arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2008.10062cs.DS
收藏

流式子模匹配满足原-对偶方法

Streaming Submodular Matching Meets the Primal-Dual Method

Roie Levin, David Wajc

摘要:

我们研究了流子模最大化匹配/ $b$匹配约束(MSM/MSbM),并提出了改进的上限和下限为这些问题。在上界方面,我们给出了原始-对偶算法,实现了以下近似比。$\bullet$ $3+2\sqrt{2}\approx 5.828$用于单调MSM,提高了以前最好的$7.75$比例。$\bullet$ $4+3\sqrt{2}\approx 7.464$用于非单调MSM,提高了以前最好的$9.899$比例。$\bullet$ $3+ε$用于最大权重b匹配,改进了以前的最佳比率$4+ε$。在下界方面,我们改进了MSM的$\frac{e}{e-1}\approx 1.582$的最佳下界,并给出了基于ETH的多时间单调MSM流算法的$\approx 1.914$的下界。我们最重要的贡献是算法技术。我们发现,(随机)原始-对偶方法,起源于最大权重匹配(MWM)的研究,也是有用的MSM的上下文中。据我们所知,这是第一次使用基于原始-对偶的分析进行流子模块优化。我们还展示了如何重新解释以前的算法MSM在我们的框架,因此,我们希望我们的工作是一个步骤,统一新老技术流子模块最大化,它铺平了道路,进一步的新成果。

英文摘要:

We study streaming submodular maximization subject to matching/$b$-matching constraints (MSM/MSbM), and present improved upper and lower bounds for these problems. On the upper bounds front, we give primal-dual algorithms achieving the following approximation ratios. $\bullet$ $3+2\sqrt{2}\approx 5.828$ for monotone MSM, improving the previous best ratio of $7.75$. $\bullet$ $4+3\sqrt{2}\approx 7.464$ for non-monotone MSM, improving the previous best ratio of $9.899$. $\bullet$ $3+ε$ for maximum weight b-matching, improving the previous best ratio of $4+ε$. On the lower bounds front, we improve on the previous best lower bound of $\frac{e}{e-1}\approx 1.582$ for MSM, and show ETH-based lower bounds of $\approx 1.914$ for polytime monotone MSM streaming algorithms. Our most substantial contributions are our algorithmic techniques. We show that the (randomized) primal-dual method, which originated in the study of maximum weight matching (MWM), is also useful in the context of MSM. To our knowledge, this is the first use of primal-dual based analysis for streaming submodular optimization. We also show how to reinterpret previous algorithms for MSM in our framework; hence, we hope our work is a step towards unifying old and new techniques for streaming submodular maximization, and that it paves the way for further new results.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑