Fuzzy neural network with independent classification performance and sample input order

Abstract

The present invention provides a fuzzy neural network with independent classification performance and sample input order. Based on an existing fuzzy neural network, a sample input order is changed into parallel input from serial input, when each sample enters into a classifier, through the calculation of a similarity matrix value, the similarity value between the sample and all samples is calculated rather than the calculation of a similarity value between samples which enter into the classifier previously, the expansion and compression between samples are determined by a calculation result, thus a finally formed classification interface is obtained according to the similarity calculation among all samples, the process has nothing to do with a sample order, and the sample dependence is avoided. The present invention provides an algorithm, through the introduction of the concept of a similarity matrix, a condition that training samples passes a fuzzy neural network in series one by one in an original fuzzy neural network is changed into a condition that all samples pass in parallel in one time, and a problem that the classification performance of the fuzzy neural network strongly depends on the sample input order is solved fundamentally.

Description

technical field [0001] The invention relates to the technical field of pattern classification, in particular to a fuzzy neural network whose classification performance has nothing to do with the input order of samples. Background technique [0002] Pattern classifiers are an important part of solving many engineering problems such as pattern recognition, image processing, and signal detection. However, in such problems, the definitions of patterns and categories often have ambiguity. The most commonly used tool for pattern classification problems. [0003] Fuzzy neural network combines the advantages of both fuzzy logic and neural network. It is a new type of model that not only has the learning ability and optimization ability of neural network, but also has the rule representation of fuzzy system IF-THEN form. It can be applied to model Classification, pattern clustering, function approximation and many other fields have been extensively studied by researchers. [0004] ...

AI Technical Summary

Problems solved by technology

[0019] However, there is an important problem in the application process of the FMM model and learning algorithm: the classification performance of the classifier is strongly dependent on the order of the sample input

If two super-boxes of different categories overlap, there will be a phenomenon that the patterns in the overlapping area belong to two different categories, and the membership degree values ​​are all equal to 1, resulting in the result that the patterns in the overlapping area cannot be correctly classified

Therefore, it is necessary to perform compression operations on existing overlapping hyperboxes, and the hyperboxes established earlier have a higher probability of being compressed, so the training results are strongly dependent on the order of input samples

Images