Hundred and fifth KGS Computer Go Tournament

Sunday August 3, 2014

These results also appear on an official KGS page.


formatDouble round robin
board size13×13
time19 minutes plus 10/30s


The first round started at 08:05 UTC.

Result table

1pachi [-]1 1/22/22/22/22/29Winner
2AyaMC [2d]21/2 2/21/21/22/27
3coldmilk [-]30/20/2 2/22/22/26
4NiceGo19N [-]40/21/20/2 2/22/25
5HiraBot [1d]50/21/20/20/2 2/23
6Orego12 [6k]60/20/20/20/20/2 0

Black won 9 games and White won 13.

Players and format

Six players registered. One of them was 'HiraBot', a new program to these KGS tournaments. HiraBot is the creation of Kiyoshi Fukumoto.

I had previously stated that if there were six or fewer players, I would use a "double round robin" format for the tournament, instead of the usual "Swiss". So I cancelled the Swiss tournament with 911 at the end of its KGS URL (leaving two weak players in it for absurd technical reasons), and ran the double round robin with 912 at the end of its URL and starting five minutes later.

An unfortunate effect of using double round robin is that the crosstable above is copied directly from KGS, and does not link to the games played. When I run a Swiss, I use a perl script to create the crosstable, with links to game records, that I normally display with these reports. But I am not willing to modify my script so that it can also handle round robins.


pachi vs coldmilk
Move 79

After round 1, Orego12 disconnected and did not return. It therefore forfeited its remaining nine games. Its creator and operator Peter Drake has apologised for this. He has set it up wrong so that it disconnected after its first game.

In round 5, pachi and coldmilk had an enjoyable and confusing semeai with a ko, started by the move shown to the right. As far as I could tell, it was played competently. Pachi won the game.

Mid-way through the tournament, when each pair of players had played one game, the scores were


Each player had beaten all those below it in the list, except that NiceGo19N had beaten AyaMC.

HiraBot vs NiceGo19N
Moves 38 & 39

In the round 6 game between HiraBot and NiceGo19N, HiraBot played move 38 as shown to the left. This surprised me. It seems to me that it must be better for White to play 38 at 39, Black captures one stone, White now plays at 38, Black connects. This way White has the same shape as in the line actually played, while Black has lost an eye. I know that programs do not use such thinking; but in this case the result is so clear that I am surprised MC roll-outs did not generate the same result.

coldmilk vs pachi
Move 85

In the round 10 game between coldmilk and pachi, pachi played move 85 as shown to the right. Its creator remarked "pachi keeps considering n7 ... but always switches to something else ... and winrate is going now ... there's some problem in the simulations". Indeed, a move at n7, shown as "A" in the diagram, kills the white group and must win the game.
    However, pachi won this game, and I have failed to find a point, from the move shown onwards, at which it was losing. So it's possible that it was winning throughout, and didn't care which move it made (though it did report its own chance of winning as 32%).

Annual points

Players receive points for the 2014 Annual KGS Bot Championship as follows:


Details of processor numbers, power, etc.

Aya, MC version, running on one machine: 980X 3.3GHz 6 cores.
coldmilk, running on a 16-core Xeon, 2.9GHz.
HiraBot, running on i7-3770k@4.2GHz 4 cores 8 threads.
oakfoam, running on i7-4790K + i7-2600K + i7-920, 3*8 threads.
Orego, running on 2x Opteron 6282SE (31 threads) with 128GiB RAM.
pachi, running on 2x Opteron 6282SE (31 threads) with 128GiB RAM.