Design of stabilized polynomial-based ensemble fuzzy neural networks based on heterogeneous neurons and synergy of multiple techniques

Abstract

Listen

Translate article

In this study, a novel category of polynomial-based ensemble fuzzy neural networks (PEFNNs) are proposed. The study is focused on the development of advanced design methodologies to improve the performance (prediction accuracy) of the model when dealing with nonlinear regression problems. In contrast to the conventional fuzzy polynomial-based models, we adopt a hybrid network structure composed of heterogeneous neurons. The first layer of PEFNNs consists of fuzzy regular polynomial neurons optimized by clustering method. In the consecutive layers, we engage two types of polynomial neurons, which are selected and optimized by evolutionary algorithms. Moreover, an enhanced topology based on fuzzy module and enhanced interconnection (FM&EI) is designed to strengthen the characteristics of fuzzy feature information as well as increase the number and diversity of neurons. Multiple techniques are used synergistically to reinforce the performance of PEFNNs. First, a coefficient-based performance compromise algorithm (CPC) is designed to select neurons by considering the performance and complexity of the neuron. Second, L2-norm regularization is considered to improve the performance of the model. Third, evolutionary algorithm is employed to adjust the structural parameters of PEFNNs. Furthermore, FM&EI and hybrid network structure which consist of heterogeneous neurons are considered as one of the multiple approaches to construct the ensemble model. The performance and stability of PEFNNs are evaluated with a diversity of datasets. A thorough comparative analysis also is covered.