A Hyperbelief Space Approach to Computing Optimal Policies for POMDPs
Thursday, December 5, 2013, at 15:00, Aula Magna
Typically, the location of a robot is described by a configuration. If there is uncertainty in the configuration, one can lift the configuration space into the belief space, which is essentially the set of possible probability functions for the robot's position; in this case, a belief is merely an a posteriori pdf for the robot's configuration. If a robot system plans into the future, the future beliefs will depend on future sensor measurements, which cannot be known at planning time. To deal with this, we can lift the belief space to the hyperbelief space, which is merely the set of all possible pdfs on the belief space. In the hyperbelief space, policies have deterministic effects: for a specific control policy, the transition from one stage of execution to the next is deterministic. In this higher-dimensional hyperbelief space, one need not (explicitly) worry about the representation of uncertainty. Thus, it can be convenient to model partially observed Markov decision processes (POMDPs) in the hyperbelief space. In this talk, we present results for sampling-based anytime algorithms that determine nearly optimal policies for POMDPs by representing the system evolution in the hyperbelief space.
Seth Hutchinson received his PhD from Purdue University in 1988, and joined the University of Illinois in 1990, where he is currently a Professor in the Department of Electrical and Computer Engineering. Dr. Hutchinson has served as Editor-in-Chief for the IEEE Trans. on Robotics and was the Founding Editor-in-Chief of the Conference Editorial Board of the IEEE Robotics and Automation Society. He has published approximately 200 papers on the topics of robotics and computer vision, and is coauthor of the books "Principles of Robot Motion: Theory, Algorithms, and Implementations," published by MIT Press, and "Robot Modeling and Control," published by Wiley. Hutchinson is a Fellow of the IEEE. http://www.uiuc.edu/~seth