Dear Don, Don Dailey: <[email protected]>: >I like the idea of just using the bayeselo ratings - it's always going to be >the most accurate unless a program is changing.
Discounting by time may help, like KGS. The period could be a few months for 50%, just my intension. Seeking the list of ratings, most programs have version number in their name. Strong exceptions are Erica and Enos but Aja told me he will stop improving Erica soon (or stopped already) and I believe Enos will use version number as a part of its name (no evidence :-). >But the problem with periodical copying is that your program can suddenly >change its rating and it may even move around when you are not even playing >games. This should never be a problem if the new ratings are more accurate. >I think there is a way to solve this. The bayeselo ratings are updated >every 6 hours. So what I could do is just continue to use the incremental >rating system but after each program is incrementally rated I could push the >ratings a bit in the direction of the bayeselo system - in this way they >will always be in close agreement as long as the programs are active. Is this better than mine? This will just makes all things slower, won't it? If the implementaion is easier and can be done quickly, it might have a value to do, though. Best, Hideki >On Tue, Oct 12, 2010 at 8:36 PM, Hideki Kato <[email protected]> wrote: > >> David Fotland: <028001cb6a5a$3522dff0$9f689f...@com>: >> >Don, >> > >> >Can the CGOS 19x19 starting rank be changed from 1200 to 1800? Almost all >> >programs are now stronger than gnugo on 19x19. With the 1200 start it >> takes >> >a very large number of games to climb over 2200, and by then the K value >> is >> >so small that the top programs take a huge number of games to accurate >> >ratings. The bayeselo and regular ratings are very different, I think >> >mainly due to this problem. >> >> This will only reduce the number of earlier games. There are two >> problems in current rating algorithm of CGOS, I think, and I'd like to >> propose periodical copy of the bayeselo rating values to current >> ones. >> >> The first is a big rating gap between top level programs and the >> anchor, ie, 2700 or 2800 (according to bayeselo but 2500 in daily >> rating) and 1800 of GNU Go. Assuming there are only two programs, Zen >> and GNU Go, for example, Zen needs about ten wins (I have no exact >> number) to get one Elo point so that Zen will take several months to >> reach 2800 Elo. >> >> The other is a strong newcomer program. If a strong program loses a >> game against the newcomer, its rating significantly drops and never be >> recovered, AFAIK. KGS has a recovery mechanism for this. (If CGOS >> also has this, I'm sorry.) >> >> Those two problems are not solved (perhaps, relaxed a little) by just >> changing the initial rating from 1200 to 1800. >> >> Current rating system of CGOS has been working fine for many years on >> (human) Go clubs, right? The strength of a (human) player changes >> much more slowly than a program and human players stay much longer >> than programs because programs frequently get version-up. >> >> Hideki >> -- >> Hideki Kato <mailto:[email protected]> >> _______________________________________________ >> Computer-go mailing list >> [email protected] >> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >> >---- inline file >_______________________________________________ >Computer-go mailing list >[email protected] >http://dvandva.org/cgi-bin/mailman/listinfo/computer-go -- Hideki Kato <mailto:[email protected]> _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
