- Title: Bayesian reinforcement learning
- Speaker: Christos Dimitrakakis (University of Oslo)
- Time and date: 1pm to 2pm, November 10th, 2021 (Wednesday)
- Room: Virtual (Zoom)
The Game AI Research Group is glad to announce a (virtual) talk by Christos Dimitrakakis on Wednesday November 10th, 2021, at 13:00.
All welcome (especially students), no pre-booking required
Abstract
I will give a short introduction to the reinforcement learning problem and its optimal solution from a Bayesian perspective. Then I will outline three basic algorithms: Tree Search, Bayesian policy gradient and value function methods. For the former, I will describe some approximation methods that increase tractability [1]. For the latter, I will point out a fundamental conceptual flaw in most such methods used in practice [2]. Finally, I will give a brief overview of useful Bayesian models for this problem.
Bio
Christos Dimitrakakis’s main interests are reinforcement learning, fairness, differential privacy, and Bayesian inference. He is currently a professor at the University of Oslo, teaching courses on reinforcement learning, statistical decision theory and social and scientific aspects of machine learning. Other past and current appointments include EPFL, the University of Amsterdam, University of Frankfurt, Chalmers University of Technology, Tokyo Institute of Technology, Harvard University, and the University of Lille. He obtained his PhD in 2006 at IDIAP, from EPFL, on ensemble methods for speech recognition and reinforcement learning. He also was a developer for TORCS the open racing car simulator.