Solving Sokoban Optimally with Domain-Dependent Move Pruning
Renato Reis Leme, André Grahl Pereira, Marcus Ritt, Luciana Salete Buriol · 2015
Move pruning increases the efficiency of heuristic search techniques by not expanding parts of the state space. In this article, we propose an admissible domain-dependent move pruning (DDMP) technique for Sokoban. When exploring a node DDMP analyzes and selects a subset of successor nodes required to be generated to preserve all optimal solutions. DDMP has low space and time overhead. It reduces the number of successor nodes that need to be generated and thus the branching factor. Reducing the number of successor nodes is especially important for Sokoban due to the highly costly heuristic functions and deadlock detection techniques. In addition, the subset of selected successor nodes is, in general, the "worst successor nodes available" which increases the chance of deadlocks early detecting. We define DDMP formally and prove its admissibility. When applied to the standard set of instances DDMP reduces the branching factor, detects more deadlocks, and decreases the effort to solve instances in the number of explored nodes and total time. DDMP has a positive synergy with recent deadlock detection techniques. Combined they increase the number of optimally solved instances compared to previous methods.