The intersection of learning theory, game theory and mechanism design is becoming increasingly relevant: i) data input to machine learning algorithms are either owned or generated by self-interested parties, ii) machine learning is used to optimize economic systems (e.g. auction platforms) or to learn how to optimally act in strategic settings, iii) machine learning models used in critical systems are becoming prone to adversarial attacks, iv) several machine learning approaches can be framed as finding the equilibrium of a game, as opposed to the minimizer of an objective function. The theoretical foundations of these problems lie at the intersection of learning theory, game theory and mechanism design.
Already, online learning and game theory have played a key role in some landmark advances in Machine Learning. Online learning has provided some of the most successful optimization methods used in training large-scale deep neural networks. Game-theoretic modeling has enabled the design of Generative Adversarial Networks (GANS) and inspired approaches for training deep neural network classifiers that are robust to adversarial attacks. Finally, min-max tree search and regret minimization algorithms are central in solving Go and Texas Hold’em. More broadly, the world is moving towards the co-existence of multiple AIs that learn from their interaction, which might be collaborative, strategic or adversarial.
The objective of the program is to further advance the interaction between learning and games and rethink its mathematical foundations. The semester will study the foundations of: 1) Min-Max Optimization, 2) Multi-Agent Reinforcement Learning, 3) Dynamical Systems and Learning, 3) Behavioral Game Theory, 4) Econometrics and Learning, 5) Mechanism Design and Learning. Moreover, it will address practical challenges in application domains where these techniques seem most appropriate, such as: 1) generative adversarial networks, 2) adversarial robustness, 3) learning with humans in the loop, 4) learning as a model of strategic behavior, 5) interactions of multiple learners.
The semester aims to bring together members of different communities, including machine learning, economics, operations research, theoretical computer science, and social computing.
sympa [at] lists.simons.berkeley.edu (body: subscribe%20games2022announcements%40lists.simons.berkeley.edu) (Click here to subscribe to our announcements email list for this program).
Vasilis Syrgkanis (Microsoft Research; chair), Constantinos Daskalakis (Massachusetts Institute of Technology), Dylan Foster (Massachusetts Institute of Technology), Michael Jordan (UC Berkeley), Christos Papadimitriou (Columbia University), Georgios Piliouras (Singapore University of Technology and Design), Éva Tardos (Cornell University)
Long-Term Participants (tentative, including organizers):
Yang Cai (Yale University), Nicolo Cesa-Bianchi (Università degli Studi di Milano), Suchi Chawla (University of Wisconsin - Madison), Constantinos Daskalakis (Massachusetts Institute of Technology), Alex Dimakis (University of Texas, Austin), Dylan Foster (Massachusetts Institute of Technology), Nika Haghtalab (Cornell University & UC Berkeley), Michael Jordan (UC Berkeley), Ramesh Johari (Stanford University), Kevin Leyton-Brown (University of British Columbia), Katrina Ligget (Hebrew University of Jerusalem), Panagiotis Mertikopoulos (French National Center for Scientific Research), Jamie Morgenstern (University of Washington), Denis Nekipelov (University of Virginia), Noam Nisan (Hebrew University of Jerusalem), Christos Papadimitriou (Columbia University), Georgios Piliouras (Singapore University of Technology and Design), Vasilis Syrgkanis (Microsoft Research), Éva Tardos (Cornell University), Mengdi Wang (Princeton University)
Those interested in participating in this program should send an email to the organizers games2022 [at] lists.simons.berkeley.edu (at this address).