On Semeai Detection in Monte-Carlo Go
A frequently mentioned limitation of Monte-Carlo Tree Search (MCTS) based Go programs is their inability to recognize and adequately handle capturing races, also known as semeai, especially when many of them appear simultaneously. The inability essentially stems from the fact that certain group status evaluations require deep lines of correct tactical play which is directly related to the exploratory nature of MCTS. In this paper we provide a technique for heuristically detecting and analyzing semeai during the search process of a state-of-the-art MCTS implementation. We evaluate the strength of our approach on game positions that are known to be difficult to handle even by the strongest Go programs to date. Our results show a clear identification of semeai and thereby advocate our approach as a promising heuristic for the design of future MCTS simulation policies.
KeywordsKernel Density Estimation Terminal Position Game State Bandwidth Parameter Score Cluster
We like to thank Ingo Althöfer for pointing the community to the potential of score histograms and thereby encouraging our work, as well as for kindly commenting on a preliminary version of this paper. We also thank Rémi Coulom for comments on an early version of this paper as well as for kindly providing an automated analysis by his Go program Crazy Stone of one of the discussed example positions. We further thank Shih-Chieh (Aja) Huang and Martin Müller for the creation and sharing of their regression test suite.
- 1.Baudiš, P., Gailly, J.: PACHI: state of the art open source Go program. In: van den Herik, H.J., Plaat, A. (eds.) ACG 2011. LNCS, vol. 7168, pp. 24–38. Springer, Heidelberg (2012)Google Scholar
- 4.Coulom, R.: Computing Elo ratings of move patterns in the game of Go. ICGA J. 30, 198–208 (2007)Google Scholar
- 5.Coulom, R.: Criticality: a Monte-Carlo heuristic for Go programs. Invited Talk at the University of Electro-Communications, Tokyo, Japan (2009)Google Scholar
- 7.Gelly, S., Wang, Y., Munos, R., Teytaud, O.: Modifications of UCT with patterns in Monte-Carlo Go. Technical report 6062, INRIA (2006)Google Scholar
- 9.Huang, S.-C., Müller, M.: Investigating the limits of Monte Carlo tree search methods in computer Go. In: van den Herik, H.J., Iida, H., Plaat, A. (eds.) CG 2013. LNCS, vol. 8427. Springer, Heidelberg (2014) Google Scholar
- 10.Pellegrino, S., Hubbard, A., Galbraith, J., Drake, P.D., Chen, Y.-P.: Localizing search in Monte-Carlo go using statistical covariance. ICGA J. 32(3), 154–160 (2009)Google Scholar
- 12.Silver, S.: Reinforcement learning and simulation-based search in computer Go. Ph.D. thesis, University of Alberta (2009)Google Scholar
- 13.Silver, D., Tesauro, G.: Monte-Carlo simulation balancing. In: International Conference on Machine Learning, pp. 945–952 (2009).Google Scholar
- 15.Wistuba, M., Schaefers, L., Platzner, M.: Comparison of Bayesian move prediction systems for computer Go. In: Proceedings of the IEEE Conference on Computational Intelligence and Games (CIG), pp. 91–99, September 2012Google Scholar