Thanks for the replies. After studying some of the losses I realized the problem: I was letting gnugo do the scoring and it was using Japanese rules! So libego filled in all its territory while gnugo kept passing.
So now I'm starting gnugo like in [1] and getting results much more in line with my expectations. >> Libego played at old CGOS with name sth like UCT-107-???k >> >> 100k was about 1550k >> 200k about 1650k > http://cgos.boardspace.net/public/cross-UCT.107.100k.html > http://cgos.boardspace.net/public/cross-UCT.107.10k.html > http://cgos.boardspace.net/public/cross-UCT.107.200k.html > http://cgos.boardspace.net/public/cross-UCT.107.20k.html > http://cgos.boardspace.net/public/cross-UCT.107.50k.html Thanks. Here are the results against gnu-go (which I believe is basically the same version that was the 2000-anchor in Don's scaling experiment [2]): 50K: ri_gnugo_3_7_10 7/37 18.9 100K: ri_gnugo_3_7_10 6/38 15.8 200K: ri_gnugo_3_7_10 7/31 22.6 I.e. libego should win the odd game, but lose most of them. It also suggests that Don's enhancements in the Lite version [3], are worth quite a bit of strength. Darren [1]: export BLACK="gnugo --mode gtp --level 1 --chinese-rules --capture-all-dead" export WHITE="/path/to/ego_opt -uctplayouts 100000" export TWOGTP="/usr/local/src/gogui-0.9.1/bin/twogtp" time $TWOGTP -black "$BLACK" -white "$WHITE" -games 20 -size 9 -alternate -auto -komi 6.5 -sgffile GE [2]: http://greencheeks.homelinux.org:8015/~drd/public/study.jpg [3]: Don wrote: The lite version I am testing has one enhacement that makes it not uniformly random. There is code to make it prefer captures in a probabilistic way. I think I just test N moves, and if one of them is a capture I play it, otherwise I choose a move randomly which doesn't fill an eye. But this enhancement isn't needed to get it up to the level of gnugo 3.6 ...I never allow suicide, I test only for simple ko in the play-outs. _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
