LEARNING AND BUCKET BRIGADE DYNAMICS IN CLASSIFIER SYSTEMS

Classifier systems are rule-based adaptive systems whose learning capabilities emerge from processes of selection and competition within a population of rules (classifiers). These processes are ruled by the values of numerical variables which measure the fitness of each rule. The system's adaptivity is ensured by a fitness reallocation mechanism (the bucket brigade algorithm) and by genetic algorithms which are responsible for the internal dynamics of the system. In this paper we discuss classifier systems as dynamical systems, the main focus being on the asymptotic dynamics due to the bucket brigade, abstracting from the action of the genetics. This topic is discussed with reference to a specific task domain, in which the system is used as a detector of statistical properties of periodic or fluctuating external environments. We also describe a major consequence of the genetics on the bucket brigade dynamics, namely the proliferation of individual rules into subpopulations of equivalent classifiers; we then show that this can eventually lead to undesired stochastic behavior or to the destabilization of correct solutions devised by the system. © 1990.