Research On Self-Evolution Construction Method Of Fuzzy Min-Max Neural Network

Fuzzy neural network is a kind of neural network which can process fuzzy information.It combines the advantages of fuzzy logic and neural network.Fuzzy min-max neural network is a kind of fuzzy neural network composed of a large number of interactively connected neurons,which has the advantages of online learning.However,it still has the following problems:(1)The generation of hyperbox is strongly dependent on the order of sample input;(2)Network complexity is affected by extension parameters;(3)The error rate of single-layer structure is high;(4)Contraction rule eliminates large range of hyperbox.The three models proposed in this paper overcome these defects and their main contributions are as follows:1.A density-sorting-based convolutional fuzzy min-max neural network is proposed.Before the operation of fuzzy min-max neural network,a step is proposed to sort samples based on density,which reduces the influence of sample input order in the process of hyperbox training.Then,the improved model is combined with the convolutional neural network,and the classification part of the convolutional neural network is replaced by the improved model,so that the convolutional neural network has online learning ability,and the training time is reduced while the classification accuracy is improved.2.A self-adapting fuzzy min-max neural network is proposed.The model is structurally modified to use a deep structure for training and a top-down way to divide the hyperbox.In the initialization step,the range where all samples reside is treated as a hyperbox,which solves the problem that the hyperbox creation depends on the sample input order.Secondly,several segmentation operations were carried out to cut the hyperbox according to the sample category and gradually refine the hyperbox set.This hyperbox partitioning method does not require the participation of extended parameters,which reduces the network complexity and improves the classification accuracy.3.A self-evolving fuzzy min-max neural network is proposed.The model uses a bottom-up approach in which hyperboxes are merged in a multi-layer structure to generate the final set of hyperboxes.A custom edge length parameter is proposed to directly construct the initial hyperbox,which eliminates the influence of sample input order in the process of hyperbox construction.Through the multi-layer optimization operation,the hyperbox is fused to achieve better performance and reduce the network complexity.