Title: Strong Wolfe condition-based variable stacking length multi-gradient parameter identification algorithm

Authors: Yiqiao Shi; Shaoxue Jing

Addresses: School of Intelligent Manufacturing, Jiangsu Vocational College of Electronics and Information, Huaian, Jiangsu 223003, China ' School of Physics and Electronic Electrical Engineering, Huaiyin Normal University, Huaian, Jiangsu 223300, China

Abstract: This paper considers the acceleration of the gradient algorithm for the linear models. Traditional stochastic gradient algorithm requires less computation, but it converges to the true parameter slowly. To accelerate gradient algorithm, a novel gradient algorithm using several gradients is proposed. One important issue of the proposed algorithm is how to determine the stacking length. The stacking length defines the number of gradients used in each recursion. A variable stacking length based on the strong Wolfe condition is presented to enable the algorithm to converge faster. The stacking length obtained by using the SW condition can ensure that the proposed multi-gradient algorithm converges faster. Several experiments are made to validate the proposed algorithm.

Keywords: parameter estimation; stochastic gradient; multi-gradient; strong Wolfe condition; convergence speed.

DOI: 10.1504/IJMIC.2022.128315

International Journal of Modelling, Identification and Control, 2022 Vol.41 No.4, pp.289 - 294

Received: 08 Dec 2021
Accepted: 18 Jan 2022
Published online: 17 Jan 2023 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article