Collaborative Research: AF: Small: RUI: Data Science from Economic Foundations

该项目旨在发展一种适应经济环境特殊挑战的计算学习理论。这样的环境通常有战略参与者，他们知道平台设计人员将使用他们的数据来制定未来的决策。当参与者预测到这些决定时，他们可能会改变自己的行为；消费者可能会等待优惠券购买，承包商可能会以高价格开始，以便于以后的谈判。战略操作反过来又需要新的学习算法来正确地解释数据并谨慎地使用它。感兴趣的应用范围从电子商务和管理决策（公司根据数据优化运营）到社会安全网的设计（接受者的选择影响经济上最弱势群体的生计）。最终目标是理论结论，可以指导设计和监管这些系统。在线学习的标准理论在具有战略意义的环境中是不够的。设计师现在需要控制一个微妙的反馈循环：数据的使用方式决定了将首先提供什么数据。此外，当学习者可以根据数据改变算法时，了解哪些学习算法是稳定的也很重要。为了解决这些问题，本项目寻求使用和扩展计算机科学中的数据隐私技术和经济学中的动态博弈理论。本研究将以两种主要方式应用这些工具。第一个是比较的：理解许多不同的应用程序将突出有助于或阻碍算法学习的结构特征。这些包括来自契约理论和授权的标准经济模型、定价和社会效益目标。第二种方法是应用比较分析的见解来设计新系统。目标是为经过充分研究的战略环境产生新的算法，或者了解政府或公司如何修改这些环境以使学习（或减轻其危害）成为可能。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This project seeks to develop a theory of computational learning adapted to the particular challenges of economic environments. Such environments often have strategic participants who are aware that their data will be used by a platform designer to make future decisions. When participants anticipate these decisions, they may change their behavior; a consumer may wait for a coupon to purchase, and a contractor may start their price high to facilitate later negotiations. Strategic manipulation in turn necessitates new learning algorithms which interpret the data correctly and use it carefully. The applications of interest range from e-commerce and managerial decision-making, where firms optimize operations from data, to the design of the social safety net, where the selection of recipients impacts the livelihoods of the most economically vulnerable. The end goal is theoretical conclusions that can guide both designers of and regulators for such systems.The standard theory of online learning is insufficient in settings with strategic concerns. A designer now needs to control a subtle feedback loop: the way data will be used dictates what data will be provided in the first place. It is additionally important to understand which learning algorithms are stable when the learner can change their algorithm in response to the data. To tackle these issues, this project seeks to use and expand on techniques from data privacy in computer science and the theory of dynamic games in economics. The research will apply these tools in two primary ways. The first is comparative: understanding many distinct applications will highlight the structural features that help or hinder algorithmic learning. These include standard economic models from contract theory and delegation, pricing, and targeting of social benefits. The second approach is to apply the insights from the comparative analysis to design new systems. The goal is to produce new algorithms for well-studied strategic environments, or understand how these environments can be modified by a government or firm to enable learning (or mitigate its harms).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.