[Computer-go] Computer Go tournaments in European Go Congress
nick at maproom.co.uk
Sat Aug 7 03:56:55 PDT 2010
Last Wednesday, August 4th, there were two Computer Go Tournaments in
Tampere, Finland, as part of the 2010 European Go Congress. You can
read the short official report at http://www.egc2010.fi/results.php
I would like to congratulate the winners, Many Faces of Go (19x19) and
Valkyria (9x9); and to thank those who made it possible, Jukka Jylänki
(organiser of the events) and William Shubert (creator and maintainer of
KGS, also see below).
The results were
19 x 19
1st Many Faces of Go 6 wins 120 Euros prize
2nd Valkyria 4 wins 60 Euros prize
3rd Pachi 2 wins 40 Euros prize
4th Leela 0 wins
9 x 9
1st Valkyria 6 wins 81 Euros prize
2nd Many Faces of Go 4 wins
3rd Leela 1 win
4th Pachi 1 win
Magnus, David, and Petr should hear from the organisers of the EGC
sometime, about their prize money.
The rest of this report may be of little interest except to those who
were directly involved.
The rules of the events required the programs to be present in the room
where the events were "held", in a department of Tampere university,
several miles away from the city centre and from the main congress
venues. Entrants could either use the Windows machines provided, or
bring their own hardware. They could either attend in person, or
entrust the running of their programs to Jukka or to me. David Fotland
entrusted Many Faces of Go, and Gian-Carlo Pascutto entrusted Leela, to
me, Magnus Persson entrusted valkyria to Jukka, and Petr Baudiš ran
pachi himself, on a Unix laptop he brought.
The actual tournament was to be held on KGS, using its tournament
scheduling system but our own tie-break system (which was one extra
tie-break round, otherwise the toss of a coin).
Our plan was to get all the competitors running, and hold a small test
tournament, on the day before the tournaments. I had considerable
difficulty with Many Faces of Go (its installation changes the Windows
registry, the account I had been assigned on the machines provided did
not give the power to change the registry, Many Faces is protected by a
key related to a hash which it generates when you try to run it, and
this key depends not only on the physical hardware it is running on but
on other aspects of its environment). I failed to get it working before
David, in California, had to go to bed, so I was unable to test it on
Tuesday morning. I believe that someone more competent than me would
have found it easier, but not trivial.
I did get Leela running without difficulty (but see below), and Jukka
got valkyria running with no problems at all.
We then agreed to test the two bots we had running, using a very small
and fast test tournament on KGS. This went well at first, with Leela
failing to join the first-round game and valkyria therefore winning on
time. I realised that I had Leela logged in to KGS with the wrong
account, closed it, edited its config file, and relaunched it. I then
found that the KGS tournament scheduler was no longer running, and the
KGS tournaments page https://www.gokgs.com/tournList.jsp was displaying
an error message. As we planned to hold the tournaments on KGS using
its tournament scheduler, this was a serious problem.
That evening (in Finland), when David in California got up, he was able
to give me yet more unlock codes, until I got Many Faces running.
However the KGS tournament system was still broken when I left the
building for the night.
On Wednesday morning I was very pleased to find that the KGS tournament
system was working again. The KGS admin 'sweety' had informed William
Shubert about the problem, and he had fixed it. Petr arrived well
before the scheduled 10:00 start, and got his pachi working. We held a
second test tournament with very fast time limits, this time with all
four players present, and it all went well.
The tournaments themselves went perfectly smoothly, with the results
listed above. All four players seemed keen to get involved in
interesting fights, which made things interesting for us and for the
When I closed down the machines we had been using, removing Many Faces
of Go and Leela from them, I discovered a problem. The Many Faces
systems were fine, but the Leela systems were very slow. Using the
Windows task manager showed that they each still had several "Leela"
processes running on them, even after I had closed kgsGtp, and that
these processes were between them using all the system cycles, even
though they had no game to think about.
So Leela's performance may have been impaired in its games by having to
compete for processor cycles with its own zombie processes. I accept
very little of the blame for this. I would of course have killed any
zombies each time I relaunched it, if only I had known they were
Nick Wedd nick at maproom.co.uk
More information about the Computer-go