Using Rprop for on-line learning of inverse dynamics

In this paper, the Rprop algorithm is compared with Backpropagation in the on-line learning of inverse dynamics using Kawato´s feedback error learning structure. Since Rprop is a batch learning algorithm a window of NM samples is used. The samples are selected to avoid unnecessary adaptation of weights. Three nonlinear plants are used as a testbed, for each plant three trajectories are considered to compare the training methods. It is shown that the selection of an appropriate learning rate for Backpropagation is a difficult task avoided using Rprop. Also, the proposed scheme shows an improved performance in terms of training time over Backpropagation.