Ranking-based fuzzy min-max classification neural network

Abstract

The performance of fuzzy min-max classification neural network (FMM) is affected by the input sequence of training set patterns. This paper proposes a ranking-based fuzzy min-max Classification Neural Network (RFMM) to overcome this shortcoming. RFMM improves FMM through the following three aspects. First, RFMM ranks the input order of the training set patterns according to their membership degree to the center point of same class, so that the finally constructed network is fixed and does not depend on the input order of the training set. Second, a new membership function based on Manhattan distance is constructed, which overcomes the problem that the membership degree obtained by the membership function in the FMM will not decrease steadily with the increase of the distance between the input pattern and the hyperbox. At last, RFMM uses the method based on individual contour coefficient to classify the patterns in overlapping regions, which overcomes the problem that when the FMM eliminates the overlapping region by shrinking hyperboxes, the membership degree of the patterns in the contracted region to the class they belong is changed. Experimental results show that RFMM has better learning ability, and compared with other FMM methods, RFMM shows higher classification accuracy and lower network complexity.