回避行動の再利用メカニズムを備えた強化学習のための関数近似器修正手法と多関節ロボットへの応用  [in Japanese] A Modification Algorithm of Function Approximator for the Reinforcement Learning with Reusing Mechanism of Avoidance Actions : Proposal and its Application to Motion Learning of Multi-Link Robot  [in Japanese]

Search this Article

Author(s)

Abstract

強化学習などの学習手法をロボットの運動学習に適用する際に問題となる学習コスト(転倒によるダメージなど)を軽減する一手法として,我々はあるタスクの学習中に回避行動を分離して学習しこれをほかのタスクの学習で再利用する手法を強化学習の枠組で提案,4リンク程度の土台非固定型ロボットへの応用を行ってきた[1].本稿では分離学習における分離性能を向上させることを目的として基底関数を修正する手法を提案し,運動学習における有効性を示す.さらに回避行動を再利用することによって運動学習における転倒ダメージが軽減するかを検討する.

Applying a learning method, such as reinforcement learning, to learning motions of multi-link robots requires large cost, such as damage from falling down. To overcome this problem, we proposed a reusing mechanism for reinforcement learning where the avoidance actions, such as not to fall down, are learned separately from primary actions, then they are reused in learning new tasks [1]. A method to apply it to learning whole-body motions of 4-link robot whose base is not fixed to a ground was also developed. In this paper, we propose a new method to modify basis functions of a function approximator of an action value function to improve the separative performance, and demonstrate the method works effectively in learning whole-body motions of a multi-link robot. Furthermore, we investigate a learning cost of damage from falling down in learning whole-body motions is reduced by reusing avoidance actions.

Journal

  • IEICE technical report

    IEICE technical report 107(410), 87-92, 2007-12-15

    The Institute of Electronics, Information and Communication Engineers

References:  7

Codes

  • NII Article ID (NAID)
    110006549440
  • NII NACSIS-CAT ID (NCID)
    AN10091178
  • Text Lang
    JPN
  • Article Type
    ART
  • ISSN
    09135685
  • NDL Article ID
    9330283
  • NDL Source Classification
    ZN33(科学技術--電気工学・電気機械工業--電子工学・電気通信)
  • NDL Call No.
    Z16-940
  • Data Source
    CJP  NDL  NII-ELS 
Page Top