A reinforcement learning fuzzy controller for set-point regulator problems

Intelligent control refers to controllers that can analyze their performance and make necessary changes to their behavior in order to satisfy certain predefined control goals. This paper describes a self-learning controller model that can efficiently learn the control law for complex systems through reinforcement learning techniques and dynamic programming-like algorithms. The controller is applied to a class of problems called general set-point regulator problems in which the objective is to drive the system to the set-point while optimizing some performance objective function, making no a priori assumptions about the dynamics of the plant or its optimal trajectory. The relevant tasks for a self-learning controller are discussed. Learning is accomplished via incremental, online dynamic programming-like algorithms. Both temporal differences and Q-learning are used in the learning algorithm. Experimental results with both are reported on the inverted pendulum balancing problem, the power system stabilization problem, and the tethered satellite system retrieval problem.