Predicting chaotic time series by ensemble self-generating neural networks

In this paper, we introduce ensemble self-generating neural networks (ESGNNs) for chaotic time series prediction. ESGNNs are combined the ensemble averaging method with SGNNs. ESGNNs create self-generating neural trees (SGNTs) to shuffle the order of given training data independently, and the network output is averaged of all SGNTs output. We investigate improving capability of ESGNNs for three chaotic time series and compare with the backpropagation neural networks (BPNNs). Experimental results show that using various SGNTs through ensemble averaging method significantly improves the predictive performance of ESGNNs on diverse chaotic time series.