LQR controller design for affine LPV systems using reinforcement learning (Q6577130)

A multiple-input multiple-output (MIMO) continuous-time linear parameter-varying (LPV) system is considered. The objective is to find a data-driven sub-optimal state feedback policy that minimizes a performance index and stabilizes the system over the entire parameter domain, using only available input-state data from suitable vertex systems. First, under suitable assumptions, a model-based reinforcement learning algorithm is developed to design a sub-optimal control policy. The stability guarantee and convergence properties of the algorithm are discussed. Next, an on-policy data-driven algorithm is presented. Finally, an off-policy algorithm is developed for offline learning scenarios. A numerical example and an air vehicle system are used to demonstrate the effectiveness of the proposed approach.