Generation of Fuzzy Classification Rules by Non-Overlapping Input Partitioning

Abstract

The paper proposes a new method for generating fuzzy classification rules from numerical data. The main idea of the method consists in separating the input feature space into a number of non-overlapping hyperboxes, which contain input data from one classification class only, and a consequent generation of fuzzy rules and membership functions for each hyperbox. An appropriate fuzzy inference mechanism is proposed for classifying new input data into the output classification space. The proposed method formalizes the synthesis of fuzzy rule-based systems and could also be used for function approximation and design of fuzzy control systems. The method is numerically compared to some existing fuzzy classification methods using the Fisher iris data. The comparison results show that it outperforms most of them and can successfully be used for the development of fuzzy classifiers