Learning and Control in Multi-Agent Games on Networks

网络多智能体博弈中的学习和控制

• 批准号: RGPIN-2018-04551
• 负责人: Pavel, Lacra
• 金额: $ 3.35万
• 依托单位: University of Toronto
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2019
• 资助国家: 加拿大
• 起止时间: 2019-01-01 至 2020-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=689309
• 关键词: Learning; Control; Multi; Agent; Games

Game theory has recently become an indispensable tool enabling the analysis, modeling and design of traditional as well as emerging complex networks. It is a powerful mathematical framework used to analyze strategic interactions of cooperative or competing decision-makers (called players or agents), while taking into account their preferences and requirements. The past decade has witnessed an explosion of interest in issues that intersect game theory and networks. There are many network-based applications in the fields of communication networks, the smart electric grid, the Internet of Things, social networks, network security, mobile service markets or even epidemic control. While there has been substantial progress, there are major challenges in applying traditional game theory to the understanding and design of complex modern networks. For example, the agents' goals (and decisions) interfere and this means that there is need for coordination. However, centralized coordination is often impractical in large networks (e.g. ad-hoc or peer-to-peer networks); or it consumes excessive bandwidth and energy. The proposed research program focuses on the extension of Nash's game theory to multi-agent networks. Our goal is to design nearest-neighbor interaction rules that achieve a desirable collective configuration (e.g. Nash equilibrium), rely on locally available information, minimize superfluous communication and processing overhead, and are provably convergent in general classes of games, and on general symmetric/asymmetric networks. We will focus on the following themes: (i) learning about the other agents, (ii) learning about the game (reinforcement learning), (iii) achieving robustness, resilience and dealing with malicious behaviour of the agents, and (iv) hybrid learning. Long-term our program will generate a theoretical framework that contributes to a general unified theory of games on networks. By combining a system-theoretic approach with operator theory, graph theory and networks, we will generate novel methodologies and algorithms, advancing the state-of-the-art in game theory on networks. Some of the by-products of this research will be general rules to deal with asymmetric and delayed networked information in optimizing network performance. This is important in the complex and highly heterogeneous networks around us. Applications range from wireless networks and social networks, to demand-response management in smart-grids and the design of networks of autonomous agents. It will also lead to insights and innovative technologies in the design of new network protocols, and in understanding networks such as social networks. Five PhD students, five Master students and one post-doctoral fellow will be trained.

博弈论最近已经成为分析、建模和设计传统以及新兴复杂网络的不可或缺的工具。它是一个强大的数学框架，用于分析合作或竞争决策者（称为参与者或代理人）的战略互动，同时考虑到他们的偏好和需求。过去十年见证了人们对博弈论和网络相关问题的兴趣激增。在通信网络、智能电网、物联网、社交网络、网络安全、移动服务市场甚至疫情控制等领域都有很多基于网络的应用。虽然已经取得了实质性的进展，但在应用传统博弈论来理解和设计复杂的现代网络方面仍存在重大挑战。例如，代理的目标（和决策）相互干扰，这意味着需要协调。然而，集中协调在大型网络（例如ad-hoc或点对点网络）中通常是不切实际的；或者消耗过多的带宽和能量。提出的研究计划侧重于将纳什博弈论扩展到多智能体网络。我们的目标是设计最近邻交互规则，实现理想的集体配置（例如纳什均衡），依赖于本地可用信息，最小化多余的通信和处理开销，并且在一般游戏类别和一般对称/非对称网络中证明是收敛的。我们将重点关注以下主题：(i)学习其他代理，（ii）学习游戏（强化学习），（iii）实现鲁棒性，弹性和处理代理的恶意行为，以及（iv）混合学习。从长远来看，我们的计划将产生一个理论框架，有助于网络博弈的一般统一理论。通过将系统理论方法与算子理论、图论和网络相结合，我们将产生新的方法和算法，推动网络博弈论的最新发展。本研究的一些副产品将是在优化网络性能时处理不对称和延迟网络信息的一般规则。这在我们周围复杂且高度异构的网络中非常重要。应用范围从无线网络和社交网络，到智能电网的需求响应管理和自主代理网络的设计。它还将在新网络协议的设计和理解网络（如社交网络）方面带来见解和创新技术。培养博士研究生5人，硕士研究生5人，博士后1人。