Classifier systems learn by means of a market economy of instructions (classifiers) and data (messages). The system is computationally parallel since all classifiers may be applied to all messages simultaneously. Classifier systems discover by means of a genetic algorithm which creates promising new classifiers from successful ones. This genetic procedure conducts a parallel search for the building blocks of problem solutions. Classifier systems already exhibit successful learning and discovery, including the emergence and stability of both multilevel default hierarchies and chains of classifiers relevant to means of software packages developed for expressing problematic environments and for formulating classifier programs. The project will study the different roles of strength (wealth), the roles of generality and specificity in hierarchies and classifier reproduction, and the emergence of building blocks. Adaptive systems which develop their own goals will also be covered. Our computer experiments are guided by a theory of adaptation. The fundamental theorem characterizes the rate at which the genetic algorithm samples the building blocks of useful models of the environment. We will extend this theorem to cover the discovery of regularities in rapidly changing environments.
