Sokoban: Evaluating Standard Single-Agent Search Techniques in the Presence of Deadlock

摘要

Single-agent search is a powerful tool for solving a variety of applications. Most of the academic application domains used to explore single-agent search techniques have the property that if you start with a solvable state, at no time in the search can you reach a state that is unsolvable (it may, however, not be minmal). In this paper we address the implicatiosn that arise when states in the search are unsolvable. These so-called deadlock states are largely responsible for the failure of our attempts to solve positions in the game of Sokoban.