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 > -----Original Message----- > From: [email protected] [mailto:computer-go- > [email protected]] On Behalf Of Petr Baudis > Sent: Monday, September 06, 2010 2:05 AM > To: [email protected] > Subject: [Computer-go] 40-core Pachi on KGS > > 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 > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
