Robust Decentralized Control of Large-Scale Networked Systems: Fundamental Limits and Data-Driven Feedbacks

感知和反馈的概念一直是工程科学、生物系统以及社会科学和经济学的许多领域的重要组成部分。网络科学在这些学科中的出现揭示了许多感知和反馈的新机会。与提供大量有用数据的各种传感器的先进性和可用性相结合，联网系统提供了允许执行非常复杂的任务的能力。网络化网络物理系统开发中的一个主要挑战是建立一个统一的数学框架，该框架允许将物理系统层与计算层、网络层以及潜在的人类层集成在一起。控制系统将在任何这样的框架中发挥不可避免的作用。另一个相当大的挑战来自于难以获得大规模互联控制系统的准确模型，以及需要在获得进一步的感官信息后改进这些模型；从这个意义上说，未来的控制输入需要以数据驱动的方式进行调整。最后，研究建议的控制体系结构对故障或恶意行为的脆弱性，并通过弹性设计进行预防，是安全关键的网络物理系统的重要组成部分。为了克服这一问题的高度跨学科性质，拟议的项目在三个密切相关的主要方面进行了讨论。第一个重点是说明反馈下放所造成的根本限制。至关重要的是，在可控制性、稳定性和复原力的背景下，从数学上描述权力下放的缺点。第二个目标是设计健壮的、数据驱动的分布式算法。通过感官观察获得大量数据为减少系统建模中不可避免的不确定性提供了机会；这导致了可以实时实施的数据驱动的解决方案，其中关键部分的信息只有在事后才能获得。网络控制系统中动态和反馈的存在，将在这些环境下设计数据驱动算法的挑战与机器学习文献中的在线优化的支柱框架区分开来。最后，第三个重点针对由于分散反馈而产生的性能和安全问题；分布式控制中的主要假设是所有代理合作以实现全局目标，而项目这部分的目标是调查这些算法在失败、延迟或恶意行为方面的脆弱性，并提出健壮的解决方案。该计划将支持2名博士后研究人员、6名博士生、7名硕士学生和2名USRA的培训，并将有助于加强加拿大在这些关键投资领域的地位。

The concepts of sensing and feedback have been essential parts of engineering sciences, biological systems, and many areas of social sciences and economics. The emergence of network sciences within these disciplines has revealed numerous new opportunities for sensing and feedbacks. Integrated with the advancement and availability of a variety of sensors, which supply an abundance of useful data, networked systems have provided capabilities that allow for the execution of tasks that are remarkably complex. One major challenge in the development of networked cyber-physical systems is to establish a unifying mathematical framework that allows for the integration of the physical system layer with computing, networking, and potentially human layers. Control systems will play an inevitable role in any such framework. The other considerable challenge stems from the difficulty of obtaining accurate models for large-scale interconnected control systems, and the need to improve these models upon obtaining further sensory information; in this sense, future control inputs need to be adjusted in a data-driven fashion. Finally, studying the vulnerability of the proposed control architectures to failure or malicious behaviours and prevention through resilient designs is an essential part of safety-critical cyber-physical systems. To overcome the highly interdisciplinary nature of this problem, the proposed project has been discussed in three intimately related main thrusts. The first thrust aims at characterizing fundamental limitations imposed by the decentralization of feedbacks. It is crucial to mathematically characterize the shortcomings of decentralization in the context of controllability, stabilizability, and resilience. The second thrust aims at designing robust and data-driven distributed algorithms. The availability of massive data through sensory observations provides the opportunity for reducing the inevitable uncertainties in system modelling; this leads to data-driven solutions that can be implemented in real-time, with a crucial part of information only available at hindsight. The presence of dynamics and feedback in networked control systems differentiates the challenges in designing data-driven algorithms in these settings from the pillar framework of online optimization in machine learning literature. Finally, the third thrust aims at performance and security issues that arise due to the decentralization of feedbacks; the predominant assumption in distributed control is that all agents cooperate to achieve a global objective, and the goal of this part of the project is to investigate the vulnerabilities of these algorithms with respect to failure, delays, or malicious behaviours, and to propose robust solutions. This program will support the training of 2 postdoctoral researchers, 6 PhD students, 7 MSc students, and 2 USRAs and will contribute to strengthen Canada's position in these key areas of investment.