Q-Learning and Double Q-Learning with Sliding Mode Control to stabilize a Ball-and-Beam System

In this paper, Q-learning and Double Q-learning reinforcement learning algorithms were used to fine-tune sliding mode controller parameters to balance the Ball-and-Beam system. Each algorithm was experimented with a 5-action action set and a 12-action action set, and the results were compared. The result is evaluated based on the system response after a disturbed reset. The experiment result shows that Double Q-learning performs better with larger action sets, putting the ball at the center of the beam only after 1.8 seconds. However, Q-learning performs better with smaller action sets, taking up to 2.3 seconds. Double Q-learning generally performed better than Q-learning, taking approximately 0.3 seconds less overall.