Toward a better understanding of rule initialisation and deletion

Kovacs, Tim; Bull, Larry

Abstract

A number of heuristics have been used in Learning Classifier Systems to initialise parameters of new rules, to adjust fitness of parent rules when they generate offspring, and to select rules for deletion. Some have not been studied in the literature before. We study the interaction of these heuristics in an attempt to improve performance and detect any unnecessary methods. We evaluate the two published methods for initialisation of new rules in XCS and find the one based on parental values results in better evolutionary search but larger population sizes than the one based on population means. In preliminary work we demonstrate that when the difficulty of the 6 multiplexer is increased by reducing the population size limit and turning off subsumption we can improve performance by discounting the fitness of both parents and children. Copyright 2007 ACM.

• Presentation Conference Type: Conference Paper (published)
• Conference Name: Proceedings of GECCO 2007: Genetic and Evolutionary Computation Conference, Companion Material
• Publication Date: Aug 27, 2007
• Peer Reviewed: Peer Reviewed
• Pages: 2777-2780
• Book Title: Proceedings of the 2007 GECCO Conference on Genetic and Evolutionary Computation
• DOI: https://doi.org/10.1145/1274000.1274060
• Keywords: learning classifier systems, parent rules, heuristics
• Public URL: https://uwe-repository.worktribe.com/output/1026424
• Additional Information: Title of Conference or Conference Proceedings : 2007 GECCO Conference