Using Neural Reinforcement Controllers in Robotics

Reinforcement Learning is a promising paradigm for the training of intelligent controllers when only a minimum of a priori knowledge is available. This is especially true for situations, where uncertainties and nonlinearities in sensors and actuators make the application of conventional control design techniques impossible or at least a very hard task. The article describes a neural control architecture based on the asynchronous dynamic programming approach (Barto et al., 1993). Its capabilities are shown on two applications from the domain of robotics.