THE METHOD OF DIAGNOSTIC MODEL SYNTHESIS BASED ON RADIAL BASIS NEURAL NETWORKS WITH THE SUPPORT OF GENERALIZATION PROPERTIES

Abstract

Urgent problemof automation of radial basis neural network synthesis based on a set of precedents for decision-making in the diagnosis is
solved in the paper. The method for the synthesis of radial basis neural network is proposed. It forms at the beginning one class pattern, which,
if necessary, supplemented with new patterns formed on the basis of wrongly recognized instances, and then operates with the distance from
the instances to the patterns of the clusters. On the basis of the obtained pattern coordinates it further automatically synthesize structure and adjust the weights of the network, which is further optimized to improve the generalizing and interpretability properties by weights contrasting. The proposed method does not require the user specify the number of clusters, has no uncertainty in selection the number of neurons in the first layer and in the choice of the initial values of the network weights, seeks to minimize the size of the network, and characterized by an acceptable time of learning through the use of network optimization procedure allows to obtain nonredundant, contrast, and interpretable neural models. The software implementing proposed method has been developed. The experiments confirming efficiency of developed software have been conducted. They allow to recommend the proposed method for use in practice in solving the problems of diagnostic model constructing by precedents to automate the decision-making in technical and biomedical diagnostics.