Third 2011 Slow KGS Computer Go Tournament

August 21st-26th 2011

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


format15-round Swiss
board size19×19
timeFour hours each, sudden death


The first round started at 22:00 UTC on the evening of Sunday 21st.

Result table

Zen pachi AyaMC gomor ManyF Stone Orego
B13R W17R B112R B14R W18R B114R W15R B110R W115R W11R B111R W16R B113R B12R 15114114Winner
2pachi2 W03R B07R W012R
W12R B111R W115R B11R W113R B14R W18R B19R W110R W15R B114R 1211974
3AyaMC W04R B08R W014R B02R W011R B015R
B13R W16R B07R B11R W112R W19R B113R 812334
4gomorra3 B05R W010R B015R W01R B013R W03R
B09R W112R B114R W12R B17R W16R B18R 811934
5ManyFaces1 B01R W011R W04R B08R B06R W17R W19R B012R W014R
B15R W115R B13R W110R 811234
6StoneGrid B06R W013R W09R B010R W01R B012R B02R W07R W05R B015R
B14R W111R 51108
7Orego12 W02R B05R W014R B09R W013R B06R W08R W03R B010R W04R B011R

In the table above,
   0 indicates a loss
   1 indicates a win
   a superscript indicates the round in which a game was played
   a subscript shows how the result was determined:
      R for resignation
      T for time
      F for forfeit
      a number for the points difference after counting.
All the 0s, 1s and Js are links to the game record.

The numbers in this table may not add up as you expect. This is because the "Wins" column incudes byes.


The results were clear, as shown by the cross-table above. Zen19S won all its games, while pachi2 won all its games except those against Zen19S.

When pachi2 is playing on KGS, it is possible to "chat" to it, by saying winrate. It responds by stating its current assessment of its probability of winning. However KGS does not allow ordinary users to chat to a player who in a tournament, only admins can do this. So, as a KGS admin, I sometimes asked pachi2 for its winrate, and reported the result in the game chat.

In round 5, pachi2 had a won position against Orego12 when I did this. It responded In 534069 playouts at 64 machines, white O11 can win with 100.51% probability. As an assessment of its chances in that game, this was accurate enough; but it shows that it does not calculate these percentages quite as you might expect. Its operator Jean-loup Gailly has explained "There can be several reasons for this. On a single machine, statistics are updated by multiple threads in parallel without locking because locking is too costly. In distributed mode, statistics received from other machines are also added without locking."

AyaMC vs ManyFaces1
Moves 105-114.

In its round 6 game with AyaMC, ManyFaces1 as Black attacked a weak white group, as shown to the right. AyaMC responded effectively, saving its group and killing the attacking stones. This produced a big gain, and AyaMC won the game.

StoneGrid vs pachi2
Moves 50-66.

In pachi2's round 9 game with StoneGrid, it played out part of a losing ladder, as shown to the left. As it did so, its "winrate" dropped steadily, from 57% after move 49, to 39% after move 65. However it was able to recover from this large loss, and win the game.

Round 9 was also Zen19S's first (and only) bye. Its operator Hideki Kato had been waiting for a bye, as an opportunity to integrate a new 4-core computer with Zen19S's existing hardware. But as the draw for round 9 was made at midnight in Japan, he may have preferred to sleep. I believe he did add the new computer some time before round 12.

In round 12, Zen19S did not show up for its game with pachi2. While we were waiting for Zen19S, I asked pachi2 "winrate", and it answered "In 0 playouts at 46 machines, none can win with 100.00% probability". The "0", the "none", and the "100.00%" are all reasonable in the circumstances of a game which has not yet started; but the "46" is worrying. Pachi2 was meant to be using 64 machines. Its operator Jean-loup Gailly later learned that some of them had stopped working properly.

Zen19S eventually joined this game, with only 32½ minutes left on its clock, whicl pachi2 still had four hours. Zen19S was still able to play well enough to win.

In the round 15 game between Zen19S and gomorra3, the kibitzers' view was that gomorra3 was ahead for some of the game.

The players receive points for the 2011 Annual KGS Bot Championship as follows:

Many Faces of Go2

Details of processor numbers, power, etc.

Aya, running on a 980X 3.3GHz 6 cores
Gomorra, running on a cluster using 8x6 cores. Cores are running at 2.67 Ghz.
Many Faces of Go, running on an 8 core XEON SMP.
Orego, running on one of the five nodes of a custom Linux cluster. The node has two AMD Six Core Dual Opteron 2427 2.2 GHz (12 cores total), 8 GB RAM, Centos Linux.
pachi, running on 64 unspecified 22-core platforms.
StoneGrid, running on an i7-2600.
Zen running on one dual Intel Xeon X5680/4 GHz, one Intel Xeon W3680/4 GHz, two Intel i7 920/3.2 GHz, and one Intel Q9550/3 GHz.