System identification using modular neural network with improved learning

Abstract

This paper addresses the problem of the identification of nonlinear dynamic systems using modularly structured neural network with the new learning algorithm for the learning of both gating and expert networks weights. Here we start with the standard learning procedure for such networks in the sense that the problem of learning is formulated and treated as a mixture estimation problem in which the log-likelihood function should be maximised. But, as opposite to the established methods for modular networks we combine the gradient type of learning for gating weights with a least-squares algorithm for the learning of expert networks weights, and the experts are simple one-layer nets with a single linear output unit, The very result of such an approach is a simple structured modular network with improved learning and it seems with good capabilities for the identification of general nonlinear dynamic systems. The identifications of a discrete-time nonlinear system corrupted with noise and of a real world system are presented. Later process represents the identification of the positioning of a car engine throttle valve from real data set (3000 samples of measured noisy data).