Parallel Large Neighborhood Search

Industrial optimization applications must be robust: they must provide good solutions to problem instances of different size and numerical characteristics, and must continue to work well when side constraints are added. A good testbed for constructing such a robust solver is a set of network design problem instances recently made public by France Telecom. Together, these instances comprise the desired diversity in the aforementioned qualities of scale, numerical attributes, and additional constraints. We apply Constraint Programming to the above problems. In order to provide a robust solver, however, we forgo traditional depth first search in favor of another search method typically providing more predictable performance. Large Neighorhood Search (LNS) is a method using the full power of constraint programming, while maintaining the robustness benefits of local search. To further enhance robustness we parallelize our search in a variety of ways. Most interesting is a method based on a portfolio of algorithms which is shown to outperform all previously known CP-based methods for this problem set.