CAREER: Cooperative Control Under Communication Constraints

职业：通信约束下的协同控制智力价值：协同控制和网络化控制领域正处于技术革命的边缘。在国家安全、交通、通信和商业领域的新兴应用要求分布式网络能够进行多用户通信、协同信号和信息处理、多模态数据的传感器融合以及分布式计算、驱动和控制。这些信息丰富的应用程序对网络提出了特定的要求：网络必须优雅地扩展到大量代理；代理可能是地理上分布的，通常需要在有噪声的、带宽有限的信道上进行通信；网络可能由异构组件组成，包括计算能力有限的嵌入式系统；而且，网络架构可能是分散的，需要代理之间的局部协调。最近在探索基于图的多智能体协作控制方面有相当多的活动。在本提案中，PI将研究如何添加具有噪声和延迟的真实通信信道来影响协同控制性能。对于有噪声的信道情况，假设每个智能体都完全了解其邻居的状态是不合理的。我们将研究本地知识与全球知识的问题。这个项目将建立在pisi的基础上，广泛研究控制与通信约束、图形模型和信息论。在具有大量弱耦合介质的系统中，可以使用统计力学工具，如mean-field近似来确定系统的定性行为。如果介质之间有较强的耦合，那么在这些工程系统中就有可能存在多相。我们开展这项研究的主要目标是：(1)基本限制和权衡。项目的这一方面考虑了大型协作控制系统的定性缩放行为。此外，PI将考虑通信质量和由此产生的控制性能之间的基本限制和权衡。(2)算法的发展。提出的研究将开发用于解决大因子马尔可夫决策问题的强化学习技术。(3)教育发展。更广泛的影响：为了发展一种有用的理论来解释合作控制的进步，我们需要把沟通和控制看作一枚硬币的两面。一方面，香农理论告诉我们什么是可以沟通的，另一方面，控制理论告诉我们什么是应该沟通的。来自信息论、图形模型、概率论和统计力学的工具将用于开发处理随机合作控制问题的正式框架。这个框架将使我们能够真实地模拟代理之间的通信，并使我们能够理解信息和控制之间的相互作用。拟议中的研究项目融合了许多学科的工具。PI计划鼓励本科生、研究生和实践工程师为这个跨学科的研究项目做出贡献。课程、讲座和项目的设计将鼓励这种参与。为了促进对正在开发的理论和算法的统一看法，并为评估思想提供验证平台，我们将重点关注以下驱动应用：传感器网络中的分布式目标和驱动、分布式优化、多智能体决策、消息传递算法和多用户信息理论。本科生和研究生将被要求进行各种实验和模拟来验证算法。

CAREER: Cooperative Control Under Communication ConstraintsSummary StatementIntellectual Merit: The fields of cooperative and networked control are on the verge of a technologicalrevolution. Emerging applications in national security, transportation, communication, andcommerce require distributed networks to be capable of multi-user communication, collaborativesignal and information processing, sensor fusion of multi-modal data, and, distributed computation,actuation, and control.These information rich applications place specific demands upon networks: the networks mustscale gracefully to a large numbers of agents; the agents may be geographically distributed, oftenrequiring communication over noisy, bandwidth-limited channels; the networks may consist of heterogeneouscomponents, including embedded systems with limited computational power; and, thenetworks architecture may be decentralized, requiring local coordination amongst the agents.There has been quite a bit of recent activity in exploring graph-based, multi-agent cooperativecontrol. In this proposal the PI will examine how the addition of realistic communication channels,with noise and delay, can effect the cooperative control performance. For noisy channel situationsit is no longer reasonable to assume that each agent has perfect knowledge of its neighbors states.We will examine the question of local versus global knowledge. This project will build on the PIsextensive research on control with communication constraints, graphical models, and informationtheory. In systems with large numbers of weakly coupled agents statistical mechanics tools, likemean-field approximation, can be used to determine the qualitative behavior of the system. Ifthere is stronger coupling between agents then there is the possibility of multiple phases in theseengineered systems.Our main objectives in carrying out this research are: (1) Fundamental limits and tradeoffs.This aspect of the project considers the qualitative scaling behavior of large cooperative controlsystems. In addition, the PI will consider the fundamental limits and tradeoeffs between the qualityof the communication and the resulting control performance. (2)Algorithm development. Theproposed research will develop reinforcement learning techniques for solving large factored Markovdecision problems. (3) Educational development.Broader Impact: In order to develop a useful theory that can explain the advancements incooperative control one needs to view communication and control as two sides of a coin. On theone hand Shannon theory tell us what can be communicated and on the other hand control theorytheory tell us what should be communicated. Tools from information theory, graphical models,probability theory, and statistical mechanics will be used to develop a formal framework for treatingstochastic cooperative control problems. This framework will allow us to realistically model thecommunication between agents and allow us to understand the interaction between informationand control.The proposed research project blends tools from many disciplines. The PI plans to encourageundergraduates, graduate students, and practicing engineers to contribute to this interdisciplinaryresearch program. Course curriculum, talks, and projects will be designed to encourage this participation.To promote a unified view of the theory and algorithms under development and to provide atestbed for the evaluation of ideas we will focus on the following driving applications: distributedestimation and actuation in sensor networks, distributed optimization, multi-agent decision making,message-passing algorithms, and multi-user information theory. Undergraduate and graduatestudents will be called on to perform various experiments and simulations to validate the algorithms.