Improving the performance of fuzzy classification systems by membership function learning and feature selection

We have formulated a method for generating fuzzy if-then rules from numerical data for pattern classification problems. In the previous study, we fixed the shape of the membership functions for each variable, i.e., we evenly divided each axis into a number of fuzzy sets and did not change the shape of membership functions. The number of the fuzzy sets for each axis was typically specified as five so that each fuzzy set can be interpreted as a linguistic value. We examine the performance of a fuzzy classification system with the ability of adjusting the membership functions. We use an error-correction learning method for adjusting membership functions. We also examine the effect of selecting a small number of features on the performance of the fuzzy classification system. We use class information entropy to measure how well an axis is divided. In computer simulations, we examine the performance of our methods on several real-world pattern classification problems with continuous attributes. From the simulation results, we show that our methods perform well