Improving Initial Transients of Online Learning Echo State Network Control System via Feedback Adjustment

Abstract

Echo state networks (ESNs) have gained popularity in online learning control systems due to their easy training. However, online learning ESN controllers often undergo slow convergence and produce unexpected outputs during the initial transient phase. Existing solutions, such as prior training or control mode switching, can be complex and have drawbacks. This work offers a simple yet effective method to address these initial transients by integrating a feedback proportional-differential (P-D) controller. Simulation results show that the proposed control system exhibits fast convergence and strong robustness against plant dynamics and hyperparameter changes. This work is expected to offer practical benefits for engineers seeking to implement online learning ESN control systems.