Data-Driven Model Predictive Control With Reinforcement Learning for Linear Time-Invariant Systems

We propose a data-driven reinforcement learning model predictive control (DD-RLMPC) scheme for linear time-invariant (LTI) systems. The scheme integrates reinforcement learning (RL) and data-driven model predictive control (DD-MPC) through value iteration. Also, a value function approximation technique is applied to approximate the terminal cost, thereby providing a direct method based on behavioral systems theory, thus using historical operation data to bypass the system identification step. The scheme first operates offline to derive the optimal approximated value function, and then operates online for controller design. Furthermore, the proposed DD-RLMPC scheme offers flexibility in selecting the prediction horizon, thus provides a potential to significantly reduce the computational burden compared to the terminal equality-constrained DD-MPC methods. We demonstrate the convergence, stability, and feasibility of the proposed DD-RLMPC scheme, with properties verified by simulation results.