Earticle

초록

영어

Analysing Go games with AIpowered engines is now common practice among amateur and professional Go players. These engines rely on deep artificial neural networks (ANN) and MonteCarlo tree search (MCTS) to evaluate board positions. For a given ANN, the engine’s accuracy increases with the number of positions evaluated during the tree search, or number of visits (often called the number of playouts). Although more playouts lead to more accurate analyses, the relationship between playouts and engine accuracy has not been fully explored. In this study, we analysed thousands of games from amateur and professional players, using different numbers of playouts per move, ranging from 5 to 15,000. A statistical analysis of the results shows that a constant number of playouts per move leads to varying accuracy during the game (high accuracy during the opening and endgame, low accuracy during the middle game). Quantitative guidelines to choose the number of playouts depending on one’s rank are derived. Looking at AI prediction of human win rate, the playouts needed to reach 99% accuracy are 140 for 1d, 235 for 5d, and 1,500 for 9d Fox players and 5,500 for a 9p professional player.

목차

Abstract
Ⅰ. Introduction
Ⅱ. Methods
Ⅱ.1 Go games datasets
Ⅱ.2 Evaluation of Katago’s accuracy
Ⅲ. Results
Ⅲ.1 From Opening to Endgame
Ⅲ.2 Impact of Player’s Rank anAI perspective
Ⅲ.3 Impact of Player’s Rank aHuman perspective
Ⅳ. Balancing AI Accuracy and Computational Power
Ⅴ. Conclusions
References

키워드

저자

• Quentin Rendu [ Hamburg University of Technology (TUHH), Hamburg, Germany ]

참고문헌

발행기관

간행물

표지보기 관심저널 등록

• 간행물명

  바둑학연구 [Journal of Go Studies]

• 간기

  반년간

• pISSN

  1738-3730

• 수록기간

  2004~2026

• 십진분류

  KDC 691 DDC 794