Delta rule-based neural networks for inverse kinematics: error gradient reconstruction replaces the teacher

Abstract

Control tasks which feed back a scalar error signal (critic) to a controlling neural network form a more general class than those which provide a teacher that is ordinarily required by delta-rule-based networks like backpropagation or CMAC networks. The author introduces an interface that builds teacher vectors from critic values by reconstruction of the gradient of the critic function. Backpropagation networks have been trained by this method to learn the inverse kinematics of simulated planar manipulators. Different strategies for efficient sampling of critic values with respect to restrictions imposed by a real robot arm are proposed, and simulation results are reported