Apprentissage de modèles à contraintes concis à partir de données sans erreurs : études sur l’acquisition d’équations arithmétiques booléennes et de modèles d’ordonnancement à court terme
Ramiz Gindullin · HAL (Le Centre pour la Communication Scientifique Directe) · 2024
Using constraint logic programming, the goal of this thesis is to develop several constraint acquisition techniques for the situations where we have error-free data. Such situations render majority of ML techniques unusable and new approaches are required. The proposed constraint acquisition techniques are applied for two use cases: search for new sharp bounds conjectures for eight combinatorial objects and the constraint acquisition from a single valid short-term production schedule. The contributions of the thesis include (i) a constraint model to acquire Boolean-arithmetic expressions from data, (ii) an automatically generated database of anti-rewriting constraints that prevent the generation of simplifiable Boolean-arithmetic equations, (iii) a number of formulae synthesis techniques which can acquire a single formula combining several learning biases, (iv) the acquisition of a variety of scheduling constraints such as temporal, resource, calendar and shift constraints, and in this later case (v) the generation of a MiniZinc scheduling model.