Initial Temperature T to a high value and generate initial parameters cij and σji , randomly and compute the cost function (Fold ). 2. Generate a set of new parameters cij and σji and compute the new cost function (Fnew ). Obtain the change in the cost function δ = Fnew − Fold . If δ < 0, memorize the new set of membership functions and proceed until the termination criterion is reached. Otherwise, go to Step 3. 3. If δ > 0 and probability of accepting the new set of membership functions P (δ) = exp(−δ/T ) ≤ random[0, 1]), the center and width values are not changed. Now, go to Step 2 by reducing the temperature T to the half of its previous value. 4. Repeat Steps 2 and 3 until an acceptable solution has been found or until a speciﬁed number of iterations has been reached.

3 Evolution of Fuzzy Controllers and Applications

67

3.4 Summary Fuzzy logic controllers have proved their worth and are popular nowadays to solve real-world complex problems. As the performance of an FLC depends on its KB, several attempts had been made to design a suitable KB. Several methods had been tried by various investigators, to solve the problem. There is a chance of further improvement and much further work is necessary. Both the Linguistic as well as Precise Fuzzy Modeling have been used separately, to solve a variety of problems and some satisfactory results have been obtained. Linguistic fuzzy modeling ensures better interpretability, but precise fuzzy modeling aims to achieve higher accuracy. It is obvious that as interpretability of the fuzzy model increases, its accuracy will decrease and vice-versa. Thus, depending on the physical problem, a particular type of fuzzy modeling is chosen. It is challenging to obtain a proper balance between interpretability and accuracy of a fuzzy model. These two properties are inversely related and it is important to investigate as to whether a pareto-optimal front exists.

References 1. Zadeh, L.A.: Fuzzy sets. Information and Control 8 (1965) 338–353 2. Mamdani, E.H., Assilian, S.: An experiment in linguistic synthesis with a fuzzy logic controller. International Journal on Man-Machine Studies 7 (1975) 1–13 3. Wang, L.X.: Analysis and design of hierarchical fuzzy systems. IEEE Trans. on Fuzzy Systems 7 5 (1999) 617–624 4. Lee, M.L., Chung, H.Y., Yu, F.M.: Modeling of hierarchical fuzzy systems. Fuzzy Sets and Systems 138 (2003) 343–361 5. Pratihar, D.K., Deb, K., Ghosh, A.: A genetic-fuzzy approach for mobile robot navigation among moving obstacles. International Journal of Approximate Reasoning 20 (1999) 145–172 6. Takagi, T., Sugeno, M.: Fuzzy identiﬁcation of systems and its application to modeling and control. IEEE Trans. on Systems, Man and Cybernetics SMC-15 (1985) 116–132 7. Pham, T.D., Valliappan, S.: A least square model for fuzzy rules of inference. Fuzzy Sets and Systems 64 (1994) 207–212 8. Bustince, H., Calderon, M., Mohedano, V.: Some considerations about a least square model for fuzzy rules of inference. Fuzzy Sets and Systems 97 (1998) 315–336 9. Goldberg, D.E.: Genetic algorithms in search, optimization, machine learning. Addison-Wesley, Reading, Mass, USA (1989) 10. Karr, C.: Genetic algorithms for fuzzy controllers. AI Expert (1991) 38–43 11. Thrift, P.: Fuzzy logic synthesis with genetic algorithms, Proc. of Fourth International Conference on Genetic Algorithms. (1991) 509–513 12. Cordon, O.: Gomide, F., Herrera, F., Hoﬀmann, F., Magdalena, L.: Ten years of genetic-fuzzy systems: current framework and new trends. Fuzzy Sets and Systems 141 (2004) 5–31

68

D.K. Pratihar and N.B. Hui

13. Hoﬀman F., Pﬁster, G.: Evolutionary design of a fuzzy knowledge base for a mobile robot. Intl. Jl. of Approximate Reasoning 17 4 (1997) 447–469 14. Pham, D.T., Karaboga, D.: Optimum design of fuzzy logic controllers using genetic algorithms. Journal of Syst. Engg. 1 (1991) 114–118 15. Ishibuchi, H., Nakashima, T., Murata, T.: Performance evaluation of fuzzy classiﬁer systems for multidimensional pattern classiﬁcation problems. IEEE Trans. on Systems Man and Cybernetics 29 (1999) 601–618 16. Cordon, O., DeeJesus, M.J., Herrera, F., Lozano, M.: MOGUL: a methodology to obtain genetic fuzzy rule-based systems under the iterative rule learning approach. Intl. Jl. of Intelligent Systems 14 11 (1999) 1123–1153 17. Gonzalez, A., Perez, R.: SLAVE: a genetic learning system based on an iterative approach. IEEE Trans. on Fuzzy Systems 7 2 (2001) 176–191 18. Furuhashi, T., Miyata, Y., Nakaoka, K., Uchikawa, Y.: A new approach to genetic based machine learning and an eﬃcient ﬁnding of fuzzy rules-proposal of Nagoya approach. Lecture notes on Artiﬁcial Intelligence 101 (1995) 178–189 19. Wang, L., Yen, J.: Extracting fuzzy rules for system modeling using a hybrid of genetic algorithms and Kalman ﬁltering. Fuzzy Sets and Systems 101 (1999) 353–362 20. Farag, W.A., Quintana, V.H., Lambert-Torres, G.: A genetic-based neuro-fuzzy approach for modeling and control of dynamical systems. IEEE Trans. on Neural Networks 9 (1998) 576–767 21. Ishibuchi, H., Nil, M., Murata, T.: Linguistic rule extraction from neural networks and genetic algorithm-based rule selection, Proc. of IEEE Intl. Conf. on Neural Networks. Houston, TX (1997) 2390–2395 22. Abdessemed, F., Benmahammed, K., Monacelli, E.: A fuzzy-based reactive controller for a non-holonomic mobile robot. Robotics and Autonomous Systems 47 (2004) 1–22 23. Yupu, Y., Xiaoming, X., Wengyuan, Z.: Real-time stable self learning FNN controller using genetic algorithm. Fuzzy Sets and Systems 100 (1998) 173–178 24. Angelov, P.P., Buswell, R.A.: Automatic generation of fuzzy rule-based models from data by genetic algorithms. Information Science 50 (2003) 17–31 25. Nandi, A.K.: Pratihar, D.K.: Automatic design of fuzzy logic controller using a genetic algorithm-to predict power requirement and surface ﬁnish in grinding. Journal of Materials Processing Technology 148 (2004) 288–300 26. Nawa, N.E., Hashiyama, T., Furuhashi, T., Uchikawa, Y.: A Study on fuzzy rules discovery using pseudo-bacterial genetic algorithm with adaptive operator, Proc. IEEE Int. Conf. Evolutionary Computation. Indianapolis, USA, (1997) 13–16 27. Ishibuchi H., Nakashima, T.: Eﬀect of rule weights in fuzzy rule-based classiﬁcation systems. IEEE Trans. on Fuzzy Systems 9 4 (2001) 506–515 28. Ghosh A., Nath, B.: Multi-objective rule mining using genetic algorithms. Information Sciences 163 (2004) 123–133 29. Pratihar, D.K., Deb, K., Ghosh, A.: Optimal path and gait generations simultaneously of a six-legged robot using a GA-Fuzzy approach. Robotics and Autonomous Systems 41 1 (2002) 1–20 30. Marichal, G.N., Acosta, L., Moreno, L., Mendez, J.A., Rodrigo, J.J., Sigut, M.: Obstacle avoidance for a mobile robot: a neuro-fuzzy approach. Fuzzy Sets and Systems 124 (2001) 171–170 31. Song, K.T., Sheen, L.H.: Heuristic fuzzy-neuro network and its application to reactive navigation of a mobile robot. Fuzzy Sets and Systems 110 (2000) 331–340

3 Evolution of Fuzzy Controllers and Applications

69

32. Li, W., Ma, C., Wahl, F.M.: A Neuro-fuzzy system architecture for behavior based control of a mobile robot in unknown environments. Fuzzy Sets and Systems 87 (1997) 133–140 33. Hui, N.B., Mahendar, V., Pratihar, D.K.: Time-optimal, collision-free navigation of a car-like mobile robot using a neuro-fuzzy approach. Fuzzy Sets and Systems 157 16 (2006) 2171–2204 34. Jang, J.S.R., Sun, C.T., Mizutani, E.: Neuro-Fuzzy and Soft Computing. Prentice-Hall of India Pvt. Ltd., New Delhi (2002) 35. Hui, N.B., Pratihar, D.K.: Automatic design of fuzzy logic controller using a genetic algorithm for collision-free, time-optimal navigation of a car-like robot. International Journal of Hybrid Intelligent Systems 5 3 (2005) 161–187 36. Latombe, J.C.: Robot motion planning. Kluwer Academic Publishers (1991) 37. Casillas, J., Cordon, O., Herrera, F.: Learning fuzzy rules using ant colony optimization algorithms, Proc. of 2nd Intl. Workshop on Ant Algorithms. Brussels, Belgium (2000) 13–21 38. Berenji, H., Khedkar, P.: Learning and tuning fuzzy controllers through reinforcements. IEEE Transactions on Neural Networks 3 5 (1992) 724–740 39. Denna, M., Mauri, G., Zanaboni, A.M.: Learning fuzzy rules with Tabu search– an application to control. IEEE Transactions on Fuzzy Systems 7 2 (1999) 295–318 40. Bagis, A.: Determining fuzzy membership functions with tabu search – an application to control. Fuzzy Sets and Systems 139 (2003) 209–225 41. Kim, D., Rhee, S.: Design of a robust fuzzy controller for the arc stability of CO2 welding process using the Taguchi method. IEEE Transactions on Systems, Man and Cybernetics–Part B 32 2 (2002) 157–162 42. Alfaro, H.M., Garcia, S.G.: Mobile robot path planning and tracking using simulated annealing and fuzzy logic control. Expert Systems with Applications 15 (1998) 421–429

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design: An Application to Promoter Recognition in DNA Sequences Romesh Ranawana and Vasile Palade University of Oxford Computing Laboratory, Oxford, United Kingdom January 16, 2006 [email protected] [email protected] Summary. This chapter presents a novel methodology for the customization of neural network based multi-classiﬁers used for the recognition of promoter regions in genomic DNA. We present a framework that utilizes genetic algorithms (GA’s) for the determination of optimal neural network parameters for better promoter recognition. The framework also presents a GA based method for the combination of the designed neural networks into the multi-classiﬁer system.

4.1 Introduction All sequenced genomes1 are extremely data-rich. The information contained within these sequences includes the answer for a lot of the problems faced by humans, but, due to their complexity, it is near impossible to analyze them using traditional methodologies. Thus, to fully realize the value of any sequenced genome and to fully understand it, eﬃcient and intelligent computational methods are needed in order to identify the biological relevant features in the sequences and to provide insight into their structure and function. Due to the diﬃculty involved in deﬁning these features in a manner understandable to a computer algorithm, designers are being forced into using machine learning techniques for their implementation. Machine learning algorithms allow us to design eﬃcient programs for pattern recognition tasks which do not have well deﬁned data sets. Some of the recently used methodologies include the usage of neural networks [13, 30, 31], decision trees [27], genetic programming [10], hidden markov models [2, 9, 11, 12] and fuzzy logic [17, 32].

1

The genome of an organism is its set of chromosomes, containing all of its genes and associated DNA. Please see Section 4.2 for further details.

R. Ranawana and V. Palade: A Neuro-Genetic Framework for Multi-Classiﬁer Design: An Application to Promoter Recognition in DNA Sequences, Studies in Computational Intelligence (SCI) 66, 71–94 (2007) c Springer-Verlag Berlin Heidelberg 2007 www.springerlink.com

72

R. Ranawana and V. Palade

One observation that has been made with the use of these types of methodologies for gene recognition is the diﬀerence in results obtained through the use of each type of algorithm. Although each type of system provides similar results in terms of the accuracy displayed, slight diﬀerences in the correct and incorrect genes recognized have been observed. Thus, it can be concluded that each methodology focuses the system to specialize in speciﬁc types of genes. It is due to this reason that hybrid multi-classiﬁers have now become increasingly popular in the implementation of pattern recognition systems. These hybrid multi-classiﬁer systems are implemented by building a series of classiﬁers to solve the same task, but by using a diﬀerent implementation algorithm for each of them. The results of each of these individual classiﬁers are then combined to obtain the ﬁnal result of the classifying system. Due to their robustness and their remarkable ability in deriving meaning from complicated or imprecise data, neural networks can be used to extract patterns and detect trends that are too complex to be noticed by either humans or other computer based techniques. Thus, they have been used extensively in the solving of problems related to bioinformatics. Nevertheless, one of the central dogmas involved in the design of a neural network for the solving of any problem lies in the decision on the characteristics or parameters that deﬁne the network. In this chapter, we present a framework which can be utilized for the automatic design and implementation of a neural network based multi-classiﬁer system for the identiﬁcation of protein coding regions (i.e. genes) in strings of DNA.

4.2 Biological Background 4.2.1 Genes and Promoters Gene expression involves the usage of the information contained within a gene to produce a protein, which in turn leads to the production of the phenotype determined by that particular gene. The production of these proteins is a multi-step process, which begins with transcription and translation and is followed by folding, post-translation modiﬁcation and targeting. During transcription, the DNA is copied into RNA2 by a polymerase called RNA polymerase (RNAP). This transcription yields mRNA, which is the ﬁrst step of the protein production process. Promoters within genomes sequences are sequences of DNA located upstream of protein coding regions. They are usually short sequences and are bounded by speciﬁc transcription factors3 . RNAP has the ability to recognize 2

3

Is a nucleic acid, and is used as intermediatory during the production of proteins from genomic information. A transcription factor (TF) is a protein which binds onto speciﬁc DNA sequences on a genome sequence.

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design Promoter

upstream

5’

73

-35

-10

+1

T

3’

mRNA

Fig. 4.1. A simple Prokaryotic Gene model

promoters of protein coding genes. Thus, RNAP uses the successful identiﬁcation of promoters for the identiﬁcation of protein coding regions. This paper uses the same method for the identiﬁcation of genes in genomic DNA. The proposed system works on the premise that the successful identiﬁcation of a promoter region leads to the successful identiﬁcation of the start position of a gene. Figure 4.1 displays a simple model of a Prokaryotic gene. The ‘−35’ positions and ‘−10’ positions located within the promoter. The ‘+1’ position marks the beginning of the gene and the ‘T’ signiﬁes the ‘terminator’ which marks the end of the gene. It is the DNA between these two positions which is transcribed in to mRNA, which is turn used for the production of the protein. ’In prokaryotes, the sequence of a promoter is recognized by the Sigma(σ) factor of the RNA polymerase. These sites are represented by the ‘−35’ and ‘−10’ symbols in Figure 4.1. E.Coli promoters are composed of two of these sites4 . These are the locations to which E.Coli polymerase, the RNAP used by E.Coli, binds onto in order to begin the transcription of protein. These two binding sites are always located at the points known as the −35 hexamer box and the −10 hexamer box. Consensus sequences have been identiﬁed for each each of these two locations. A consensus sequence is the most probably sequence to occur at a certain position. The spacer between the −10 hexamer box and the transcriptional has a variable length, the most probable length being 7. The spacer between the −10 site and the −35 site is also of variable length and can vary between 15 and 21 bases. An “ideal” E.Coli promoter would look something like the sequence displayed in Figure 4.2. It displays the most probable sequences that could occur at the two hexamer boxes. It is this variation that can make the recognition of these promoters diﬃcult with traditional methodologies. Many promoter sequences have the pyrimidine5 at position +1 (one nucleotide upstream of the transcription start site or the gene), and purine6 at the transcriptional start site. In addition to the more obvious features described, a few more non-obvious features have been observed by [5] and [16]. [13] used the software available at http://www-lecb.ncifcrf.gov/toms/ delia.html to display the sequence logos of 438 E.Coli promoters that were aligned according to their transcriptional start sites. Their results are displayed 4 5 6

These sites are also known as binding sites. C or T A or G

74

R. Ranawana and V. Palade

5’

TTGACA -35 site

17 bp

TATAAT

7 bp

+1

3’

-10 site

Fig. 4.2. A simple Prokaryotic Gene model

Fig. 4.3. A ﬂow-chart of the framework utilized for the development of the system

in Figure 4.3. This software used the Shannon Entropy to measure the nonrandomness of the individual bases at each position independently. We use these observations later in the paper to display the ability of our neural networks to learn these distributions automatically. 4.2.2 The Importance of Studying the E.Coli Gene E.Coli K-12 and its close relatives have been extensively studied in genetic, molecular and biochemical studies from the dawn of modern microbiology and molecular biology. The development of a ‘virtual cell’ maybe far into the future, but it is a well known concept that the place to start the consideration of what will work and how to set about it lies in the development of a simple but adaptable system. In this regard, E.Coli K-12 provides an opportunity to test out approaches to modelling cellular and organismal systems. The Bioinformatics Initiative7 lists a set of reasons giving out reasons as to why E.Coli is currently the organism of choice for such a study, both as a reference organism for many studies on prokaryotic systems and as a source of information on proteins and metabolic pathways that are shared by eukaryotics as well. 4.2.3 Methods Used for the Identiﬁcation of E.Coli Promoters Most of the methods that have been used for the detection of E.Coli promoters involve the alignment of all the given sequences at the transcriptional start site 7

http://www.ecoli.princeton.edu/E coli Bioinformatics Document.pdf

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

75

(i.e. the ‘+1’ position) and then scanning them for the consensus sequences. They look for the primary consensus sequences at positions ‘−35’ and ‘−10’. Some of the more successful methods scan for some of the weaker motifs that have been identiﬁed at positions ‘+1’, ‘−22’ and the ‘−44’ position. Some examples of the use of these methods are listed out in [20] and [26]. Each of these systems programmed known knowledge about the E.Coli promoter into the system in order correctly classify unknown sequences. The main problem associated with these systems and other systems which use this technique is the uncertainty involved in the placement of the consensus sequence within the promoter. This can be caused due to the consensus sequence varying from the typical or by the consensus sequence being placed at a slightly diﬀerent position. The spacer between consensus sequences is also variable. Another method involves the use of a scoring matrix to calculate the variation of the diﬀerent bases within the sequence. An example of such an implementation can be found at [4]. Neural networks have also been successfully utilized for the recognition of these E.Coli promoters [13, 14, 30]. Statistical methods like Hidden Markov Models and the EM algorithm have also been successfully utilized for this purpose. The system presented in this paper also attempts to solve this problem, but utilizes a hybridization of a few of the methods mentioned above to achieve the objective. We built nine neural networks, each of which was individually trained and customized for better performance. These neural networks were then combined using a genetic algorithm. The results show us that the customized neural networks were able to identify the consensus sequences and identify important subsections of the DNA sequences shown to them. Manual entry of the consensus sequence information was not required. 4.2.4 The Data Set The data set used consisted of 473 E.Coli sequences and 600 non-promoter sequences. The positive data sequence was obtained from the PromEC8 website. Each of the obtained E.Coli sequences were 101 nucleotides in length and were aligned at the transcriptional start site. Each of the transcriptional start sites were positioned such that they each appeared at position 76 of the string. Thus, each of the positive sequences consisted of nucleotides from 101 nucleotides starting from the −75 position, 75 nucleotides upstream of the transcriptional start site, and ending at the +25 position, 25 nucleotides downstream of the transcriptional start site. The negative data set was obtained from the database compiled by the E.Coli GenomeProject9 , at the University of Wisconsin-Madison. These negative data consisted of E.Coli genes with the preceding promoter region deleted. Using the above mentioned data set, we constructed three data sets: a training sequence, a positive test sequence and a negative test sequence. 8 9

http://bioinfo.md.huji.ac.il/marg/promec/ http://www.genome.wisc.edu/sequencing/k12.htm#seq

76

R. Ranawana and V. Palade

The training set consisted of 314 positive sequences and 429 negative sequences. The remaining positive and negative data were divided up into a positive test set of 159 sequences and a negative test set of 171 sequences. For the purpose of initial training and testing the originating neural networks, all three data sets were encoded using −2, −1, 1 and 2 to represent the bases A, T, C and G, respectively.

4.3 Multi-Classiﬁer Systems 4.3.1 Classiﬁers Classiﬁcation, in computer science terminology, is the act of placing a given data item within a set of categories based on the properties of that particular data item. A classiﬁer is a computer based agent who would perform this classiﬁcation. For any pattern classiﬁcation application research area currently under investigation, there are many possible classiﬁcation algorithmic solutions which could be utilized. Any learning problem involves the usage of a set of training examples of the form {(x1 , y1 ), ..., (xm , ym )} for the approximation of a function f (x). The x values are usually of the form < xi,1 , xi,2 , ..., xi,n >, and are composed of either real or discrete values. The y values are the expected outputs for the corresponding x vales, and are usually drawn from discrete sets of classes. Thus, when presented with the training set, the learning paradigm attempts to approximate the function f (x) and produces a classiﬁer. The produced function is an approximation of the function f (x) as the training set, more often than not, contains random noise. Thus, with the use of multiple learning paradigms, one obtains a set of classiﬁers h1 , h2 , ..., hL , each of which attempts to approximate the function f (x). Due to diﬀerences between the learning paradigms, the way that the approximation is evolved also diﬀers. Thus, the results obtained through each classiﬁer can also be diﬀerent. 4.3.2 Why use Multiple Classiﬁers? An ensemble of classiﬁers or a multi classiﬁer, is a collection of classiﬁers whose individual decisions are combined in some way to obtain a combined decision. [8] claims that such an ensemble can only be more successful than the individual classiﬁers if the individual classiﬁers disagree with each other. As mentioned in [33], there are two main groups of methodologies for the combinations of classiﬁers, namely, feature-vector-based methods10 and syntactic-and-structural methods11 . In addition to this distribution, each group is composed of a large variety of identiﬁable subgroups. For many of the 10 11

E.g. Neural networks E.g. Fuzzy rule bases

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

77

classiﬁcation problems encountered, it has been experimentally proven that diﬀerent classiﬁers exhibit diﬀerent degrees of success. No classiﬁer has yet has proven to be totally perfect in the classiﬁcation of any problem domain. Often, it has also been observed that diﬀerent classiﬁers can disagree with each other on classiﬁcation decisions when presented with the same training data and testing data. Thus, Individual classiﬁer models are recently being challenged by combined pattern recognition systems, which often show better performance. This is due to the reason that, when a system that incorporates many methodologies is used, the inadequacies of one methodology are usually nulliﬁed by the characteristics of another. That is, a multiple classiﬁer can nullify the predictive inaccuracy obtained through the user of singular classiﬁcation systems. Thus, genes not recognized by one classiﬁer will be recognized by another and genes incorrectly recognized by one classiﬁer will be rejected by another (Figure 4.4). The challenge in building such a system is the selection of a combination method for the results provided by each classiﬁer in order to come up with the optimal and most popular result. Here, the performance of the entire system can be proven to be never much worse than that of the best expert [8]. [3] also states three causes as to why multiple classiﬁers exhibit better performance when compared to singular systems, and then goes on to state that the use of multiple classiﬁcation solutions and their combinations leads to the nulliﬁcation of many of the problems encountered through the use of singular classiﬁers. In a situation where a number of classiﬁers are available, the simplest approach would be to select the best performing classiﬁer and use it for the classiﬁcation task [18]. This approach, although simple and easy to implement,

Classifier 2 Classifier 1

Classifier 3

Input Space

True Positives True Positives identified by the classifier

Fig. 4.4. A Classiﬁer problem with diﬀerent classiﬁers being used for the same data set. The shaded region represents the true positives and the ellipses represent the space classiﬁed by each classiﬁer as being true positives

78

R. Ranawana and V. Palade

does not guarantee good performance [22]. It is highly probable that a combination of classiﬁers would outperform a singular classiﬁer [25]. On the other hand, diﬀerent and worse performing classiﬁers might only add to the complexity of the problem and provide even worse result than the worst classiﬁer. Thus, it is a well known fact that if a multi-classiﬁer system is to be successful, the diﬀerent classiﬁers should have good individual performances and be suﬃciently diﬀerent from each other [28]. But, neither individual performances [21, 35] nor the diversity [24, 29] provides an ideal measure of how successful the combination of the classiﬁers will be. As explained in [25], the core element of classiﬁer selection is the selection criteria. The most natural choice being the combined performance, which will also the criterion for selection of the combiner. The only drawback of this methodology is the exponential complexity of testing out all possible combinations of a given set of classiﬁers. It has been proven that the usage of an ensemble of neural networks for certain classiﬁcation problems can improve classiﬁcation performance, when compared to the use of singular neural networks. [19] and [23] provide results tested on protein secondary structure prediction. [1] also lists out an overview of applications in molecular biology. Of the diﬀerent methodologies used for the combination of multiple neural net classiﬁers, majority voting, neural networks, Bayesian inference and the Dempster-Shafer theories have proven the most popular [15, 33, 34]. The Dempster-Schafer method has proven to be successful, but has a considerable dependency on the function used for the alignment of probability. For the type of output produced by neural networks, posterior class-conditional probabilities can be calculated. The calculation of these probabilities becomes relatively simple, specially when the number of output classes is small. In this Chapter, we test two methods (LAP and LOP) that can be used for the combination of results obtained through multiple neural network classiﬁcations and are compared with the use of a variation of the two methods, the LOP2 method. These two methods, the LAP and LOP methods, were introduced in (Hansen and Krogh, 1999). 4.3.3 The LAP and LOP Methods for Combining Classiﬁers Introduction to the LAP and LOP Methods [7] presents a general composition method that considers the average of results of several neural networks to obtain a combined result that would improve the overall classiﬁcation performance. The method that we used for the combination of the classiﬁers designed in this research was a variation of the LOP (Logarithmic Opinion Pool) [7] method, LOP2. We compare this method with the LAP (Linear Average Predictor) method, which is more commonly used. The LOP method is a general ensemble method which shows how the error or an ensemble can be written

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

79

as the average error between members (called the ensemble ambiguity). [7] claims that this proves that ensembles always improve average performance. Description of the Ensemble Composition Method used in this Research An ensemble consists of M predictors fi , which are combined into the combined predictor F. Let each predictor produce an output in the form of a probability vector (fi1 ,...,fiN ), where is the estimated probability that input belongs to class cj . There is also a coeﬃcient M associated with each ensemble M of the form (α1 ,...,αM ) where i=1 αi = 1. The LAP Method The combined predictor for the LAP of the ensemble is then deﬁned in Equation (4.1). j = FLAP

M

αi fij

(4.1)

i=1

The LOP Method The combined predictor for the LOP of the ensemble is then deﬁned in Equation (4.2), where Z is a normalization factor given by Equation (4.3). M

1 j j α1 logfi (4.2) FLOP = exp Z i=1 M N

j Z= exp α1 logfi (4.3) j=1

i=1

The LOP2 Method The LOP algorithm was changed by removing the condition (α1 ,...,αM ) M where i=1 αi = 1 , with α also being given the capability to take on negative values. We found that this slight alteration allowed the system to nullify small errors created by the more successful classiﬁers. This research attempted to utilize this methodology for the combination of the classiﬁers designed, and utilized genetic algorithms for the optimization of the unknown coeﬃcients (α1 ,...,αM ).

80

R. Ranawana and V. Palade

4.4 Design of the System 4.4.1 Justiﬁcation for the Development of the Framework The development of this system was promoted through the development of the MultiNNProm System [http://web.comlab.ox.ac.uk/oucl/work/romesh. ranawana/RP2004a.pdf] and the problems encountered during its implementation. MultiNNProm provided us with very promising results in terms of the accuracy and precision displayed. The main problem encountered during the development was the large amount of time taken to properly conﬁgure the neural networks and to determine the optimal encoding methods for the given data set. All these design decisions had to be made on an ad-hoc basis, with the testing of many diﬀerent conﬁgurations and encoding methods being required in order to obtain an optimal set of parameters. As each test run required the training and testing of a neural network, this process was extremely time consuming. In order to negate these problems, we successfully designed and implemented a framework which would automatically determine the optimal parameters required for the design of the multi-classiﬁer system. Within this framework, we utilized a genetic algorithm based method for the determination of the optimal number of layers and the optimal number of neurons on each layer. We also implemented a component of this framework to automatically determine an optimal set of encodings for the presented DNA data set. Thus, when presented with a DNA data set, the framework successfully designed and trained a set of neural networks for inclusion in a multi-classiﬁer system. The following sub-sections describe the implementation of the framework and also lists out the results obtained. 4.4.2 Determination of Optimal Neural Networks Conﬁgurations for the given DNA Data Set Determination of the Optimal Number of Hidden Layers In order to obtain an optimal neural network conﬁguration for the DNA data considered in this paper, we ran a series of simulations, each testing out the accuracy of the system for diﬀerent numbers of hidden layers. We began the simulation run with a neural network that contained one hidden layer and 10 neurons on that hidden layer. The network was trained until an error rate of 1e − 4 was reached. The resulting network was then tested on a separate testing set. A similar set of training/testing runs were then conduced on separate networks, each containing one neuron more than the amount present on the previously trained/tested network. Figure 4.5 graphically illustrates the normalized error of each conﬁguration ﬁtted onto a third order polynomial curve. This curve identiﬁes the fact that as the number of neurons on the ﬁrst layer increases, the generalization capability of the system in turn decreases. Each of these networks demonstrated an accuracy in excess of 96% when presented with the training set.

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

81

Fig. 4.5. The ﬁrst order polynomial curve ﬁt of the error of the system with respect to the number of neurons on the hidden layer of a one hidden layer neural network

Fig. 4.6. The polynomially ﬁt curve of the errors for the one hidden layer and two hidden layer neural networks

We then performed a similar set of simulations on networks with two hidden layers. We maintained a constant number of neurons on the ﬁrst hidden layer and varied the number of units on the second hidden layer. Figure 4.6 shows the accuracy of the networks that contained 50 neurons on the ﬁrst hidden layer and a varying number of neurons on the second hidden layer. It was also observed that the performance of the system with either one or two hidden layers was similar in terms of the accuracy with respect to the number of neurons on the layer before the output layer.

82

R. Ranawana and V. Palade

Fig. 4.7. A comparison of the errors for the systems with 1, 2, 3, 4 and 5 hidden layers Table 4.1. Neural network conﬁgurations for the types of neural network trained and tested. Here, ∗ indicates the variable value Type One hidden Layer Two hidden layer Three hidden layer Four hidden layer Five hidden layer

Conﬁguration ∗

50:∗ 100:50:∗ 100:80:50:∗ 125:100:80:50:∗

A marked improvement in the generalization of the system was observed once the number of hidden layers exceeded 2. Figure 4.7 compares the accuracy of the systems trained and tested using 1, 2, 3, 4 and 5 with the number of neurons on the ﬁnal layer being varied. Table 4.1 lists the conﬁgurations used as the base for the conﬁguration of each of these networks. It was observed that the performance did not vary drastically once the number of layers exceeded 3. While the performance of the systems with 3 hidden layers diminished by small increments, the performances of the remaining two systems remained almost constant. We accredited this increase in performance along with the increase in the number of hidden layers to the networks being able to identify the complexity of the given promoters more successfully. We concluded that neural networks with a higher number of hidden layers were able to recognize relationships of a higher order better than networks with one or two hidden layers. That is, they were able to correctly learn and classify second and third order relationships among the bases of the DNA string, which in turn leads to more genes being classiﬁed.

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

83

Table 4.2. The best performing neural network conﬁgurations for varying numbers of hidden layers with the error percentage observed Number of hidden layers 1 2 3 4 5

Optimal Conﬁguration 18 50:6 100:50:5 100:80:50:18 125:100:80:50:24

Error(%) 12.84 12.73 11.84 11.55 11.43

Listed in Table 4.2 are the minimum errors obtained for each type of neural network along with the number of neurons present on the ﬁnal hidden of that optimal network. Thus, it was concluded that the optimal network with regard to accuracy, generalization and complexity was the network that contained three hidden layers, with 10 neurons on the ﬁnal hidden layer. Although the network with 5 hidden layers did perform moderately better, the network with 3 hidden layers was deemed to be better due to the its reduced complexity and the lesser time required for training. Determination of the Number of Units on Each Layer Once the optimal number of neural network hidden layers for the dataset was determined, the next task was the determination of the number of units required on each layer in order to enhance generalization. For this purpose, we initialized a random set of vectors to represent the number of neurons on each layer. Each vector consisted of 12 integer values, where the ﬁrst four values varied between 1 and 50, the second four between 1 and 20 and the remainder between 1 and 10. The sum of the ﬁrst, the second and ﬁnal four values were used to represent the number of neurons on the ﬁrst, second and third hidden layers respectively. We divided the number on neurons in each layer within each chromosome in order to maximize the eﬀect of mutation and crossover when used within the neural network. If each unit count was used by itself, individual chromosomes would only have included three values, and would have consequently nulliﬁed most of the advantageous eﬀects of the genetic algorithm. Four initial populations, each consisting of 60 vectors were initialized. Each of these populations was then exercised through a genetic algorithm in order to obtain an optimal set of vectors representing the conﬁguration of the network. The 15 best performing vectors of each population was then extracted to create a new population, and was exercised through a genetic algorithm in order to obtain an optimal set of conﬁgurations for the neural network. This stage was required for the nulliﬁcation of any local minimums that the genetic algorithm would have encountered. The ﬁtness function utilized for the selection of vectors was the classiﬁcation error exhibited by the neural network in terms of the percentage of

84

R. Ranawana and V. Palade

Table 4.3. The best performing neural network conﬁgurations with respect to the number of units on each hidden layer and the error displayed by each system. Label a b c

Hidden Layer 1 103 75 56

Hidden Layer 2 63 19 43

Hidden Layer 3 22 17 54

Error(%) 9.82 8.36 8.36

false positives and false negatives with respect to the total amount of testing data presented12 . Thus, the genetic algorithm attempted to minimize the total error displayed by the system with respect to the test data. Each neural network was trained until the error reached a value of 1E − 3 or the number of training epochs reached 100. The genetic algorithm was implemented using a precision of 0.9, which meant that the 6 best performing vectors were included in their entirety within the next generation. Also, a crossover probability of 0.7 and a mutation probability of 0.1 was used along with roulette wheel selection [6] for its implementation. Each evolutionary function was iterated 60 times before the selection of the winning vectors was made. Table 4.3 lists out the 3 best performing conﬁgurations along with the error displayed by each of them. The best performing system during the determination of the number of layers displayed an error rate of 11.43%, whereas with the introduction of the genetic algorithm and better unit number selections, the recognition rate was increased to produce a minimum error rate of 8.36%. Thus, it can be inferred that the correct conﬁguration of the layers leads to a better rate of recognition. When the true positives and true negatives of conﬁgurations b and c were compared, it was found that the sequences correctly classiﬁed by each network diﬀered by around 3%. This lead us to the conclusion that diﬀerent network conﬁgurations can lead diﬀerent networks specializing in speciﬁc classes of promoters and non-promoters. 4.4.3 Method used for the Determination of Optimal Encoding Methods To test the hypothesis that the encoding method can aﬀect the performance of the system, we ran a series of genetic algorithms which tested diﬀerent encoding methods. We initialized ten random populations of 60 vectors, where each vector was represented by four random real values which ranged from −10 to +10. Here, the four values on the vector corresponded to the four DNA bases. Each population was then exercised through a genetic algorithm which used the accuracy of the system of the system with respect to the testing set as the ﬁtness function. Each population was iterated 60 times through the evolutionary algorithm with a mutation probability of 0.8, a precision of 0.8. The algorithm used roulette wheel selection for cross-overs selections. The 12

The inverse of the Precision

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

85

Error

10 best performing vectors from each population were then extracted to form a new population and was then exercised through a genetic algorithm which identical attributes. This step was performed in order to obtain an optimal set of encoding schemes by using the best chromosomes identiﬁed by the individual populations and also assists with the negating of negative eﬀects caused through local minimums. The training graph for the best performing vectors of each generation of the ﬁnal optimized population is displayed in Figure 4.8 by open circles. Here, the average error of each generation is represented by a cross. These results show us that the accuracy of the system was optimized from an error rate of around 8% to a value close to 4%. Thus, the total accuracy of the system was increased from 92% to 96% through the utilization of the genetic algorithm. Table 4.8 lists out the three pest performing encoding methods along with the accuracy displayed by the usage of each. Displayed in Figure 4.9(a) and (b) are value comparisons of the 3 best performing vectors. Figure 4.9(a) shows displays the distribution of each vector where it can be seen that the distribution of the values for the diﬀerent bases displays a very distinct pattern. The numerical values of A and C remain close to each other and vary between 0 and 1, whereas, the values of T and C remain numerically distanced from each others values. It was also observed that the values of T and C always remain

Population Number

Fig. 4.8. Training graph for the encoding optimization genetic algorithm Table 4.4. The three best performing encoding methods along with the error rate displayed by each Label e1 e2 e3

A 0.4788 0.1309 0.2702

T 1.3868 1.9883 0.0693

C 0.1692 0.6046 0.4255

G −0.3659 −0.0513 1.5572

Error(%) 0.0436 0.0489 0.0409

86

R. Ranawana and V. Palade 2.5 2 1.5 Value

A T

1

C 0.5

G

0 2

1

3

-0.5

2.5 2

Values

1.5 Vector 1 Vector 2

1

Vector 3 0.5 0 AT

CG

-0.5 Bases

Fig. 4.9. A comparison of the best performing encoding vectors

either larger or smaller, conversely, to the values of A and C. Thus, the curve always remains either convex or concave. The line never assumes a zigzagged value. These conclusions are substantiated by Figure 4.9(b), which shows the distribution of the diﬀerent bases for the three wining vectors, demonstrating the more or less constant values maintained by both A and C. The best performing network with the optimized number of neurons on each layer displayed an error rate of 8.36%, whereas, with the introduction of an optimized set of encoding methods, the error rate was reduced to 4.36%. As with the results obtained through the optimization of the network conﬁguration, the results obtained through the use of each encoding method produced slightly varying results with relation to the promoters classiﬁed. Thus, the underlying conclusion being that each encoding method helped the network specialize on diﬀerent types of promoters present within the data set. 4.4.4 An Overview of the System As shown in Figure 4.10, the system is a neural-network based multi-classiﬁer system. Through the optimization process of the framework, we obtained three optimal neural network conﬁgurations; a, b and c, and three optimal sequence encoding methodologies; e1, e2 and e3. By combining each encoding method with each conﬁguration, we developed and trained 9 neural networks, NNe1a, NNe1b, NNe1c, NNe2a, NNe2b, NNe2c, NNe3a, NNe3b and NNe3c. Each

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

87

A T C

DNA string

G T ... ... ... A G C C T

Fig. 4.10. NNGenProm Architecture

neural network was initially trained to produce an output of -1 if it predicted the sequence to be a non-promoter and an output of 1 if it predicted a promoter. The outputs of the individual neural networks were then passed onto a probability builder function which assigned probabilities as to whether the presented sequence was an E.Coli promoter or not. Finally, the outputs of the probability functions were passed onto a result combiner which combined the results and presented a ﬁnal result as to whether the given sequence was an E.Coli promoter or not. The ﬁnal output was of the ‘yes’(1) or ‘no’(0) form. The probability function used in this implementation was identical to that used during the implementation of the MultiNNPRom system. 4.4.5 The Result Combiner The results were combined using the LOP2 method described in Section 4.3.3.2. The system was tested on three combination methods, namely, classical majority voting, LAP, and LOP2. The resulting observations showed us that the results obtained through the LOP2 method provided us with a much better recognition rate in terms of the test data, both positive and negative. The comparison of results between these three combinatorial methods is listed out in Section 4.5. The LOP2 method was implemented as follows. We let the outputs of the nine neural networks be symbolized by Oi , where 1 ≤ i ≤ 9. We also deﬁned 9 coeﬃcients αi , where 1 ≤ i ≤ 9. Then, the combined predictor was deﬁned by Equations (4.4) and (4.5) where (4.6). 9

1 αi (log(Oi .P ositive)) (4.4) O.P ositive = exp Z i=1

88

R. Ranawana and V. Palade

9

1 O.N egative = exp αi (log(Oi .N egative)) Z i=1 Z = exp

9

αi (log(Oi .P ositive))

i=1

+ exp

9

(4.5)

αi (log(Oi .N egative))

i=1

(4.6) It is one again obvious that O.P ositive + O.negative = 1. The ﬁnal conclusion on whether the given sequence was an E.Coli promoter or not was reached using equation (4.7). C(O.P ositive, O.N egative) =

Y es ; O.P ositive > O.N egative N o ; Else

(4.7)

The methodology used to determine values for the coeﬃcients is described in the next section. The coeﬃcient values were determined using a method identical to that used during the design and implementation of MultiNNProm.

4.5 Results 4.5.1 Evaluation of the Performance of the Individual Classiﬁers Each neural network performed perfectly on the training set and displayed a recognition rate of approximately 100% when presented with them after training. The test sequences were exercised through the 9 neural networks and the results obtained are listed out in Table 4.11. These performances are graphically in Figure 4.11. It can be observed that while some networks perform better with respect to the recognition of true negatives, others specialize in the recognition of true positives. For example, nne2b has a better Table 4.5. The Speciﬁcity, Sensitivity and Precision of the nine trained neural networks Network nne1a nne2a nne3c nne1b nne2b nne3b nne1c nne2c nne3c

True Negatives 97.66 97.66 98.83 98.23 98.25 98.83 94.15 98.83 94.74

True Positives 91.20 93.71 92.45 88.68 91.82 91.20 92.45 92.45 93.71

Precision(%) 94.43 95.69 95.64 93.46 95.03 95.01 93.30 95.64 94.22

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

89

Accuracy

100.00 98.00

nne1a

96.00

nne2a nne3a

94.00

nne1b

92.00

nne2b

90.00

nne3b

88.00

nne1c

86.00

nne2c

84.00

nne3c

82.00 True Negatives

True Positives

Precision

Fig. 4.11. Comparison of the true positives and true negatives recognized by each neural network Table 4.6. The results obtained when the test data sets were exercised through the combined system Attribute Speciﬁcity Sensitivity Precision

Value 0.9882 0.9748 0.9815

recognition rate with respect to true negatives, whereas nne2c has a better recognition rate with respect to true positives. nne3a, nne3b and nne2c provide the best recognition of true negatives, whereas the best recognition rate of true positives was displayed by nne2a. 4.5.2 Performance evaluation of the Combined System The speciﬁcity, sensitivity and precision of the neural network results combined using the LOP2 method are listed out in Table 4.6. These results are compared with the results obtained for the individual neural networks in Figure 4.12. These results indicate that the combination of these nine neural networks provides us with a far better recognition rate in terms of the precision, speciﬁcity and sensitivity. Thus, the performance of the system has been improved by a considerable margin. It was observed that the combination of the nine neural networks using the LOP2 method did not diminish the performance by either the true positive accuracy or the true negative accuracy of the overall precision. Whereas the true negative recognition rate remains equal to the recognition rate of the best negative classiﬁer after the application of the combinatorial function, the true positive recognition rate has been increased by a considerable fraction. Thus, as claimed by [3], the performance of the combination was not worse than the performance of the best classiﬁer.

90

R. Ranawana and V. Palade 100.00

nne 1a

98.00

nne 2a

Accuracy

96.00

nne 3a

94.00

nne 1b

92.00

nne 2b

90.00

nne 3b

88.00

nne 1c nne 2c

86.00

nne 3c

84.00

Combination

82.00 True Negatives

True Positives

Precision

Fig. 4.12. Comparison of the speciﬁcity, sensitivity and precision of the four neural networks and the combined system Table 4.7. Comparison of the speciﬁcity, sensitivity and precision of the research obtained using diﬀerent combinatorial methods Combinatorial method Majority Voting LOP2

Sp 0.9882 0.9882

Sn 0.9559 0.9748

P 0.9720 0.9815

Table 4.8. Comparison of the speciﬁcity, sensitivity and precision of MultiGenProm with previous research on the same dataset System NNGENProm Ma et al.,2001 [13] Mahadevan and Ghosh, 1994 [14]

Sp 0.9882 0.9176 0.9020

Sn 0.9748 0.9920 0.9800

P 0.9815 0.9194 0.9040

We attribute this increase in accuracy to the diﬀerent specializations exhibited by each network. The combination of the networks lead to a committee of classiﬁers that performed better than the individual classiﬁers. Thus, the inaccurate predictions of one network were negated by the correct predictions of another. The results of this combination were also compared with the results obtained through the usage of majority voting. Table 4.8 lists out this comparison. These results point our the fact that, although the combination of the 9 neural networks using either method provides us with better results than the individual neural networks, the results obtained through the LOP2 method are better. Table 21 compares the results obtained through this rerearch with previous work done on the same problem. Thus, our system is seen as a considerable improvement compared with recent research on the recognition of E.Coli promoters.

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

91

Table 4.9. The performance of the system after each step of the framework was completed Step Lowest Error (%) Determination of the number of hidden layers 11.43 Determination of the number of units on each hidden 8.36 layer Determination of the optimal encoding methods for the 4.36 DNA data set Design and implementation of the networks using opti4.36 mal conﬁgurations and encodings Combination of the neural networks 1.85 0.6

0.5

Error

0.4

0.3

0.2

0.1

Hidden layer and unit count Encoding Determination Combination

0

Fig. 4.13. The performance of the system through the diﬀerent steps of the framework

4.5.3 Summary of the Results Obtained through the use of the Presented Framework As tabulated in Table 4.9, the diﬀerent steps presented within the framework gradually help decrease the error in the recognition rate displayed by the system. These observations are graphically illustrated in Figure 4.13. The spikes on the graph are due to random chromosomes being used for each initial population. This framework presents a methodology that assumes the determination of all the unknown parameters to be an optimization problem, and uses genetic algorithms in order to determine them.

4.6 Discussion and Future Work In this chapter, we presented a novel approach for the recognition of E.Coli promoters in strings of DNA. The proposed system showed a substantial improvement in the recognition rate of these promoters, speciﬁcally in the

92

R. Ranawana and V. Palade

recognition of true positives, i.e. rejection of non-promoters. This caused our system to display a higher speciﬁcity than all other systems developed thus far. We conclude the reasons for this improvement to be threefold. Firstly, it was the use of larger neural networks than those that have been used thus far. This led to a far better rate of recognition and generalization. We also noted that an increase in the number of hidden layers on the neural network also led to a better recognition rate. Secondly, the use of multiple neural networks, each accepting the same set of inputs in diﬀerent forms. We observed the fact that the false positives and false negatives of one network could be wiped out by true positives and true negatives in other networks. Thus, the combination of the opinions of more classiﬁers led us to a system that performed far better than the individual components. Our results substantiates Dietterich’s (Dietterich, 1997) conclusion. Finally, the use of diﬀerent encoding methods, which were customized to suite the particular data set, increased the recognition capabilities of the neural networks in terms of the data presented. The above three reasons, coupled with the fact that each optimized parameter was used in conjunction with permutations of other optimized parameters lead to a system, when combined, provided us with far better accuracy when compared to singular systems or other multi-classiﬁer systems using similar parameters. As was pointed out earlier, one of the major obstacles in the development of the presented system was the time taken for the determination of optimized parameters for the development of the system. This problem was negated by the introduction of the neuro-genetic framework, which provides the designer with each of the optimized parameters necessary for the development of a system specially customized towards the given data set. The neuro-genetic framework provided us with the optimal number of layers to include within the neural networks, the number of neurons to include with each layer and also a set of optimized encoding methods specially customized for the given data set.

References 1. P. Baldi and S. Brunak. Bioinformatics - The Machine Learning Approach, volume 1. MIT Press, 1998. 2. E. Birney. Hidden markov models in biological sequence analysis. IBM Journal of Research and Development, 45(3/4), 2001. 3. T.G. Dietterich. Machine-learning research: Four current directions. The AI Magazine, 18(4):97–136, 1998. 4. D Frishman, A Mironov, HW Mewes, and M Gelfand. Combining diverse evidence for gene recognition in completely sequenced bacterial genomes [published erratum appears in Nucleic Acids Res 1998 Aug 15;26(16):following 3870]. Nucl. Acids Res., 26(12):2941–2947, 1998.

4 A Neuro-Genetic Framework for Multi-Classiﬁer Design

93

5. D.J. Galas, M. Eggert, and M.S. Waterman. Rigorous pattern-recognition methods for dna sequences: Analysis of promoter sequences from E.Coli. Journal of Molecular Biology, 186(1):117–128, 1985. 6. D.E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley Longman Publishing Co., Inc., 1989. 7. J.V. Hansen and A. Krogh. A general method for combining in predictors tested on protein secondary structure prediction. In Proceedings of Artiﬁcial Neural Networks in Medicine and Biology, pages 259–264. Springer-Verlag, May 2000. 8. L. K. Hansen and P. Salamon. Neural network ensembles. IEEE Trans. Pattern Anal. Mach. Intell., 12(10):993–1001, 1990. 9. J Henderson, S Salzberg, and K Fasman. Finding genes in dna with a hidden markov model. Journal of Computational Biology, 4(2):127–141, 1997. 10. John R. Koza and David Andre. Automatic discovery of protein motifs using genetic programming. In Xin Yao, editor, Evolutionary Computation: Theory and Applications. World Scientiﬁc, Singapore, 1996. 11. A Krogh. Two methods for improving performance of a hmm and their application for gene ﬁnding. In Proceedings of the 5th International Conference on Intelligent Systems for Molecular Biology, pages 179–186. AAAI Press, 1997. 12. D. Kulp, D. Haussler, M. G. Reese, and F. H. Eeckman. A generalized hidden markov model for the recognition of human genes in DNA. ISMB-96, pages 134–141, 1996. 13. Q. Ma, J.T.L. Wang, D. Shasha, and C.H. Wu. Dna sequence classiﬁcation via an expectation maximization algorithm and neural networks: A case study. IEEE Transactions on Systems, Man, and Cybernetics, part C: Applications and Reviews, Special Issue on Knowledge Management, 31(4):468–475, November 2001. 14. I Mahadevan and I Ghosh. Analysis of E.Coli promoter structures using neural networks. Nucl. Acids Res., 22(11):2158–2165, 1994. 15. E.J. Mandler and J. Schurmann. Combining the classiﬁcation results of independent classiﬁers based on the dempster/schafer theory of evidence. Pattern Recognition and Artiﬁcial Intelligence, X:381–393, 1988. 16. G. Mengeritsky and T.F. Smith. Recognition of characteristic patterns in sets of functionally equivalent dna sequences. Comput. Appl. Biosci., 3(3):223–227, 1987. 17. Lucila Ohno-Machado, Staal A. Vinterbo, and Griﬃn Weber. Classiﬁcation of gene expression data using fuzzy logic. Journal of Intelligent and Fuzzy Systems, 12(1):19–24, 2002. 18. D. Partridge and W.B. Yates. Engineering multiversion neural-net systems. Neural Comput., 8(4):869–893, 1996. 19. S.K. Riis and A. Krogh. Improving prediction of protein secondary structure using neural networks and multiple sequence alignments. Journal of Computational Biology, 3:163–183, 1996. 20. K. Robison. A comprehensive library of dna-binding site matrices for 55 proteins applied to the complete escherichia coli k-12 genome, 1998. 21. G. Rogova. Combining the results of several neural network classiﬁers. Neural Netw., 7(5):777–781, 1994. 22. F. Roli and G. Giacinto. Hybrid Methods in Pattern Recognition, chapter Design of multiple classiﬁer systems, pages 199–226. Worldwide Scientiﬁc Publishing, 2002.

94

R. Ranawana and V. Palade

23. B. Rost and C. Sander. Prediction of protein secondary structure at better than 70% accuracy. Journal of Molecular Biology, 232(2):584–599, July 1993. 24. D. Ruta and B. Gabrys. Analysis of the correlation between majority voting error and the diversity measures in multiple classiﬁer systems. In Proceedings of the 4th International Symposium on Soft Computing, paper 1824–025, 2001. 25. D. Ruta and B. Gabrys. Classiﬁer selection for majority voting. Special issue of the journal of INFORMATION FUSION on Diversity in Multiple Classiﬁer Systems, 2004. 26. H Salgado, A Santos, U Garza-Ramos, J van Helden, E Diaz, and J ColladoVides. RegulonDB (version 2.0): a database on transcriptional regulation in Escherichia coli. Nucl. Acids Res., 27(1):59–60, 1999. 27. Steven Salzberg, Arthur L. Delcher, Kenneth H. Fasman, and John Henderson. A decision tree system for ﬁnding genes in dna. Journal of Computational Biology, 5(4):667–680, 1998. 28. A.J.C. Sharkey and N.E. Sharkey. Combining diverse neural nets. Knowl. Eng. Rev., 12(3):231–247, 1997. 29. C.A. Shipp and L.I. Kuncheva. An investigation into how adaboost aﬀects classiﬁer diversity. In Proc. IPMU 2002, pages 203–208, 2002. 30. E.E. Snyder. Identiﬁcation of protein coding regions in genomic dna. Journal of Molecular Biology, 248:1–18, 1995. 31. EC Uberbacher and RJ Mural. Locating Protein-Coding Regions in Human DNA Sequences by a Multiple Sensor-Neural Network Approach. PNAS, 88(24):11261–11265, 1991. 32. PJ. Woolf and Y Wang. A fuzzy logic approach to analyzing gene expression data. Physiol. Genomics, 3(1):9–15, 2000. 33. L. Xu, A. Krzyzak, and C.Y. Suen. Several methods for combining multiple classiﬁers and their applications in handwritten character recognition. IEEE Trans. on System, Man and Cybernetics, SMC-22(3):418–435, 1992. 34. L. Xu, A. Krzyzak, and C.Y. Suen. Associative switch for combining multiple classiﬁers. Journal of Artiﬁcial Neural Networks, 1(1):77–100, 1994. 35. G. Zenobi and P. Cunningham. Using diversity in preparing ensembles of classiﬁers based on diﬀerent feature subsets to minimize generalization error. Lecture Notes in Computer Science, 2167:576–587, 2001.

5 Evolutionary Grooming of Traﬃc in WDM Optical Networks Yong Xu1 and Kunhong Liu2 1

2

School of Medicine, University of Birmingham, Edgbaston, Birmingham, B15 2TT, UK. Email: [email protected] Intelligent Computing Lab, Hefei Institute of Intelligent Machines, Chinese Academy of Sciences, P.O. Box 1130, Hefei Anhui 230031, China. Email: [email protected]

Summary. The widespread deployment of WDM optical networks posts lots of new challenges to network designers. Traﬃc grooming is one of the most common and interesting problems. Eﬃcient grooming of traﬃc can eﬀectively reduce the overall cost of the network. But unfortunately, it has been shown to be NP-hard. Therefore, new heuristics must be devised to tackle them. Among those approaches, metaheuristics are probably the most promising ones. In this chapter, we present a thorough and comprehensive discussion on various metaheuristic approaches to the grooming of traﬃc in both static and dynamic patterns in WDM optical networks. Some future challenges and research directions are also discussed in this chapter.

Keywords: WDM Optical network, Traﬃc grooming, Genetic algorithm, Tabu search, Simulated annealing.

5.1 Introduction Wavelength-division multiplexing (WDM) technology will undoubtedly dominate next generation’s backbone transport networks. In WDM optical networks, hundreds of wavelength channels can be established in a single ﬁber to make the best use of its huge bandwidth. In such a network, a set of lightpaths on the optical layer, which is used for transmitting optical signal between its end nodes, deﬁne the virtual topology and the interconnections of them are implemented via optical cross-connects (OXCs). Each OXC can switch the optical signal coming in on a wavelength of an input ﬁber to the same wavelength in an output ﬁber. Such network architecture makes it ﬂexible to various services and protection requirements. The deployment and design of such WDM optical networks are currently the subjects of intensive research and development eﬀorts. Y. Xu and K. Liu: Evolutionary Grooming of Traﬃc in WDM Optical Networks, Studies in Computational Intelligence (SCI) 66, 95–137 (2007) c Springer-Verlag Berlin Heidelberg 2007 www.springerlink.com

96

Y. Xu and K. Liu

Currently, the typical transmission rate on a wavelength is 10 Gbit/s (OC-192) and higher rate (40 Gbit/s or OC-768) is also commercially available. On the contrary, the requirement of end-users, such as Internet service providers (ISPs), universities, and industries is generally several to a few hundreds megabits, much lower than that rate. The huge gap between these two requires that a number of such low rate traﬃc streams be multiplexed onto high rate wavelengths to make full use of their capacity and at the same time to minimize the network’s cost. However, multiplexing/demultiplexing a traﬃc stream to/from a wavelength requires the wavelength to drop at the two end nodes of it at which an electronic equipment must be used to process such operation. In SONET/WDM networks, for example, a SONET add/drop multiplexer (ADM) is required to electronically combine such lowrate traﬃc streams onto a high-rate circuit (wavelength). Furthermore, such electronic equipment is often very expensive for high-rate traﬃc transformations. Therefore it is beneﬁcial if each wavelength drops only at those nodes onto which the low-rate streams are congregated in order to reduce the cost. This is realized by the addition of a wavelength ADM (WADM) at each node. With the WADM, each wavelength is allowed to either drop at a node or optically bypass it with no electronic equipment requirement at that node for that wavelength. A question that arises in such a system is how to groom low-rate traﬃc streams onto each wavelength so that the amount of electronic equipment needed in the network is minimized while as few wavelengths as possible are used. This problem, often referred to as the traﬃc-grooming (TG) problem, has been proved to be NP-complete [1] and drawn increasing interest in the literature recently [1–6]. The topic of TG has always been one of the popular problems in optical communication ﬁeld since it was ﬁrst proposed in 1998. It has been the main focus in the recent special issue on WDM-Based Network Architectures in IEEE Journal on Selected Areas in Communications [7]. The First and Second Workshop on TG in WDM Networks have been held recently [8]. Comprehensive review and speciﬁc review have also been done in this subject [4,9,10]. Due to the NP-complete property of this problem, research showed that evolutionary approaches are a very good option for traﬃc-grooming problems. In this chapter of the book we will provide a detailed discussion on how evolutionary approaches can be used to various grooming problems. Some future challenges and research directions are also discussed in this chapter. One purpose of this chapter is to provide the researchers and practitioners in evolutionary computation community a new application area and those in traﬃc grooming community a new tool to improve the grooming performance. The rest of this chapter is organized as follows. Section 5.2 gives the basic concept of traﬃc grooming in WDM optical networks. Sections 5.3 and 5.4 give a thorough discussion on evolutionary grooming of both static and dynamic traﬃc patterns. An application of evolutionary approaches based traﬃc grooming technique will be discussed in Section 5.5. Some future

5 Evolutionary Grooming of Traﬃc in WDM Optical Networks

97

challenges and research directions about traﬃc grooming are discussed in Section 5.6. Finally, conclusions are drawn in Section 5.7.

5.2 Traﬃc Grooming in WDM Optical Networks TG is one of the most important traﬃc engineering techniques which is deﬁned as the allocation of sub-wavelength traﬃc tributaries onto full wavelength channels in order to achieve eﬃcient utilization of network resources, e.g., minimizing cost or blocking probability, or maximizing revenue. This is a complicated combinatorial optimization problem and has been shown to be NP-complete even in a single-hub unidirectional ring network with nonuniform traﬃc requests between each pair of nodes [1]. It was ﬁrst introduced in 1998 [11–13] and has aroused much interest from the researchers in both academia and industrial societies due to its high academic and commercial values. It is also one of the most attractive issues in journals related to optical networks. In WDM optical networks, each ﬁber transmits a number of wavelengths simultaneously in a WDM manner and at the same time every wavelength can carry several traﬃc streams in a Time Division Multiplexing (TDM) manner. By using the ADM, low rate traﬃc can be multiplexed onto high rate circuits. 4 OC-3 traﬃc streams, for instance, can be multiplexed onto an OC-12 circuit and 16 OC-3’s onto an OC-48, etc. The number of low rate traﬃc streams a wavelength can accommodate is referred to as traﬃc granularity. Since high speed ADM is very expensive, the use of ADM’s usually dominates the total cost of today’s backbone networks. Hence TG is often used to avoid each wavelength dropping at too many nodes on the network to reduce such cost. The beneﬁt of TG in WDM optical networks can be shown by the simple example given below. Consider a 4-node unidirectional ring. The traﬃc request between each pair of nodes is 8 OC-3’s and the wavelength capacity is OC-48. Thus each wavelength can carry two node pairs’ amount of traﬃc. Since there are 6 pairs of traﬃc on the ring, 3 wavelengths are needed. Fig. 5.1(a) shows the traﬃc assignment with no grooming, where wavelength λ1 accommodates traﬃc requests between node pairs 1 ↔ 2 and 3 ↔ 4, λ2 2 ↔ 3 and 1 ↔ 4, λ3 1 ↔ 3 and 2 ↔ 4. In this conﬁguration, every wavelength drops at every node so that 3 ADM’s are needed at every node. 12 ADM’s are needed on the ring. Figure 5.1(b) gives another conﬁguration with proper grooming. In this scenario, nodes 1, 2, and 4 equip with a WADM, which allows only wavelengths that carry traﬃc to or from that node to be added/dropped at it and bypasses all the other wavelengths. The assignment of traﬃc is as follows: λ1 : 1 ↔ 2 and 1 ↔ 3; λ2 : 2 ↔ 3 and 2 ↔ 4; λ3 : 1 ↔ 4 and 3 ↔ 4. There are also 3 wavelengths on the ring but since each wavelength drops at only three nodes, 9 ADM’s are needed this time, 3 ADM’s are saved compared to the scenario in Fig. 5.1(a).

98

Y. Xu and K. Liu OC-3's SONET ADM SONET ADM

λ1

SONET ADM OC-3's

SONET ADM

λ2 SONET ADM

λ2

SONET ADM

λ3

WDM

λ3

λ1 SONET

1

ADM

λ1, λ2, λ3 4

WDM

2

λ1

WDM

λ3

3

λ1 SONET ADM

WDM

λ2

λ2

SONET ADM

OC-3's

SONET ADM

λ3 SONET ADM

SONET ADM OC-3's

(a) A possible grooming scenario with 12 ADM’s OC-3's SONET ADM

λ1

SONET ADM

OC-3's

SONET ADM

λ2

SONET ADM

λ3

WADM

λ3

λ1

1

λ1, λ2, λ3 4

WADM

2

WADM

SONET ADM

λ2

SONET ADM

OC-3's

3

λ3 SONET ADM

WDM

λ2

λ1 SONET ADM

SONET ADM

OC-3's

(b) Another possible grooming scenario with only 9 ADM’s

Fig. 5.1. A grooming example on a 4-node unidirectional ring with two diﬀerent scenarios

In these two grooming scenarios, the traﬃc is carried by the same number of wavelengths and every wavelength works at their full capacity, but the number of ADM’s required is diﬀerent. This shows that after appropriate grooming, not only can the burden of the electric devices be alleviated but the amount of them can be reduced. In large-scale networks, the reduction of

5 Evolutionary Grooming of Traﬃc in WDM Optical Networks

99

the amount of AMDs due to good grooming will bring about a remarkable saving to the cost of constructing them. Therefore, traﬃc grooming is one of the problems that must be taken into serious account in the design of today’s high-speed telecommunication networks. According to whether traﬃc changes or not, there are basically two grooming categories: static grooming and dynamic grooming. The former is to groom a known and unchanged traﬃc pattern in order to use the minimum number of ADM’s to support it. The latter is to groom a set of changing traﬃc patterns, either known or unknown, in order to use the minimum number of ADM’s to support this changing traﬃc. It was shown that traﬃc grooming is NP-hard even for a unidirectional ring with only one egress node and nonuniform traﬃc. The grooming of arbitrary traﬃc on general network topology is even diﬃcult. Therefore, evolutionary approach is a good candidate for it. The following sections will discuss the evolutionary approaches to the grooming of static and dynamic traﬃc.

5.3 Evolutionary Grooming of Static Traﬃc This section presents some important algorithms for grooming of static traﬃc in WDM networks. Grooming of static traﬃc has been one of the most interested topics in the literature. Although it is more practical to groom dynamic traﬃc, the grooming of static traﬃc can not only serve as an approximate to the grooming of dynamic traﬃc but also provide a basic methodology to it. All the major evolutionary techniques, i.e., genetic algorithm (GA), tabu search (TS), and simulated annealing (SA), have been applied to this kind of grooming. Up to now, the main eﬀorts are devoted to ring and mesh topologies, and we will discuss each of these approaches to this kind of grooming in the following subsections. 5.3.1 Grooming of Static Traﬃc in WDM Ring Networks It is reasonable to groom traﬃc in ring networks due to its widespread use in today’s infrastructural networks. Although many researchers have already worked in this ﬁeld right from the beginning of the TG problem being proposed, and great achievements have been reached [1–3, 5, 6, 10–23], there are still some regions remained to be explored. Today, many researchers still try to develop new heuristics for solving TG problem in WDM ring networks eﬃciently [24, 25]. In this part, we explain the applications of evolutionarybased approaches for the static traﬃc grooming in WDM ring networks. Although there are many diﬀerent approaches, we mainly focuse on two typical approaches: circle construction based two-step model and directly traﬃc matrix based operation.

100

Y. Xu and K. Liu

5.3.1.1 Circle Construction Based Two-Step Model The circle construction based model was ﬁrst proposed in [2]. In this model, all the traﬃc connections are ﬁrst divided into the basic traﬃc rate, and then as few circles as possible are constructed to include all the requested connections, where each circle consists of multiple nonoverlapping (i.e., link disjoint) connections. The circle construction heuristic is shown in Fig. 5.2. In this algorithm, it searches the traﬃc matrix from the beginning to the end, and then selects the traﬃc requests in the order of the longest to the shortest stride to construct circles. The connection will be added into the circle under the condition that at least one end node of it already existing in the circle, otherwise, it will be put into a so-called GapMaker list which is initially empty. When this process ends, a greedy algorithm is employed to check the GapMaker list to construct as few circles as possible with minimum number of end nodes. For example, given a 5-node ring network with a traﬃc matrix shown in Fig. 5.3(a). Let i and j denote the source and the destination node of a traﬃc request and rij represent the traﬃc request from node i to node j, a possible circle may be constructed by the traﬃc stream 0-4 and 4-0. But as rij is not necessarily equal to rji , only ﬁve circles can be constructed for r04 and r40 , and two basic traﬃc requests belonging to r40 will be put into the GapMaker list. After the circles have been constructed, another heuristic will be used to groom circles onto wavelengths with overlapping as many end nodes belonging to diﬀerent circles as possible so as to reduce the number of ADM’s. If the granularity of a wavelength is 8, eight circles can be groomed onto a wavelength. This model was widely applied to the grooming of static traﬃc in ring networks and two SA based and one GA based approaches were proposed based on it in [3, 15, 20] respectively. For the purpose of comparison, only the GA approach is shown below. The GA approach proposed in [15] aimed to minimize the total number of ADM’s. The operation of the GA can be illustrated as follows. A binary coding method is used and each bit xi w represents an ADM needed on wavelength w at node i if it equals “1” or indicates no ADM necessary if it is “0”. The ﬁtness function can be presented as minimizing: W N w w=1 i=1 xi (where W and N are the total number of wavelengths and nodes in the ring respectively). The traﬃc request is ﬁrst divided into basic elements (the lowest speed rate), then the circle construction heuristics is applied to construct C basic circles based on the traﬃc request. Finally a GA is employed to grooming circles onto wavelengths. A seed obtained from the greedy circle construction grooming algorithm shown in Fig. 5.2 is inserted into the initial population and the rest chromosomes are generated by random permutation of the C circles. In decoding process, two methods, next-ﬁt (NF) and best-ﬁt (BF), are applied to assign circles to wavelengths in accordance with the order of them in the chromosome.

5 Evolutionary Grooming of Traﬃc in WDM Optical Networks

101

Fig. 5.2. Circle construction heuristics (source: reference [2])

NF algorithm distributes circles evenly among all wavelengths but BF tries to fully pack every wavelength except possibly the last one. The corresponding algorithms are shown in Fig. 5.4. An exponential ranking selection scheme is adopted with the probability pi = (r −1)∗ rM −i /(rM −1) for the ith individual to be selected in M total individuals, where r is set to be 0.99. Two crossovers, partially mapped crossover (PMX) and order crossover (OX), are used to produce oﬀspring. In mutation, the elements of an individual are right-shifted by a random number of positions in a cyclic manner.

102

Y. Xu and K. Liu n 0 1 2 3 4 0 1 2 3 4

0 4 1 2 7

5 0 3 3 2

3 1 0 6 4

(a)

1 3 5 0 3

5 4 3 5 0

n 0 1 2 3 4

n 0 1 2 3 4

0 1 2 3 4

0 1 2 3 4

0 3 5 6 3

4 0 4 2 7

6 3 0 5 1

(b)

6 3 3 0 2

2 4 4 3 0

0 4 5 6 7

5 0 4 3 7

6 3 0 6 4

6 3 5 0 3

5 4 4 5 0

(c)

Fig. 5.3. Three sample traﬃc matrices in a WDM ring network with 5 nodes (source: reference [19])

Fig. 5.4. NF and BF algorithms (source: reference [15])

5.3.1.2 Direct Traﬃc Matrix Based Grooming Method The GA based circle construction model may not be very eﬃcient for large scale networks. Therefore, another GA-based approach directly operating on

5 Evolutionary Grooming of Traﬃc in WDM Optical Networks

103

traﬃc matrix and grooming the traﬃc in a comprehensive way was proposed in [16]. The framework of the GA is described below. First, the traﬃc matrix is converted into a vector V, and a random permutation of the traﬃc items is used as a chromosome, V = (x0 , . . ., xk , . . ., xn(n−1) − 1) if there are n nodes in a ring. The position of each item is used to determine the decoding order. The connection between the component’s subscript k of the vector V and those of the traﬃc matrix can be formulated as: ⎧ k = (2n − i − 1) × i/2 + j − i, 0 ≤ i < n − 1, i + 1 ≤ j < n ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ or 0 ≤ k < n(n − 1)/2 (5.1) n (n − 1) + i (i − 1) ⎪ k= + j, 1 ≤ i < n, 0 ≤ j < i ⎪ ⎪ ⎪ 2 ⎪ ⎩ or n(n − 1)/2 ≤ k < n(n − 1) The algorithm in Fig. 5.5 can be used to calculate i, j from k: The basic ideas behind the decoding algorithm proposed in this approach can be described as follows. If a wavelength drops at π nodes, it can at most accommodate π(π − 1) traﬃc requests. If all the π(π − 1) traﬃc requests are assigned to the wavelength, the numbers of ADM’s and wavelengths used on the ring must be minimized. Based on this observation, a ﬁrst-ﬁt heuristic incorporated with a greedy local improvement was proposed to decode each chromosome, as is shown in Fig. 5.6. The algorithm assigns the ﬁrst encountered traﬃc items onto each wavelength until one link load on the wavelength exceeds its capacity. Then the local improvement algorithm is called to assign such a traﬃc item to it with no additional node added. The objective of this algorithm is to minimize the number of ADM’s. Therefore, the fewer ADM’s a chromosome uses, the ﬁtter it is. If two chromosomes require the same number of ADM’s, the one with smaller number of wavelengths is assigned higher ﬁtness value. The order-mapped crossover (OMX) proposed in [26] is used to produce oﬀspring and the inversion of a section of some randomly selected genes is used as mutation in this algorithm. if(k

Our partners will collect data and use cookies for ad personalization and measurement. Learn how we and our ad partner Google, collect and use data. Agree & close