Block-structured echo state network based on error reduction mechanism

Abstract

The echo state network (ESN) is a special recurrent neural network, which is a powerful method of time series prediction. However, the traditional ESN with single reservoir cannot fully mine the feature information of complicated time series. In this article, a block-structured echo state network (BESN) with cascaded modules is proposed to solve this problem based on error reduction mechanism. In BESN, the external inputs and the outputs of the previous module form the inputs of the next adjacent module, and the prediction errors of the previous module are defined as the target outputs of the next module. Meanwhile, the number of modules is determined by a self-organizing method for BESN. Finally, the performance of BESN is tested on two benchmarks.