Multiobjective Evolution of the Explainable Fuzzy Rough Neural Network with Gene Expression Programming

The fuzzy logic-based neural network usually forms fuzzy rules via multiplying the input membership degrees, which lacks expressiveness and flexibility. In this paper, a novel neural network model is proposed via integrating the gene expres- sion programming to the interval type-2 fuzzy rough neural network to generate fuzzy rules with more expressiveness via various logic operators. The network training is regarded as a multiobjective problem via simultaneously considering network precision, explainability, and generality. Though the fuzzy rule is straightforward, to further increase the network explainability, the network complexity is minimized to generated concise and few fuzzy rules. For settlement, inspired by the extreme learning machine and the broad learning system, an enhanced distribut- ed parallel multiobjective evolutionary algorithm is proposed. The evolutionary algorithm can flexibly explore the forms of fuzzy rules, and the weight refinement of the final layer via pseudoinverse computation can significantly improve precision and convergence. Experimental results show that the proposed evolutionary network framework is superior in both effectiveness and explainability.