摘要：

We develop a learning-based automated assume-guarantee (AG) reasoning framework for verifying ω-regular properties of concurrent systems. We study the applicability of non-circular (AG-NC) and circular (AG-C) AG proof rules in the context of systems with infinite behaviors. In particular, we show that AG-NC is incomplete when assumptions are restricted to strictly infinite behaviors, while AG-C remains complete. We present a general formalization, called LAG, of the learning based automated AG paradigm. We show how existing approaches for automated AG reasoning are special instances of LAG. We develop two learning algorithms for a class of systems, called ∞-regular systems, that combine finite and infinite behaviors. We show that for ∞-regular systems, both AG-NC and AG-C are sound and complete. Finally, we show how to instantiate LAG to do automated AG reasoning for ∞-regular, and ω-regular, systems using both AG-NC and AG-C as proof rules.

展开