Condition Number-based Evolving ESN

Abstract

Echo state network (ESN) is a kind of recurrent neural network without involving gradient problem. However, the reservoir of ESN often contains hundreds of neurons, whose corresponding high-dimensional state matrix may result in ill-conditioned solution problem. To solve it, the condition number-based evolving ESN (CNEESN) is proposed, whose sub-reservoir is generated by condition number analysis and differential evolution algorithm (DE). Firstly, the influence of condition number on output weight matrix is analyzed. Secondly, the randomly generated singular values are optimized by condition number and DE based optimize strategy. Finally, simulation result on a benchmark dataset has shown the superiority of the proposed CNEESN.