A New Fuzzy Membership Function with Applications in Interpretability Improvement of Neurofuzzy Models

Abstract

Local model interpretability is a very important issue in neurofuzzy local linear models applied to nonlinear state estimation, process modelling and control. This paper proposes a new fuzzy membership function with desirable properties for improving the interpretability of neurofuzzy models. A learning algorithm for constructing neurofuzzy models based on this new membership function and a hybrid objective function is derived as well, which aims to achieve optimal balance between global model accuracy and local model interpretability. Experimental results have shown that the proposed approach is simple and effective in improving the interpretability of Takagi-Sugeno fuzzy models while preserving the model accuracy at a satisfactory level.