DOI QR코드

DOI QR Code

Full-board position evaluation of 50 AlphaGo vs AlphaGo games, using influence function

세력 함수를 활용한 알파고 간의 50개 대국에 대한 형세 판단

  • Lee, Byung-Doo (School of Artificial Intelligence, Yong-In University)
  • Received : 2021.05.10
  • Accepted : 2021.06.13
  • Published : 2021.06.20

Abstract

Full-board position evaluation in Go is a measurement of judging the advantages and disadvantages between black and white players during a game playing, and through this, the proper tactics and strategies would be undertaken in the near future. In this paper, we tried to evaluate the full-board positions of the 50 AlphaGo vs AlphaGo games using influence function that halved according to the distance. According to the experimental results, there is a limit to making accurate evaluation when the full-board position is assessed only by influence function. In order to overcome this, it is necessary to solve life-and-death problems to deal with dead stones, and it showed that if this is reinforced, we can precisely evaluate the full-board position in Go.

바둑에서의 형세 판단은 현재 대국 중인 흑백 대국자 간의 유불리를 판단하는 척도가 되며, 이를 통해 곧바로 적절한 전술과 전략을 구사하게 된다. 본 논문에서는 거리에 따라 반감하는 세력 함수를 활용하여 알파고 간의 50개 대국의 형세 판단을 하고자 했다. 실험 결과에 따르면 단지 세력 함수만을 사용하여 형세 판단을 하게 되면 정확한 판단을 함에 한계가 있음이 밝혀졌다. 이를 극복하기 위해 사석 처리를 위한 사활문제 해결이 필요하며, 이를 보강하게 되면 바둑에서의 정밀한 형세 판단을 할 수 있음을 보였다.

Keywords

Acknowledgement

이 논문은 2020년도 용인대학교 학술연구조성비 재원으로 수행된 연구임

References

  1. B.D. Lee, "Contour Tracing to Solve Life-and-Death Problem in Go", Journal of Korea Game Society, Vol. 20, No. 2, pp. 91-100, 2020. https://doi.org/10.7583/JKGS.2020.20.2.91
  2. M. Muller, "Counting the Score: Evaluation in Computer Go", from https://webdocs.cs.ualberta.ca/~mmueller/ps/goeval.pdf, 2021.
  3. Baduk Time, "How to compute full-board evaulation", from https://baduktime.com/g2/bbs/board.php?bo_table=research2&wr_id=950, 2021.
  4. B.D. Lee, "Multi-Strategic Learning, Reasoning and Searching in the Game of Go", PhD thesis, Auckland University, 2005.
  5. A. Hwang et al., "Move Evaluation in Go Using Deep Convolutional Neural Network", ICLR conference paper, pp. 1-8, 2015.
  6. DeepMind, "AlphaGo vs AlphaGo", from https://deepmind.com/alphago-vs-alphago, 2021.
  7. J. Burmeister and J. Wiles, "AI Techniques Used in Computer Go", from http://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=&ved=2ahUKEwid3riq1ZjwAhUGGqYKHQqgDOMQFjABegQIAhAD&url=ftp%3A%2F%2Fftp.cse.ucsc.edu%2Fpub%2Fcompgo%2FPAPERS%2Fcomp-go.AI.ps&usg=AOvVaw2EeByNtMJvaoMWfH8my7PA, 2021.
  8. Sensei's Library, "Influence function", from https://senseis.xmp.net/?InfluenceFunction, 2021.
  9. A. Zobrist, "Feature Extraction and Representation for Pattern Recognition and the Game of Go", PhD thesis, University of Wisconsin, 1970.
  10. Wikipedia, "Breadth-first search", from https://en.wikipedia.org/wiki/Breadth-first_search, 2021.
  11. B.D. Lee, "The Best Sequence of Moves and the Size of Komi on a Very Small Go Board, using Monte Carlo Tree Search", Journal of Korea Game Society, Vol. 10, No. 5, pp. 77-82, 2018.
  12. J.U. Kim, "Understanding the Rules of Baduk", PUBPLE Press, 2017.