[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Bug-gnubg] rating estimation bug?
From: |
Joern Thyssen |
Subject: |
Re: [Bug-gnubg] rating estimation bug? |
Date: |
Wed, 26 Jun 2002 07:46:57 +0000 |
User-agent: |
Mutt/1.2.5.1i |
On Tue, Jun 25, 2002 at 07:50:35PM +0200, address@hidden wrote
> Hi Joern,
>
> First of all thanks for the fast implementation.
>
> > > The ratings do not seem to correspond with the classifications
> > > very well. Bug in formula implementation?
>
> > The classification of the players into ET, world class, expert etc. is
> > based on the error rate per move, whereas the rating is based on the
> > total error rate.
>
> I don't understand how that makes sense. Two disagreeing estimates of
> ones playing strength??
Yes!
Suppose the classification is world-class. This indicates that if you
played a world-class player you would be about 50-50 to win the match.
The rating diff says that if you played that particular game with the
same error rate against a perfect opponent your rating difference would
be xxx. Of course, against a world-class player your rating diff would
be zero.
Still, if as a world-class player you played a 10000 move game against a
perfect opponent, you would still lose a lot of rating.
> I expected the estimated rating to be just a non-linear function of the
> error rate per move, thus providing a numerical quantification of the
> existing classification.
No, it's a non-linear function of error rate pr move, match score, and
number of moves in the game.
> If you know the average equity loss per move, you can estimate the game
> winning chance against a perfect opponent and compute a rating based on
> this. I think this is the "rating" estimate we want.
For this to be true you would have to decide on a match score and the
number of moves in the game. Then the only parameter left is the error
rate pr. move.
Jørn
--
Joern Thyssen, PhD
Vendsysselgade 3, 3., 9000 Aalborg
+45 9813 2791 (private) / +45 2077 2689 (mobile) / +45 9633 7010 (work)