Eighteenth KGS Computer Go Tournament

Sunday September 3rd, 2006

These results also appear on official KGS pages: Formal Division, Open Division which link to the game records.


Both divisions:
19x19 board
Chinese rules, komi 7½
58 minutes each absolute time


Both divisions: four-round Swiss.


The first round started at 16:00 UTC for the Formal and 16:10 for the Open division, subsequent rounds started at two-hour intervals.


As usual, the tournament was held in two divisions, Formal and Open, with more restrictive entry conditions for the Formal division.

Formal Division   19x19


Open Division   19x19


The "real" names of the bots listed above, and of their programmers, are listed here: programs which have registered for KGS Computer Go Tournaments.

Formal division

In round 1 we saw that the Monte Carlo programs have a tenedncy to win by narrow margins – MoGoBot beat the much weaker PyBot by 3½, while (in the Open division) StoneCrazy beat the much weaker SimpleBot by ½. My guess is that there is a simple reason for this. It has been found that getting a MC program to maximise its chance of winning is more effective than getting it to maximise its margin of victory. So towards the end of the game, when these MC programs are 100% confident of winning, they don't care how many points they give away as long as they are confident they are still ahead. Their opponents are happy to accept the points they are randomly offered, and may be able to catch up until the game is very close.

In the round 2 game MoGoBot/GNU, MoGoBot chased GNU's stones in a ladder for which GNU had a ladder-breaker. It was unable to recover from this.

After seven moves of its game with PyBot, AyaBot disappeared from the server, and eventually lost on time.

In round 3, AyaBot was still missing. It was lucky to get a bye, so its absence did not harm its score. Eventually its owner connected to the server and realised that it needed to be restarted.

The end of the PyBot/Orego game was strange. Both players passed, then they agreed on the status of all the groups (there was nothing at all difficult here). Then Orego vanished from the server.
         This gave me, as organiser, a problem. If Orego have vanished before the end of the game, it would have lost on time. It was clearly ahead on the board, but it had not actually finished the game. I decided to treat the game as I would have done if the players had failed to agree on the status of a group, and adjudicated it myself, as a win for Orego.

In round 4, Orego was still missing. It was drawn as black against GNU Go, and as it never appeared to make its first move, it eventually lost on time.
         However, there is something that worries me about this game. GNU was present at the start of the game, and waited for its opponent to make a move. Eventually GNU was removed from the game by the server. If, after GNU had been removed, Orego had arrived and made a move, I fear that GNU might have lost on time, through no fault of its own. Or would the server have invited GNU to rejoin, if Orego had appeared?
         GNU's kgsGtp log file, sent to me by its operator Gunnar Farnebäck, complete from the end of the round 3 game until the end of the round 4 game, is in the panel to the right. The warning lines, shown here in bold, are the most relevant ones.

I still feel that Monte Carlo techniques lead to play that is rather dull for the spectators. It is more interesting to watch programs that use non-MC methods, such as AyaBot. The figure to the right shows the game CrazyStone-AyaBot.sgf after White (CrazyStone) has played move 254.

White's left-centre group has only one eye, and no plausible way to make another. As long as this white group remains dead, Black is well ahead. White's best chance must be to kill Black's centre group. But it is Black's move, and it has several simple ways to make a second eye for its centre group.

However, Aya was not content to win the game. Like many human 10-kyus, it decided it would like to kill even more stones, and played at the marked point. This led to a sequence with many blunders and missed opportunities, in which White killed the black centre group, and Black brought it back to life and managed to kill the white centre-right group as well. While this may be bad Go, it is a better spectator sport.

Formal division

SimpleBot did well in losing its first-round game with StoneCrazy by only half a point. It did worse after that, not making a single live group. Its owner Aloril has explained the problem to me. It keeps a list of points not to be played on because they are inside secure territory. The version that played in this tournament was failing to clear this list between games.

Forfeited games

The following "forfeited games" record are available by clicking on them:
Round 1, Open division: MoGoBot19-WeakBot50k.sgf
Round 3, Formal division: PyBot-Orego.sgf

Other game records are available from the KGS pages: Formal Division. Open Division.