Multilayer feedforward potential function network

Abstract

The authors present a multilayer feedforward network, called the Gaussian potential function network (GPFN), performing association or classification based on a set of potentially fields synthesized over the domain of input space by a number of Gaussian potential function units (GPFUs). A GPFU as a basic component of the GPFN is designed to generate a Gaussian form of a potential field. A weighted summation of Gaussian potential fields generated by a suitable number of GPFUs provides an arbitrary shape of a potential field over the domain of input space. The authors also present a detailed learning algorithm for the GPFN. Learning consists of the determination of the minimally necessary number of GPFUs and the adjustment of the locations and shapes of the individual Gaussian potential fields defined by GPFUs as well as the summation weights. The learning of the minimally necessary number of GPFUs is based on the control of the effective radius of GPFUs, while the parameter learning is based on the gradient descent procedure.