Zeroth-Order Algorithms for Smooth Saddle-Point Problems
نویسندگان
چکیده
Saddle-point problems have recently gained increased attention from the machine learning community, mainly due to applications in training Generative Adversarial Networks using stochastic gradients. At same time, some only a zeroth-order oracle is available. In this paper, we propose several algorithms solve smooth (strongly) convex-concave saddle-point oracles and estimate their convergence rate its dependence on dimension $n$ of variable. particular, our analysis shows that case when feasible set direct product two simplices, for term by $\log n$ factor worse than first-order methods. We also consider mixed setup develop 1/2th-order methods use minimization part maximization part. Finally, demonstrate practical performance problems.
منابع مشابه
Domain Decomposition Algorithms for Saddle Point Problems
In this paper, we introduce some domain decomposition methods for saddle point problems with or without a penalty term, such as the Stokes system and the mixed formulation of linear elasticity. We also consider more general nonsymmetric problems, such as the Oseen system, which are no longer saddle point problems but can be studied in the same abstract framework which we adopt. Several approach...
متن کاملFrank-Wolfe Algorithms for Saddle Point Problems
We extend the Frank-Wolfe (FW) optimization algorithm to solve constrained smooth convex-concave saddle point (SP) problems. Remarkably, the method only requires access to linear minimization oracles. Leveraging recent advances in FW optimization, we provide the first proof of convergence of a FW-type saddle point solver over polytopes, thereby partially answering a 30 year-old conjecture. We a...
متن کاملMultilevel Gradient Uzawa Algorithms for Symmetric Saddle Point Problems
In this paper, we introduce a general multilevel gradient Uzawa algorithm for symmetric saddle point systems. We compare its performance with the performance of the standard Uzawa multilevel algorithm. The main idea of the approach is to combine a double inexact Uzawa algorithm at the continuous level with a gradient type algorithm at the discrete level. The algorithm is based on the existence ...
متن کاملResidual reduction algorithms for nonsymmetric saddle point problems
In this paper, we introduce and analyze Uzawa algorithms for non-symmetric saddle point systems. Convergence for the algorithms is established based on new spectral results about Schur complements. A new Uzawa type algorithm with optimal relaxation parameters at each new iteration is introduced and analyzed in a general framework. Numerical results supporting the efficiency of the algorithms ar...
متن کاملذخیره در منابع من
با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید
ژورنال
عنوان ژورنال: Communications in computer and information science
سال: 2021
ISSN: ['1865-0937', '1865-0929']
DOI: https://doi.org/10.1007/978-3-030-86433-0_5