BUILDING A FULLY DEFINED NEURO-FUZZY NETWORK WITH A REGULAR PARTITION OF A FEATURE SPACE BASED ON LARGE SAMPLE

Abstract

The problem of model synthesis automation for pattern classification on the features is solved. The method of neuro-fuzzy modelsynthesis on precedents is proposed. It is able to construct completely defined neural models based on the regular partition of a feature space. The method calculates the coordinates of the cluster centers as coordinates of the centers of rectangular blocks in the space of feature intervals and clusters membership to classes determine on the training set: for clusters containing observations the membership is determined by the maximum frequency of instances of the corresponding classes in the cluster, and for clusters that do not contain observations the membership is determined by the maximum potential induced on it by the clusters with known class membership. The resulting set of clusters-rules is mapped to the structure of Mamdani neuro-fuzzy network and its parameter values are calculated on the base of parameters of feature set partition and cluster centers. The proposed method does not require loading the entire training sample in the computer memory and speeds up the process of model synthesis providing an acceptable level of data generalization by obtained neural model. The software that implements the proposed method is developed. The experiments confirming the performance of developed mathematical support are conducted. They allow to recommend the method for the construction of neuro-fuzzy models based on a large samples.