Redefined Fuzzy Min-Max Neural Network

The classical fuzzy min-max (FMM) neural network easy to cause the overlap of hyperboxes from different classes, which affect the pattern classification performance. In this paper, we propose a redefined fuzzy min-max (RFMM) neural network to solve this problem. The main contribution is to modify the basic architecture of FMM by adding a redefined hyperbox layer. The proposed RFMM is a four-layer feedforward neural network. The generated hyperbox layer and the redefined hyperbox layer are connected through the proposed hyperbox filter, hyperbox optimization and hyperbox combination. The RFMM learning algorithm is an expansion/contraction/redefinition process. The effectiveness of RFMM is evaluated based on ten benchmarks. Experimental results indicate that RFMM leads to better classification performance than various FMM-based, support vector machine-based models and lower sensitivity to the maximum size of expansion coefficient.

I. Introduction

Artificial neural networks(ANNs) and fuzzy logic are the two most widely used models in pattern classification [1]. In 1992, by combining ANN and fuzzy set theory, fuzzy min-max (FMM) neural network was proposed by Simpson, forming a general framework for solving pattern classification problems [2]. FMM is a supervised learning neural network classifier that uses fuzzy sets as a pattern class, with an incremental learning paradigm, which performs a single-pass learning process through data samples [2].