Using the Global Constraint Seeker for Learning Structured Constraint Models: A First Attempt
Résumé
Considering problems that have a strong internal structure, this paper shows how to generate constraint models from a set of positive, flat samples (i.e., solutions) without knowing a priori neither the constraint candidates, nor the way variables are shared within constraints. We describe two key contributions to building such a model generator: (1) First, learning is modeled as a bicriteria optimization problem over ranked constraint candidates returned by the Constraint Seeker, where we optimize both the compactness of the model, and the rank (or appropriateness) of the selected constraints. (2) Second, filtering out irrelevant candidate models is achieved by using meta data of the global constraint catalog that describe links between constraints. Some initial experiments on a proof-of-concept implementation show promising results.