Robust Learning-based Predictive Control for Discrete-time Nonlinear Systems with Unknown Dynamics and State Constraints

Model predictive control (MPC) has been widely employed as an effective method for model-based constrained control. For systems with unknown dynamics, reinforcement learning (RL) and adaptive dynamic programming (ADP) have received notable attention to solving adaptive optimal control problems. Recently, works on the use of RL in the framework of MPC have emerged, which can enhance the ability of MPC for data-driven control. However, the safety under state constraints and the closed-loop robustness are difficult to be verified due to approximation errors of RL with function approximation structures. Aiming at the above problem, we propose a data-driven robust MPC solution based on incremental RL, called data-driven robust learning-based predictive control (dr-LPC), for perturbed unknown nonlinear systems subject to safety constraints. A data-driven robust MPC (dr-MPC) is firstly formulated with a learned predictor. The incremental Dual Heuristic Programming (DHP) algorithm using an actor-critic architecture is then utilized to solve the online optimization problem of dr-MPC. In each prediction horizon, the actor and critic learn time-varying laws for approximating the optimal control policy and costate respectively, which is different from classical MPCs. The state and control constraints are enforced in the learning process via building a Hamilton-Jacobi-Bellman (HJB) equation and a regularized actor-critic learning structure using logarithmic barrier functions. The closed-loop robustness and safety of the dr-LPC are proven under function approximation errors. Simulation results on two control examples have been reported, which show that the dr-LPC can outperform the DHP and dr-MPC in terms of state regulation, and its average computational time is much smaller than that with the dr-MPC in both examples.