Automatic generation of fuzzy control rules by machine learning methods

This paper presents a multi-strategy learning technique for automatic generation of fuzzy control rules. In order to eliminate irrelevant input variables and to prioritize relevant ones according to their influences on the output value(s), the ID3 algorithm is adopted to classify the given set of training I/O data. The resulting decision tree can be easily converted into IF-THEN rules, which are then fuzzified. The fuzzy rules are further improved by tuning the parameters that define their membership functions using the gradient-descent approach. Experimental results of applying the proposed technique to nonlinear system identification have shown improvements over previous work in the area. In addition, it has been successfully applied to mobile robot control in unknown environments.