Artificial neural networks are suitable for the prediction of chaotic time series. A modified back-propagation algorithm with neuron splitting is used to train feed-forward multilayer perceptron networks for prediction. There are two ways of parallelizing: distributing the training set for batch learning or distribute the vector-matrix-operations for on-line training. Three implementation are compaired: PVM on a workstation cluster and Parix and the new PVM/Parix on a Transputer system. Results about the quality of forecasting an examplary time series and speedups of the parallel programs are presented.
HTML, Postscript (187k)