ボルツマン選択を用いたDeep Q Network

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: 2017/12/01

タイトル(英語): A Deep Q Network with Boltzmann Selection

著者名: 北　悠人（千葉工業大学大学院情報科学研究科情報科学専攻），山口　智（千葉工業大学情報科学部情報工学科）

著者名(英語): Yuto Kita (Graduate School of Information and Computer Science, Chiba Institute of Technology), Satoshi Yamaguchi (Dept. of Computer Science, Chiba Institute of Technology)

キーワード: 強化学習，深層学習，Deep Q Network，ボルツマン選択，ε-グリーディ法　　Rinforcement Learning，Deep learning，Deep Q network，Boltzmann Selection，ε-greedy Selection

要約(英語): The reinforcement learning is a method of training for an agent for accomplishing task by selecting suitable action from the current state. Deep Q network is combining convolutional network with Q-learning. By using the Convolutional Neural Network, Deep Q Network can apply to large dimentional input state tasks without special pre-processing. However Deep Q Network needs a large iteration for getting excellent outputs. The reason of that the Deep Q Network is using ε-greedy for action selection, and the ε is set to high value (close to one) in initial stage in learning. High ε value means that the agent selects action randomly in the learning. Hence, the agent needs large number of iteration of learning for accomplishing a task. In this paper adopts the Boltzmann selection to Deep Q Network. Finally, our algorithm has been applied to 2 kinds of arcade learning environment tasks, and results showed that our algorithm is better than ordinary Deep Q Network.

本誌: 電気学会論文誌C（電子・情報・システム部門誌） Vol.137 No.12 （2017） 特集Ⅰ：電気・電子・情報関係学会東海支部連合大会　特集Ⅱ：国際会議ICESS2016

本誌掲載ページ: 1676-1683 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/137/12/137_1676/_article/-char/ja/