Data-Driven MPC for Linear Systems using Reinforcement Learning

This paper proposes a novel scheme to solve the optimal control problem for unknown linear systems in a data driven manner. The method doesn't require any prior knowledge of the system, and only utilizes past input-output trajectories to describe the system features implicitly and realize the prediction on the basis of behavioral systems theory. Meanwhile, we adopt reinforcement learning to update the terminal cost function online to ensure the closed-loop stability. The merit of the proposed scheme is the avoiding of the system identification process and the complex design process of terminal cost, terminal set and terminal controller in the standard MPC. We verify the performance of the algorithm by simulation.