Decentralized nonlinear estimation and control of multi-agent systems

多智能体系统的分散非线性估计与控制

• 批准号: 0601661
• 负责人: Randy Freeman
• 金额: $ 18万
• 依托单位: Northwestern University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2006
• 资助国家: 美国
• 起止时间: 2006-05-01 至 2010-04-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0601661&HistoricalAwards=false
• 关键词: Decentralized; nonlinear; estimation; control; multi

Decentralized Nonlinear Estimation and Control of Multi-Agent SystemsIntellectual merit: This project involves the design and analysis of control and estimation algorithms for nonlinear dynamic systems consisting of multiple interacting agents. Each agent has its own dynamics which are affected by the agent's local control action and possibly also by disturbances or by uncontrolled interactions with other agents. The salient features of the systems considered are that 1) the controllers are decentralized, namely, each agent decides on a control action based solely on information it receives through its sensors or through direct communication with neighboring agents, and 2) whether or not two agents are neighbors is a function of their states, so that the topology of the network of interconnected agents changes with time as the states of the agents evolve. The focus of the proposed work is on the simultaneous control and estimation of these networked systems: decisions about local control actions are based on current estimates of the parameters of the entire collection of agents, so that the controller and estimator dynamics affect each other through feedback interactions. These complex nonlinear interactions prevent the application of any simple separation principle; as a result, the estimators are taken into account in the design of the controllers (and/or vice versa). This project will provide specific algorithms for the combined decentralized control and estimation of these multi-agent systems, with performance properties demonstrated both through mathematical analysis and through simulation. Applications of the proposed algorithms may include single-team objectives, such as formation control or the optimal distribution of mobile sensors, or multi-team objectives such as multi-player pursuit/evasion.Before being modified to work with estimators, the controllers must be designed first to work under full information. An important class of local controllers are the gradient controllers in which each agent applies a control action in the steepest descent direction relative to the agent's individual cost (which may or may not be the same as the costs of the other agents). One goal of the proposed work is to identify classes of cost functions for which these gradient schemes have appropriate global convergence properties (for example, any equilibrium corresponding to an undesirable configuration should be unstable). Another goal is to incorporate auxiliary control objectives into the basic descent strategies, such as collision/obstacle avoidance (in the case of mobile agents) or the maintenance of network connectivity. For decentralized implementation, local control algorithms will rely on estimators for information needed about the collection but not available from local sensors. Such estimators must be dynamic in the sense that the signals they estimate are changing with time as the states of the agents evolve. Inputs to the estimators come from local sensors or from direct communication with neighboring agents. Goals of the proposed work include the design and analysis of appropriate estimators, with an emphasis on their convergence properties, estimator gain optimizations, and adaptation schemes.Broader impacts: The proposed work will provide tools for the design of multi-agent systems, potentially contributing to a variety of applications in the fields of mobile robots, sensor networks, manufacturing (e.g., self-assembly of smart parts), and multi-player games (e.g., coordinated automated battlefield or search-and-rescue scenarios). Graduate students funded through this proposal will benefit from the interdisciplinary perspective gained through their involvement with the Northwestern Institute on Complex Systems, where they share ideas with researchers from diverse areas (e.g., economics, medicine, physics, chemistry, and engineering). In addition, a sophomore-level undergraduate student will participate in the project through a Fellow Assistant Researcher Award sponsored by Northwestern's Residential College program. This undergraduate student will benefit from the opportunity to make a significant contribution to a research project, an important step in preparing the student for a scientific career.

多智能体系统的分散非线性估计与控制智力优势：本项目涉及由多个交互智能体组成的非线性动态系统的控制和估计算法的设计和分析。每个智能体都有自己的动态，这些动态受到智能体局部控制行为的影响，也可能受到干扰或与其他智能体之间不受控制的相互作用的影响。所考虑的系统的显著特征是：1)控制器是分散的，即每个智能体仅根据其通过传感器接收的信息或通过与相邻智能体的直接通信来决定控制动作；2)两个智能体是否相邻是它们状态的函数，因此相互连接的智能体网络的拓扑结构随着智能体状态的演变而随时间变化。所提出的工作的重点是对这些网络系统的同时控制和估计：关于局部控制动作的决策是基于对整个代理集合参数的当前估计，因此控制器和估计器动态通过反馈交互相互影响。这些复杂的非线性相互作用阻碍了任何简单分离原理的应用；因此，在控制器的设计中考虑了估计器（和/或反之亦然）。该项目将为这些多智能体系统的分散控制和估计提供具体的算法，并通过数学分析和仿真证明其性能。所提出算法的应用可能包括单队目标，如编队控制或移动传感器的最佳分配，或多队目标，如多人追击/逃避。在修改控制器以与估计器一起工作之前，必须首先将控制器设计为在完全信息下工作。一类重要的局部控制器是梯度控制器，其中每个代理在相对于代理的个人成本（可能与其他代理的成本相同，也可能不相同）的最陡下降方向上应用控制动作。提出的工作的一个目标是确定这些梯度方案具有适当的全局收敛性质的成本函数的类别（例如，任何与不希望的配置相对应的平衡都应该是不稳定的）。另一个目标是将辅助控制目标整合到基本下降策略中，例如碰撞/避障（在移动代理的情况下）或维护网络连接。对于分散实现，本地控制算法将依赖于收集所需信息的估计器，但这些信息无法从本地传感器获得。这样的估计器必须是动态的，因为它们估计的信号随着代理状态的演变而随时间变化。估计器的输入来自本地传感器或与邻近代理的直接通信。提出的工作目标包括设计和分析适当的估计器，重点是它们的收敛特性、估计器增益优化和自适应方案。更广泛的影响：提议的工作将为多智能体系统的设计提供工具，可能有助于移动机器人、传感器网络、制造（例如，智能部件的自组装）和多人游戏（例如，协调自动化战场或搜索和救援场景）领域的各种应用。通过该提案资助的研究生将受益于通过参与西北复杂系统研究所获得的跨学科视角，在那里他们与来自不同领域（例如，经济学，医学，物理学，化学和工程学）的研究人员分享想法。此外，一名大二的本科生将通过西北大学住宿学院项目赞助的研究员助理研究员奖参与该项目。该本科生将受益于为研究项目做出重大贡献的机会，这是为学生的科学事业做准备的重要一步。