A self-organization reconstruction method of ESN reservoir structure based on reinforcement learning

The dynamic reservoir of the randomly generated Echo State Network(ESN) contains numerous redundant neurons, resulting in collinearity in the high-dimensional state space matrix. This collinearity impacts the prediction performance of the network. In order to address this issue, this paper introduces a self-organizing ESN structure optimization model that is based on reinforcement learning, called SR-ESN. The SR-ESN model employs pruning methods to reconstruct the reservoir using contribution and decision mechanisms. To mitigate potential instability caused by high coupling among neurons in a single reservoir, the concept of ensemble learning is applied to create multiple initial reservoir pools, thereby enhancing screening diversity. Simultaneously, the model utilizes reinforcement learning's decision mechanism to identify effective neurons. Neurons with low contribution are pruned, while those with high contribution are retained for self-organizing reconstruction. This optimization of the network structure enhances its prediction performance. Based on both artificial and real datasets, the proposed SR-ESN model demonstrates superior prediction performance with minimal structural complexity compared to other prediction models.