Sokoban puzzle is very challenging problem for both humans and computers. It also illustrates differences between human and artificial intelligence – different problems are difficult for humans and for computers. Whereas algorithmic techniques for Sokoban solving have been intensively studied by previous research, factors determining difficulty for humans have not been sufficiently explained so far. We describe two methods for difficulty rating of Sokoban puzzle – a problem decomposition metric and a computational model which simulates human traversal of a state space. We evaluate these metrics on large scale data on human solving (2000 problems solved, 785 hour of problem solving activity).