Maybe it's not too bad an idea to train a neural net for scoring a Mahjong hand as a fast estimate in place of an exhaustive pattern matching?