Refining Interface Alphabets for Compositional Verification

Techniques for learning automata have been adapted to automatically infer assumptions in assume-guarantee compositional verification. Learning, in this context, produces assumptions and modifies them using counterexamples obtained by model checking components separately. In this process, the interface alphabets between components, that constitute the alphabets of the assumption automata, are fixed: they include all actions through which the components communicate. This paper introduces alphabet refinement, a novel technique that extends the assumption learning process to also infer interface alphabets. The technique starts with only a subset of the interface alphabet and adds actions to it as necessary until a given property is shown to hold or to be violated in the system. Actions to be added are discovered by counterexample analysis. We show experimentally that alphabet refinement improves the current learning algorithms and makes compositional verification by learning assumptions more scalable than non-compositional verification.