A Novel Method for Training an Echo State Network with Feedback-Error Learning
The optimal choice for the level of internal noise in the reservoir was significantly different for the two training methods. The figures show the mean position error during testing for different noise levels. (a) The networks were trained with the original method with the training sequence consisting of 5 repetitions of the YMCA movement. (b) The corresponding results when the networks were trained with the new method with and the movement sequence repeated once. All experiments were run 10 times, and the number of internal nodes was 200 in all the networks. Based on the results we chose noise level 0.03 for the original method and 0.2 for the new method.