Playing Game 2048 with Deep Convolutional Neural Networks Trained by Supervised Learning

• Item type: SIG Technical Reports(1)
• 公開日: 2018-06-22
• タイトル: Playing Game 2048 with Deep Convolutional Neural Networks Trained by Supervised Learning
• タイトル / 言語: en
• タイトル: Playing Game 2048 with Deep Convolutional Neural Networks Trained by Supervised Learning
• 言語: eng
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Graduate School of Engineering, Kochi University of Technology
• 著者所属: School of Information, Kochi University of Technology
• 著者所属(英): en
• 著者所属(英): Graduate School of Engineering, Kochi University of Technology
• 著者所属(英): en
• 著者所属(英): School of Information, Kochi University of Technology
• 著者名: Naoki, Kondo; Kiminori, Matsuzaki
• 著者名(英): Naoki, Kondo; Kiminori, Matsuzaki
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Game 2048 is a stochastic single-player game and development of strong computer players for 2048 has been based on N-tuple networks trained by reinforcement learning. Some computer players were developed with (convolutional) neural networks, but they did not perform well. In this study, we develop computer players for 2048 based on deep convolutional neural networks (DCNNs). We increment the number of convolutional layers from two to nine, while keeping the number of weights almost the same. We train the DCNNs by applying supervised learning with a large number of play records from existing strong computer players. The best average score achieved is 86,030 with five convolutional layers, and the best maximum score achieved is 401,912 with seven convolutional layers. These results are better than existing neural-network-based players, while our DCNNs have less weights.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Game 2048 is a stochastic single-player game and development of strong computer players for 2048 has been based on N-tuple networks trained by reinforcement learning. Some computer players were developed with (convolutional) neural networks, but they did not perform well. In this study, we develop computer players for 2048 based on deep convolutional neural networks (DCNNs). We increment the number of convolutional layers from two to nine, while keeping the number of weights almost the same. We train the DCNNs by applying supervised learning with a large number of play records from existing strong computer players. The best average score achieved is 86,030 with five convolutional layers, and the best maximum score achieved is 401,912 with seven convolutional layers. These results are better than existing neural-network-based players, while our DCNNs have less weights.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AA11362144
• 書誌情報: 研究報告ゲーム情報学（GI）; 巻 2018-GI-40, 号 2, p. 1-6, 発行日 2018-06-22
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8736
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会