Implementierung eines Klassifizierungssystems
Author: | Marcus Hutter (1991) |
Comments: | 72 pages with listing, in German |
Keywords: | Artificial Intelligence; Genetic Algorithms; Classifier System |
Reference: | Diplomarbeit, TU-München, Lehrstuhl für KI, 1991 |
Paper: | PostScript (485kb) or
PDF (330kb) |
C-Source: | cfssim.c (49kb) -
cfsexmple.c (7kb) |
Note: | Contains the first proof (in Section 2.2.4) of the equivalence of ranking
and tournament selection (defined in Section 2.1.4). |
Abstract: A classifier system is a massively parallel rule based system,
whose components (classifier) can exchange messages, whose behavior is
is assessed by a teacher (reinforcement), and which is able to learn by
means of credit assignment and a genetic algorithm. For an introduction
we have to refer to the, meanwhile extensive, literature; see especially
Goldberg (1989). The concept of a classifier system was first developed
by Holland (1986), but meanwhile a multitude of variants and extensions
exist (Booker et. al, 1989). So far it is impossible to
compare these variants in their performance, statements on the
quality of the various approaches are, hence, hard to impossible.
The program developed in this diploma thesis allows, for the first time,
a direct comparison of the most important variants.
The thesis describes the program, in which we have taken special attention
to an efficient implementation.
Zusammenfassung: Ein Klassifizierungssystem (CFS, engl. Classifiersystem)
ist ein massiv paralleles regelbasiertes System, dessen Komponenten (Classifier)
Nachrichten (Messages) austauschen können, dessen
Verhalten von einem Lehrer beurteilt wird (Reinforcement) und
das mittels Credit-Assignment und genetischen Algorithmen fähig
ist zu lernen. Für eine einführende Darstellung muß
auf die inzwischen sehr umfangreiche Literatur, insbesondere Goldberg
(1989), verwiesen werden. Das Konzept des CFS wurde zuerst von
Holland (1986) entwickelt, inzwischen gibt es aber
eine Vielzahl von Varianten und Erweiterungen (Booker et. al,
1989). Bisher ist es nicht möglich, diese Varianten in ihrer
Performance zu vergleichen, eine Aussage über die Güte der
verschiedenen Ansätze ist somit kaum oder überhaupt nicht möglich.
Das in dieser Diplomarbeit erstellte Programm gestattet erstmals bzgl.
der wichtigsten Varianten einen direkten Vergleich. In den
folgenden Kapiteln wird dieses Programm, bei dem besonders auf eine
effiziente Implementierung geachtet wurde, beschrieben.
BibTeX Entry
@MastersThesis{Hutter:92cfs,
author = "Marcus Hutter",
school = "Theoretische Informatik, TU M{\"u}nchen",
title = "{I}mplementierung eines {K}lassifizierungs-{S}ystems",
year = "1991",
pages = "72 pages with C listing, in German",
url = "http://www.hutter1.net/ai/pcfs.htm",
ps = "http://www.hutter1.net/ai/pcfs.ps",
pdf = "http://www.hutter1.net/ai/pcfs.pdf",
abstract = "A classifier system is a massively parallel rule based system,
whose components (classifier) can exchange messages, whose behavior is
is assessed by a teacher (reinforcement), and which is able to learn by
means of credit assignment and a genetic algorithm. For an introduction
we have to refer to the, meanwhile extensive, literature; see especially
Goldberg (1989). The concept of a classifier system was first developed
by Holland (1986), but meanwhile a multitude of variants and extensions
exist (Booker et. al, 1989). So far it is impossible to
compare these variants in their performance, statements on the
quality of the various approaches are, hence, hard to impossible.
The program developed in this diploma thesis allows, for the first time,
a direct comparison of the most important variants.
The thesis describes the program, in which we have taken special attention
to an efficient implementation.",
note = "72 pages with C listing, in German",
}