Improved linearization of constraint programming models

Constraint Programming (CP) standardizes many specialized “global constraints” allowing high-level modelling of combinatorial optimization and feasibility problems. Current Mixed-Integer Linear Programming (MIP) technology lacks both a modelling language and a solving mechanism based on high-level constraints. MiniZinc is a solver-independent CP modelling language. The solver interface works by translating a MiniZinc model into the simpler language FlatZinc. A specific solver can provide its own redefinition library of MiniZinc constraints. This paper describes improvements to the redefinitions for MIP solvers and to the compiler front-end. We discuss known and new translation methods, in particular we introduce a coordinated decomposition for domain constraints. The redefinition library is tested on the benchmarks of the MiniZinc Challenges 2012–2015. Experiments show that the two solving paradigms have rather diverse sets of strengths and weaknesses. We believe this is an important step for modelling languages. It illustrates that the high-level approach of recognizing and naming combinatorial substructure and using this to define a model, common to CP modellers, is equally applicable to those wishing to use MIP solving technology. It also makes the goal of solver-independent modelling one step closer. At least for prototyping, the new front-end frees the modeller from considering the solving technology, extracting very good performance from MIP solvers for high-level CP-style MiniZinc models.