If you look at the rating table on the cgos web page you will see that 600 ELO difference corresponds to about 97% winning percentage. At the levels I tested against gnugo a single game per 100 could swing it 50 ELO. Since it did not lose a single game you could assume that it was either lucky, or that it's self-play rating was pretty accurate. It seems like I played 100-200 games or so.
I think the version of gnugo I used tests around 1800 on CGOS. The graph implies about 1000 ELO difference which is consistent with winning every game. However, I don't remember if I calibrated the graph or whether it's arbitrary. It seems like I had a version of gnugo as an anchor, but I don't see it in the graph. I could have simply extrapolated from CGOS for one of the version. http://cgos.boardspace.net/index.html - Don Don Dailey wrote: >> Don, the data was derived from self-play, wasn't it? >> > > Yes, it was derived from self play. > > I also did a study at one time where I tried these doublings against a > stable gnugo version and got very similar results - the program went > from being crushed by gnugo (rarely winning a game) to the point where > gnugo could not win a single game. The ratings differences were very > similar at the 3 or 4 levels near gnugo, but at the extreme ranges they > were inflated (you can't accurately rate a program that always wins.) > > To get the most accurate ratings you want to play a good percentage of > opponents that are near your own strength. > > - Don > > > >> Chrisotph >> >> _______________________________________________ >> computer-go mailing list >> computer-go@computer-go.org >> http://www.computer-go.org/mailman/listinfo/computer-go/ >> >> > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ > > _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
