Further experiments in Fuzzy Classifier Systems for mobile robot control

Abstract

We present a Michigan-style Fuzzy Classifier System architecture, which operates at the level of individual rules. We compare performance of this architecture with a Pittsburgh Fuzzy Classifier System investigated previously it operates at the level of whole rule-bases. The application domain is mobile robotics, and the problem is autonomous acquisition of an "investigative" obstacle avoidance competency. In our work, both Fuzzy Classifier System types are restricted to operating on the rules of fuzzy controllers with pre-defined fuzzy membership functions (MFs). Generally, both approaches were capable of producing fuzzy controllers with competencies that exceeded that of the hand-coded fuzzy controller presented in this paper. The previously investigated Pittsburgh architecture yielded a more stable convergence on high performance solutions than the Michigan architecture that was created to operate at the level of individual rules.