In Constraint Programming, the ability to reduce the search space through propagation is a key strength. However, propagation is a double-edged sword, as greater pruning power often comes with increased computation time. The best propagator for each problem constraint depends on the specific instance and may change during search. Literature shows that Machine Learning (ML) techniques and activity-based heuristics have been applied to statically choose propagators for a batch of problems and dynamically adapt propagation strength. We propose merging these efforts by using an oracle function, derived via ML, to determine whether to run complex propagators for a target constraint. Our approach, through a combination of design choices, is flexible and easily embedded in state-of-the-art solvers. This paper focuses on investigating the feasibility of building an oracle for the Energetic Reasoning propagator. Our experiments demonstrate that high prediction accuracy can be achieved, provide suggestions for classification features, and highlight critical issues to address when constructing such an oracle.
Blogger's Review: The integration of machine learning to optimize propagator selection in constraint programming illustrates the potential of combining AI with traditional algorithms, particularly in enhancing solving efficiency. Despite challenges related to computational overhead, the findings indicate feasibility and offer new directions for future research.