Artificial General Intelligence and AI in Games

Marc Lanctot
Google DeepMind

Where: Aud 4
When: Monday April 3rd, 15:00 - 16:00

Please note that DeepMind prefers not to have outside media attend the talk or that university communication departments write an article about it. If you invite other people, please let them know about this.

Abstract:
As we scale up artificial general intelligence (AGI), our agents will interact both with humans and with other agents. This introduces a number of challenges that are not considered in (single-agent) reinforcement learning, such as nonstationary environments, coordination, modeling / anticipating of other agents’ actions or plans, communication, and fast-growing complexity in the number of agents. In this talk, I will give an overview of the multiagent learning research at DeepMind. I will start with an overview of deep learning and reinforcement learning (RL), highlighting some of the key contributions from the past few years. Then I will talk about the specific challenges of extending RL to multiagent systems, and our efforts to solve these problems. Specifically, I will talk about multiagent reinforcement learning in AlphaGo, Fictitious Self-Play for imperfect information games (with applications to Poker), and techniques for analyzing environmental outcomes of independent learners in sequential social dilemmas.

Bio:
Marc Lanctot is a research scientist at Google DeepMind in London, United Kingdom. He received his Ph.D. degree in Computer Science with a focus on Artificial Intelligence from the Department of Computer Science, University of Alberta, Edmonton, Canada, in 2012. His doctoral research work focused on Monte Carlo methods for regret minimization in imperfect information games and game-tree search in stochastic perfect information games. Previously, he was a post-doctoral research fellow at the Department of Data Science and Knowledge Engineering, Maastricht University, where he studied Monte Carlo tree search variants in games, and extensions to simultaneous move games and imperfect information games. He is currently working on general multiagent reinforcement learning.