Generating fuzzy rules from data

This paper introduces an effective method of developing fuzzy rules from continuous valued data. The fuzzy rules may be used for control applications without tuning. The fuzzy rules are created by exploiting the properties of decision trees, as embodied by the C4.5 decision tree learning system. A crisp decision tree is created by creating a discrete set of fuzzy output classes and providing a set of training examples to C4.5. Fuzzy rules are then extracted from the decision tree. The fuzzy rule learning system has been applied to chemical plant start-up control and the Box-Jenkins gas furnace prediction problem. Comparisons are made to fuzzy rule sets created by others for these problems. The learned rules are able to provide smooth control.