A Dueling-Double-Deep Q-Network Controller for Magnetic Levitation Ball System

Deep reinforcement learning technique has attracted extensive attention in the field of control with the rapid development of artifical intelligence. Model-free control methods perform more effcient than traditional model-based control method for complex nonlinear system since they don't require accurate models. In this study, by using the Dueling-Double-DQN (Dueling-DDQN, 3-DQN) algorithm, a self-learning controller is developed for the magnetic levitation ball system. The numerical simulation demonstrates the effectiveness of our controller. Moreover, we compare the above controller with the DQN-based self-learning controller. Numerical simulation shows that the attraction domain of the Dueling-DDQN controller is larger that the latter.