Dynamic Threshold Selection for Sequential Learning in Radial Basis Function Networks

Abstract

For sequential learning in Radial Basis Function (RBF) Networks, there is a requirement for dynamic selection of threshold because a constant threshold is inadequate to accommodate functions of varying amplitudes. In this paper, a new criterion is defined for the dynamic selection of the Euclidean output deviation threshold. Its effect on the learning process experienced by RBF networks with regard to functions of variable amplitude is shown. This improved network can automatically select a suitable threshold for its own supervised learning depending on the objective parameters set to achieve certain accuracy level of the desired output. This paper also proposes further automation to neuron growing and pruning within Radial Basis Function (RBF) neural networks. The proposed dynamic threshold selection technique has shown significant improvement in achieving stable neuron growth rate in dealing with signal amplitude variation.