Neural network modeling of dynamical systems

The authors present a class of discrete-time, neural network-based, nonlinear models suitable for such applications in a system identification framework. Two applications are presented. In the first one, which deals with the prediction of a non-Gaussian time series, the authors were able to build the optimal predictor, namely, the conditional mean estimator. The authors also describe a Schmidtt trigger composed of two neurons. The latter problem served as a test bed for a gradient based learning scheme, which despite converging often to local minima, proved effective for training recurrent systems. The positive results obtained on these and a number of other nonlinear processing tasks suggest that this method might prove useful in many practical applications.