High generalization capability artificial neural network architecture based on RBF-network

Abstract

This paper describes the issue of error level fluctuations due to training set shrinking in RBF-networks. An architecture of artificial neural network (ANN) based on RBF-network is presented with a learning algorithm to train it. The presented architecture is multi-layer, unlike original RBF-network thus has a potential in deep learning. Numeric results lead to a conclusion about error level fluctuations being significantly lower for the presented architecture compared to RBF-network in case of training set shrinking. This displays a greater generalization ability of the presented architecture. The paper contains an application of ANN to the task of restoring the dielectric parameters for subject placed in waveguide.