CAREER: Integrating Machine Learning with Game Theory for Multiagent Communication and Coordination

职业：将机器学习与博弈论相结合以实现多智能体通信和协调

• 批准号: 2046640
• 负责人: FEI FANG
• 金额: $ 46.4万
• 依托单位: Carnegie-Mellon University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-04-01 至 2026-03-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2046640&HistoricalAwards=false
• 关键词: CAREER; Integrating; Machine; Learning; Game

Many societal challenges we are facing involve multiple decision-makers (agents), each with their own goals or preferences. More importantly, these agents often need to communicate and coordinate with each other to achieve their goals or satisfy their preferences. For example, in security, public safety, and environmental sustainability domains, law enforcement agencies defend against attackers and poachers. These agencies often work together with the local community to combat the actions of their opponents using communication and coordination, e.g., through community policing programs, or relying on justice collaborators. Game theory is an established paradigm for reasoning about strategic interactions among multiple decision-makers. It consists of mathematical models with the common assumptions that the players act rationally and they will try to make the best decisions to obtain their own best possible outcome. Several game-theoretic models and algorithms have been successfully deployed in the field to help law enforcement agencies allocate their limited resources in the presence of opponents. However, the problem of communication and coordination in complex environments is still underexplored. This research aims to design new game-theoretic models for multiagent communication and coordination. In addition, this research attempts to develop novel machine learning-enhanced computational frameworks for solving these games. These will findings be applied to the real-world problems of wildlife protection and food bank operations.This research seeks to establish theoretical foundations of multiagent communication and coordination in settings with varying commitment power (i.e., some agents can commit to a strategy first), make algorithmic advances, and make a transformative real-world impact. The research will provide answers to the following questions: (i) How to find the best communication and coordination strategies in large-scale multiagent interaction? (ii) How to account for the bounded rationality of human agents? (iii) How to deal with the uncertainties in the environment, e.g., noise in communication? The research consists of three thrusts for three critical classes of interactions: defender-attacker-community interaction, platform-users interaction, and mediators-agents interaction. In each thrust, the researchers attempt to answer the three questions by (i) propose new game-theoretic models and solution concepts; (ii) theoretically analyze the behavioral and computational aspects of the games and characterize the impact of coordination and communication; (iii) build human behavior models from data; (iv) propose efficient algorithms based on mathematical programming, deep learning, and multiagent reinforcement learning to compute close-to-equilibrium strategies given the human behavior models and uncertainties. The results will enrich the body of knowledge in computational game theory and transform the thriving line of work into new research topics that integrate game theory with machine learning and other research areas in and outside Artificial Intelligence.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.

我们面临的许多社会挑战涉及多个决策者（代理人），每个人都有自己的目标或偏好。更重要的是，这些代理人经常需要相互沟通和协调，以实现他们的目标或满足他们的偏好。例如，在安全、公共安全和环境可持续性领域，执法机构防御攻击者和偷猎者。这些机构经常与当地社区合作，通过沟通和协调（例如通过社区警务计划或依靠司法合作者）来打击对手的行为。博弈论是推理多个决策者之间的战略互动的既定范式。它由数学模型组成，其共同假设是参与者理性行事，他们将尝试做出最佳决策以获得自己的最佳可能结果。几种博弈论模型和算法已在该领域成功部署，以帮助执法机构在对手面前分配有限的资源。然而，复杂环境中的沟通和协调问题仍然没有得到充分探索。本研究旨在设计用于多智能体通信和协调的新博弈论模型。此外，这项研究尝试开发新颖的机器学习增强计算框架来解决这些游戏。 这些研究结果将应用于野生动物保护和食品银行运营的现实问题。这项研究旨在为具有不同承诺权力的环境中的多主体沟通和协调建立理论基础（即，一些主体可以首先承诺一项策略），取得算法进步，并对现实世界产生变革性影响。该研究将为以下问题提供答案：（i）如何在大规模多主体交互中找到最佳的沟通和协调策略？ (ii) 如何解释人类主体的有限理性？ (iii) 如何应对环境中的不确定性，例如通讯中的噪音？该研究包括针对三类关键交互类别的三个主旨：防御者-攻击者-社区交互、平台-用户交互以及中介者-代理交互。在每个主旨中，研究人员都试图通过以下方式回答这三个问题：（i）提出新的博弈论模型和解决方案概念； (ii) 从理论上分析游戏的行为和计算方面，并表征协调和沟通的影响； (iii) 根据数据建立人类行为模型； (iv) 提出基于数学编程、深度学习和多智能体强化学习的有效算法，以在给定人类行为模型和不确定性的情况下计算接近均衡的策略。这些成果将丰富计算博弈论的知识体系，并将蓬勃发展的工作转化为新的研究课题，将博弈论与机器学习以及人工智能内外的其他研究领域相结合。该奖项反映了 NSF 的法定使命，并被视为值得通过使用基金会的智力优点和更广泛的影响审查标准进行评估来支持。

项目成果

Navigates Like Me: Understanding How People Evaluate Human-Like AI in Video Games

像我一样导航：了解人们如何评估视频游戏中的类人人工智能

• DOI: 10.1145/3544548.3581348
• 发表时间: 2023-03-02
• 期刊: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems
• 作者: Stephanie Milani;Arthur Juliani;I. Momennejad;Raluca Georgescu;Jaroslaw Rzepecki;Alison Shaw;Gavin Costello;Fei Fang;Sam Devlin;Katja Hofmann
• 通讯作者: Katja Hofmann

Multi-defender Security Games with Schedules

多后卫安全游戏及时间表

• 发表时间: 2023-12
• 期刊: Springer
• 作者: Zimeng Song; Chun Kai
• 通讯作者: Chun Kai

Bandit Data-Driven Optimization for Crowdsourcing Food Rescue Platforms

Bandit 数据驱动的众包食品救援平台优化

• DOI: 10.1609/aaai.v36i11.21475
• 发表时间: 2022-06
• 期刊: Proceedings of the AAAI Conference on Artificial Intelligence
• 作者: Shi, Zheyuan Ryan;Wu, Zhiwei Steven;Ghani, Rayid;Fang, Fei
• 通讯作者: Fang, Fei

Robust reinforcement learning under minimax regret for green security

绿色安全的极小极大遗憾下的鲁棒强化学习

• 发表时间: 2021-01
• 期刊: Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence
• 作者: Xu, Lily;Perrault, Andrew;Fang, Fei;Chen, Haipeng;Tambe, Milind
• 通讯作者: Tambe, Milind

NewsPanda: Media Monitoring for Timely Conservation Action

NewsPanda：媒体监测及时采取保护行动

• DOI: 10.48550/arxiv.2305.01503
• 发表时间: 2023-04-30
• 作者: Sedrick Scott Keh;Z. Shi;David J. Patterson;N. Bhagabati;Karun Dewan;Areendran Gopala;Pablo R. Izquierdo;D. Mallick;Ambika Sharma;Pooja Shrestha;Fei Fang
• 通讯作者: Fei Fang