Fifty-seventh KGS Computer Go Tournament

Sunday March 7th 2010

These results also appear on an official KGS page which links to the records of all the games.


format8-round Swiss
board size19×19
time29 minutes plus 25/20s


The first round started at 16:00 UTC.

Result table


The numbers in this table may not add up as you
might expect. This is because byes, and games
against absent opponents, are both treated as wins.


There were thirteen players registered, so I could have removed IdiotBot to make the numbers even before round 1. However, break19 had sometimes not appeared in the past, so I left IdiotBot in the draw in case break19 failed to play.

After round 1 had started, break19 appeared late, and eventually lost its game with WeakBot50k on time.

SimpleBot played its first few moves against PNUGo very slowly, taking 20 of its 29 minutes for its first six moves. It then began to speed up, but not by enough, and eventually lost on time, only playing 24 of its 25 stones in the first 30-second byo-yomi period.

PueGo vs ManyFaces1
Moves 101, 102, 106-108.
(103-105 tenuki).
Moves 140-142.

I removed IdiotBot from the draw before round 2.

Normally I do not feel able to comment on these games, as I am a 2-kyu player and many of the programs are stronger than me. But here is an example where a human 2-kyu (or even a 5-kyu), with some understanding of eye shape and in particular the "bent four in the corner" shape, can "outread" both players.
       In round 2, the position shown to the right appeared in the game between PueGo and ManyFaces1 SGF. Move 101 is not good, but it is not a disaster. Move 102 is good, and leaves the black corner group unsettled. Both players ignored the unsettled corner for the next three moves; then Puego played 106, which kills the corner.
       However, PueGo did not understand what it had achieved. Later, when ManyFaces1 played 141, it responded with 142, which leaves the corner almost totally alive (it is a multi-move approach ko in Black's favour). If instead it had played 142 on the corner point, the black group would still have been dead.

In the round 3 game betwen pachi2 and Manyfaces1 SGF, pachi2 started well, and had chances of killing a central group. Even afer this group made two eyes, pachi2 was ahead, but was getting low on time. It blundered with move 268, when it had only 33 seconds of main time left, and had no chance of a win after that.

In the round 4 game betwen pachi2 and Zen19 SGF, pachi2 started well, and was ahead until move 240. By then it had less than three minutes left, while Zen19 still had seven minutes; and again pachi2 started to blunder under time pressure, and again lost.

In Conclusion

Many Faces of Go won this event convincingly. I felt this was well deserved. It was using 16 cores, and its author David Fotland has recently removed a bug which had been preventing it from taking full advantage of multiple cores. He has also been working on other improvements.

I get the impression that Hiroshi Yamashita's Aya continues to improve steadily.

Yamato's Zen did not do as well as expected, though it continues to maintain a solid 2d rating on KGS. Maybe it was unlucky in this event. I hope we can one day see it again with the "Zengg" changes made by Hideki Kato.

Erik van der Werf's Steenvreter has recently acquired a rated account (stv1) on KGS, and is now also rated at 2d there, though currently over only 11 games.

Pachi shows that it can play well. But its author Petr Baudiš suspects it has a bug in its time management. I will not be surprised if it improves significantly over the next few months.

I think it is fair to say that there are now several unrelated programs (all Monte-Carlo based) which are 2-dan, on the KGS scale as calculated from 19×19 games against human opponents. Much work has been done using MC methods since UCT was first published in September 2006 by Kocsis and Szepesvári. The use of RAVE MC tree search methods is now used as well as the original UCT, and there appears to be scope for further programming improvements in the use of MC methods. We can also expect gains from use of more processor power, as MC programs, unlike their predecessors, are able to make effective use of increased processor speed and numbers.

Details of processor numbers, power, etc.

Aya, running on Xeon 2.6GHz 8cores
break, probably running on a single processor Intel(R) Celeron(R), 1.7Ghz
MoGo, running on double-core AMD Athlon(tm) 64 X2 Dual Core Processor 4800+ (2.5GHz).
Many Faces of Go, running on 16 cores: four Core2 Quad 2.3 GHz.
Orego, unspecified but probably running on 2 x 3 GHz Dual-Core Intel Xeon
pachi, running on 22 cores (2.4 GHz) of three Mac Pro machines, with eight cores each.
GNU Go, unspecified platform
Fuego r1098, running with eight threads on an i7 920.
running on one processor of a 4GiB RAM, AMD Athlon(tm) 64 X2 Dual Core Processor 4000+, shared with WeakBot50k
Steenvreter, running on an Intel Core2, Q6600 (4 cores)
running on one processor of a 4GiB RAM, AMD Athlon(tm) 64 X2 Dual Core Processor 4000+, shared with SimpleBot
running on a Mac Pro 8 core, Xeon 2.26GHz.