Ensemble Separable Recurs ive Techniques for MLP Networks

Novel hybrid or separable recursive training strategies are proposed for the training of feedforward neural networks which include switching modules and ensemble between them. This new technique for updating weights combines nonlinear recursive training algorithms for the optimization of nonlinear weights with recursive least square (RLS) type algorithms for the training of linear weights in one integrated routine. The proposed new variant of hybrid weight update includes two form of switching mechanism based on the condition of input data to the system (correlated or random). The performance switching modules and dual ensembling approach on switching mechanism is illustrated by the simulation studies. Simulation results demonstrate the superiority of the new proposed hybrid variant training schemes compared to traditional recursive prediction schemes using various chaotic non-linear benchmark problems.