Radial-Basis Function Neural Network Synthesis on the Basis of Decision Tree

Abstract

The problem of neural network synthesis on the precedents is addressed. The aim of the study is to create methods for radial-basis neural network model constructing having high levels of generalization and accuracy, which do not require user participation in the process of model building. The method of decision tree transforming into a neural network model is proposed. For a given sample, a decision tree is built, on the basis of which leaf nodes the cluster centers' are allocated, after which the structure of the radial-basis network is synthesized by associating of selected clusters with neurons of the first layer, the cluster centers coordinates are placed into the weights of neurons of the first layer, and then the weights are adjusted. The method for transforming a regression tree into a radial-basis network is proposed. It allocates clusters for solved problem as a regression tree, but to improve accuracy for each cluster, it builds a particular linear regression model of the output feature dependence from the neuron???s output determining belonging to the corresponding cluster. The method of converting a random forest into a neural network model is proposed. Trees of the forest, built on a given sample, are transformed into separate neural network models, which are combined into joint network. The experiments on practical problems solving were carried out. Their results were confirmed the efficiency of the proposed methods.