A neural network system with reinforcement learning to control a dynamic arm model

A neural network system is presented for controlling a two-link dynamic arm model where the task is to move the arm from any initial position to any final position in the sagittal plane. The controller produces joint torque-time profiles that begin and end with equilibrium values at the initial and final positions, respectively. A memory type neural network is trained by supervised learning methods to predict the joint's static equilibrium torque values corresponding to joint angles. A reinforcement learning network is used to determine the parameters needed for synthesis of the torque-time profiles for each joint. The reinforcement signal is computed based on the distance between the desired end point position and velocity and the states achieved based on the generated torque profiles. The general pattern of the torque-time plots is decided a priori according to the literature. The methods of training and an illustrative example of the algorithm's performance are presented.