A growing and pruning method for a history stack neural network based adaptive controller

This paper presents a neural network based control strategy for adaptive control of a robotic manipulator. The neural network learns the inverse dynamics of the robotic manipulator while controlling the robot on-line without any a priori knowledge of the manipulator inertial parameters or the equation of dynamics. The only assumptions that must be made about the target system are the number of inputs and outputs to the system. A history stack algorithm is used to facilitate simultaneous control and learning. Learning performance is improved by growing and pruning neurons from the neural network based on the magnitude of the trajectory error. Simulation of a two degree of freedom serial link manipulator allows verification of the effectiveness of the algorithm. Results show improved performance in comparison to a controller using the history stack alone.