Online Learning and Economics

A Google TechTalk, presented by Frederico Fusco, 2025-02-25 Google Algorithms Seminar, ABSTRACT: It is common in economic theory to model the intrinsic uncertainty of reality with random variables whose laws are perfectly and publicly known. However, practical applications often call for a more realistic and data-driven approach, as the relevant features of the problem at hand must be obtained on the field. A natural step in this direction is provided by the online learning framework, where no initial information is assumed, and data is collected online by repeatedly interacting with the environment. Online learning is particularly suited for economic applications as it captures the exploration-exploitation tradeoff, as well as the intriguing feedback models that may naturally arise. This talk surveys recent works on online learning and economics, focusing on bidding, auction design, and bilateral trade. The goal is to provide a perspective on the technical challenges faced in designing and analyzing learning algorithms and to offer the tools to address them. Part of the talk is devoted to the construction of “hard instances” as a way of proving tightness results. The talk is partly based on the recent papers “The Role of Transparency in Repeated First-Price Auctions with Unknown Valuations” (STOC 24), “No-Regret Learning in Bilateral Trade via Global Budget Balance” (STOC 24), and “Selling Joint Ads: A Regret Minimization Perspective” (EC 24). ABOUT THE SPEAKER: Federico Fusco is an Assistant Professor at the Department of Computer, Control, and Management Engineering at Sapienza University of Rome. Previously, he was a PostDoc and completed his PhD under the supervision of Stefano Leonardi in the same institution. During his Ph.D., he was hosted by Paul Duetting at Google Research Zurich as a research intern and then as a student researcher. Federico's research interests span Algorithmic Game Theory, Online Learning, and Submodular Maximization.