Hybrid Search Methods

Key to making computational optimization support more accessible is to increase the robustness of solvers. By developing hybrid approaches that complement the strength of different methodologies we aim to boost solution efficiency while making solvers less sensitive to poor modeling choices. Today, standard constraint programming solvers follow a general yet weak approach of integrating constraints with one another (namely by having them exchange information via variable domains only). At the same time, in order to achieve a reasonable solution efficiency, users of the system are required to choose if not even develop problem tailored heuristics that guide the search. We propose a radical departure from this approach: First, by exploiting methods from mathematical programming and constraint programming we devise methods for the tight integration of constraints that can be employed by solvers automatically, for example via the exchange of dual information. Our second goal then is to devise new and more robust solution techniques that automatically adapt to the problem or problem instance that needs to be solved. Particularly, in the Cornflower project we investigate how the efficiency of exact solvers can be boosted by heuristics, randomization, and learning.



Related Publications