cs.AIMay 14, 2026

Synthesizing POMDP Policies: Sampling Meets Model-checking via Learning

Authors: Debraj ChakrabortyAnirban MajumdarPrince MathewSayan MukherjeeJean-François Raskin

Organizations: Nanyang Technological University, Singapore · Tata Institute of Fundamental Research, Mumbai, India · Université Libre de Bruxelles, Brussels, Belgium · IITB Trust Lab, Department of CSE, IIT Bombay, Mumbai, India

Abstract

Partially Observable Markov Decision Processes (POMDPs) are the standard framework for decision-making under uncertainty. While sampling-based methods scale well, they lack formal correctness guarantees, making them unsuitable for safety-critical applications. Conversely, formal synthesis techniques provide correctness-by-construction but often struggle with scalability, as general POMDP synthesis is undecidable. To bridge this gap, we propose a synthesis framework that integrates sampling, automata learning, and model-checking. Inspired by Angluin's LL^* algorithm, our approach utilizes sampling as a membership oracle and model-checking as an equivalence oracle. This enables the synthesis of finite-state controllers with formal guarantees, provided the sampling-induced policy is regular. We establish a relative completeness result for this framework. Experimental results from our prototypical implementation demonstrate that this method successfully solves threshold-safety problems that remain challenging for existing formal synthesis tools. We believe our algorithm serves as a valuable component in a portfolio approach to tackling the inherent difficulty of POMDP synthesis problems.

Explore similar work

CardsList
  1. Robust Parameter Learning for Uncertain MDPs

    May 2, 2026Yannik Schnitzer, Alessandro Abate, David ParkerMarkov Decision ProcessesConfidence Intervals