Mechanical system modelling using recurrent neural networks via quasi-Newton learning methods

Abstract

This paper describes a set of efficient learning algorithms for recurrent neural networks (RNN) to facilitate the nonlinear modelling of mechanical systems. These learning algorithms are based on the quasi-Newton methods that estimate the inverse Hessian of an objective function from the gradient to enable Newton-like optimization algorithms. The simulation results with two Boolean functions indicate that the new algorithms based on the classical quasi-Newton methods are about two orders of magnitude faster than the steepest descent method. Furthermore, the learning algorithms based on the quasi-Newton with initial scaling and self-scaling are even more efficient than the classical quasi- Newton methods. For instance, the self-scaling method is three orders of magnitude faster than the steepest descent method. To validate the usefulness of the RNNs in nonlinear mechanical system modelling, RNNs are trained to emulate the step response of a robot arm and identify an adequate model of a 20 HP screw compressor from its operating data.