A discrete-time neural network model for systems identification

Neural networks can be powerful tools for nonlinear signal processing and systems modeling. The authors present a class of discrete-time, neural-network-based, nonlinear models suitable for such applications in a systems identification framework. The parameters for these models include a local interconnection neighborhood size, a time constant characterizing the neurons, a weight matrix, and input output connection matrices. The prediction error identification method trains the net to solve two problems: the prediction of a chaotic time series generated by the logistic function, and the demodulation of FSK (frequency-shift keying) signals. The proposed approach allows a speedup of 1 to 2 orders of magnitude while preserving important properties of its continuous-time analog, and its speed permits global minima to be determined by simulated annealing. The model allows multichannel applications for control or prediction