Generating Intrinsic Rewards by Random Recurrent Network Distillation

• Item type: SIG Technical Reports(1)
• 公開日: 2021-02-22
• タイトル: Generating Intrinsic Rewards by Random Recurrent Network Distillation
• タイトル / 言語: en
• タイトル: Generating Intrinsic Rewards by Random Recurrent Network Distillation
• 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: Department of Computer Science, Graduate School of Engineering, Nagoya Institute of Technology
• 著者所属: Department of Computer Science, Graduate School of Engineering, Nagoya Institute of Technology
• 著者所属: Department of Clinical Engineering, College of Life and Health Sciences, Chubu University
• 著者所属: Department of Computer Science, Graduate School of Engineering, Nagoya Institute of Technology
• 著者所属: Department of Computer Science, Graduate School of Engineering, Nagoya Institute of Technology
• 著者所属(英): en
• 著者所属(英): Department of Computer Science, Graduate School of Engineering, Nagoya Institute of Technology
• 著者所属(英): en
• 著者所属(英): Department of Computer Science, Graduate School of Engineering, Nagoya Institute of Technology
• 著者所属(英): en
• 著者所属(英): Department of Clinical Engineering, College of Life and Health Sciences, Chubu University
• 著者所属(英): en
• 著者所属(英): Department of Computer Science, Graduate School of Engineering, Nagoya Institute of Technology
• 著者所属(英): en
• 著者所属(英): Department of Computer Science, Graduate School of Engineering, Nagoya Institute of Technology
• 著者名: Zefeng, Xu; Koichi, Moriyama; Tohgoroh, Matsui; Atsuko, Mutoh; Nobuhiro, Inuzuka
• 著者名(英): Zefeng, Xu; Koichi, Moriyama; Tohgoroh, Matsui; Atsuko, Mutoh; Nobuhiro, Inuzuka
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Exploration in sparse reward environments pose significant challenges for many reinforcement learning algorithms. Rather than solely relying on extrinsic rewards provided by environments, many state-of-the-art methods generate intrinsic rewards to encourage the agent explore the environments. However, we found that existing models fall short in some environments, where the agent must visit a same state more than once. Thus, we improve an existing model to propose a novel type of intrinsic exploration bonus which will reward the agent when a new sequence is discovered. The intrinsic reward is the error of a recurrent neural network predicting features of the sequences given by a fixed randomly initialized recurrent neural network. Our approach performs well in some Atari games where conditions must be fulfilled to develop stories.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Exploration in sparse reward environments pose significant challenges for many reinforcement learning algorithms. Rather than solely relying on extrinsic rewards provided by environments, many state-of-the-art methods generate intrinsic rewards to encourage the agent explore the environments. However, we found that existing models fall short in some environments, where the agent must visit a same state more than once. Thus, we improve an existing model to propose a novel type of intrinsic exploration bonus which will reward the agent when a new sequence is discovered. The intrinsic reward is the error of a recurrent neural network predicting features of the sequences given by a fixed randomly initialized recurrent neural network. Our approach performs well in some Atari games where conditions must be fulfilled to develop stories.
• 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: AN10505667
• 書誌情報: 研究報告数理モデル化と問題解決（MPS）; 巻 2021-MPS-132, 号 15, p. 1-6, 発行日 2021-02-22
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8833
• 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