Information orientation-based modular Type-2 fuzzy neural network

Abstract

Type-2 fuzzy neural networks (T2FNNs) have gained popularity due to their processing ability for high uncertainty. However, concerned with the high-dimensional problems of nonlinear systems, the interpretability of individual T2FNNs is weak due to the exponential growth of fuzzy rules. To deal with this problem, an information orientation-based modular T2FNN (IO-MT2FNN) is developed to improve its interpretability in this paper. First, an information entropy-based decomposition method is designed to divide the original input space into three sub-spaces, namely edge, local and global regions. Then, the information with different attributes is separated to provide an unambiguous interpretation. Second, the independent module describing these regions with type-2 fuzzy sets is embedded in the membership function layer of IO-MT2FNN to represent the coupling relationship between regional information in an interpretable way. Third, an information mapping strategy is introduced with low-order Gaussian kernel matrices, instead of a high-order mapping matrix, to extract the features from the allocated information in each module, which enables IO-MT2FNN to achieve a compact topology through dimensionality reduction. Finally, the simulations demonstrate that the proposed IO-MT2FNN can compete with the advanced approaches in terms of interpretability for the prediction of high-dimensional and complex systems.