Motion Planning under Uncertainty and Partial Observability
- 14:00 26th October 2017 ( week 3, Michaelmas Term 2017 )Lecture Theatre B, Wolfson Building
The subject of this talk are motion planning problems where agents move inside environments that are subject to uncertainties and potentially not fully observable. The goal is to compute a strategy or a set of strategies for an agent that is guaranteed to satisfy certain safety or performance specifications. Such problems are naturally modeled by Markov decision processes (MDPs) or partially observable MDPs (POMDPs).
We discuss several technical approaches, ranging from the computation of permissive strategies that guarantee safe reinforcement learning in unknown environments, a game-based abstraction framework for POMDPs, as well as the utilization of parameter synthesis for Markov chains to compute randomized strategies for POMDPs.
We also consider preliminary work on actively including humans into verification and synthesis processes, and what challenges arise.