Performance comparison of dynamic neural networks as applied to robot inverse kinematic computations

Abstract

This paper compares the performance of two dynamic neural structures, namely the recurrent neural network and the dynamic neural processor. The recurrent neural structure consists of a single layer feedforward (static) network included in a feedback configuration with a time delay. On the other hand, the dynamic neural processor (DNP), proposed by the authors, consists of two nonlinear dynamic neurons coupled to function as antagonistic neural subpopulations with adaptable self- and inter-subpopulation feedback weights. The DNP is developed based on the neuro-physiological evidence that nervous activity of any complexity depends on the interaction of the excitatory and inhibitory neural subpopulations. An algorithm to modify the DNP's weights is developed. The task of computing the inverse kinematic transformations of a two-linked robot is used for comparing the performance of the recurrent neural network and the DNP.