Exploiting global constraints for search and propagation.

摘要

This thesis focuses on Constraint Programming (CP), that is an emergent paradigm to solve complex combinatorial optimization problems. The main contributions revolve around constraint filtering and search that are two main components of CP. On one side, constraint filtering allows to reduce the size of the search space, on the other, search defines how this space will be explored. Advances on these topics are crucial to broaden the applicability of CP to real-life problems.;For what concerns search, we introduce algorithms to count the number of solutions for two important families of constraints: occurrence counting constraints, such as alldifferent, symmetric_alldifferent and gcc, and sequencing constraints, such as regular. These algorithms are the building blocks of a new family of search heuristics, called constraint-centered counting-based heuristics. They extract information about the number of solutions the individual constraints admit, to guide search towards parts of the search space that are likely to contain a high number of solutions. Experimental results on eight different problems show an impressive performance compared to other generic state-of-the-art heuristics.;Finally, we experiment on an already known strong form of constraint filtering that is heuristically guided by the search (quick shaving). This technique gives mixed results when applied blindly to any problem. We introduced a simple yet very effective estimator to dynamically disable quick shaving and showed experimentally very promising results.;For what concerns constraint filtering, the contribution is twofold: we firstly propose an improvement on an existing algorithm of the relaxed version of a constraint that frequently appears in assignment problems ( soft_gcc). The algorithm proposed outperforms the previously known in terms of time-complexity both for the consistency check and for the filtering and in term of ease of implementiation. Secondly, we introduce a new constraint (both hard and soft version) and associated filtering algorithms for a recurrent sub-structure that occurs in assignment problems with heterogeneous resources (hierarchical_gcc). We show promising results when compared to an equivalent decomposition based on gcc.