Nonlinear time series prediction with a discrete-time recurrent neural network model

Summary form only given. Discusses the application of a discrete-time recurrent neural network model to signal processing and time series prediction. This network constitutes a black box model for input-output nonlinear system identification. Two samples are considered, which consist of (i) predicting a deterministic time series generated by the Mackey-Glass equation and (ii) a stochastic nonGaussian time series. In the deterministic case, a 9-neuron network converged to a solution with prediction error comparable to that of feedforward networks, with faster learning than backpropagation, and absolutely no windowing or prior knowledge about the time series. In the stochastic case, a 25-neuron network was trained and converged to a solution close to the conditional mean also with no prior information or ad hoc assumptions. Ongoing research into the relation of neural networks to the Volterra-Wiener theory of nonlinear systems is also discussed