Exploiting Entropy in Constraint Programming

The introduction of Belief Propagation in Constraint Programming through the CP-BP framework makes possible the computation of an estimation of the probability that a given variable-value combination belongs to a solution. The availability of such marginal probability distributions, effectively ranking domain values, allows us to develop branching heuristics but also more generally to apply the concept of entropy to Constraint Programming. We explore how variable and problem entropy can improve how we solve combinatorial problems in the CP-BP framework. We evaluate our proposal on an extensive set of benchmark instances.