Exploiting Problem Data to Enrich Models of Constraint Problems

In spite of the effectiveness of Constraint Programming languages and tools, modeling remains an art and requires significant involvement from a CP expert. Our goal is to alleviate the load of the human user, and this paper is a first step in this direction. We propose a framework that enriches a ‘generic’ constraint model of a domain area with a set of constraints that are applicable to a particular problem instance. The additional constraints used to enrich the model are selected from a library; and a set of rules determines their applicability given the input data from the instance at hand. We address application domains where problem instances slightly vary in terms of the applicable constraints, such as the Building Identification (BID) problem and we use Sudoku puzzles as a vehicle to illustrate the concepts and issues involved. To evaluate our approach, we apply it to these domains, using constraint propagation on the generic model to uncover additional information about the problem instance. Our initial results demonstrate our ability to create customized models whose accuracy is further improved with the use of constraint propagation. We also discuss results obtained by solving the newly inferred models, showing that the combination of rule-based constraint inference and constraint propagation is a step towards precise modeling. Finally, we discuss domains that can benefit from our approach (e.g., timetabling and machine translation) and present directions for future work.