車速追従制御のための強化学習における転移可能な方策の学習手法

カテゴリ: 論文誌(論文単位)

グループ名: 【C】電子・情報・システム部門

発行日: 2021/12/01

タイトル(英語): Learning Transferable Policy in Reinforcement Learning for Vehicle Velocity Tracking Control

著者名: 夏　有輝也（横浜国立大学大学院理工学府），濱上　知樹（横浜国立大学大学院理工学府），菅家　正康（(株)明電舎），吉田　健人（(株)明電舎），庭川　誠（(株)明電舎）

著者名(英語): Yukiya Natsu (Graduate School of Engineering Science, Yokohama National University), Tomoki Hamagami (Graduate School of Engineering Science, Yokohama National University), Masayasu Kanke (MEIDENSHA CORPORATION), Kento Yoshida (MEIDENSHA CORPORATION), Mak

キーワード: 深層強化学習，車両速度追従制御，階層的方策，State Abstraction　　deep reinforcement learning，vehicle velocity tracking，hierarchical policy，state abstraction

要約(英語): We propose the control system for driving robot using Hierarchical Reinforcement Learning. Driving Robots are playing an active role in test driving for evaluating fuel consumption and exhaust gas of automobiles. We can consider Reinforcement Learning as one of the control methods for driving robot. The control system using Reinforcement Learning has the advantage that there is no need to adjust parameters manually. However, Reinforcement Learning suffer from poor sample efficiency because it requires a lot of trials. In this research, we propose the control system for driving robot using the algorithm for learning hierarchical policy. Moreover, we introduce State Abstraction in Hierarchical Reinforcement Learning. By using abstract state, each low-level policy specialize in distinct behavior. The advantage of this method is that we can improve the sample efficiency by transferring low-level policies learned using multiple vehicles. The experimental result shows that the proposed method improve the sample efficiency in vehicle velocity tracking task.

本誌: 電気学会論文誌C（電子・情報・システム部門誌） Vol.141 No.12 （2021） 特集Ⅰ：電気・電子・情報関係学会東海支部連合大会　特集Ⅱ：研究会優秀論文

本誌掲載ページ: 1492-1499 p

原稿種別: 論文／日本語

電子版へのリンク: https://www.jstage.jst.go.jp/article/ieejeiss/141/12/141_1492/_article/-char/ja/