Skip to main content

XCS's Strength-Based Twin. Part I

Tim Kovacs, XCS's Strength-Based Twin. Part I. Chapter in Learning Classifier Systems. ISBN 3-540-20544-6, pp. 61–80. January 2003. PDF, 474 Kbytes.

Abstract

Wilson's XCS has rapidly become the most popular classifier system of all time, and is a major focus of current research. XCS's primary distinguishing feature is that it bases rule fitness on the accuracy with which rules predict reward, rather than the magnitude of the reward predicted (as traditional, strength-based systems do).

XCS is a complex system and differs from other systems in a number of ways. In order to isolate the source of XCS's adaptive power, and, in particular to study the difference between strength and accuracy-based fitness, we introduce a system called Strength-Based XCS (SB--XCS), which is as similar to the accuracy-based XCS as we could make it, apart from being strength-based.

This work provides a specification of SB--XCS and initial results for it and XCS on the 6 multiplexer and woods2 tasks. It then analyses the solutions found by the two systems and finds that each prefers a particular type of solution. A sequel paper provides further analysis.

Bibtex entry.

Contact details

Publication Admin