Re: [Bug-gnubg] FIBS rating

[Joern Thyssen]

On Sun, Jul 20, 2003 at 04:09:48AM -0300, Albert Silver wrote
> I've had time to familiarize myself with the new FIBS rating
> difference display and must say I don't like it. No offense, but
> having to remember both our ratings and have to do mental arithmetic
> to arrive at a rating performance is onerous. Why not put back the old
> rating system but just lower the GNU rating to 2050 as the reference.
> 2200 is just too much, and this seems more appropriate. Sure it could
> peak at 2100, but it can just as easily drop below this, and I think
> that is the proper rating for it. It was a feature I enjoyed.

I'm sorry, but there has been too much noise about the FIBS ratings. The
problem is that I chose some arbitrary linear interpolation based on the
normalised error rate per move, which lead to all sorts of strange
behaviour. Obviosuly my own fault, I should have done proper research
before I implemented it. I've removed the absolute fibs ratings until
someone (not me!) does a scientific investigation of how the fibs
ratings and relate abs. fibs ratings to any of the numbers output by
gnubg.

Following the recent discussion on rec.games.backgammon regarding luck
adjusted resulsts, you should also know that using the normalised error
rate per move (or any other error rate) as basis for the fibs rating
leads to biased ratings. The best we can do is to input the luck
adjusted result of the match into the FIBS rating formula. This gives
unbiased rating estimates. The problem is that it's only possible to
provide relative fibs ratings.

Jørn