This work explores the implementation of reinforcement learning (RL)-based approaches to replace model predictive control (MPC) in cases where practical implementations of MPC are infeasible due to excessive computation times. Specifically, with the use of externally enforced stability guarantees, an RL-based controller that is trained to optimize the same cost function as the MPC with a long horizon that achieves the desirable closed-loop performance can serve as a potentially more appealing real-time option as opposed to using the same MPC with a shorter horizon. A benchmark nonlinear chemical process model is used to demonstrate the feasibility of this RL-based framework that simultaneously guarantees stability and enables improvements in computational efficiency and potential control quality of the closed-loop system. To explore the influence of the RL training method, two RL algorithms are explored, with one imitation learning method used as a reference.
Khodaverdian et al. (Mon,) studied this question.