Reduction of Generalization Error in Fuzzy System Modeling

This paper proposes a technique to reduce the overfitting of the fuzzy models to the training data set during the supervised training phase. Typically a training data set is employed in extraction of the unsupervised fuzzy rule base (FRB) of a fuzzy model (FM), and in supervised training of the FRB to reduce the output error of FM for the training data set. However, recently developed optimization tools usually results in the overfitting of the FM to the training data set, which causes unacceptable rise in the output error for the verification data set. The proposed approach is based on dynamic construction of synthetic training data sets with similar statistical features to the verification data set. The proposed technique is tested on simple single-input and several multi-input benchmark data sets for the commonly used TS fuzzy inference method. The test results indicated that the proposed method is successful in reducing the verification error.