SEER: ML Controls Energetic Reasoning in Constraint Programming

This research addresses a fundamental challenge in Constraint Programming: balancing the pruning power of propagators with their computational cost. While powerful propagators can significantly reduce the search space, they often come with a high computational overhead, and their effectiveness varies by problem instance and search stage. The paper introduces SEER, a method that merges machine learning (ML) with activity-based heuristics to dynamically control propagation strength. SEER proposes using an ML-trained "oracle function" to predict whether running complex propagators, specifically focusing on Energetic Reasoning, will be beneficial for a given constraint. The design choices make SEER flexible and embeddable in existing state-of-the-art solvers. Experimental results demonstrate that high prediction accuracy can be achieved, offering insights into effective classification features and highlighting key challenges for building such an oracle. This approach aims to optimize the trade-off between search space reduction and computation time.