Simple fuzzy rule-based classification systems perform well on commonly used real-world data sets

Abstract

Real world pattern classification problems usually involve many attributes. Thus, it is often claimed that fuzzy rule based systems with grid type fuzzy partitions are not applicable to such pattern classification problems due to the exponential increase of the number of fuzzy if-then rules (i.e., the curse of dimensionality). When we use K antecedent fuzzy sets for each attribute of an n dimensional pattern classification problem, the total number of possible fuzzy if-then rules is K/sup n/, which is intractably huge for a large value of n. Thus we can not directly apply grid type fuzzy partitions to high dimensional pattern classification problems. If a few attributes can be selected from a large number of attributes for a high dimensional pattern classification problem, we can use a grid type fuzzy partition. The point is whether grid type fuzzy partitions based on a few attributes have high classification ability or not. The aim of the paper is to examine the performance of such fuzzy partitions by computer simulations on real world pattern classification problems with many attributes. Simulation results clearly show that a few attributes have high generalization ability for some real world pattern classification problems.