Deep reinforcement learning-based model predictive control of uncertain linear systems

This paper introduces deep reinforcement learning (DRL)-based model predictive control (MPC) for constrained linear systems with bounded additive disturbances. The essential idea of this work lies in a novel integration of MPC into DRL. In the proposed method, rigid tube MPC is utilized to obtain predicted state and action sequences. Then the predicted sequences are employed in the soft actor-critic (SAC) algorithm to form temporal difference (TD) targets and to update the policy online. By utilizing generated TD targets in the DRLMPC framework, the SAC algorithm improves the optimality without requiring a large number of real datasets for pre-training while ensuring safety during the control process. Moreover, the proposed method has a comparable computational complexity to rigid tube MPC. Finally, numerical simulations and comparisons are provided to demonstrate that our framework results in a better performance at the cost and convergence.