Reconciling Copying and Trailing for Constraint Programming Systems

Backtracking search is a well-known problem solving technique, which has proven to be successful in particular for constraint solvers. Traditionally, there are two techniques for backtracking: recomputation with copying and trailing. These approaches are basically dual: upon backtracking, recomputation redoes some steps in the search, while trailing undoes some steps. Both techniques have particular advantages and disadvantages. For instance, while recomputation with copying easily scales to concurrent systems, trailing can be significantly more efficient for specific types of variables, such as Boolean variables, and enables advanced techniques like conflict clause learning. In this thesis, we investigate how trailing and copying can be combined in a single system, in a best-of-both-worlds approach. We present a preliminary design for incorporating trailing in Gecode, a constraint solver based on recomputation and copying. We discuss which additions and modifications to the current design of Gecode are necessary to support trailing. The results show that the hybrid recomputation and trailing system considerably outperforms the original recomputation-only system for particular problem classes.