Heuristics for Dynamically Adapting Propagation

Building adaptive constraint solvers is a major challenge in constraint programming. An important line of research towards this goal is concerned with ways to dynamically adapt the level of local consistency applied during search. A related problem that is receiving a lot of attention is the design of adaptive branching heuristics. The recently proposed adaptive variable ordering heuristics of Boussemart et al. use information derived from domain wipeouts to identify highly active constraints and focus search on hard parts of the problem resulting in important saves in search effort. In this paper we show how information about domain wipeouts and value deletions gathered during search can be exploited, not only to perform variable selection, but also to dynamically adapt the level of constraint propagation achieved on the constraints of the problem. First we demonstrate that when an adaptive heuristic is used, value deletions and domain wipeouts caused by individual constraints largely occur in clusters of consecutive or nearby constraint revisions. Based on this observation, we develop a number of simple heuristics that allow us to dynamically switch between enforcing a weak, and cheap local consistency, and a strong but more expensive one, depending on the activity of individual constraints. As a case study we experiment with binary problems using AC as the weak consistency and maxRPC as the strong one. Results from various domains demonstrate the usefulness of the proposed heuristics.