Towards a Compositional SPIN

This paper discusses our initial experience with introducing automated assume-guarantee verification based on learning in the SPIN tool. We believe that compositional verification techniques such as assume-guarantee reasoning could complement the state-reduction techniques that SPIN already supports, thus increasing the size of systems that SPIN can handle. We present a "light-weight" approach to evaluating the benefits of learning-based assume-guarantee reasoning in the context of SPIN: we turn our previous implementation of learning for the LTSA tool into a main program that externally invokes SPIN to provide the model checking-related answers. Despite its performance overheads (which mandate a future implementation within SPIN itself), this approach provides accurate information about the savings in memory. We have experimented with several versions of learning-based assume guarantee reasoning, including a novel heuristic introduced here for generating component assumptions when their environment is unavailable. We illustrate the benefits of learning-based assume-guarantee reasoning in SPIN through the example of a resource arbiter for a spacecraft. Keywords: assume-guarantee reasoning, model checking, learning.