Two Representative Computer Go Games

These two games illustrate the state of the art in 2001.

Case Study 1: Ing Cup 1999, Go4++ vs Goemate

Figure 25: Ing Cup 1999: Goemate (B) - Go4++ (W), moves 1-50

Figure 26: Ing Cup 1999: Goemate (B) - Go4++ (W), moves 51-100

Figure 27: Ing Cup 1999: Goemate (B) - Go4++ (W), moves 101-200

Figure 28: Ing Cup 1999: Goemate (B) - Go4++ (W), moves 201-248. 231 captures below 230, 245 passes

Figures 25 to 28 show the deciding game of the 1999 Ing Cup, played in Shanghai on November 13, 1999. Playing White and receiving a komi* of 8 points, Go4++ by Michael Reiss won by 13 points over Goemate, developed by Chen Zhixing as the successor program of Handtalk. This game develops in a tight territorial fashion typical of most current top programs, with little fighting going on. Up to move 16, both programs follow standard opening principles by first surrounding the corners and then expanding to the sides. The standard joseki* moves from 16 to 23 are most likely contained in the opening book of both programs. With move 30, Go4++ starts reducing the large framework that Black has built on the left side. Black invades strongly at 31 and 39, and even though later White can connect the result is not bad for Black in both cases. White 52 is a strange shape move, but it succeeds in splitting up Black's left side. Around move 63 the game has already become an endgame contest. Black 75 is too passive. White 76 threatens to destroy the bottom left side, but Black fails to defend, allowing successive moves at 94 and 112. However, in return Black captures some white stones in the center in the sequence from 83, and until move 130 the game remains very close. 131 is an inexplicable retreat and must have been caused by a programming bug. Of course Black should just connect at 132. Up to 137 Black loses more than 10 points. The remaining endgame is uneventful, and White achieves a safe win.

The performance of both programs in this game is respectable. Their play is rather simple and safe, mostly surrounding territory. While there is still a large number of less-than-optimal moves, there are few really big mistakes. Both programs demonstrate an understanding of many aspects of Go. For example, they can build safe territory as well as large frameworks, and can react early to reduce an opponent's sphere of influence. The programs clearly incorporate important principles of Go: they don't just apply rote patterns as the early Go programs typically did. This is evident in situations such as move 52, where the program has insufficient detail knowledge to produce a stylish move, but is still able to find the right idea and select a reasonable play in that general area. Top programs are very careful to avoid getting weak groups, and can play a normal endgame. In this game, White was especially skillful in eliminating the opponent's potential sente* moves.

This game shows current computer Go at its best. However, it cannot be denied that the style of play, which is typical of recent tournament games, hides much of the inherent complexity of Go. In contrast, in the next case study another top program is subjected to a more severe test.

Case Study 2: Giving Many Faces a 29 Stone Handicap

In August 1998, at the US Go congress in Santa Fe, New Mexico, a 29 stone handicap game was played between David Fotland's program The Many Faces of Go and the author, a 6 dan amateur player. Like Go4++ and Goemate, Many Faces is one of the strongest Go programs in the world, and a few months after this game it won the 1998 Ing Cup. Many Faces is regarded as one of the best programs when it comes to tactical fighting. However, in this case its aggressiveness backfires. Despite the huge handicap, the game ends with a six point win for the human.

Figure 29: Many Faces (B, 29 stones handicap) - Martin Müller (W)

Figure 30: Many Faces (B, 29 stones handicap) - Martin Müller (W), moves 1 - 100. Move 28 is played back at 21, 39 at 31, 48 at 32, 50 captures below 31, 55 at 31, 57 at 43

Figure 31: Moves 101-200: 165 at 157, 196 connects below 184

Figure 32: Moves 201-279: 236 at 223, 264 connects right of 256, 275 passes

Figures 29 to 32 show the starting position and the game record. In the beginning, White sprinkles some stones around the board to probe for weaknesses, but Black defends well. In the bottom left corner, White uses a confused ko* fight to secure one group, then continues to create complications from this basis. By move 85, Black's group in that corner has been reduced to only one eye*. Black invests too many moves in a failed attempt to rescue this group and in a counterattack against the white stones floating in the center. With move 207 (7 in Figure 32), White isolates another black group in the lower right corner, and kills it a bit later by a combination exploiting a hidden dependency between two seemingly safe eye areas. This second big capture makes the game very close, and White easily overtakes Black in the remaining endgame. Throughout this game, most computer moves are quite reasonable, but there are just enough mistakes to allow White to grind out a win.

Conceptually, Black's main problem seems to be that the program tries to fight it out with a stronger opponent on even terms, instead of preserving some of its huge initial advantage by playing slow, ultra-safe moves. One might argue that a program playing a safer, more territorial style will be harder to overcome. However, a few weeks before this game, in an exhibition match held at the AAAI conference, professional player Janice Kim had already given Handtalk a similar huge handicap and won [52].

File translated from TEX by TTH, version 2.25.
On 15 Aug 2001, 11:15.
Edited by Martin Müller.