Invited Talk
Table Host
Chair: Anna Lukina
Decidable Problems for Partially Observable MDPs
Abstract
In this very short talk I will try to convince you to study partially observable Markov decision processes (POMDPs). I will introduce the model and highlight some successful techniques to approaching POMDPs (almost all interesting decision problems are undecidable, so nontrivial approaches are required). Towards the end, I will introduce a recently proposed restriction on how much information is lost because of the partial observability (we call it the “revelations” mechanism) and tell you a bit about how it allows us to recover decidability for omega-regular objectives. I will then close with a selection of the many open problems that remain in this area.