CAREER: Identifying and Exploiting Multi-Agent Symmetries

科学家们普遍认为，我们的宇宙遵循一定的对称模式和原则，这导致了诸如守恒定律等深刻的含义。人工智能(AI)可以而且已经从利用这些对称性中获得了巨大的好处。这个项目寻求识别和利用合作人工智能任务中普遍存在的对称性，在合作人工智能任务中，一组多个自主的顺序决策者或代理计划并学习以最大化他们的综合利益。例如，考虑自适应交通信号控制的应用，其中每个交叉口可以被建模为以适应实时交通条件以减少拥堵的方式控制其交通信号的代理。当路网的拓扑是规则的，如4连通的网格，且道路条件是均匀的时，存在一定的对称性。如果处理得当，这种多智能体对称性可以被识别和利用，以极大地提高当前合作人工智能解决方案的效率和有效性。该项目还将拟议的研究整合到一系列教育倡议中，在PI的大学的课程开发和本科研究经验中发挥关键作用，以及在学术界与行业从业者和社区利益相关者之间架起桥梁的外联活动。这项研究将建立一个统一的框架，并开发一套相互依赖的方法，为合作的人工智能任务制定、识别和利用多代理对称性。该研究首先采用了一种数学上严格的语言，将多智能体对称性的概念表达到对称马尔可夫博弈的框架中，揭示了其核心属性，可以通过规划和学习方法来利用。然后，研究计划具体化了如何利用几种最常见的多智能体对称类型，包括置换对称、欧几里得对称和混合类型的多智能体对称的层次。接下来，研究计划讨论对实践至关重要的问题，包括识别和利用近似多智能体对称性，以及处理部分可观测性。最后，该研究展示了几个实际应用，包括自适应交通信号控制、自动电路设计和材料设计，以评估和展示所提出的方法。该项目由Robust Intelligence和既定的激励竞争研究计划(EPSCoR)共同资助。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

It is widely believed by scientists that our universe follows certain symmetry patterns and principles, which lead to profound implications such as conservation laws. Artificial intelligence (AI) can and has already benefited tremendously from exploiting these symmetries. This project seeks to identify and exploit symmetries that are prevalent in cooperative AI tasks, where a group of multiple autonomous sequential decision makers, or agents, plan and learn to maximize their combined benefit. As an example, consider the application of adaptive traffic signal control, where each intersection can be modeled as an agent controlling its traffic signal in a way that adapts to real-time traffic conditions to reduce congestion. There exist certain symmetries when the topology of the road network is regular, e.g., as a 4-connected grid, and the road condition is uniform. When done properly, such multi-agent symmetries can be identified and exploited to greatly improve the efficiency and effectiveness of the current solutions to cooperative AI. This project also integrates the proposed research into an array of education initiatives, playing key roles in the curriculum development and undergraduate research experiences at the PI's university, as well as outreach activities that bridge academia with industry practitioners and community stakeholders.This research will establish a unified framework and develop a set of interdependent methods that formulate, identify, and exploit multi-agent symmetries for cooperative AI tasks. The research first adopts a mathematically rigorous language to formulate the notion of multi-agent symmetry into the framework of symmetric Markov game, revealing its core property which can be exploited by planning and learning methods. Then, the research plan concretizes how to exploit several most common types of multi-agent symmetries, including permutation symmetries, Euclidean symmetries, and hierarchies of multi-agent symmetries of mixed types. Next, the research plan discusses issues that are critical for practice, including identifying and exploiting approximate multi-agent symmetries and dealing with partial observability. Finally, the research features several real-world applications, including adaptive traffic signal control, automated circuit design, and material design, to evaluate and showcase the proposed methodology. This project is jointly funded by Robust Intelligence and the Established Program to Stimulate Competitive Research (EPSCoR).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.