bug-gnubg
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Bug-gnubg] FIBS rating and error rates


From: kvandoel
Subject: Re: [Bug-gnubg] FIBS rating and error rates
Date: Thu, 21 Aug 2003 21:08:06 +0200 (CEST)

On Thu, 21 Aug 2003, Joern Thyssen wrote:

> > One issue in the rating formula is that I weigh moveErr and cubeErr by
> >
> > errTot = moveErr/X + C*cubeErr/X, with
> > X=(unforcedMoves+unforcedCubedecisions), and C=1, and assume R depends
> > on errTot only.

> I can't explain why, but I would rather see

> errTot = ( (1-C) * moveerr + C * cubeerr)/X

That's the same as the factor (1-C) can be absorbed in a(N) and you
can rename C/(1-C) to C`.

> > Some exploratory simulations indicate C<1. i.e., cube errors are not
> > as important as is assumed.For example an errTot = .015 predicts
> > R=300 but if entirely due to cube errors R = 116.
> >
> > I would like to refine the R estimate by doing some simulations where
> > I vary move and cube errors independently.
> >
> > Please point me to any relevant previous work on this subject....

> I can't recall having seen anything.
>
> Note that Snowie uses another errTot:
>
> errTot = (moveErr + cubeErr) /
>          (total moves player 1 + total moves player2)

Intuitively it seems to be  that "playing strength" should be related to
your mistakes,  your errors per decision,  so I think  the gnubg formula
makes a lot more sense.

So perhaps if I want to improve on the current rating formula I should try
a fit of the form:

R = a(N)*moveErr/unforcedMoves +b(N)*cubeErr/unforcedCubedecisions;

instead.

Kees





reply via email to

[Prev in Thread] Current Thread [Next in Thread]