Selective adjustment of rotationally-asymmetric neuron σ-widths

Abstract

Radial Basis Networks are a reliable and efficient tool for performing classification tasks. In networks that include a Gaussian output transform within the Pattern Layer neurons, the method of setting the σ-width of the Gaussian curve is critical to obtaining accurate classification. Many existing methods perform poorly in regions of the problem space between examples of differing classes, or when there is overlap between classes in the data set. A method is proposed to produce unique σ values for each weight of every neuron, resulting in each neuron having its own Gaussian ‘coverage’ area within problem space. This method achieves better results than the alternatives on data sets with a significant amount of overlap and when the data is unscaled.