Learning-Based Compositional Verification for Synchronous Probabilistic Systems

We present novel techniques for automated compositional verification of synchronous probabilistic systems. First, we give an assume-guarantee framework for verifying probabilistic safety properties of systems modelled as discretetime Markov chains. Assumptions about system components are represented as probabilistic finite automata (PFAs) and the relationship between components and assumptions is captured by weak language inclusion. In order to implement this framework, we develop a semi-algorithm to check language inclusion for PFAs and a new active learning method for PFAs. The latter is then used to automatically generate assumptions for compositional verification.