A Weight Optimization Method of Deep Echo State Network Based on Improved Knowledge Evolution

As a deep network, deep echo state network (DESN) is equipped with the reservoir as the core and has great prediction performance with lower computational cost. It has been widely used in the field of time series prediction. However, during the construction process of DESN, the weights of reservoirs such as the input weights and the internal weights are randomly initialized and not adjusted after initialization, which limits the performance of DESN. The traditional weight optimization methods can improve the performance, but they require a lot of training samples and training time. Therefore, this paper firstly introduces the knowledge evolution into the weight optimization of DESN, and then proposes a weight optimization method of DESN based on improved knowledge evolution. First, the weights of DESN are split into two parts by the weight-level splitting, namely the fit-hypothesis and the reset-hypothesis. Second, the sailfish optimizer is used to train the fit-hypothesis. Finally, simulation results on three time series datasets indicate that DESN optimized by this method has better prediction performance and requires less training samples and training time, compared with the one optimized by other existing weight optimization methods.