Generalizing Constraint Models in Constraint Acquisition
Dimos Tsouros, Senne Berden, Steven Prestwich, Tias Guns
TL;DR
This paper tackles the limitation of Constraint Acquisition (CA) methods that learn ground constraints for a single problem instance by introducing GenCon, a framework for generalizing to parameterized constraint models. GenCon learns constraint-level classifiers over a parameterized feature representation of constraints, enabling the extraction of interpretable constraint specifications (CSs) when possible and providing a generate-and-test alternative for non-interpretable classifiers. The approach is validated on multiple CA benchmarks (Sudoku, Golomb, Exam Timetabling, Nurse Rostering) under both clean and noisy conditions, showing high precision and recall and robustness to up to 20% noise; it outperforms or matches existing generalization approaches in several settings. The work demonstrates that ML-based generalization of constraint models can make CA more robust and scalable across varying problem instances, with practical implications for more reusable CP models and interactive constraint learning.
Abstract
Constraint Acquisition (CA) aims to widen the use of constraint programming by assisting users in the modeling process. However, most CA methods suffer from a significant drawback: they learn a single set of individual constraints for a specific problem instance, but cannot generalize these constraints to the parameterized constraint specifications of the problem. In this paper, we address this limitation by proposing GenCon, a novel approach to learn parameterized constraint models capable of modeling varying instances of the same problem. To achieve this generalization, we make use of statistical learning techniques at the level of individual constraints. Specifically, we propose to train a classifier to predict, for any possible constraint and parameterization, whether the constraint belongs to the problem. We then show how, for some classes of classifiers, we can extract decision rules to construct interpretable constraint specifications. This enables the generation of ground constraints for any parameter instantiation. Additionally, we present a generate-and-test approach that can be used with any classifier, to generate the ground constraints on the fly. Our empirical results demonstrate that our approach achieves high accuracy and is robust to noise in the input instances.