Visualization of the performance of Rainbow DQN in playing Atari games

• Item type: National Convention(1)
• 公開日: 2022-02-17
• タイトル: Visualization of the performance of Rainbow DQN in playing Atari games
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: 人工知能と認知科学
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_5794
• 資源タイプ: conference paper
• 著者所属: 早大
• 著者名: Renke, Liu
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Deep learning has been widely applied to various fields in a recent decade, however most of them aim at solving specific problems. Deep reinforcement learning （DRL） is a combination of deep learning and reinforcement learning, which uses neural networks to learn from a predetermined reward function based on the environmental feedback, thus it is capable of solving multiple problems at the same time. DeepMind provides a method called Rainbow DQN[1], which combines six modifications in the field of DRL, and it performs better than human masters in some Atari games in 2017. In this study, we introduce a visualization method of the neural network learning together with testing procedure based on Rainbow DQN, by employing an encoder-decoder architecture that uses a part of the original neural network as an encoder, some additional layers as embedding method, and an extra reversed neural network as the decoder. We further conduct the analysis of the gaming behavior, and observe better gaming performance in some Atari games.
• 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: AN00349328
• 書誌情報: 第84回全国大会講演論文集; 巻 2022, 号 1, p. 569-570, 発行日 2022-02-17
• 出版者 / 言語: ja
• 出版者: 情報処理学会