Multi-reservoir echo state network with five-elements cycle

Abstract

Echo state network (ESN) is reservoir computing model that effectively replace recurrent neural network (RNN). However, building reservoir in traditional ESN often has randomness, making it difficult to effectively determine the reservoir that matches a given task. Therefore, a multi-reservoir ESN based on five-elements cycle (FEC-MRESN) is proposed to design the reservoir automatically in this paper. First, FEC-MRESN designs a pruning algorithm using a top-down strategy to remove redundant neurons from the reservoir based on the generation and restriction relations between elements in the five-elements cycle. Second, based on the reservoir neurons retained by the pruning algorithm, an exponential weight assignment method is studied to achieve deterministic assignment of reservoir weights. Finally, FEC-MRESN is tested on some time series benchmark datasets. The experimental results show that FEC-MRESN not only improves prediction accuracy, but also removes redundant reservoir neurons, improves network generalization performance and training efficiency.