Time Window Width Influence on Dynamic BPTT(h) Learning Algorithm Performances: Experimental Study

The purpose of the research addressed in this paper is to study the influence of the time window width in dynamic truncated BackPropagation Through Time BPTT(h) learning algorithms. Statistical experiments based on the identification of a real biped robot balancing mechanism are carried out to raise the link between the window width and the stability, the speed and the accuracy of the learning. The time window width choice is shown to be crucial for the convergence speed of the learning process and the generalization ability of the network. Although, a particular attention is brought to a divergence problem (gradient blow up) observed with the assumption where the net parameters are constant along the window. The limit of this assumption is demonstrated and parameters evolution storage, used as a solution for this problem, is detailed.