Performance-Oriented Data-Driven Control: Fusing Koopman Operator and MPC-Based Reinforcement Learning

This letter develops the machinery of Koopman-based Model Predictive Control (KMPC) design, where the Koopman derived model is unable to capture the real nonlinear system perfectly. We then propose to use an MPC-based reinforcement learning within the Koopman framework combining the strengths of MPC, Reinforcement Learning (RL), and the Koopman Operator (KO) theory for an efficient data-driven control and performance-oriented learning of complex nonlinear systems. We show that the closed-loop performance of the KMPC is improved by modifying the KMPC objective function. In practice, we design a fully parameterized KMPC and employ RL to adjust the corresponding parameters aiming at achieving the best achievable closed-loop performance.