Using XCS to describe continuous-valued problem spaces

Wyatt, David; Bull, Larry; Parmee, Ian

Abstract

Learning classifier systems have previously been shown to have some application in single-step tasks. This paper extends work in the area by applying the classifier system to progressively more complex multi-modal test environments, each with typical search space characteristics, convex/non-convex regions of high performance and complex interplay between variables. In particular, two test environments are used to investigate the effects of different degrees of feature sampling, parameter sensitivity, training set size and rule subsumption. Results show that XCSR is able to deduce the characteristics of such problem spaces to a suitable level of accuracy. This paper provides a foundation for the possible use of XCS as an exploratory tool that can provide information from conceptual design spaces enabling a designer to identify the best direction for further investigation as well as a better representation of their design problem through redefinition and reformulation of the design space.

• Presentation Conference Type: Conference Paper (published)
• Conference Name: IWLCS 2003, IWLCS 2004, IWLCS 2005: Learning Classifier Systems
• Publication Date: Jan 1, 2007
• Journal: Learning Classifier Systems
• Publisher: Springer Verlag
• Peer Reviewed: Peer Reviewed
• Pages: 308-332
• Series Title: Lecture Notes in Computer Science
• Series Number: 4399
• Series ISSN: 0302-9743
• ISBN: 9783540712305
• DOI: https://doi.org/10.1007/978-3-540-71231-2_21
• Keywords: artificial intelligence, computation, abstract devices, XCS
• Public URL: https://uwe-repository.worktribe.com/output/1034871