Applications of Linear Adaptive Dynamic Programming (ADP) to a Non-linear Four-bar Mechanism
Emil Mühlbradt Sveen, Jing Zhou · 2023
This paper investigates the feasibility of using learning based Adaptive Dynamic Programming (ADP) for linear time-invariant systems to control the position of a highly nonlinear four-bar system. An off-line off-policy controller iteration approach, inspired by Reinforcement Learning, is used to find a near-optimal controller for the four-bar crank motor. The results show that convergence of a near-optimal linear control policy is possible under the chosen exploration strategy, and the paper further exploit the non-linearity of the four-bar to investigate the limits of the exploration strategy parameters to shed light on the non-triviality of selecting a suitable exploration strategy in ADP. The efficiency of the ADP approach is demonstrated, as just 1.4 seconds of simulation time is required to convergence to a nearontimal ADP controller.