Minimax-Bayes Reinforcement Learning

Abstract:
While the Bayesian decision-theoretic framework offers an elegant solution to the problem of decision making under uncertainty, one question is how to appropriately select the prior distribution. One idea is to employ a worst-case prior. However, this is not as easy to specify in sequential decision making as in simple statistical estimation problems.  I will present (sometimes approximate) minimax-Bayes solutions for various reinforcement learning problems to gain insights into the properties of the corresponding priors and policies. We find that while the worst-case prior depends on the setting, the corresponding minimax policies are more robust than those that assume a standard (i.e.\ uniform) prior.

Bio:
My research in artificial intelligence and machine learning includes reinforcement learning, preference elicitation, privacy, fairness, experiment design, safety, recommendation systems, job matching, crowdsourcing and social computing more generally. More recently, I have been interested in the computational, societal and statistical problems that arise when humans interact with artificially intelligent systems. I obtained my PhD in 2006 from EPFL, and I am currently a professor a the university of St. Gallen. Previously, I was a professor at the university of Oslo, an assistant professor at the university of Lille and a senior researcher at Chalmers university.