[Bug-gnubg] Absolute FIBS rating formula

[Philippe Michel]

On Sat, 26 Jul 2014, Michael Petch wrote:

> I will be building the next release of GNUbg (1.03.000) next Sunday
> evening. So any changes that one may have should be put in as soon as
> possible so testing can occur.

I have redone the simulations from Kees van den Doel that led to the 
"Error based abs. FIBS rating" in match statistics :
http://people.cs.ubc.ca/~kvdoel/tmp/ratings/

This number became quite inaccurate years ago when the definition of a 
"close cube decision" changed. The cube errors rating loss is 
underestimated by a lot. On the other hand the offset used to get the 
absolute rating is 2050 when the rating of the bots based on gnubg 2ply 
is more like 2150.

The old formula is :

R = a2(N)*EPM+b(N)*EPC,
where
a2(N) = 8798 + 25526/N,
and
b(N) = 863 - 519/N.

and what I got is :

a2(N) = 11877 + 32191/N,
and
b(N) = 4766 - 5285/N.

As you can see, the grading is tougher and the higher base rating of gnubg 
may or may not compensate for it depending on the level of play.


I'd be interested if someone who has a fair database of FIBS matches (I 
don't) could check if these new values are realistic. I'm not familiar 
with the Python API, but I suppose it should be possible to write a short 
script that loads an already analyzed match and shows the FIBS rating from 
both formulas.

I'd be interested by more accurate data from people running bots on FIBS, 
preferably if they use the 1.0 weights file : what is their average rating 
for various levels of play ? And are they cheated by droppers in a 
significant way ?