Optimal Sokoban solving using pattern databases with specific domain knowledge

A pattern database (PDB) stores shortest distances from abstract states to a set of abstract goal states. For many search problems the best heuristic function is obtained using PDBs. We aim to find optimal solutions for Sokoban using PDBs. Due to the domain-specific characteristics of the goal states a straightforward application of PDBs in Sokoban results in an ineffective heuristic function. We propose an alternative approach, by introducing the idea of an instance decomposition to obtain an explicit intermediate goal state which allows an effective application of PDBs. We also propose a domain-specific tie breaking rule. When applied to the standard set of instances this approach improves heuristic values on initial states, detects considerable more deadlocks in random states, and doubles the number of optimally solved instances compared to previous methods.