A reinforcement learning adaptive fuzzy controller for robots. (Q1398938)

A new reinforcement learning scheme for a class of serial link robot arms is presented. In previous work, it is typical for behavior to be learned through trial-and-error interactions with a dynamic environment. In the author's proposed learning scheme, an agent collects signals from a fixed gain controller, from an adaptive critic element, and a fuzzy action-generating element. The mechanism for action generation is a fuzzy approximator with a set of tunable parameters. A performance measurement mechanism sends an error measurement to the adaptive critic element in order to generate and transfer a reinforcement learning signal to the agent. A tuning algorithm, based on Lyapunov stability theory, is used to guarantee both tracking performance and stability. Simulation results are provided for two examples that demonstrate the proposed reinforcement adaptive fuzzy control scheme. The examples involve balancing a rigid pole mounted on a cart and a two-degree-of-freedom robotic manipulator.