Online designed of Echo State Network based on Particle Swarm Optimization for system identification

Complexities with existing algorithms have thus far limited supervised training techniques for Recurrent Neural Networks (RNNs) from widespread use. Echo State Network (ESN) presents a novel approach to train RNNs. Certain properties make ESN online learning unsuitable. This paper proposes a modified version of ESN structure for complex nonlinear system online prediction. The Particle Swarm Optimization (PSO) is adopted to online train the output weights of ESN, as against computing it, which greatly improve the modeling accuracy, avoid derivative calculations, and expand the scope of application. The nonlinear system, static function SinC and Mackey-Glass chaos mapping are used to verify the effectiveness of the proposed ESN+PSO approach.