Figure 5: ICGC 1988: Dragon (B) - Codan (W)
Figure 5: ICGC 1988: Dragon (B) - Codan (W)
The following three games illustrate the performance of the top Go programs in 1988, 1994 and 2000. The first game, shown in Figure 5, was the final of the International Computer Go Congress (Ing Cup), played in Taipei, Taiwan on November 11, 1988. The komi was 8 points, as in all Ing-sponsored competitions. Codan by Kazuyoshi Hayashi playing White won by 7 points against Liu Dong-Yue's Dragon. Overall, both programs play a very solid, territory-oriented game. Codan loses a group on the top edge, but gives it up early enough to avoid disaster, and wins the game by making slightly more efficient moves on average. Much of the game consists of simple, boundary-settling local sequences of play.
Figure 6: ICGC 1994: Handtalk (B) - Many Faces of Go (W)
Six years later, Ken Chen's Go Intellect won the same event on a tiebreak with five wins to one loss. The game shown in Figure 6 was played in Taipei, Taiwan on November 17, 1994 between the two other programs with five wins, Chen Zhixing's Handtalk and David Fotland's Many Faces of Go. This game is one of Handtalk's rare losses in the time from 1993 to 1997, when it dominated the computer Go scene. In contrast to the 1988 game, this game is characterized by intense fighting, with the focus on the attack and defense of weak groups. The ability to cut and connect weak groups, and the amount of knowledge about Life and Death, play the dominant roles. In this game, Handtalk makes two decisive group-related mistakes: move 99 loses the black group at the bottom. It should be played below 96, after which Black could either capture the stones 94 and 96 by playing to the left of 96, or capture the other cutting stone 98 by a move at 100. The second mistake is 141 (41 in the diagram on the right side), which lets White lead out the previously dead group in the center. After these two reversals, White has a big lead. Finally, White blunders at 158, which should be at 160 immediately, but Black misses the chance to turn the corner into ko and lets White repair the damage one move later.
Figure 7: Mind Sports Olympiad 2000: Go4++ (B) - Goemate (W)
As an example of the current state of the art, Figure 7 shows the deciding game of the computer Go tournament at the Mind Sports Olympiad, played in London on August 22, 2000. Playing Black and giving a komi of 6.5 points, Go4++ by Michael Reiss lost by just half a point to 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 26, standard opening sequences are played out to occupy all corners and sides. The moves from 27 to 33 are also a standard sequence. After that, both programs continue to surround territory, with Black simply giving up stones such as 25 and 35/47 rather than risking a big fight by running out. While the play of both programs in this game is rather simple and safe, their overall performance is very respectable. There may still a large number of less-than-optimal moves, but 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. Programs are careful to avoid getting weak groups, and play a reasonable endgame.
Comparing these three games, computer Go seems to have come full circle. Early programs knew little more than simple rules and patterns to surround territory and capture stones. The next generation, lead by Goliath, Go Intellect and Handtalk, dominated their opponents through a better knowledge of attack and defense. A typical game from this period was decided by a large margin, with the stronger program saving more of its own groups and killing more opponent groups. Go4++ lead the revolution leading to the current state, by demonstrating that a program which is not so strong in fighting but very efficient in taking territory can win a large percentage of games, even when giving up a few small groups along the way. In recent years, these two extreme approaches have converged to the point where it is hard to distinguish between playing styles. Current programs are stronger than fifteen years ago in judging group strength, life and death, and tactics, but most prefer to play a peaceful game where these strengths are not so apparent. Their style of play hides much of the inherent complexity of Go.