EAGER: Research in the Interface of Algorithmic Game Theory and Learning

近年来，机器学习和计算机科学与经济学之间的接口取得了巨大的进步。机器学习的进步是由人类正在生成和收集的海量数据推动的。现在人们普遍认为，科学创新需要发展计算方法来处理这些数据，并将其用于推理和预测。这在算法、机器学习和统计学的接口方面取得了显著的进步。与此同时，世界上的许多经济活动已经通过获得在线存在的旧市场以及直接受到在线活动的启发和推动的新市场转移到互联网上，如赞助搜索和广告拍卖。在网络经济活动日益重要的推动下，人们越来越有兴趣通过计算机科学和经济学的接口来研究其联合计算和经济特性，其中包括算法博弈论。PI和他在麻省理工学院的团队对学习和算法博弈论做出了多项贡献。这项拟议研究的目标是推动这两个领域之间的研究前沿。PI和他的团队计划实现以下四个目标。目标(1)是促进对游戏中学习动力的理解。目标(2)是设计“学习机制”来解决学习和推理任务，当获取数据的唯一途径是通过战略性数据提供者，而产生好的数据是有代价的。除了在线学习，该团队预计(1)中的进展将对算法博弈论中的基本问题产生影响，特别是在目标(3)中：改进计算近似纳什均衡的算法的最新水平。(2)中的进展将立即应用于众包，但该团队还计划调查另一个应用，受大规模在线公开课的巨大增长的推动：目标(4)是开发良好的同行评分方案。

Recent years have seen tremendous advances in Machine Learning and in the interface between Computer Science and Economics. Progress in Machine Learning has been driven by the vast amounts of data that humanity is generating and collecting. It is now widely accepted that scientific innovation necessitates the development of computational methodology to process this data and use it for inference and prediction. This has resulted in remarkable progress at the interface of Algorithms, Machine Learning and Statistics. At the same time, much of the world's economic activity has been transferred to the Internet via old markets that obtained online presence as well as new markets that are directly inspired and enabled by online activity, such as sponsored search and ad auctions. Driven by the increasing importance of online economic activity there has been much interest in investigating its joint computational and economic characteristics through research at the interface of Computer Science and Economics, which includes Algorithmic Game Theory.The PI and his group at MIT have made several contributions to both Learning and Algorithmic Game Theory. The goal of the proposed research is to push the research front in the interface between these two fields.The PI and his team plan to pursue 4 goals, as follows. Goal (1) is to advance understanding of learning dynamics in games. Goal (2) is to design "learning mechanisms'' to solve learning and inference tasks when the only access to data is through strategic data providers with a cost for producing good data. Besides online learning, the team expects that advances in (1) will have implications to fundamental problems in Algorithmic Game Theory, particularly in goal (3): improving the state-of-the-art in algorithms for the computation of approximate Nash equilibria. Progress in (2) will have immediate applications in crowd-sourcing, but the team also plans to investigate another application, motivated by the tremendous growth of Massive Online Open Courses: goal (4) is to develop good peer grading schemes.