Towards the Development of Cognitive Maps in Classifier Systems

Abstract

Classifier systems are well tested vehicles for implementing geneticalgorithms in machine learning environments. This paper presentsa novel system architecture that transforms a classifier system'sknowledge representation from message-based structures to self-organizingneural networks. These networks have been integrated with a classifiersystem to produce a Hybrid Learning System (HLS) that exhibits adaptivebehaviour when driven by low level environmental feedback. Problemsare represented within HLS as objects characterized by environmentalfeatures. Objects controlled by the system have preset goals setagainst a subset of their features and the system has to achievethese goals by developing a behavioural repertoire that efficientlyexplores and exploits the problem environment. Three types of knowledgestructures evolve during this adaptive process: a cognitive map ofuseful regularities within the environment (encoded in a self-organizingnetwork); classifier behaviour calibrated against feature statesand targets (encoded in a set of self-organizing feature maps); apopulation of complex behaviours (evolved from a gene pool suppliedas part of the initial problem specification).