Construction And Evolution Of Reinforced Fuzzy Min-max Neural Network

Fuzzy neural network haves the respective advantages of both fuzzy theory and neural network,and has a wide range of applications.Fuzzy min-max neural network has strong learning ability and good performance,and it is a classic network model,but there are still some problems:(1)Different input order has different effects on hyperboxes;(2)Expansion parameters have user-defined,which have different effects on training results;(3)Contraction rules change the sample affiliation and removes the hyperboxes to a large extent.The reinforced fuzzy min-max neural network proposed in this paper can solve the above problems,and the main innovations include the following three aspects.1.A parametric reinforced fuzzy min-max neural network is proposed.The model is a multi-layer network model,defining a custom division length parameter,putting all sample points randomly into the training model,constructing the hyperboxes by division operation,and eliminate the impact of problem one.The hyperbox is selected and classified by the judgment operation,and the hyperbox that meets the requirements is output.Repeat these two steps repeatedly so that the model training is complete.Compared with the classical original model,the classification accuracy is improved,but there is a problem of a large number of hyperboxes.2.A hyperbox reinforced fuzzy min-max neural network is proposed.The model is based on parametric reinforced fuzzy min-max neural network with the aim of reducing the number of hyperboxes.By merging the hyperboxes generated by the parametric reinforced model,the merging process will result in the overlap of different types of hyperboxes.The definition of hyperbox center of mass is proposed and use it to generate new expansion formulas for the contraction process of the overlap,which further improves the classification accuracy and reduces the number of hyperboxes compared with the parametric reinforced fuzzy min-max neural network.3.An artificial bee colony optimization reinforced fuzzy min-max neural network is proposed.In the overall framework of the reinforced fuzzy min-max neural network,the selection of different custom division length and expansion parameters have different effects on the experiments,which further affect the classification accuracy.An artificial bee colony optimization algorithm was used to optimize the operation of these two parameters.The optimal parameters are determined by selecting different nectar sources,calculating the adaptation values corresponding to different honey sources,and continuously updating the honey sources to output the optimal solution after the optimization model.