機械学習による囲碁の着手の日本語表現  [in Japanese] Japanese expression of the move of Go by machine learning  [in Japanese]

Access this Article

Search this Article

Author(s)

Abstract

近年の囲碁プログラムの強さは,プロ棋士に 4 子のハンデで勝つなど,ほとんどのアマチュアにとって充分な域に達しつつある.そのため,次の段階として人間を教える・楽しませるといった目的での研究も盛んになってきている.指導碁や接待碁で人間を来しませる要素の 1 つに 「感想戦,検討,対局中のお喋り」 があるが,このためには "形" を表現する単語 (ツケ,ハネなど) をコンピュータに表現させることが望ましい.そこで本論文では,機械学習を用いて盤面と若手から単語を導くことを目指した.まず,形の単語を約 70 種類に絞ったうえで,アマチュア高段者 6 人に棋譜を波して各着手にラベル付けをしてもらった.この際,「ハネとも言えるし,オサエとも言える」 ような手が頻繁にあるという困難さを考慮し,複数のラベルを付けることができるようなフォーマットとし評価の参考とした.学習には,着手の周囲の配石パターン以外に,呼吸点の変化や石が何線にあるかなど囲碁特有の特徴畳を用いることで性能向上を図った.人間同士であっても単語の一致率は約 82% にすぎないが,比較的単純な機械学習でもこれに近い値を出すことに成功した.着手の日本語表現によって,コンピュータとの感想戦,検討,お喋りの実現に近づくとともに,初級者の知識定着も図ることができる.: Computer Go programs have recently won against professional players with a 4-stone handicap, which is a level of strength sufficient for most amateur players. A new target for research is then to create programs able to entertain or teach Go to human players, but communication is a major obstacle, especially because moves in the game of Go are described by many specific terms such as Tsuke or Hane. In this research, our goal is to make the program able to label the moves with their associated specific term. We used machine learning to deduce the term for a move from the local patterns of stones. First, 6 strong amateur Go players recorded for each move of some game records the corresponding specific term, or possibly multiple terms, from a pre-selected list of 71 terms. Secondly, a machine learning algorithm was executed and the performance was improved by using not only the local patterns of stones but also features specific to the game of Go, such as changes of liberties or distances to the edge of the board. The human players associated the same specific term to a move at a rate of 82% and our progam succeeded to achieve a similar rate although the machine learning method was rather simple. Such derivation of the terms for moves is a first step towards Go programs able to chat with human players during game reviews or matches.

Computer Go programs have recently won against professional players with a 4-stone handicap, which is a level of strength sufficient for most amateur players. A new target for research is then to create programs able to entertain or teach Go to human players, but communication is a major obstacle, especially because moves in the game of Go are described by many specific terms such as Tsuke or Hane. In this research, our goal is to make the program able to label the moves with their associated specific term. We used machine learning to deduce the term for a move from the local patterns of stones. First, 6 strong amateur Go players recorded for each move of some game records the corresponding specific term, or possibly multiple terms, from a pre-selected list of 71 terms. Secondly, a machine learning algorithm was executed and the performance was improved by using not only the local patterns of stones but also features specific to the game of Go, such as changes of liberties or distances to the edge of the board. The human players associated the same specific term to a move at a rate of 82% and our progam succeeded to achieve a similar rate although the machine learning method was rather simple. Such derivation of the terms for moves is a first step towards Go programs able to chat with human players during game reviews or matches.

Journal

  • 研究報告ゲーム情報学(GI)

    研究報告ゲーム情報学(GI) 2015-GI-33(4), 1-7, 2015-02-26

    Information Processing Society of Japan (IPSJ)

Codes

  • NII Article ID (NAID)
    110009882471
  • NII NACSIS-CAT ID (NCID)
    AA11362144
  • Text Lang
    JPN
  • Article Type
    journal article
  • ISSN
    09196072
  • Data Source
    NII-ELS  IR 
Page Top