A Class of Diagonal QuasiNewton Methods for Large

Authors: Wah June Leong

Publish Date: 2015/04/10

Volume: 39, Issue: 4, Pages: 1659-1672

PDF Link

Abstract

We study the convergence properties of a class of low memory methods for solving largescale unconstrained problems This class of methods belongs to that of quasiNewton family except for which the approximation to Hessian at each step is updated by means of a diagonal matrix Using appropriate scaling we show that the methods can be implemented so as to be globally and Rlinearly convergent with standard inexact line searches Preliminary numerical results suggest that the methods are good alternative to other low memory methods such as the CG and spectral gradient methodsThis work was partially done while the author was visiting Institute of  Computational Mathematics and Scientific/Engineering Computing Academy of Mathematics and Systems Science During the visit the author was supported by the joint Chinese Academy of SciencesAcademy of Sciences for the Developing World CASTWAS Fellowship no 3240157252

Keywords: