Reformulating planning as probabilistic inference---where it helps and where not
, University of Stuttgart
Date: Tuesday, April 08, 2014
Time: 11:00 AM to 12:00 PM Note: all times are in the Eastern Time Zone
Location: 32-G882 (Hewlett Room)
Host: Tomas Lozano-Perez, MIT CSAIL
Contact: Teresa Cataldo, email@example.com
Speaker URL: None
TALK: Marc Toussaint
Reformulating planning problems as probabilistic inference problems is interesting, but does not necessarly solve fundamental problems. In this talk I will review three variations of the theme where the reformulation has lead to novel theoretical insights and efficient algorithms. These are in the context of stochastic optimal control and model-free Reinforcement Learning, for multi-agent POMDPs, and for relational MDPs. I will conclude with some questions and first steps on a problem we currently work on: how to efficiently plan in the case of uncertainty over existence of objects.
Since 2013 Marc is professor at University of Stuttgart, heading the Machine Learning and Robotics group. Before that he stayed at TU and FU Berlin as an assistant professor, at U Edinburgh as a postdoc with Chris Williams and Sethu Vijayakumar, and initially studied physics. His general research is in combining decision theory and machine learning methods, mostly with applications in robotics. Specific research topics include Reinforcement Learning, probabilistic relational domains, exploration and active learning, and optimization and control in robotics.
Created by Teresa Cataldo at Thursday, April 03, 2014 at 2:39 PM.