Robust Distributed Online Convex Optimization

该奖项的研究目标是研究在联网的多代理系统上进行在线优化的分布式算法。在线优化是指在信息是动态的，而不是先验的，并且随着时间的推移越来越多地揭示的情况下，最佳地利用有限的代理资源。当前在线优化方法的一个基本假设是在中心位置有信息可用。在信息在代理之间分发的联网场景中，这一假设变得有问题。将所有数据传输到中央位置可能成本高昂或效率低下，并会引发隐私问题和信息泄露的可能性。这项研究将导致设计出健壮的分布式策略，以处理现实世界应用程序中存在的多个中断来源。研究方法从通过鞍点动力学综合在线分布式算法到发展具有可证明的正确性保证的严格的数学分析。交付成果包括可证明正确的在线分布式策略目录、用于评估性能和复杂性的新概念和工具、通过计算机模拟进行的演示和验证、研究结果的文档记录以及工程专业学生教育。联网的机器人系统在海洋勘探、灾难恢复、环境监测和监视等广泛的现代应用中具有巨大的影响。科学和军事领域目前的趋势是在不确定的情况下部署这些系统，在这种情况下，通信受到限制，容易出现故障和干扰。这项研究的成功结果将使联网系统能够在这些高度分散的情况下稳健而有效地运行。这项研究将广泛传播，并对各种教育活动产生积极影响，包括高中、本科和研究生教育，本科生参与研究，研究生监督，以及保留少数族裔在STEM学科。

The research objective of this award is to investigate distributed algorithms for online optimization over networked multi-agent systems. Online optimization refers to the best use of limited agent resources in scenarios where information is dynamic, not a priori available, and increasingly revealed over time. An underlying assumption of present online optimization approaches is the availability of information at a central location. This assumption becomes problematic in networked scenarios, where information is distributed among agents. Transmitting all data to a central location might be costly or inefficient, and raises privacy concerns and the possibility of information leakage. The research will result in the design of robust, distributed strategies that can deal with multiple sources of disruption present in real-world applications. The research approach progresses from the synthesis of online distributed algorithms via saddle-point dynamics to the development of rigorous mathematical analysis with provably correct guarantees. Deliverables include a catalog of provably correct online distributed strategies, novel concepts and tools for the evaluation of performance and complexity, demonstration and validation via computer simulations, documentation of research results, and engineering student education.Networked robotic systems have an enormous impact in a wide range of modern applications, including oceanographic exploration, disaster recovery, environmental monitoring, and surveillance. The current trend in scientific and military domains points towards the deployment of these systems in uncertain scenarios where communication is limited and subject to failure and interference. A successful outcome of this research will enable the robust and efficient operation of networked systems in these highly decentralized scenarios. The research will be broadly disseminated and positively impact a variety of educational activities, including high-school, undergraduate, and graduate education, involvement of undergraduates in research, supervision of graduate students, and retention of minorities in STEM disciplines.