TY - JOUR
T1 - An extended delayed weighted gradient algorithm for solving strongly convex optimization problems
AU - Andreani, R.
AU - Oviedo, H.
AU - Raydan, M.
AU - Secchin, L. D.
N1 - Funding Information:
The first author was financially supported by FAPESP, Brazil (Projects 2013/05475-7 and 2017/18308-2 ), CEPID-CeMEAI, Brazil ( FAPESP 2013/07375-0 ) and the National Council for Scientific and Technological Development – CNPq, Brazil (Projects 301888/2017-5 and 306988/2021-6 ). The second author was financially supported by FGV, Brazil (Fundação Getulio Vargas) through the excellence post-doctoral fellowship program. The third author was financially supported by the Fundação para a Ciência e a Tecnologia (Portuguese Foundation for Science and Technology) through the projects UIDB/00297/2020 and UIDP/00297/2020 (Center for Mathematics and Applications). The fourth author was financially supported by FAPES (Fundação de Amparo à Pesquisa e Inovação do Espírito Santo) (grant 116/2019 ) and CNPq (Project 309136/2021-0 ). We would like to thank two anonymous referees and the Principal Editor for their comments and suggestions that helped us to improve the final version of this paper.
Publisher Copyright:
© 2022 Elsevier B.V.
PY - 2022/12/15
Y1 - 2022/12/15
N2 - The recently developed delayed weighted gradient method (DWGM) is competitive with the well-known conjugate gradient (CG) method for the minimization of strictly convex quadratic functions. As well as the CG method, DWGM has some key optimality and orthogonality properties that justify its practical performance. The main difference with the CG method is that, instead of minimizing the objective function on the entire explored subspace, DWGM minimizes the 2-norm of the gradient vector on the same subspace. The main purpose of this study is to extend DWGM for solving strongly convex nonquadratic minimization problems while keeping a low computational cost per iteration. We incorporate the scheme into a tolerant line search globalization strategy, and we show that it exhibits q-linear convergence to the unique global solution. We compare the proposed extended DWGM with state-of-the-art methods for large-scale unconstrained minimization problems. We use some well-known strongly convex test problems, but also solve some regularized logistic regression problems that appear in machine learning. Our numerical results illustrate that the proposed scheme is promising and exhibits a fast convergence behavior. Moreover, we show through numerical experiments on CUTEst problems that the proposed extended DWGM can be very effective in accelerating the convergence of a well-established Barzilai–Borwein-type method when the iterates get close to minimizers of non-convex functions.
AB - The recently developed delayed weighted gradient method (DWGM) is competitive with the well-known conjugate gradient (CG) method for the minimization of strictly convex quadratic functions. As well as the CG method, DWGM has some key optimality and orthogonality properties that justify its practical performance. The main difference with the CG method is that, instead of minimizing the objective function on the entire explored subspace, DWGM minimizes the 2-norm of the gradient vector on the same subspace. The main purpose of this study is to extend DWGM for solving strongly convex nonquadratic minimization problems while keeping a low computational cost per iteration. We incorporate the scheme into a tolerant line search globalization strategy, and we show that it exhibits q-linear convergence to the unique global solution. We compare the proposed extended DWGM with state-of-the-art methods for large-scale unconstrained minimization problems. We use some well-known strongly convex test problems, but also solve some regularized logistic regression problems that appear in machine learning. Our numerical results illustrate that the proposed scheme is promising and exhibits a fast convergence behavior. Moreover, we show through numerical experiments on CUTEst problems that the proposed extended DWGM can be very effective in accelerating the convergence of a well-established Barzilai–Borwein-type method when the iterates get close to minimizers of non-convex functions.
KW - Conjugate gradient methods
KW - Gradient methods
KW - Large-scale optimization
KW - Strongly convex functions
UR - http://www.scopus.com/inward/record.url?scp=85133180918&partnerID=8YFLogxK
U2 - 10.1016/j.cam.2022.114525
DO - 10.1016/j.cam.2022.114525
M3 - Article
AN - SCOPUS:85133180918
SN - 0377-0427
VL - 416
JO - Journal of Computational and Applied Mathematics
JF - Journal of Computational and Applied Mathematics
M1 - 114525
ER -