情報処理学会 第81回全国大会講演要旨

Evaluation function is a very important factor for policy decision but always hard to define exactly especially in imperfect information games. Most of the studies on the game of Mahjong utilize concrete game rules and use traditional AI methods with artificially well-designed function blocks. For the benchmark of agreement rate on tile discard, traditional baseline is 62.1%. Our past proposal designed a new model with deep convolutional neural networks and raised this result by 6.7%. However the model is still not perfect. In this paper, we make a comparison of the behavior on learning past-made-action information in different ways, and explore better approaches for including off-table messages such as ranks and scores for model learning.