Theory and Applications of Exploitation-oriented Learning XoL in Multi-agent Systems

2016 Fiscal Year Final Research Report

• Project/Area Number: 26330267
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Intelligent informatics
• Research Institution: National Institution for Academic Degrees and Quality Enhancement of Higher Education
• Principal Investigator: Miyazaki Kazuteru 独立行政法人大学改革支援・学位授与機構, 研究開発部, 准教授 (20282866)
• Project Period (FY): 2014-04-01 – 2017-03-31
• Keywords: マルチエージェントシステム / 経験強化型学習 / 強化学習 / 機械学習 / 人工知能 / ソフトコンピューティング

Outline of Final Research Achievements

This research has achieved several progresses about theory and applications of Exploitation-oriented Learning XoL in multi-agent learning. In multi-agent learning, it is important to avoid the concurrent learning problem that occurs when multiple agents learn simultaneously. Firstly, we have proposed a method to avoid the problem. Secondly, we have focused on a positive effect of an indirect reward which is given to the agent that does not receive a reward directly. Especially, we have proposed a method to reduce the perceptual aliasing problem caused by imperfect perception. We have also described the relationship between our previous multi-agent learning theorem and the positive effect. Lastly, we have extended application areas to show the effectiveness of XoL in multi-agent learning through experiments to Keepaway tasks like soccer games. We believe that these results contribute to claim that XoL surpasses traditional reinforcement learning methods in multi-agent learning.

Free Research Field

機械学習