Research Article

A Novel Method for Training an Echo State Network with Feedback-Error Learning

Figure 7

The plots illustrate why the original method without repetitions (experiment 1) fails. Compared to the true target (a), the estimated target in the first epoch (c) is very noisy. It has the general shape of the true target, but when training the initial, random ESN (b) with this noisy estimate, the result is a network which outputs mostly noise (d). This only gets worse in the succeeding epochs. Plotted are motor commands (joint angle velocities) for the 4 DOFs at each time step in the training sequence.
(a) True target
(b) Initial ESN output
(c) Estimated target
(d) ESN output after trained with (c)