Abstract
This paper studies knowledge finding for fuzzy inference system with multiple inputs. Recently many research studies on Classifier System (CS) which is a new paradigm of machine learning have been reported. This system is a production system which generates production rules using genetic algorithms. Fuzzy Classifier System (FCS), which uses a fuzzy inference system and a fuzzy rule base in place of the production system and the rule base of the CS, can handle continuous variables. The FCS uses a chromosome into which a fuzzy rule of fuzzy rule table is encoded, and it does not increase the size of the chromosome exponentially in dealing with multi-input systems. However, the FCS was applied to a single input-single output function approximation problem, and only its capability to handle continuous variables was shown. No study on methods of credit apportionment which is very important for multi-input systems has been reported.This paper shows that the feature of the FCS is its capability to find fuzzy rules by applying the FCS to a multi-input system. Simulations of collision avoidance of a ship are done to show that the FCS can find fuzzy control rules from payoffs based on only successes/failures of the steering. This paper also studies a method of credit apportionment for this problem. With this method, the fuzzy rules which describe knowledge in mutually related variables can be obtained.
