Sokoban: improving the search with relevance cuts

Humans can effectively navigate through large search spaces, enabling them to solve problems with daunting complexity. This is largely due to an ability to successfully distinguish between relevant and irrelevant actions (moves). In this paper we present a new single-agent search pruning technique that is based on a move's influence. The influence measure is a crude form of relevance in that it is used to differentiate between local (relevant) moves and non-local (irrelevant) moves, with respect to the sequence of moves leading up to the current state. Our pruning technique uses the m previous moves to decide if a move is relevant in the current context and, if not, to cut it off. This technique results in a large reduction in the search effort required to solve Sokoban problems.