Data Generation Feedback Relearning Control for Unmodeled Nonlinear Systems

Yong Zhang, Chaoxu Mu, Dongbin Zhao

Abstract: Reinforcement learning (RL) algorithms require continuous interaction with the controlled plant to optimize the objective function and control policy. However, the security risk and poor real-time performance limit the application of RL algorithms. In this article, a data generation feedback relearning (DGFR) control algorithm is developed to avoid these limitations. The DGFR control algorithm improves the control performance by interacting with a proposed delayed neural network-based data generation model. Considering the time-delay characteristics of industrial production, the data generation model is different from the identifier of the system model, but it is used to generate the target data required by the DGFR control algorithm. The interaction between the algorithm and the data generation model avoids the risk of trial-and-error and improves real-time performance. Finally, the superiority of the proposed control algorithm is verified by giving comparative experiments in conjunction with a power grid system and a permanent magnet synchronous motor system.