Éditeur(s) : HAL CCSD
Résumé : National audience
EIAH Proceedings

hal-00634758
https://hal.archives-ouvertes.fr/hal-00634758

Éditeur(s) : HAL CCSD ACM
Résumé : International audience
In this study, we use an adapted version of Pitsburgh-like learning classifier system to perform over classification tasks. The Adapted Pittsburgh Classifier System, enhanced with a new mechanism, allows us to consider the classification problems and their treatment by a given LCS in a different manner. Our aim is to exhibit elements in order to prove that, using the action covering mechanism, this system is able to build an inner map of a given classification learning sample. In the context of this paper, this map is built using an intrisic property of Pittsburgh-like CS: the use of various collections of classifiers amongst a unique population.
Proceedings of the 12th annual conference comp on Genetic and evolutionary computation
Portland,Oregon, United States

hal-00520609
https://hal.archives-ouvertes.fr/hal-00520609
DOI : 10.1145/1830761.1830823

Éditeur(s) : HAL CCSD ACM
Résumé : International audience
In 2008 a scientific paper written by Butz and al. investigated Matching in Learning Classifier Systems. Matching represents at least 65% of the computational time when executing a classifier system as reported in Llora and al. We propose to modify that encoding using a fast and accurate matching replacing standard matching algorithm for character-based classifier systems.
Proceedings of the 12th annual conference comp on Genetic and evolutionary computation
Portland, Oregon, United States

hal-00520610
https://hal.archives-ouvertes.fr/hal-00520610
DOI : 10.1145/1830761.1830820

Éditeur(s) : HAL CCSD ACM New York, NY, USA
Résumé : International audience
This paper focuses on the study of the behavior of a genetic algorithm based classifier system, the Adapted Pittsburgh Classifier System (A.P.C.S), on maze type environments containing aliasing squares. This type of environment is often used in reinforcement learning litterature to assess the performances of learning methods when facing problems containing non markov situations. Through this study, we discuss on the performances of the APCS on maze type environments and also of the efficiency of an improvement of the APCS learning method inspired from the XCS : the covering mechanism. We manage to show that, without any memory mechanism, the APCS is able to build and to keep accurate strategies to produce regular sub-optimal solutions to these maze problem. This statement is shown through a comparison of the results obtained by the XCS, XCSM and XCSMH on distinct maze problems with these obtained by the APCS.
Proceedings of the 2008 GECCO conference companion on Genetic and evolutionary computation
Atlanta, GA, United States

hal-00520607
https://hal.archives-ouvertes.fr/hal-00520607
DOI : 10.1145/1388969.1389013

Éditeur(s) : HAL CCSD
Résumé : J'étais à l'époque dans le laboratoire I3S de Sophia-Antipolis. Je suis dorénavant au laboratoire GRIMAAG en Guadeloupe.
We focused on the communication problem within multi-agents systems, inspirited by nature and more precisely by the ants that are a perfect example of living integrating learning (ontogenesis) to genetic inheritance (phylogenesis): the Baldwin Effect.
We worked with agents represented by classifiers systems using genetic algorithm. We introduced a new fixed knowledge based classifiers system, the simplified Pittsburgh classifier system, which we studied in a spontaneous coordination multi-agents problem. Then, we successfully allowed “homogeneous” agents to communicate their “knowledge” using distributed elitism.
Finally, “heterogeneous” agents efficiently exchanged information thanks to a minimal communication model we defined, bound to reality and extended.
Nous nous sommes intéressés à la problématique de la communication au sein des systèmes multi-agents en nous inspirant de la nature et plus particulièrement des fourmis qui sont un parfait exemple de l'intégration de l'apprentissage du vivant (ontogenèse) au capital génétique de celui-ci (phylogénie) : l'effet Baldwin.
Nous avons choisi de travailler avec des systèmes de classeurs à algorithme génétique pour représenter les agents. Nous avons introduit un nouveau type de système de classeurs à base de connaissance de taille fixe, les systèmes de classeurs de type Pittsburgh simplifié, dont nous avons étudié les propriétés dans un contexte multi-agents de coordination spontanée.
Nous avons ensuite proposé avec succès aux agents "homogènes" de communiquer pour échanger leur "connaissance" à l'aide de l'élitisme distribué.
Enfin, les agents "hétérogènes" ont échangé efficacement des informations grâce à un modèle minimal de communication que nous avons ancré dans la réalité puis étendu.
https://tel.archives-ouvertes.fr/tel-00192372

tel-00192372
https://tel.archives-ouvertes.fr/tel-00192372
https://tel.archives-ouvertes.fr/tel-00192372/document
https://tel.archives-ouvertes.fr/tel-00192372/file/These.pdf

Éditeur(s) : HAL CCSD
Résumé : International audience
This paper focuses on the study of the influence of a newly implemented mechanism on a Pittsburgh-like classifier system. The Adapted Pittsburgh Classifier System is a learning classifier system that uses genetic algorithms to evolve its ruleset. The new mechanism discussed is inspired from Wilson work on the eXtended Classifier System (XCS): it allows the concerned LCS to adapt its rule set when facing a new signal by modifying an existing rule. This mechanism is called covering mechanism due to the fact that the rule is covered up by a new rule which is sensitive to this new signal. Effects of this covering mechanism are first measured on the performances of APCS on well known multi-step environments: maze-type environments (Woods 101 and E2). In addition, further measures presented in this paper indicate that we could possibly rely on the behavior of this covering mechanism to automate the research of a correct cognitive capacity needed by the APCS to solve a given multistep problem.
2008 Eighth International Conference on Hybrid Intelligent Systems
Barcelone, Spain

hal-00520608
https://hal.archives-ouvertes.fr/hal-00520608
DOI : 10.1109/HIS.2008.105

Éditeur(s) : HAL CCSD springer
Résumé : International audience
This paper focuses on the study of the behavior of a genetic algorithm based classier system, the Adapted Pittsburgh Classier System (A.P.C.S), on maze type environments containing aliasing squares. This type of environment is often used in reinforcement learning literature to assess the performances of learning methods when facing problems containing non markovian situations. Through this study, we discuss on the performance of the APCS upon two mazes (Woods 101 and Maze E2) and also on the eciency of an improvement of the APCS learning method inspired from the XCS: the covering mechanism. We manage to show that, without any memory mechanism, the APCS is able to build and to keep accurate strategies to produce regular sub-optimal solution to these maze problems. This statement is shown through a comparison between the results obtained by the XCS on two specic maze problems and those obtained by the APCS.
Lecture notes in computer science

hal-00542922
https://hal.archives-ouvertes.fr/hal-00542922
DOI : 10.1007/978-3-642-17508-4_8

Éditeur(s) : HAL CCSD
Résumé : International audience
This paper focuses on the study of the behaviour of a singular classifier system, the Adapted Pittsburgh Classifier System (A.P.C.S), on environments containing aliasing situations. Maze type environments are often used in reinforcement learning literature to assess the performances of learning methods when facing problems containing non Markovian situations. Those situations are often encountered when performing reinforcement learning on aliased data samples issued from meteorological simulations.
International Journal of Artificial Intelligence

hal-00520611
https://hal.archives-ouvertes.fr/hal-00520611