Abstract:

Game-playing has been of classical interest to the field of artificial intelligence since the creation of the field. At first, expert programs defeated grandmasters in chess, and many similar successes of super-human performance followed in many other games: Go, Poker, Dota 2, Starcraft, Diplomacy, and more. With a few notable exceptions, these systems were designed to excel at playing a single game. In other words, depth (very strong performance, ideally "super human") took precedence over the ability to play many games well (breadth). In this talk, I will cover some recent work on general game-playing. First, I will describe our work on Kaggle Game Arena which outlines current performance of frontier models. Then, I will discuss recent research addressing the ultimate goal of developing a fully general game-playing agent. By leveraging the capabilities of transformers, language models, and search, recent techniques can be used to train a variety of agents ranging from achieving super-human performance in chess to creating a meta-agent that uses code generation to play any game directly from a natural language description of its rules. I will conclude by showing a recent study of our most general agent that was evaluated on over 400 games.

Speaker's Bio:

Marc Lanctot is a research scientist who worked at Google DeepMind for 12 years (2014 - 2026). His research interests include multiagent reinforcement learning, computational game theory, multiagent systems, and game-tree search. In the past few years, Marc has investigated general agent evaluation based on computational social choice, game-theoretic approaches to multiagent reinforcement learning, and both training and evaluating language models in general game-playing. Marc received a Ph.D. degree in artificial intelligence from the Department of Computing Science, University of Alberta in 2013. Before joining Google DeepMind, Marc completed a Postdoctoral Research Fellowship at the Department of Knowledge Engineering, Maastricht University, in Maastricht, The Netherlands on Monte Carlo tree search methods in games.