The 1st World AI Go Open, 2017 CITIC Securities Cup, was held in Ordos City, Inner Mongolia, China, during 16–18 August 2017. Twelve teams in total participated in the 1st World AI Go Open where four were from China, three from Japan, and the remaining five from Taiwan, Korea, Belgium, France and the USA, respectively, as listed in Table 1.
The board size was set to be 19 by 19. Chinese rules (Chinese Weiqi Association, 2002) were applied. The tournament was separated into two stages, the preliminary stage and the knockout stage, held on the first and second days respectively. During the preliminary stage, the Swiss system was used with five rounds. The best eight entered the knockout stage. As for thinking time, in the preliminary stage, each program had 30 minutes of thinking time per game. If a program ran out of time, it lost immediately. In the knockout stage, each program had 30 minutes with 10 times 20-second byo-yomi (countdown).
The participants
Program Name
Country
Program Name
Country
FineArt
China
AQ
Japan
Abacus
China
DolBaram
Korea
Tianrang
China
Leela
Belgium
OracleWQ
China
Golois
France
DeepZenGo
Japan
MuGo
USA
Rayn
Japan
CGI
Taiwan
Table 2 shows the results for the preliminary stage. CGI, developed by Computer Games and Intelligence (CGI) Lab, Department of Computer Science, National Chiao Tung University, Taiwan, won all of its five games and ranked first. FineArt, developed by Tencent Holdings Limited, the largest Internet company in China, won the second by winning four games and losing one against CGI. The third was DeepZenGo, which has won several computer Go champions in the past. The fourth to the eighth were TianRang, Rayn, DolBaram, AQ and Leela respectively.
The detailed scores of the preliminary stage
No.
Program Name
Score
SOS
Rank
Game I
Game II
Game III
Game IV
Game V
Opp.
Score
Opp.
Score
Opp.
Score
Opp.
Score
Opp.
Score
1
FineArt
4
15
2
9
1
6
1
7
1
12
0
5
1
2
Abacus
2
9
10
5
0
11
1
8
0
10
0
4
1
3
Tianrang
3
15
4
7
1
8
1
12
0
5
0
10
1
4
OracleWQ
0
11
12
8
0
7
0
6
0
11
0
2
0
5
DeepZenGo
3
16
3
2
1
12
0
9
1
3
1
1
0
6
Rayn
3
14
5
10
1
1
0
4
1
8
1
12
0
7
AQ
2
12
7
3
0
4
1
1
0
9
1
8
0
8
DolBaram
3
10
6
4
1
3
0
2
1
6
0
7
1
9
Leela
2
12
8
1
0
10
1
5
0
7
0
11
1
10
Golois
2
11
9
6
0
9
0
11
1
2
1
3
0
11
MuGo
1
11
11
12
0
2
0
10
0
4
1
9
0
12
CGI
5
14
1
11
1
5
1
3
1
1
1
6
1
Fig. 1 shows the results at the knockout stage. In the quarterfinal, CGI defeated Leela, Tianrang defeated Rayn, DeepZenGo defeated Dolbaram, and FineArt defeated AQ. In the semifinal, CGI defeated Tianrang and DeepZenGo defeated FineArt. Finally, DeepZenGo won against CGI and got the title of “The 1st World AI Go Open Champion.” A human vs. AI match was then held in the third day, where the champion DeepZenGo played against world champion KongJie 9p, who was partnered with the second place, CGI, as his assistant. The winner of the match was DeepZenGo.
The results of the knockout stage.
Selected games
This report comments on two games: one is the game where CGI defeated FineArt in the preliminary stage and the other is the game where DeepZenGo defeated CGI in the final. In the first game, CGI won with a critical life and death problem in shoban (endgame). In the second game, CGI made some mistakes during the endgame and lost. The following game records are represented in Smart Game Format (SGF), where Color[xy] denotes the player (Color is B for black and W for white) playing at position on the 19 ×19 board.
In the beginning, two AI programs played equally strong. In fact, since AI programs have surpassed human Go knowledge recently, it is hard to comment on the moves played by AI programs. However, the key that led to White winning is easily observed even for amateur players. For the 241st move near the endgame, Black seemed to misjudge a life-and-death group near the upper right corner of Fig. 2 (left). It is like a classical bent four in the corner problem, which can generally be killed by ko. However, in this game, White had more outside liberties, thus it could live without ko. This life-and-death case can be solved by dan amateur players, while strong AI Go programs may still miscalculate occasionally. In this game, CGI judged the life-and-death correctly and thus got more territory in the upper right corner of Fig. 2 (right). Hence, CGI won the game at the end.
Game 2: CGI (Black) vs. DeepZenGo (White) b+resign
There are several interesting moves during the game. First, Black played the 13th move focusing on the right side of the lower right corner of Fig. 3 (left) and abandoning the corner. For this move, human players often play hane (a move that goes around one or more of the opponent’s stones) at 15 and make a tiger mouth at 19 first in order to make the territory of the corner stable. However, after Black played at 29 and 31, the size of the black territory in the lower space became big enough to cover the loss of the lower right corner. Second, the cut of Black move 67 split white stones into two groups without an eye. After several attacking and defensing moves, Black took 10 stones being ahead of White. Third, the moves around 150 were too complicated for even professional players to explain precisely. Finally, CGI made two blunders near the endgame. At the 203rd move, CGI only took one stone while the advantage of playing at 204 seemed to be much bigger. In addition, for the 223rd move, CGI should have played hane at 224 instead. For these blunders, CGI lost the game. Fig. 4 shows the win rates during the game which were generated by CGI (DeepZenGo also showed similar trends). In the middle game, CGI was in advantage. However, the win rates dropped dramatically near the endgame.
Win rates during Game 2 generated by CGI.
Final remarks about computer go
Since the program AlphaGo, developed by DeepMind, defeated Lee Sedol, a top professional player (Silver et al., 2016), computer Go has entered a new era. One of the most important techniques that led to the success was deep reinforcement learning. In late 2017, DeepMind published a more advanced version of AlphaGo without human knowledge, named AlphaGo Zero (Silver et al., 2017), and announced they were not continuing the development of computer Go. However, several other teams continued to develop computer Go programs based on the techniques used in Zero. For example, Leela called for volunteers to build the Go program based on Zero (Pascutto, 2017), and Facebook released an open-sourced Go program called ELF OpenGo (Tian et al., 2017), which uses AlphaGo Zero’s architecture. These are open for further research on computer Go. In the future, it is expected that tournaments like CITIC will provide developers with more opportunities to exchange experiences and research results.
I-Chen Wu and Ti-Rong Wu from CGI (left) and Hideki Kato from DeepZenGo (right).
References
1.
“Bent four in the corner.” From https://senseis.xmp.net/?BentFourInTheCornerIsDead.
2.
Chinese Weiqi Association (2002). Rules of Weiqi (Go), retrieved from http://home.snafu.de/jasiek/c2002.pdf.
3.
Tian, Y., Gong, Q., Shang, W., Wu, Y. & Zitnick, C.L. (2017). ELF: An extensive, lightweight and flexible research platform for real-time strategy games. In Advances in Neural Information Processing Systems (NIPS) (pp. 2656–2666).
4.
Pascutto, G.-C. (2017). Leela Zero. From https://zero.sjeng.org/.
5.
“Smart Game Format.” From https://www.red-bean.com/sgf/index.html.
6.
Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., et al. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587), 484–489. doi:10.1038/nature16961.
7.
Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., et al. (2017). Mastering the game of Go without human knowledge. Nature, 550(7676), 354. doi:10.1038/nature24270.
8.
“The 1st World AI Go Open 2017.” From http://waigo2017.chinaweiqi.net/EN/index.html.
