The computer Othello 10 × 10 tournament was held as part of the 2018 Computer Olympiad, which took place in Taipei, Taiwan, from July 9th to 10th, 2018. A total of six teams participated in the tournament. The participating teams and the final standings are listed in Table 1. All programs, with the exception of Othello LTBeL, were based on tree search with Alpha-beta pruning. Othello LTBeL was implemented following the AlphaZero approach. Some programs use deep learning to play several moves for the opening book.
Othello 10 × 10 was first included as an event in the 2017 ICGA Computer Olympiad with a total of three participants. This year there were six participants, many of them strong programs. The rules of Othello are very simple. Two players take turns placing a piece on the size 10 × 10 board. For each ply, a player can only capture the opponent stones or pass. The piece played on the four corner cannot be captured. When no players can play, the game is over. The player who has more pieces (disks) on the board wins. In the tournament, the games were played according to a round-robin system in which one program played twice against all other programs, playing as black and white once each. For each game, every program was required to complete all of its moves in 30 minutes. The winner scored 2 points and the loser scored zero. For a draw, both scored 1.
The participants and final standings
Ranking
Program
Author
Points
1
Decapus
Fabien Letouzey
20
2
Deep Nikita
Andrew Lin
15 (playoff)
3
Othello LTBeL
Ching-Zong Chang
14
4
Curiosity10
Wei-yuan Hsu
8
5
Persona
Surag Nair, Nai-Yuan Chang and Shun-Shii Lin
4
6
Uncle Darius
Tai-Xiang Wang and Zi-Yu Xiao
0
Decapus won the gold of the 2018 Computer Olympiad Othello 10 × 10 tournament by winning all its games. Deep Nikita won the second place after playoff against Othello LTBeL. Othello TLBeL earned the bronze medal with 14 points. The cross table is listed in Table 2.
The cross table
Program
Decapus
DeepNikita
Curiosity10
UncleDarius
OthelloLTBeL
Persona
Total score
Decapus
–
4
4
4
4
4
20
Deep Nikita
0
–
4
4
2
4
14
Curiosity10
0
0
–
4
0
4
8
Uncle Darius
0
0
0
–
0
0
0
Othello LTBeL
0
2
4
4
–
4
14
Persona
0
0
0
4
0
–
4
This report provides commentary on two games between Decapus (gold) and Deep Nikita (silver). In the first game, Deep Nikita played first; the move sequences are f7 g5 f4 e7 d6 g7 d7 e4 d5 g6 d4 c8 d8 e8 g4 c7 f8 e9 h8 h7 h5 c5 b6 c4 c6 a5 b4 b5 h6 a3 g8 c3 b3 d3 e3 c2 a4 d2 e2 f3 e1 c1 g3 f1 g1 f2 g9 f9 f10 h4 i5 h2 g2 h1 d1 h3 i4 i3 j3 j6 j5 i6 i7 j8 j7 h10 d10 e10 g10 c10 d9 c9 a7 a6 b7 h9 i8 a8 b8 j9 j10 i10 i9 j4 b9 a10 b10 a9 j2 j1 i1 i2 a2 a1 b1 b2.
In Fig. 1, the game state was about equal. Then, Deep Nikita (black) played e3. After that move, black had many disks next to an empty square (the frontier), losing mobility in the long run. For this reason, c2 would have been a better move.
In the second game, Decapus played first; the move sequences are f7 g5 f4 e7 d6 g7 d7 e4 d5 g6 d4 c8 d8 e8 g4 c7 f8 e9 h8 h7 h5 c5 b6 c4 c6 a5 b4 b5 h6 a3 g8 c3 b3 d3 e3 c2 a4 d2 e2 f3 e1 c1 g3 f1 g1 f2 g9 f9 f10 h4 i5 h2 g2 h1 d1 h3 i4 i3 j3 j6 j5 i6 i7 j8 j7 h10 d10 e10 g10 c10 d9 c9 a7 a6 b7 h9 i8 a8 b8 j9 j10 i10 i9 j4 b9 a10 b10 a9 j2 j1 i1 i2 a2 a1 b1 b2. After a good opening and a few small mistakes, Deep Nikita (white) was in a difficult position. At move 36 in Fig. 2, it played d4, which was the losing move; c6 would have been a better choice.
Round 4, game 1: Deep Nikita (black) – Decapus (white).
Decapus uses tree search with Alpha-beta pruning. On top of that, it combines a fast and accurate evaluation with heavy forward pruning. The pruning heuristics are a simplified variant of ProbCut Buro (1997) and late-move reductions. The evaluation function is centered around patterns (tuple networks). The board is divided into overlapping regions (the patterns (tuple networks)) and every possible instance (a configuration) in a region gets a score, computed using machine learning. The evaluation value is the sum of local scores. The pattern-configuration scores were learned as a preprocessing step using linear regression because there is a game score (disk difference). A large database of self-play games was used as the training set (supervised learning).
Round 4, game 2: Decapus (black) – Deep Nikita (white).
References
1.
Buro, M. (1997). Experiments with Multi-ProbCut and a new high-quality evaluation function for Othello. In Games in AI Research (pp. 77–96).
