[Computer-go] 40-core Pachi on KGS

[Petr Baudis]

Hi!

I have left the distributed 40-core pachi engine running overnight
after the last night KGS tournament and reserved it for few high-dan
players in the evening. Pachi was third in the 9x9 tournament, this
might give some perspective on the level of the games played among
the top programs:

rank wins-losses
8d 4-0
7d 1-2
6d 5-3
3d 7-1
2d 4-3
AGA 3d 9-2 odnihs[-] declares himself as AGA 3d
vsbots 0-1 guest account
(no losses against other players)
http://www.gokgs.com/gameArchives.jsp?user=pachi

* During the night, I attempted to fix misbehaving resign code to make
the program more human-friendly; in the process, I introduced another
bug that would make pachi resign in some won scoring positions - so in
the game list, many losses were actually won games where Done instead
of Resign would produce a positive score.

** In the Computer Go tournament, time was 4:00+25/1:00. For the first
two games against famine 8d, time was 9:00+25/1:00, then it was changed
to 9:00+3x0:20. After the game with miracle 6d, it was changed to
6:00+3x0:10 and the challenge was opened to general public.

The most spectacular game was perhaps the ko fight in:

http://files.gokgs.com/games/2010/9/5/michi2009-pachi.sgf

The games show that computers can play at fairly high level even on
fairly small clusters, successfully beating even high dans in very
advanced fights, however especially in the lower-dan games, it confirms
that systematic errors still haunt them and they simply always get
various specific situations wrong.

(Also, opening book is essential for current-level 9x9 programs. Pachi
does not have one and I think the performance suffers a lot because of
that. I think at least 6 of the 12 losses could be prevented - or at
least a less obviously bad variation chosen - by even a minimalistic
opening book. Three of the losses resulted from exactly the same
resign-in-30-moves sequence played by Pachi repeatedly.)

--
Petr "Pasky" Baudis
The true meaning of life is to plant a tree under whose shade
you will never sit.
_______________________________________________
Computer-go mailing list
Compu...@dvandva.org
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

[David Fotland]

I also left ManyFaces1 on a 16-core cluster running overnight, and I was
surprised at how well the human Dan players did. Typically they would lose
a few games, then learn how to beat the program, and go on to beat it
several games in a row. This indicates that the programs are not as strong
as has been claimed, based on single games against top dan or pro players.

For example gojin [3d] lost his first three games, then won two in a row.
Haruhiko [3d] won one, lost one, then won three in a row taking black.

ManyFaces also lost a few games to 1 kyu and 2 kyu players.

David

[David Fotland]

I should add the time limits were 9x0:15 (9 periods of 15 seconds, so 15
seconds per move with an additional two minutes). This is somewhat more
time than we used in the computer tournament.

[Olivier Teytaud]

MoGoTW has somehow a long experience in 9x9 games against pros,
and the most recent records are (if I don't forget something, I'm not sure...):
- one loss as black against Catalin Taranu (5P)
- one win and two losses against Fan Hui (2P)
(the human refused to play the last of the 4th scheduled games,
so we have had this
performance by playing white only once)
- one win and one loss against M. Chou (9P)
- two losses against M. Chou (4P, the same name as above but not the same pro)
- one win and one loss against a Korean 8p in Cadiz

i.e. 3 wins out of 9 games and it is better if restricted to long time
settings if I remember well
(the game against Catalin, and the games against both Chou players
were shorter than others).
Interestingly we have in fact more wins as black, in spite of the very
big winning rate of MoGoTW
as white against MoGoTW as black...
But the results were far from that good when we had no opening book at all.
Olivier

[David Fotland]

Oops...

Usually I have Many Faces play on-line using Japanese rules, since most
people prefer them. I forgot to switch last night, so the engine was
scoring Chinese, but the server was scoring Japanese, and that accounts for
many of the losses.

Also, many of the games were played with 0.5 komi rather than 6.5, so they
should be counted differently.

So please ignore my previous email.

Petr, did you check your games to make sure they are at 7.5 komi rather than
0.5?

David

> -----Original Message-----
> From: computer-...@dvandva.org [mailto:computer-go-

> bou...@dvandva.org] On Behalf Of David Fotland
> Sent: Monday, September 06, 2010 11:10 AM
> To: compu...@dvandva.org

[Petr Baudis]

On Mon, Sep 06, 2010 at 11:30:27AM -0700, David Fotland wrote:
> Petr, did you check your games to make sure they are at 7.5 komi rather than
> 0.5?

Yes, all of them. I usually make sure I turn off the rank when I switch
Pachi to non-19 boards since the challenges would tend to produce
non-sensical default handicaps.

Petr "Pasky" Baudis

[Robert Finking]

Thanks for the info. I almost put my info on the Sensei's page, but in
the end, thought it was too informal so started a separate wiki.

Thanks also to whoever it was who corrected my 9x9/19x19 copy and paste
error!

Regards

Raffles