TS-fuzzy modeling based on ε-insensitive smooth support vector regression

Abstract

This paper establishes a connection between Takagi-Sugeno TS fuzzy systems and e-insensitive smooth support vector regression e-SSVR, a smooth strategy for solving e-SVR. In previous e-SVR-based fuzzy models, the form of membership functions is restricted by the Mercer condition. The e-SSVR formulation puts no restrictions on the kernel. Therefore, the proposed e-SSVR-based TS-fuzzy modeling method relaxes the restriction on membership functions. By applying the reduced kernel technique, the number of fuzzy rules is reduced without scarifying the generalization ability. The computational complexity is also reduced by the reduced kernel technique. The performance of our method is illustrated by extensive experiments and comparisons.