Reinforcement Learning-based Data-driven Control Design for Motion Control Systems

Motion control systems are widely used in many fields of industry. Conventional control schemes are highly dependent on the system model to be designed. The performance of design would be greatly reduced, when the system exists unknown disturbances or uncertainty. Therefore, some scholars pointed out that the dependency on the system models can be eliminated by data-driven design schemes. In this paper, the reinforcement learning-based methods are included, which appeal to attentions gradually. The disturbances rejection problem for motion control systems is studied based on reinforcement learning. Considering the continuity of state space and action space, a method based on deep reinforcement learning algorithm is proposed to reject the periodic disturbances. Proposed deep deterministic policy gradient (DDPG) and twin delayed deep deterministic policy gradient (TD3) based algorithms are compared in simulation. The simulation results show that the periodic disturbances of the motion control systems can be rejected effectively with the proposed reinforcement learning controller.