Strategic Play with Learning Agents

When: June 25, 2024, 2-4 PM Eastern time.

Where: Virtual (as part of EC 2024's Preview Week). The talks will be given over Zoom; links to the Zoom calls can be accessed at the password-protected page here.

All registrants (whether virtual or in-person) for EC should have been sent the password to access the above page. If you have registered but do not have the password, please contact either of us (Yoav or Jon) and we can help. If you are a student who does not plan on registering for EC but would still like to attend the virtual tutorial, please contact Karina Ciulla (karina.ciulla@yale.edu); she can provide a code for free access to the virtual portion of the conference.

Brief description: Autonomous learning agents are increasingly involved in many online platforms and economic systems. In this tutorial, we present two recent threads of work on the interplay between automated learning agents, strategic players, and the mechanisms through which they interact. 

The first part of the tutorial focuses on the dynamics between learning algorithms and their strategic implications. We review known and recent results regarding the dynamics between regret-minimizing agents and then discuss the incentives and meta-games that these dynamics induce for users of such agents. The second part of the tutorial focuses on how to design non-manipulable learning agents in strategic environments. We show that many standard learning algorithms (e.g. multiplicative-weights) can be manipulated by a strategic opponent, but conversely low-swap regret algorithms are robust to such manipulation.

Prerequisite knowledge: This tutorial is accessible to all listeners with a mathematical background (e.g., from economics, computer science, or other technical fields). We assume familiarity with some basic concepts from game theory, such as normal form games, best responses, dominated strategies, and Nash equilibria. Familiarity with online learning and regret-minimization is useful, but not necessary.

Tutorial plan:
The tutorial will be divided into two 45-minute talks, each followed by a 15 minute break for questions and discussion. Both talks will be self-contained so feel free to attend either of the two talks individually (although we strongly encourage you to attend both!).

• Part 1 (2:00 PM - 2:45 PM, Yoav Kolumbus): Learning dynamics, strategic users, and meta-games. (Slides)

• Break / questions (2:45 PM - 3:00 PM).

• Part 2 (3:00 PM - 3:45 PM, Jon Schneider): Strategizing against learners, and learning in strategic settings. (Slides)

• Break / questions (3:45 PM - 4:00 PM).

Organizers: 

Yoav Kolumbus, Cornell University

Yoav is an Assistant Research Professor at the Center for Data Science for Enterprise and Society and the Computer Science and Economics Departments at Cornell. His research interests lie at the interface of machine learning, algorithmic game theory, dynamical systems, and networks, with a focus on learning in games and interactions between learning and strategic play. Prior to joining Cornell, Yoav completed his Ph.D. in Computer Science at the Hebrew University of Jerusalem, advised by Noam Nisan, and his M.Sc. in Physics, advised by Sorin Solomon.

Email: yoav.kolumbus@cornell.edu

Jon Schneider, Google Research:

Jon is a Senior Research Scientist at Google Research in New York City. His research interests include online learning, game theory, contract design, and convex optimization. Before joining Google Research, Jon completed his Ph.D. in Computer Science at Princeton University under the supervision of Mark Braverman. 

Email: jschnei@google.com