Designing a Fuzzy Adaptive Neural Network Leveraging a Feature Self-Enhancement Unit and Statistical Selection Methods

Abstract

In this study, we propose an advanced category of a fuzzy adaptive neural network (FANN) based on a feature self-enhancement unit (FSU) and statistical selection methods (SSMs). Undoubtedly, the raw data contain a large amount of information with varying importance. One of the most important tasks for regression model design is to avoid losing these important details. However, the raw data cannot participate in the whole training process due to the data fuzzification unit in the structure of conventional fuzzy neural networks (FNNs). Meanwhile, the polynomial-based neuron also has its limitations as a common node in FNNs. For example, in the polynomial neuron, the complexity of the neurons increases exponentially with the increase in network size. Consequently, overfitting and insufficient raw data information are two primary drawbacks in the structure of conventional FNNs. To address these limitations, we designed the FSU and the SSM as effective vehicles to reduce data dimensionality and select significant raw information. The proposed FANN also demonstrates the capability to improve modeling accuracy in neural networks. Moreover, this is the first instance of integrating statistical methods and feature self-enhancement techniques into a fuzzy model. To validate and showcase the superiority of the proposed FANN, the model is applied to 16 machine learning datasets, outperforming other comparative models in 81.25% of the datasets utilized. Additionally, the proposed FANN model outperformed the latest FNN models, achieving an average 5.1% increase in modeling accuracy. The comparison experiment section not only includes classical machine learning models but also references the experimental results from two recent related studies.