Automatic streamlining for constrained optimisation

Abstract

Augmenting a base constraint model with additional constraints can strengthen the inferences made by a solver and therefore reduce search effort. We focus on the automatic addition of streamliner constraints, which trade completeness for potentially very significant reduction in search. Recently an automated approach has been proposed, which produces streamliners via a set of streamliner generation rules. This existing automated approach to streamliner generation has two key limitations. First, it outputs a single streamlined model. Second, the approach is limited to satisfaction problems. We remove both limitations by providing a method to produce automatically a portfolio of streamliners, each representing a different balance between three criteria: how aggressively the search space is reduced, the proportion of training instances for which the streamliner admitted at least one solution, and the average reduction in quality of the objective value versus the unstreamlined model. In support of our new method, we present an automated approach to training and test instance generation, and provide several approaches to the selection and application of the streamliners from the portfolio. Empirical results demonstrate drastic improvements both to the time required to find good solutions early and to prove optimality on three problem classes.