Simulation of Varying Parameter Recurrent Neural Network with application to matrix inversion

Abstract

A class of adaptive recurrent neural networks (RNN) for computing the inverse of a time-varying matrix with accelerated convergence time is defined and considered. The proposed neural dynamic model involves an exponential gain time-varying term in the nonlinear activation of the finite-time Zhang neural network (FTZNN) dynamical equation. Individual models belonging to the proposed class are defined by means of corresponding error functions. It is shown theoretically and experimentally that usage of the exponential nonlinear activation accelerates the convergence rate of the error function compared to previous dynamical systems for solving the time-varying (TV) and time-invariant (TI) matrix inversion.