Neural network based dynamic models

This paper discusses the use of neural networks for modelling linear or nonlinear dynamic systems using the input and output temporal signals of the system to be modelled. The characteristics of a particular model are discussed, and some improvements are suggested. Sztipanovits proposed the application of the backpropagation network and of a special linear filter component (1992), the resonator based digital filter (RBDF), developed by Peceli (1989). This filter has promising features:-the RBDF is structurally passive, provides minimum roundoff noise, can suppress zero-input limit cycles, etc. From the implementational point of view it is a highly parallel structure which provides the adaptive system with substantial advantages. On the other hand both components have drawbacks as well:-the RBDF structure is suitable for both FIR and IIR filtering problems but its application in an adaptive IIR context is not straightforward because of stability problems,-the neural network can exhibit poor convergence properties during training. The purpose of this paper is to improve the performance of the neural network based dynamic model (adaptive filter) proposed by Sztipanovits (especially during the training phase), and to extend the range of applications to IIR systems