- This event has passed.
February 6, 2019 10:00 am - 11:00 am EST
Frans Oliehoek, Associate Professor, Delft University of Technology
Generative Adversarial Networks (GANs) are a framework in which two neural networks compete with each other: the generator (G) tries to trick the classifier (C) into classifying its generated fake data as true. GANs hold great promise for the development of accurate generative models for complex distributions, and have formed the basis of new approaches to learn from demonstrations (e.g., GAIL). As such, they clearly showcase the potential of multiagent learning methods to make impact on a large variety of machine learning tasks. However, save for some special cases, most current training methods for GANs are at best guaranteed to converge to a “local Nash equilibrium” (LNE). Such LNEs, however, can be arbitrarily far from an actual (“global”) Nash equilibrium (NE).
In this talk, I will cover some recent work which proposes to model GANs explicitly as games in mixed strategies, thereby ensuring that every LNE is an NE. With this formulation, we propose a solution method that is proven to monotonically converge to a /resource-bounded/ Nash equilibrium (RB-NE): by increasing computational resources we can find better solutions. We empirically demonstrate that our method is less prone to typical GAN problems such as mode collapse, and produces solutions that are less exploitable than those produced by GANs, and more closely resemble theoretical predictions about NEs. I will also show some limitations of our current solution technique and discuss ideas to tackle these in the future.
About the Speaker
Dr. Frans A. Oliehoek (1981) is Associate Professor at Delft University of Technology. He received his Ph.D. in Computer Science (2010) and M.Sc. Artificial Intelligence (2005) both from the University of Amsterdam (UvA). He subsequently did postdocs at MIT (2010-2012), Maastricht University (MU, 2012-2013), and UvA (2014-2017), and took up roles as assistant professor (MU), Lecturer and Senior Lecturer at the University of Liverpool (2014-2018). Frans’ research interests lie in the intersection of machine learning, AI and game theory. He is considered an expert in the field of decision making under uncertainty, with emphasis on multiagent systems. He organized several workshops on topics such as Multiagent Sequential Decision Making Under Uncertainty and Multiagent Reinforcement Learning. He received the best PC-member award at AAMAS 2012, and was awarded a number of research grants, including a prestigious £1.5M ERC Starting Grant for his project ”INFLUENCE: Influence-based Decision-making in Uncertain Environments” which started February 2018.