SGER: Foundations of Multiagent Control in Complex Environments

在许多科学领域(如探测机器人)、军事领域(如无人侦察机)和日常民用领域(如空中交通)，控制在动态和随机环境中运行的系统的能力是一个关键的瓶颈。例如，微型飞行器(MAV)的自主控制将允许更好的搜索和救援(例如，在偏远山区)、灾害响应(例如，在受损建筑物内)、生态数据收集(例如，在树顶)和军事情报(例如，侦察)。这些具有技术前景和科学挑战性的问题具有挑战性，因为它们同时具有控制系统的大多数困难，特别是系统(I)具有高度非线性的动力学；(Ii)在非平稳环境中运行；(Iii)在随机环境中运行；以及(Iv)与环境的复杂相互作用，即使不是不可能的话，也很难准确建模。该奖项的研究是追求一种基于主体的方法来进行(单主体)控制，该方法依赖于局部动作，但局部动作需要仔细协调，以确保系统接收一致的控制信号。目前还缺乏能够组合多个传感器并使用分布式计算和驱动来控制复杂系统的算法，而无需携带系统模型或求助于难以实现的代理协调例程。该研究将为多智能体控制提供理论基础，并为动态和随机环境下的控制系统提供学习和协调算法。

The ability to control systems operating in dynamic and stochastic environments is a critical bottleneck in many scientific (e.g., exploration robots), military (e.g., surveillance drones) and everyday civilian (e.g., air traffic) domains. For example, the autonomous control of a Micro Air Vehicle (MAV) would allow better search and rescue (e.g., in remote mountainous areas), disaster response (e.g., inside damaged buildings), ecological data gathering (e.g., on tree tops) and military intelligence (e.g., reconnaissance). Such technologically-promising and scientifically-challenging problems are challenging because they possess most of the difficulties of control systems simultaneously, notably that the system (i) has highly non-linear dynamics; (ii) operates in non-stationary environments; (iii) operates in stochastic environments; and (iv) has complex interactions with the environment that are difficult, if not impossible, to model accurately.Research under this award is pursuing an agent-based approach to (single-agent) control that relies on local actions, but local actions need to be carefully coordinated to ensure that the system receives a coherent control signal. There is currently a lack of algorithms that can combine multiple sensors and use distributed computation and actuation to control a complex system without carrying a model of the system or resorting to difficult to implement agent coordination routines. This research will contribute to theoretical foundations of multiagent control, and provide learning and coordination algorithms for controlling systems operating in dynamic and stochastic environments.