Lookahead and latent learning in a simple accuracy-based classifier system

Abstract

Learning Classifier Systems use evolutionary algorithms to facilitate rulediscovery, where rule fitness is traditionally payoff based and assigned under a sharing scheme. Most current research has shifted to the use of an accuracy-based scheme where fitness is based on a rule's ability to predict the expected payoff from its use. Learning Classifier Systems which build anticipations of the expected states following their actions are also a focus of current research. This paper presents a simple but effective learning classifier system of this last type, using accuracy-based fitness, with the aim of enabling the exploration of their basic principles, i.e., in isolation from the many other mechanisms they usually contain. The system is described and modelled, before being implemented. © Springer-Verlag 2004.