Sarsa-based Model Predictive Control with Improved Performance and Computational Complexity

作者：Tianxiang Lu, Kunwu Zhang, Yang Shi · 发表于：Industrial Cyber-Physical Systems · 年份：2022 · DOI：10.1109/icps51978.2022.9816896 · 被引用次数：4 · 研究领域：Computer Science

This paper investigates Sarsa-based model predictive control (MPC) for constrained discrete-time linear systems with mixed uncertainties, including parametric uncertainties and external additive disturbances. With a system subject to parametric uncertainties, i.e., an inaccurate system model in hand, the Sarsa-based update policy is constructed for the intention of preserving the optimality brought by MPC. A linear state feedback control policy via MPC is applied to achieve the regulation objective. By incorporating these two techniques into the standard tube MPC framework, the Sarsa-based MPC which essentially is a data-driven reinforcement learning-based method is proposed. The MPC parameters which are subject to the Sarsa-based update policy are specified. They include the approximated model parameters, a linear feedback control gain, and an auxiliary disturbance set used to enhance the robustness of tube MPC. The proposed method is computational inexpensive and robust. This is validated by a numerical example.