New Techniques for the Control of Multi-Agent Systems in Uncertain Environments

不确定环境中多智能体系统控制的新技术研究发展了不确定环境中多智能体系统控制的新技术。这个问题领域具有重要的国家和科学意义，因为为许多关键应用部署的现代工程系统是由多个智能体组成的，这些智能体需要在不确定的环境中协调他们的行动以实现控制目标。这样的系统被用于许多应用，包括监视、环境监视、库存控制和危险环境中的控制，以及在交通、医疗诊断、土木工程、通信网络和国防系统等领域中的许多其他领域。正在开发的主要科学思想是：(1)通过关注聚合行为来驯服控制多智能体系统的复杂性的新技术，即与其说哪个智能体执行给定动作，不如说取决于有多少智能体执行给定动作，和(2)一种新的技术，基于使用代理之间的共同随机性，以提高他们对环境中的不确定性的对冲能力。正在开发的驯服复杂性的聚集技术依赖于平均场方法，该方法已被证明非常适合于揭示物理系统中的相变，但到目前为止还没有系统地用于控制。对称破缺现象是人们特别感兴趣的，它发生在同质系统的不同部分必须使用不同的控制规则来获得最佳的系统范围性能时。正在开发的对冲不确定性的技术依赖于博弈论模型，以及从环境噪声和其他代理人的随机选择中提取常见随机性的方法。

New techniques for the control of multi-agent systems in uncertain environments"ABSTRACTThe research develops new techniques for the control of multi-agent systems in uncertain environments. The problem area is of critical national and scientific importance, since modern engineered systems deployed for many critical applications are comprised of multiple agents who need to coordinate their actions to achieve a control objective in an uncertain environment. Such systems are used for many applications, including surveillance, environmental monitoring, inventory control, and control in hazardous environments, among many others, in fields such as transportation, medical diagnostics, civil engineering, communications networks, and defense systems.The main scientific ideas being developed are (1) a new technique to tame the complexity of controlling multi-agent systems by a focus on aggregate behavior, i.e. not so much on which agent carries out a given action as on how many agents carry out a given action, and (2) a new technique to based on the use of common randomness between the agents to increase their ability to hedge against uncertainty in the environment.The aggregation techniques to tame complexity that are being developed rely on mean-field methods, which have proven well-suited to exposing phase transitions in physical systems, but have so far not been systematically used for control. The phenomenon of symmetry breaking, which occurs when different portions of a homogeneous system must use different control rules to obtain optimal system wide performance, is of particular interest. The techniques to hedge against uncertainty that are being developed rely on game-theoretic models, as well as methods to extract common randomness from ambient noise and the randomized choices of other agents.