Mellouk A., Chebira A. (eds.) Machine Learning
Подождите немного. Документ загружается.
Machine Learning
Machine Learning
Edited by
Abdelhamid Mellouk
and
Abdennacer Chebira
I-Tech
IV
Published by In-Teh
In-Teh is Croatian branch of I-Tech Education and Publishing KG, Vienna, Austria.
Abstracting and non-profit use of the material is permitted with credit to the source. Statements and
opinions expressed in the chapters are these of the individual contributors and not necessarily those of
the editors or publisher. No responsibility is accepted for the accuracy of information contained in the
published articles. Publisher assumes no responsibility liability for any damage or injury to persons or
property arising out of the use of any materials, instructions, methods or ideas contained inside. After
this work has been published by the In-Teh, authors have the right to republish it, in whole or part, in
any publication of which they are an author or editor, and the make other personal use of the work.
© 2009 In-teh
www.in-teh.org
Additional copies can be obtained from:
publication@ars-journal.com
First published February 2009
Printed in Croatia
p. cm.
ISBN 978-3-902613-56-1
1. Machine Learning, Abdelhamid Mellouk and Abdennacer Chebira
Preface
Machine Learning is often referred to as a branch of artificial intelligence which deals
with the design and the development of algorithms and techniques that help machines to
“learn”. Hence, it is closely related to various scientific domains as Optimization, Vision,
Robotic and Control, Theoretical Computer Science, etc.
Based on this, Machine Learning can be defined in various ways related to a scientific
domain concerned with the design and development of theoretical and implementation
tools that allow building systems with some Human Like intelligent behavior. Machine
learning addresses more specifically the ability to improve automatically through
experience.
This book brings together many different aspects of the current research on several
fields associated to Machine Learning. The selection of the chapters for this book was done
in respect to the fact that it comprises a cross-edition of topics that reflect variety of
perspectives and disciplinary backgrounds.
Four main parts have been defined and allow gathering the 21 chapters around the
following topics: machine learning approaches, Human-like behavior and machine Human
interaction, supervised and unsupervised learning approaches, reinforcement learning
approaches and their applications.
This book starts with a first set of chapters which addresses general approaches in
Machine Learning fields. One can find discussion about various issues: how to use the
paradigm divide and conquer to build a hybrid self organized neural network tree structure,
how to move from automation to autonomy, how to take experience to a whole new level,
how to design very large scale networks based on Hamiltonian neural networks, how to
design classifiers generative with similarity based abilities, and how information-theoretic
competitive learning can force networks to increase knowledge.
In addition, the second part addresses the problem of Human-like behavior and
machine Human interaction. It contains five chapters that deal with the following scope:
Human-Knowledge poor-process of ontological information extraction, Machine learning
for spoken dialogue system optimization, Bayesian additive regression trees applied to mail
phishing detection, Composition of web services under multiple criteria and supervised
learning problems under the ranking framework.
Another set of chapters presents an overview and challenges in several areas of
supervised and unsupervised learning approaches. Subjects deal with generation method
for a person specific facial expression map, linear subspace methods in the context of
automatic facial expression analysis, nearest neighbor re-sampling method for prognostic
gene expression patterns of tumor patients, 3D shape classification and retrieval algorithm
and classification of faults in electrical power systems using a hybrid model based on neural
networks.
The last part of the book deals with reinforcement learning approaches used in Machine
Learning area. Various techniques are developed: Genetic learning programming and Sarsa
VI
learning allow the selection of appropriate stock trading rules in financial area, convergence
of the online value-iteration in dynamic programming techniques is given in the case of the
optimal control problem for general nonlinear discrete-time systems, modular reinforcement
learning with situation-sensitive ability is used for intention estimation, experience replay
technique is applied to real-words application and finally sequential modeling and
prediction allow an adaptive intrusion detection in computer system.
This book shows that Machine Learning is a very dynamic area in terms of theory and
application. The field of Machine Learning has been growing rapidly, producing a wide
variety of learning algorithms for different applications. The ultimate value of those
algorithms is to a great extent judged by their success in solving real-world problems. There
is also a very extensive literature on Machine Learning, and to give a complete bibliography
and a historical account of the research that led to the present form would have been
impossible. It is thus inevitable that some topics have been treated in less detail than others.
The choices made reflect on one hand personal taste and expertise and on second hand a
preference for a very promising research and recent developments in Machine Learning
fields.
Finally, we would to thank all contributors to this book for their research and effort.We
hope you enjoy reading this book and get many helpful ideas and overviews for your own
study.
Editors
Abdelhamid Mellouk
IUT Creteil/Vitry,
LiSSi Laboratory,
University of Paris 12
France
Abdennacer Chebira
IUT Senart/Fontainebleau,
LiSSi Laboratory,
University of Paris 12
France
Contents
Preface V
1. Neural Machine Learning Approaches: Q-Learning
and Complexity Estimation Based Information Processing System
001
Abdennasser Chebira, Abdelhamid Mellouk, Kurosh Madani and Said Hoceini
2. From Automation To Autonomy 039
Kentarou Kurashige, Yukiko Onoue and Toshio Fukuda
3. Taking Experience to a Whole New Level 053
Luis Ignacio Lopera
4. Hamiltonian Neural Networks Based Networks for Learning 075
Wieslaw Sienko and Wieslaw Citko
5. Similarity Discriminant Analysis 093
Luca Cazzanti
6. Forced Information for Information-Theoretic Competitive Learning 125
Ryotaro Kamimura
7. Learning to Build a Semantic Thesaurus from Free Text Corpora
without External Help
145
Katia Lida Kermanidis
8. Machine Learning Methods for Spoken Dialogue Simulation
and Optimization
167
Olivier Pietquin
9. Hardening Email Security via Bayesian Additive Regression Trees 185
Saeed Abu-Nimeh, Dario Nappa, Xinlei Wang and Suku Nair
10. Learning Optimal Web Service Selections in Dynamic Environments
when Many Quality-of-Service Criteria Matter
207
Stéphane Dehousse, Stéphane Faulkner, Caroline Herssens,
Ivan J. Jureta and Marcos Saerens
VIII
11. Model Selection for Ranking SVM Using Regularization Path 231
Karina Zapien, Gilles Gasso, Thomas Gärtner and Stéphane Canu
12. Generation of Facial Expression Map using Supervised
and Unsupervised Learning
245
Masaki Ishii, Kazuhito Sato, Hirokazu Madokoro and Makoto Nishida
13. Linear Subspace Learning for Facial Expression Analysis 259
Caifeng Shan
14. Resampling Methods for Unsupervised Learning from Sample Data 289
Ulrich Möller
15. 3D Shape Classification and Retrieval Using Heterogenous Features
and Supervised Learning
305
Hamid Laga
16. Performance Analysis of Hybrid Non-Supervised
& Supervised Learning Techniques Applied
to the Classification of Faults in Energy Transport Systems
325
Jhon Albeiro Calderón, Germán Zapata Madrigal
and Demetrio A. Ovalle Carranza
17. Genetic Network Programming with Reinforcement Learning
and Its Application to Creating Stock Trading Rules
345
Yan Chen, Shingo Mabu and Kotaro Hirasawa
18. Heuristic Dynamic Programming Nonlinear Optimal Controller 361
Asma Al-tamimi, Murad Abu-Khalaf and Frank Lewis
19. Implicit Estimation of Another’s Intention Based
on Modular Reinforcement Learning
381
Tadahiro Taniguchi, Kenji Ogawa
and Tetsuo Sawaragi
20. Machine Learning for Sequential Behavior Modeling and Prediction 401
Xin Xu
1
Neural Machine Learning Approaches:
Q-Learning and Complexity Estimation Based
Information Processing System
Abdennasser Chebira, Abdelhamid Mellouk,
Kurosh Madani and Said Hoceini
LISSI laboratory, University Paris 12-Val de Marne
France
1. Introduction
Real world dilemmas, and especially industry related ones, are set apart from academic ones
from several basic points of views. The difference appears since definition of the “problem’s
solution” notion. In fact, academic (called also sometime theoretical) approach often begins
by problem’s constraints simplification in order to obtain a “solvable” model (here, solvable
model means a set of mathematically solvable relations or equations describing a behavior,
phenomena, etc…) (Madani, 2008). If the theoretical consideration is a mandatory step to
study a given problem’s solvability, for a very large number of real world dilemmas, it
doesn’t lead to a solvable or realistic solution. Difficulty could be related to several issues
among which:
- large number of parameters to be taken into account (influencing the behavior) making
conventional mathematical tools inefficient,
- strong nonlinearity of the system (or behavior), leading to unsolvable equations,
- partial or total inaccessibility of system’s relevant features, making the model
insignificant,
- subjective nature of relevant features, parameters or data, making the processing of
such data or parameters difficult in the frame of conventional quantification,
- necessity of expert’s knowledge, or heuristic information consideration,
- imprecise information or data leakage.
Examples illustrating the above-mentioned difficulties are numerous and may concern
various areas of real world or industrial applications. As first example, one can emphasize
difficulties related to economical and financial modeling and prediction, where the large
number of parameters, on the one hand, and human related factors, on the other hand, make
related real world problems among the most difficult to solve. Another illustrative example
concerns the delicate class of dilemmas dealing with complex data’s and multifaceted
information’s processing, especially when processed information (representing patterns,
signals, images, etc.) are strongly noisy or involve deficient data. In fact, real world and
industrial applications, comprising system identification, industrial processes control,
systems and plants safety, manufacturing regulation and optimization, pattern recognition,
communication networks (complex routing, large communication networks management
Machine Learning
2
and optimization, etc.) (Mellouk, 2008a), are often those belonging to such class of
dilemmas.
If much is still to discover about how the animal’s brain trains and self-organizes itself in
order to process so various and so complex information, a number of recent advances in
“neurobiology” allow already highlighting some of key mechanisms of this marvels
machine. Among them one can emphasizes brain’s “modular” structure and its “self-
organizing” capabilities. In fact, if our simple and inappropriate binary technology remains
too primitive to achieve the processing ability of these marvels mechanisms, a number of
those highlighted points could already be sources of inspiration for designing new machine
learning approaches leading to higher levels of artificial systems’ intelligence (Madani, 2007).
In this chapter, we deal with machine learning based modular approaches which could offer
powerful solutions to overcome processing difficulties in the aforementioned frame. If the
machine learning capability provides processing system’s adaptability and offers an
appealing alternative for fashioning the processing technique adequacy, the modularity may
result on a substantial reduction of treatment’s complexity. In fact, the modularity issued
complexity reduction may be obtained from several instances: it may result from
distribution of computational effort on several modules; it can emerge from cooperative or
concurrent contribution of several processing modules in handling a same task; it may drop
from the modules’ complementary contribution (e.g. specialization of a module on treating a
given task to be performed).
A number of works dealing with modular computing and issued architectures have been
proposed since 1990. Most of them associate a set of Artificial Neural Networks (ANN) in a
modular structure in order to process a complex task by dividing it into several simpler sub-
tasks. One can mention active learning approaches (Fahlman & Lebiere, 1990), neural
networks ensemble concept proposed by (Hanibal, 1993), intelligent hybrid systems (Krogh
& Vedelsby, 1995), Mixture of experts concept proposed by (Bruske & Sommer, 1995) and
(Sung & Niyogi, 1995) or structures based on dynamic cells (Lang & Witbrock, 1998). In the
same years, a number of authors proposed multi-modeling concept for nonlinear systems
modeling, where a set of simple models is used to sculpt a complex behaviour
(Goonnatilake & Khebbal, 1996), (Mayoubi et al., 1995), (Murray-Smith & Johansen, 1997),
(Ernst, 1998)) in order to avoid difficulties (modeling complexity). However, it is important
to remind that the most of proposed works (except those described in the four latest
references) remain essentially theoretical and if a relatively consequent number of different
structures have been proposed, a very few of them have been applied to real-world
dilemmas solution.
The present chapter focuses those machine learning based modular approaches which take
advantage either from modules’ independence (multi-agent approach) or from self-
organizing multi-modeling ("divide and conquer" paradigm). In other words, we will
expound online and self-organizing approaches which are used when no a priori learning
information is available. Within this frame, we will present, detail and discuss two
challenging applicative aspects: the first one dealing with routing optimization in high
speed communication networks and the other with complex information processing.
Concerning the network routing optimization problem, we will describe and evaluate an
adaptive online machine learning based approach, combining multi-agent based modularity
and neural network based reinforcement learning ((Mellouk, 2007), (Mellouk, 2008b)). On
the side of complex information processing, we will describe and evaluate a self-organizing