Learning from Data Using XCS

Abstract In this paper, we present first of all the working principles of an accuracy based learning classifier system. We also discuss the use of learning classifier systems for learning from data by considering a sample application. The sample application, the Terrain Reasoner Weight Adapter (TRWA), is a system that learns near optimal weights to be used by a path planner while generating routes. Manually generated weights are used to generate a sample data set for training the TRWA. We detail the TRWA and the significant improvements made to the usual XCS strategies in order to achieve our goal of using a supervised learning technique for the TRWA. A reward assignment scheme is developed. The use of tournament selection instead of roulette wheel selection for selecting two parents in the GA is also analyzed. The results obtained show the efficiency of the method.