A hybrid technique to enhance the performance of recurrent neural networks for time series prediction

The recurrent neural networks trained by the real time recurrent learning (RTRL) algorithm is used for time series prediction. When there is a strong nonlinear relationship connecting the adjacent samples of the time series which the network is trying to predict, the prediction performance of the network deteriorates. A scheme is proposed to overcome this drawback. This scheme incorporates cascade-correlation into the recurrent network learning after the network has been trained using RTRL. Fahlman's quickprop algorithm is incorporated into the RTRL learning to make the network converge faster. Simulation results with the above enhancements are presented. The improvement in the prediction performance is found to be considerable.<
>