LEARNING FUZZY CLASSIFIER SYSTEMS: ARCHITECTURE AND EXPLORATION ISSUES

Abstract

Introducing fuzzy logic in knowledge representation is a general technique to improve flexibility and performances of knowledge based and control software. Many researchers propose to introduce fuzzy logic representation in learning algorithms. Interesting features arise when fuzzy sets substitute the interval-based classification of input in a learning system; some of them imply an improvement in performance others an increased structural complexity in the architecture of the system and in the learning process. Focusing on Learning Classifier Systems, the introduction of fuzzy logic produces some new interesting features in this class of learning algorithms from many points of view: a new approach to classifier competition, the birth of competition vs. cooperation dilemma, and the introduction of an appropriate fuzzy interface with external world. In this paper, we discuss a fuzzyfication of the classical architecture of a learning classifier system (Holland's approach) and the improvements deriving from the use of fuzzy logic. In this work we especially discuss the competition vs. cooperation dilemma, analyzing the influence of exploration policy on the performance of crisp and fuzzy versions of learning classifier systems. We mainly focus on the use of fuzzy classifier systems to implement behaviors for reactive autonomous agents in the mobile robotics domain.