An Evolutionary Approach for Generating a Learning Classifier System Reward Policy : Review and Prospects

In this paper we review the evolutionary approach we proposed in previously published papers, regarding the emergence of a Learning Classifier System (LCS) reward policy. The idea behind our approach is to induce the emergence of a LCS reward policy, through the evolution of a population of LCS based agents. The present review intends to shed light on some aspects that were not sufficiently emphasized in previous papers and, on other hand, to prospect future work regarding this approach. First, we describe a simple, but generic architecture of an evolutive LCS based agent. The couple of modules constituting the architecture are a (LCS based) control model, generating the agent behaviour, and a biological model regulating the biological aspects of the agent life. Second, we perform an analysis of the factors influencing the outcome of reward policy evolution, like the reward regimes to adopt, or the genetic operators that one should use. Finally, we evaluate the requirements to extend our approach to Special Classifier Systems (XCS) based evolutive agents.