New Framework Combines Sampling and Model-Checking for POMDP Synthesis

There's a new approach for dealing with Partially Observable Markov Decision Processes (POMDPs) that mixes sampling, automata learning, and model-checking to develop finite-state controllers with guaranteed correctness. This method is influenced by Angluin's L* algorithm, using sampling as a way to check membership and model-checking to verify equivalence. It achieves a level of completeness as long as the sampling-driven policy is regular. A prototype has been created to demonstrate this approach, aiming to bridge the gap between scalable, unverified sampling techniques and more formal but restrictive synthesis methods. You can find the paper on arXiv.