[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [computer-go] 9x9 Ratings
On Mon, 21 Mar 2005 18:22:32 -0500, Chris Fant <chrisfant@xxxxxxxxxxxxxxxxx> wrote:
> I realized another source of error in the ranking of weak bots (such
> as mine). People try to not just win the game, but to control the
> entire board. This sometimes results in a loss even though the person
> could have easily defeated the bot if they would have played more
> conservatively.
Yes, this would definitely be an error source. I believe the only way
to get a somewhat reliable 9x9 rating is to incorporate it into a
server, making it less inviting for people to do these things. But as
stated before, there are a number of things that makes it somewhat
difficult.
Regarding the "offline" rating calculations, some of the more obvious
sources of error can be avoided by using for example GnuGo to estimate
scores and override the resignation results. BUT I've been checking
the results of many bots and I'm not so sure these things really make
that much difference. I estimate these games to account for somewhere
in the region of 5-15% of all the games for a bot when only rated
players ( including [n?] ) are accounted for. The best way to avoid
these games/players seems to be to write some rating information in
the game setup.
I've been messing around some more with my rating calculations and
this is the current "standings".. :)
Handle Est.Rating Games used
---------------------------------
gnugo3pt6 14k 200
viking5 15k 200
botnoid 19k 200
go81 20k 200
tlsbottest 21k 200
neuron 22k 200
fstoned 23k 200
simplebot 30k 200
legobot 16k 21
nio 28k 2
dumbbot 29k 4
Don't pay too much attention to the ratings for the last 3 bots on the
list, they have far too few games. Legobot may well be 16k, but I'd
like to see it play some more. :)
/Christian
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/