Chaos and periodic dynamics in adaptive motion control systems under unknown environment

Predicting and controlling dynamical systems in a previously unknown environment are difficult but challenging problems for adaptive learning and control. We examine a neuron based reinforcement learning algorithm for prediction and control of ball games such as tennis, where dynamical equations, environment, and the behavior of the opponent player are a priori unavailable. We show that stochastic reinforcement learning with a feedforward RBF network is efficient for real time learning and control. Furthermore, reinforcement learning can adapt and control not only periodic motion but also quasi-periodic, and even chaotic orbits of the ball dynamics