Efficient Self-Learning Tracking Control for Switched Systems: A Triggered-Learning Model Predictive Control Method

Existing model predictive control (MPC) methods lack online learning capability in complex environments. Reinforcement learning (RL) requires a lot of data and computing resources to obtain optimal control. This paper uses efficient self-learning to solve this problem. “Efficient” refers to the use of triggered-learning mechanism (TLM) to manage computing resources on demand. This paper proposes a triggered-learning model predictive control (TL-MPC) method for switched systems. The proposed TL-MPC endows MPC with learning capabilities through the TLM. TLM includes a Deep Deterministic Policy Gradient (DDPG) based control incremental self-learning tuning strategy and a performance-driven event-triggering strategy. The first strategy is to give the MPC controller a control increment to optimize control effect. The second strategy is to realize the on-demand learning and reduce computational resources by comparing two cost functions that characterize the system performance. In addition, the stability of switched systems under TL-MPC is analyzed using the Lyapunov function and the average dwell time technique. Finally, the effectiveness of the proposed method is verified by simulation.