Extraction of Fuzzy Rules for Classification Based on Partitioned Hyperboxes

In this article we propose a method for extracting fuzzy rules based on box-shaped regions hyperboxes that are defined for approximating class regions. In the proposed method, partition of hyperboxes is performed to achieve high accuracy in approximating the class regions. Excessive partition is prevented by the terminating criterion discussed. Each overlap between partitioned hyperboxes of different classes is resolved, if necessary, in a recursive fashion by means of defining new hyperboxes. Fuzzy rules are extracted based on the defined hyperboxes. We describe a mechanism for inference of these fuzzy rules. We show empirically that the proposed method has higher generalization ability for the iris data and Japanese hiragana data on Japanese license plates than either that of a fuzzy rule extraction method without partition of hyperboxes or multilayered neural networks.