Barrier subgradient method (Q633113): Difference between revisions
From MaRDI portal
Removed claim: author (P16): Item:Q327227 |
Changed an Item |
||
Property / author | |||
Property / author: Yu. E. Nesterov / rank | |||
Normal rank |
Revision as of 10:17, 13 February 2024
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Barrier subgradient method |
scientific article |
Statements
Barrier subgradient method (English)
0 references
31 March 2011
0 references
The author focusses on a class of problems of minimizing a nonsmooth convex function over a feasible set endowed by a self-concordant barrier. After studying the smoothing of the support function of a convex set by a self-concordant barrier, the author describes the corresponding barrier subgradient method (BSM). If \(\varepsilon\) is the required absolute accuracy, then the complexity bound of the BSM is \(O(\frac{1}{\varepsilon^2})\). The method is applied to maximize a non-negative concave function. If \(\delta\) is the required accuracy in this case, then the result is an approximate solution in the relative scale with the complexity bound \(O(\frac{1}{\delta^2})\). Several applications of BSM are presented: fractional covering problems, maximal concurrent flow problems, minimax problems with nonnegative components and semidefinite relaxation of Boolean quadratic problems. BSM is also applied to online optimization, as an alternative to the single-stage stochastic optimization. The practical domains of decision-making in uncertain environments, portfolio management and processes with full production cycles are taken into account. For all these problems, the rate of convergence of the method described in this paper is independent on the problem's data.
0 references
convex optimization
0 references
subgradient methods
0 references
non-smooth optimization
0 references
minimax problems
0 references
saddle points
0 references
variational inequalities
0 references
stochastic optimization
0 references
black-box methods
0 references
lower complexity bounds
0 references