Regression-type neural networks for system identification

In this paper we prove the effectiveness of using neural networks of regression type to identify time series and non linear dynamic systems. It is experimentally shown that, whenever the process generating the data is ruled by a linear model (such as an ARMA for time series), performances provided by the neural network are comparable with the optimal predictor given by the Kolmogorov-Wiener theory. On the other hand, performances outcome classical linear identification approaches when the system to be modelled is intrinsically non-linear. The work extends the one suggested by Narendra et Al. in (1990) by considering a reduced set of training data and a blackbox model for the system to be identified