Application of reinforcement learning control to a nonlinear bouncing cart

We consider a nonlinear bouncing cart motion, controlled by reinforcement learning (RL) control. The learning algorithm consists of Q-learning and advantage updating (AU) to keep the cart within desired limits. Q-learning is a RL algorithm that applies "delayed reinforcement" and performs optimal actions to maximize return values whereby the system performance is evaluated. RL is also extended through the use of AU in continuous-time. AU is another RL algorithm that stores both value function and advantage function, representing an estimate of the degree to which the expected total discounted reinforcement is increased by performing action other than the action currently considered to be best.