CIF: Small: Taming Convergence and Delay in Stochastic Network Optimization with Hessian Information

CIF：小：利用 Hessian 信息驯服随机网络优化中的收敛和延迟

• 批准号: 2110252
• 负责人: Jia Liu
• 金额: $ 31.79万
• 依托单位: Ohio State University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-10-01 至 2022-06-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2110252&HistoricalAwards=false
• 关键词: CIF; Small; Taming; Convergence; Delay

With the rapid integration of massive amounts of data and new network devices, today's network infrastructures are being stretched to their limits. As a result, recent years have witnessed a critical need for developing fast-converging distributed stochastic network control and optimization algorithms to increase throughput and reduce delay. This research program addresses the challenge of distributed control and optimization for next generation complex network systems, where the rapidly changing network states (e.g., network topologies, channel states, queueing states, etc.) necessitate fast-convergence and low-delay in distributed optimization algorithms. Based on the investigator's recent research on network control and optimization that leverages second-order Hessian information (SOHI), this research will develop a series of new distributed algorithmic techniques that offer orders of magnitudes improvements in both convergence speed and queueing delay compared to the traditional approaches, while attaining the same provable network-utility optimality. Specifically, the investigator?s research tasks in this project are organized around three inter-related research thrusts that exploit different degrees of SOHI: i) Heavy-ball-based joint congestion control and multi-path routing (partial SOHI); ii) Primal-dual interior-point second-order congestion control and multi-path routing (full SOHI); and iii) SOHI-based distributed control and optimization algorithm designs. This research project takes an integrated and holistic approach that draws techniques from areas of mathematical modeling, optimization theory, control theory, queueing theory, and stochastic analysis. The research project will not only advance the knowledge in the algorithmic design for next generation complex networks, but will also serve a critical need in the general networking research community by exploring new frontiers in SOHI-based network control and optimization.

随着海量数据和新网络设备的快速集成，当今的网络基础设施正在达到极限。因此，近年来迫切需要开发快速收敛的分布式随机网络控制和优化算法，以提高吞吐量并减少延迟。该研究项目解决了下一代复杂网络系统的分布式控制和优化的挑战，其中快速变化的网络状态（例如网络拓扑、通道状态、排队状态等）需要分布式优化中的快速收敛和低延迟算法。基于研究者最近利用二阶 Hessian 信息 (SOHI) 的网络控制和优化研究，本研究将开发一系列新的分布式算法技术，与传统的算法相比，这些技术在收敛速度和排队延迟方面都有数量级的改进方法，同时获得相同的可证明的网络效用最优性。具体来说，该项目中研究者的研究任务围绕三个相互关联的研究主旨进行组织，这些主旨利用了不同程度的 SOHI： i）基于重球的联合拥塞控制和多路径路由（部分 SOHI）； ii) 原对偶内点二阶拥塞控制和多路径路由（全SOHI）； iii) 基于SOHI的分布式控制和优化算法设计。该研究项目采用综合全面的方法，借鉴了数学建模、优化理论、控制理论、排队论和随机分析等领域的技术。该研究项目不仅将推进下一代复杂网络算法设计的知识，而且还将通过探索基于 SOHI 的网络控制和优化的新领域来满足一般网络研究界的关键需求。