Improving No-Good Learning in BinaryConstraint Satisfaction Problems
Mary D. Burke, Berthe Y. Choueiry · Lincoln (University of Nebraska) · 2013
Conflict-Directed Backjumping (CBJ) is an important mechanism for improving the performance of backtrack search used to solve Constraint Satisfaction Problems (CSPs). Using specialized data structures, CBJ tracks the reasons for failure and learns inconsistent combinations (i.e., no-goods) during search. However, those no-goods are forgotten as soon as search backtracks along a given path to shallower levels in the search tree, thus wasting the opportunity of exploiting such no-goods elsewhere in the search space. Storing such no-goods is prohibitive in practice because of space limitations. In this thesis, we propose a new strategy to preserve all no-goods as they are discovered and to reduce them into no-goods of smaller sizes without diminishing their pruning power. We show how our strategy improves the performance of search by exploiting the no-goods discovered by CBJ, and saves on storage space by generalizing them.