Classifier systems for control

Abstract

Classifier systems lie midway between neural networks and symbolic processing systems and potentially combine the benefits of both. They are parallel message-passing rule-based systems which use genetic algorithms to discover new rules as well as providing for reinforcement learning and programming. It has been proposed that a suitable application of genetic algorithms is to evolve robots. A most suitable way to use genetic algorithms to evolve the control systems for robots is within the framework provided by classifier systems. At a SERC workshop on learning systems a number of groups presented successful applications of the genetic algorithm to control problems. However, one cannot evolve complex systems with a simple genetic algorithm nor is it wise or safe to start from scratch in real applications where programmed knowledge can provide constraints for the genetic algorithm to work within. If the genetic algorithm is to be used to evolve control systems for industrial or commercial applications one of the best ways to do this is within the framework of classifier systems.