Learning Constraint Programming Models from Data Using Generate-And-Aggregate

Mohit Kumar, Samuel Kolb, Tias Guns

Abstract: Constraint programming (CP) is used widely for solving real-world problems. However, designing these models require substantial expertise. In this paper, we tackle this problem by synthesizing models automatically from past solutions. We introduce COUNT-CP, which uses simple grammars and a generate-and-aggregate approach to learn expressive first-order constraints typically used in CP as well as their parameters from data. The learned constraints generalize across instances over different sizes and can be used to solve unseen instances - e.g., learning constraints from a 4×4 Sudoku to solve a 9×9 Sudoku or learning nurse staffing requirements across hospitals. COUNT-CP is implemented using the CPMpy constraint programming and modelling environment to produce constraints with nested mathematical expressions. The method is empirically evaluated on a set of suitable benchmark problems and shows to learn accurate and compact models quickly.