Variable Default Hierarchy Separation in a Classifier System

A learning classifier system (LCS) is a machine learning system that incorporates a production-system framework and a genetic algorithm (GA) for rule discovery (Goldberg, 1989; Holland, 1975). A primary feature of LCSs is their potential to exploit overlapping sets of rules called default hierarchies. Default hierarchies increase rule set parsimony, enlarge the solution set, and lend themselves to graceful refinement by the GA (Holland, Holy oak, Nisbett, & Thagard, 1986). Traditionally, auction-based, specificity-biased credit allocation (CA) and conflict resolution (CR) schemes have been used to encourage default hierarchy formation in an LCS. Analyses presented in this paper suggest that these schemes cannot be expected to perform adequately in arbitrary LCS environments. This paper presents an alternate CA/CR that associates two measures with each classifier in place of the single, traditional strength measure. The first measure is a payoff estimate, which is tuned by the linear-update scheme usually used for strength. The second measure is a priority factor that is tuned to control the outcome of a necessity auction. In the necessity auction the winning classifier pays out the payoff estimate of its nearest competitor, rather than a fraction of its own payoff estimate. Results and analyses are presented that show that this CA/CR scheme can induce variable bid separation that responds to the demands of the LCS environment Additional analyses show that this scheme allows an LCS to adequately exploit a broader class of default hierarchies than traditional schemes. Several avenues are suggested for further study.