Combining statistical and reinforcement learning in rule-based classification

Abstract

BYPASS is a rule-based learning algorithm designed loosely after John Holland’s Classifier System (CS) idea. In essence, CSs are incremental data processors that attempt to (self-)organize a population of rules under the guidance of certain reinforcement policy. BYPASS uses a reinforcement scheme based on predictive scoring and handles conditions similar to those in standard classification trees. However, these conditions are allowed to overlap; this makes it possible to base individual predictions on several rules (just like in committees of trees). Moreover, conditions are supplied with simple bayesian predictive distributions that evolve as data are processed. The paper presents the details of the algorithm and discusses empirical results suggesting that statistical and reinforcement-based learning blend together in interesting and useful ways.