@article{oai:ipsj.ixsq.nii.ac.jp:02006864,
 author = {横山,治輝 and 田原,康之 and 清,雄一 and Haruki Yokoyama and Yasuyuki Tahara and Yuichi Sei},
 issue = {1},
 journal = {情報処理学会論文誌データベース（TOD）},
 month = {Jan},
 note = {麻雀AIの発展には，プレイヤの行動をどれだけ正確に再現できるかが鍵になる．再現度が高ければ，プロプレイヤを凌駕する強さを持つAIの開発や，特定のプロプレイヤを模倣したAIとの対戦環境の実現が期待できる．現在，麻雀における打牌行動の予測モデルは，畳み込みニューラルネットワーク（CNN）が主流となっているが，CNNは大域的な特徴を十分に扱うことが難しい．本研究では，この課題を解決するために，異なるスケールの特徴をとらえるTransformerを階層構造に統合した打牌予測モデルを提案する．さらに，モデルがとらえる特徴のスケールを麻雀特有の構成単位に合わせて設定している．実験として，対戦麻雀サイト「天鳳」の対戦ログを使用した教師あり学習を行った．その結果，モデルの入力次元数とパラメータ数を削減しつつ，打牌予測において最高の正解率を誇っていた麻雀AI「Suphx」を上回る正解率を達成した．, Accurate reproduction of player behavior is pivotal to the advancement of Artificial Intelligence (AI) in Mahjong. High-fidelity modeling can lead to the development of Mahjong AI capable of surpassing professional players. It also enables the creation of environments where players can compete against AI that imitates the playing style of specific professionals. Convolutional Neural Networks (CNNs) are currently the dominant models for tile discard prediction in Mahjong. However, they have limitations in capturing global game features effectively. To address this issue, we propose a novel discard prediction model based on a hierarchical Transformer architecture designed to capture multi-scale features. Furthermore, the model's feature scales are aligned with the structural components unique to Mahjong. We conducted supervised learning experiments using game logs from the online Mahjong platform Tenhou. As a result, our model achieved higher prediction accuracy than Suphx, previously the top AI for discard accuracy, while reducing both input dimensionality and the number of parameters.},
 pages = {55--65},
 title = {階層型Transformerを用いた麻雀における打牌予測},
 volume = {19},
 year = {2026}
}