Decentralized stochastic control: information structures, communication, and learning

拟议的研究计划的目标是开发一个系统的方法，分散系统的最优控制，所有代理人都有一个共同的目标。这种系统出现在几乎所有的现代技术中，包括网络控制系统、通信网络、传感器和监控网络、环境遥感、智能电网、交通网络、监测和诊断系统以及机器人技术。在这些不同的应用中出现分散控制问题的原因是，与单个代理共享所有相关信息在技术上或经济上并不总是可行的。虽然这是真的，但在代理的子集之间共享一些相关信息是可行的。在这个研究项目中，我们将研究这种信息共享如何简化分散系统的设计。 该提案映射了一个五年的研究计划，以解决两个基本问题：（Q1）假设，在系统开始运行之前，有资源在代理之间建立固定数量的通信链路。应该建立哪些链接？(Q2)假设，一旦系统开始运行，代理就可以与其他代理进行信息通信。代理人应在何时和向谁进行通信？为了回答这些问题，我们将从三个方面着手：（1）识别新的信息结构类别，从而实现易于处理的设计。(2)确定最佳的传输策略，确定何时代理应该向其他人传达信息。(3)在系统模型未知的情况下，针对上述两个方向制定学习和适应策略。 拟议的研究计划将提供广泛的培训，4博士和3个工程硕士学生在系统科学和控制工程的基本领域，从而为他们提供了一个坚实的基础，为他们未来的努力。在该计划中获得的结果将在网络控制系统，交通网络，智能电网，传感器和监控网络，医疗监控和机器人技术中找到应用。

The objective of the proposed research program is to develop a systematic methodology for the optimal control of decentralized systems where all agents have a common objective. Such systems arise in almost all modern technologies, including networked control systems, communication networks, sensor and surveillance networks, environmental remote sensing, smart grids, transportation networks, monitoring and diagnostic systems, and robotics. The reason that decentralized control problems arise in these varied applications is that it is not always technologically or economically feasible to share all pertinent information with a single agent. While this is true, it is feasible to share some pertinent information among a subset of agents. In this research program we investigate how such information sharing simplifies the design of decentralized systems. This proposal maps a five year research program to address two fundamental questions: (Q1) Suppose, before the system starts running, there are resources to build a fixed number of communication links between agents. Which links should be built? (Q2) Suppose, once the system starts running, agents can communicate information to other agents at a cost. When and what should the agents communicate and to whom? To answer these questions, three directions will be pursued: (1) Identify new classes of information structures that lead to a tractable design. (2) Identify optimal transmission strategies that determine when an agent should communicate information to others. (3) Develop learning and adaptation strategies for the above two directions when the system model is not known. The proposed research program will provide broad training to 4 PhD and 3 MEng students in fundamental areas of system science and control engineering, thereby providing them with a solid foundation for their future endeavors. The results obtained in this program will find applications in networked control systems, transportation networks, smart-grids, sensor and surveillance networks, healthcare monitoring, and robotics.