Effects of the minimal singular value on the performance of echo state networks

Abstract

Echo State Networks (ESNs) have attracted wide attention for their superior performance in time series predictions. The performance of ESNs is heavily dependent on their reservoir parameters, which have to be tuned by hand for the best performance. Hence, it is important for successful design of ESNs to understand the inter-relationship between the performance and the parameters. This paper investigates the effect of the minimal singular value on the performance of ESNs based on two computational experiments. The reservoir weight matrix is constructed by using singular value decomposition, where the singular value spectrum is generated randomly from variable intervals. Then, using computational experiments with nonlinear ESNs, the relationship between the minimal singular value and the performance of ESNs is analysed.