Search on Constraint Satisfaction Problems with Sparse Secondary Structure

This paper considers a variety of ways to detect relatively isolated, highly restricted subproblems and then exploit them to guide search for a solution. It introduces a local search method that, prior to search, estimates where such subproblems lie within constraint satisfaction problems. These subproblems are assembled into a secondary structure used with dynamic variable-ordering heuristics to guide search, while learning protects against the occasional inadequacies of local search. On some classes of difficult structured benchmark problems, this approach solves constraint satisfaction problems an order of magnitude faster.