Robotic System Sensitivity to Neural Network Learning Rate: Theory, Simulation, and Experiments

Selection of neural network learning rates to obtain satisfactory performance from neural network controllers is a challenging problem. To assist in the selection of learning rates, this paper investigates robotic system sensitivity to neural network (NN) learning rate. The work reported here consists of experimental and simulation results. A neural network controller module, developed for the purpose of experimental evaluation of neural network controller performance of a CRS Robotics Corporation A460 robot, allows testing of NN controllers using real-time iterative learning. The A460 is equipped with a joint position proportional, integral, and derivative (PID) controller. The neural network module supplies a signal to compensate for remaining errors in the PID-controlled system. A robot simulation, which models this PID-controlled A460 robot and NN controller, was also developed to allow the calculation of sensitivity to the NN learning rate. This paper describes the implementation of three NN architectures: the error back-propagation (EBP) NN, mixture of experts (ME) NN, and manipulator operations using value encoding (MOVE) NN. The sensitivity of joint trajectory error of three NN controllers to learning rate was investigated using both simulation and experimentation. Similar results were obtained from the robot experiments and the dynamic simulation. These results of state sensitivity to NN learning rate confirm that the MOVE NN is least sensitive to learning rate, implying that selection of suitable learning rates for this NN architecture for the system considered is accomplished more readily than other NN architectures. KEY WORDS—neural network control, robotic systems, sensitivity The International Journal of Robotics Research Vol. 19, No. 10, October 2000, pp. 955-968, ©2000 Sage Publications, Inc.