Negative correlation learning approach for T-S fuzzy models

In this paper an adaptive approach of achieving a proper model structure in data-driven T-S fuzzy models is proposed. By introducing negative correlation learning in the creation of the fuzzy model, the training error of the entire model is decomposed to individual rule errors with correlation penalty term. Fuzzy rules can be trained and evaluated separately. On the other hand, negative correlation learning minimizes the mutual information between rules, so that a set of cooperative and complementary fuzzy rules can be obtained. The correlation penalty term also provides a way of measuring the validity of each rule. Algorithms of generating and eliminating rules can be developed based on it, thus the appropriate structure of the model can be obtained independent to the initial number of rules.