A modified echo state network in chaotic time series prediction

Echo state network (ESN) is one of the most well-known types of reservoir computing because of its outstanding performance when chaotic time series prediction is conducted. However, sometimes it works poorly because the reservoir connectivity and weight structure are created randomly. To solve this problem, we propose a modified ESN based on contribution rate algorithm. By pruning unimportant connections without loss of major information, the proposed method can not only optimize the network structure, but also improve the generalization performance of network. Experimental results and performance comparisons demonstrate that the modified ESN outperforms the ESN without optimization.