Echo State Network Performance Analysis Using Non-random Topologies

Echo State Network (ESN) has been widely studied and applied to many problems due to the simplicity of its training phase. This is because since in this network only the output weights are trained, avoiding to deal with the gradient's vanishing problem presents in most of the recurrent neural networks. However, this technique has been criticized recently because of the echo property limitation and its random topology that may cause chaotic activity in the reservoir layer. In this paper, we present an application of the classic ESN model modifying the reservoir topology to a non-random approaches: clustered and complex networks, as an alternative solution to the chaotic activity problem. Further, the modified and classical models are compared considering two study cases: Rossler and Lorenz systems. Numerical experiments show that the proposed model has a better performance than the classical model.