学習時間の短縮に向けた状態価値を用いた知識転移手法

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: 2017/09/01

タイトル(英語): A Method of Knowledge Transfer by State Value for Reducing Learning Time

著者名: 小谷　直樹（大阪工業大学 情報科学部）

著者名(英語): Naoki Kotani (Faculty of Information Science and Technology, Osaka Institute of Technology)

キーワード: 強化学習，転移学習，状態価値，知識，政策　　reinforcement learning，transfer learning，state value，knowledge，policy

要約(英語): This paper proposes a knowledge transfer method based on state value for reinforcement learning (RL) agents. It has a fundamental problem that RL which is the one of machine learning techniques needs a lot of time or the number of trials because the agents acquire appropriate skills through trial and error in order to solve a task. Transfer learning (TL) allows the agent to transfer knowledge which is acquired by itself in other tasks, or previous knowledge to solve a target task. So, TL for RL is able to speed up the learning than simple RL. Our proposed method transfers both state value and a new policy which is given by state values of two selected knowledge to as initial knowledge for an unknown state. The effectiveness of the proposed method was verified with the simulation of the reaching problem for a multi-link robot arm. The proposed method has reduced the learning time 40% than the conventional method.

本誌: 電気学会論文誌C（電子・情報・システム部門誌） Vol.137 No.9 （2017） 特集：知能メカトロニクス分野と連携する知覚情報技術

本誌掲載ページ: 1171-1176 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/137/9/137_1171/_article/-char/ja/