Natural Niching for Evolving Cooperative Classifiers

Abstract

An evolutionary classifier, such as a learning classifier system (LCS) or a genetic programming boolean concept learner, must maintain a population of diverse rules that together solve a problem (e.g., classify examples). To maintain cooperative diversity while applying a selection operator to the population of rules, as in the Michigan-style LCS, the evolutionary algorithm must incorporate some kind of niching mechanism. The way to accomplish niching in an LCS is to force competing rules to share resources (i.e., rewards). The implicit or natural niching and speciation induced by such resource sharing, is shown to be robust in the face of severe selective pressure, low population sizes, and overlapping rule coverage. Specifically in this paper we analyze the two-niche (two competing/cooperating rules) case. We find closed form approximations for niche maintenance and niche convergence times, giving us the beginnings of a first predictive model for interacting (cooperating) rules in an evolving population. Finally, we make the case for niching/speciation as a basic, indirect form of cooperation that is fundamental to, and underlying, all other types of more direct cooperation, and which the LCS must therefore promote. Although we focus on the LCS as an example of a specific and well-known evolutionary classifier, all of our results are general enough to apply to any evolutionary algorithm, such as genetic programming (GP), that applies selection to a population of diverse classifiers.