Interpolation-Based Learning as a Mean to Speed-Up Bounded Model Checking (Short Paper)

Abstract

In this paper (This is a short paper accepted in the new ideas and work-in-progress section of SEFM 2017.) we introduce a technique to improve the efficiency of SAT calls in Bounded Model Checking (BMC) problems. The proposed technique is based on exploiting interpolation-based invariants as redundant constraints for BMC.

Previous research addressed the issue using over-approximated state sets generated by BDD-based traversals. While a BDD engine could be considered as an external tool, interpolants are directly related to BMC problems, as they come from SAT-generated refutation proofs, so their role as a SAT-based learning is potentially higher. Our work aims at understanding whether and how interpolants could speed up BMC checks, as they represent constraints on forward and backward reachable states at given unrolling boundaries.

Being this work preliminary, we do not address a tight integration between interpolant generation and exploitation. We thus clearly distinguish an interpolant generation (learning) phase and a subsequent interpolant exploitation phase in a BMC run. We experimentally evaluate costs, benefits, as well as invariant selection options, on a set of publicly available model checking problems.