There is a recent interest on first-order methods for linear programming (LP). In this paper, we propose a stochastic algorithm using variance reduction and restarts for solving sharp primal-dual problems such as LP. We show that the proposed stochastic method exhibits a linear convergence rate for sharp instances with a high probability, which improves the complexity of the existing deterministic and stochastic algorithms. In addition, we propose an efficient coordinate-based stochastic oracle for unconstrained bilinear problems, which has $\mathcal O(1)$ per iteration cost and improves the total flop counts to reach a certain accuracy.
翻译:最近有人对线性编程的第一阶方法(LP)感兴趣。 在本文中,我们建议采用一种随机算法,使用差异减少法和重新启动法来解决尖锐的初质问题,如LP。 我们表明,拟议的随机方法显示,极有可能的尖锐事件呈线性趋同率,这提高了现有确定和随机算法的复杂性。此外,我们建议对未受限制的双线性问题采用一种高效的基于协调的抽查法,每迭接费用为$\mathcal O(1)美元,并改进总浮点以达到一定的精确度。