Non-linear time series modeling with self-organization feature maps

A locally linear approach based on Kohonen self-organizing feature mapping (SOFM) is proposed for the modeling of nonlinear time series. This approach exploits the neighborhood preserving property of Kohonen feature maps. The key difference is that the local model fitting is performed directly over a matched neighborhood of the constructed SOFM neural field. The initial results show that this neural network scenario is an effective approach for local modeling of low dimensional nonlinear processes.