Combinatorial Search in CP-Based Iterated Belief Propagation

Behrouz Babaki, Bilel Omrani, Gilles Pesant

Abstract: Compared to most other computational approaches to solving combinatorial problems, Constraint Programming’s distinctive feature has been its very high-level modeling primitives which expose much of the combinatorial substructures of a problem. Weighted counting on these substructures (i.e. constraints) can be used to compute beliefs about certain variable-value assignments occurring in a solution to the given constraint. A recent proposal generalizes the propagation mechanism of constraint programming to one sharing such beliefs between constraints. These beliefs, even if not computed exactly, can be very revealing for search. In this paper we investigate how best to guide combinatorial search in this cp-based belief propagation framework. We empirically evaluate branching heuristics on a wide set of benchmark constraint satisfaction problems.