Training of Radial Basis Function Networks with H∞ Filter-Initial Simulation Results

Abstract

This paper analyzes the application of the H∞ filter for the optimization of the parameters of an artificial neural network with Gaussian-type radial activation functions. The analysis showed that the H∞ filter generates better estimates of the parameters of the artificial neural network than the linearized Kalman filter in problems where there is significant initial parameter uncertainty, insufficient knowledge of system/process characteristics, process noise, and measurement noise. Unlike the linearized Kalman filter, the H∞ filter does not rely on the assumption that process and measurement noises are subject to Gaussian distribution, which is a special advantage for the application of this method of artificial neural network parameter optimization in engineering problems.