A New Fuzzy Modeling Approach Based on Support Vector Regression

New interpretable kernels created by conjoining the univariate fuzzy membership functions with a t-norm operator are proposed in this paper. Based on support vector regression with presented kernel, a learning algorithm consisting of two phases is developed to construct fuzzy system. In the first phase, the support vector regression learning model provides architecture to extract support vectors for generating fuzzy rules, and then characterizes the support vector expansion in TS fuzzy inference procedure through simple equivalent transform. In the second phase, a reduced set method is employed to simplify the obtained fuzzy model, and a bottom-up strategy with relative degree of sharing is suggested to obtain a transparent rule base, at the same time preserves the accuracy and generalization performance of the fuzzy model. Finally, the performance of the proposed fuzzy model is compared with hierarchical clustering based on using a self-organizing network modeling methods.

1. Introduction

In recent years the interest in data-driven approaches to the modeling of fuzzy system has increased. The key issue in constructing a fuzzy model is to identify the model structure. So far, the two widely used and particularly focused approaches are the self-organizing neural networks and fuzzy clustering [1]. However, these fuzzy modeling algorithms are difficult to determine the number of fuzzy rules systematically and effectively, so it is hard to balance the tradeoff between the necessary accuracy of the model and its complexity. Although there are some methods for rule base simplification about complexity reduction, but most of these methods tackle the problem from an interpretability view in order to obtain a transparent rule base [2]. Few researchers pay attention to the problem of generalization of fuzzy model, and how to get a good fuzzy model which gives excellent performance.