At the recent US Go Congress in Portland I had a few conversations
with AGA Ratings Statistician Paul Matthews. While the conversations
were centered upon the difficulties of rating Go playing computer
programs, it came out at one point that there had been a very long and
heated argument in the AGA about ratings of handicap v.s no-handicap
games. Today I learned from a past president of the New Mexico Go Club
that at one point only non-handicap games were AGA rated. What Paul
said is that the AGA attempted to resolve the dispute with data
(imagine that!). Two separate rating tables were kept, one for
handicap games and another for non-handicap games. Over time it turned
out that the ratings for individuals converged (the rating in the two
tables were the same), and at that time the AGA changed its policy and
then both handicap and non-handicap games became rated.
At this time games between humans and computers are explicitly not
counted in the AGA rating system. The AGA has decided to follow the
same data-driven process when the system is put in place for setting
ratings for computer programs. At first there will be a separate table
of ratings for human-computer and possibly computer-computer games
that will not count in the human ratings. After the data is collected
and analyzed the question of having games between humans and computers
being counted in the rating system will be reconsidered.
Cheers,
David
On 4, Sep 2008, at 5:06 PM, [EMAIL PROTECTED] wrote:
I'm not surprised that the data for games with 90% winning chances
is lacking.
The McMahon pairing system is widely used in western go tournaments
to prevent mismatched games,
because most players don't like mismatched games (either as the
stronger or the weaker player).
Rating systems are relatively new in the go world and they don't
seem very popular (or even known) in the far east where 99% of the
go population lives.
So I guess go rating systems are still quite immature:
AGA, KGS and EGF all have different conversion functions to map
winning probability to dan and kyu ranks.
These conversions are probably based on different statistical data
sources and probably also on other considerations like
calculation convenience and different theoretical considerations of
their inventors.
Dave
Van: [EMAIL PROTECTED] namens Don Dailey
Verzonden: do 4-9-2008 20:55
Aan: computer-go
Onderwerp: Re: [computer-go] Kaori-Crazystone
Here is something interesting from this page:
Note how different the expectations of each system are regarding even
games between players of unequal strength. If you can win 90% of even
games against a 2 kyu player, the AGA believes you are 1.33 ranks
higher, the EGF believes you are 2.42 ranks higher, and the IGS
believes
you are 2.80 ranks higher. The lack of agreement stems from a
tradition
of playing handicap games between players of different ranks, so there
is a lack of data regarding non-handicap games between mismatched
opponents.
- Don
On Thu, 2008-09-04 at 11:02 -0700, terry mcintyre wrote:
> This page http://en.wikipedia.org/wiki/Go_ranks_and_ratings gives
a table of win probabilities versus rank differences.
>
> I haven't yet found such a table for handicap games.
>
> Terry McIntyre <[EMAIL PROTECTED]>
>
>
> "Go is very hard. The more I learn about it, the less I know." -
Jie Li, 9 dan
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
