I'm contacting you about something being discussed over on the Gnubg mailing
list. It just so happens the discussion is regarding a value we are looking
at in the whole Monty vs Mpetch series (Just coincidental).
What happened was we had a fellow contact us about a question regarding how
the Luck adjusted Value, and the Fibs Luck adjusted Rating difference are
computed. I directed the individual to the formula at http://www.cs.ubc.ca/~kvdoel/tmp/ratings/ as well as gave the person the
fibs rating formula and how to compute the MWC values from the information.
With that being said I posted this information and received an interesting
response from one Christian Anthon (one of the primary Gnubg contributors)
with this rather interesting comment that casts *potential* doubt on Gnubg
releases since March 2006, and the Luck Adjusted Values (And Luck adjusted
Ratings Difference).
What he said was:
"This doesn't work any more. The cube error rate was broken and had to
be fixed, which invalidates the formulas. New ones could be made if
somebody care. I don't. "
I asked/confirmed with him if he believed the Luck adjusted Values may not
be as good estimates as they use to be. The response was:
" Yes that is the point. The IsCloseCube function was seriously damaged.
Amongst others it counted every too good to double as a close cube
decision, regardless of how close it actually was."
If you are interested in the exact nature of the changes - the file
differences (before and after the changes) can be found here if you are
curious about the cvs commit comment, and the actual changes made you can
see this mailing list post:
"the function checking for close cube decisions was totally wrong.
Amongst other things it included all too good positions and the
threshold of 0.25 was far too large to produce a sensible output for
exporting (a money beaver could be considered close to a double).
I have fixed the function and lowered the threshold to 0.16 (very bad).
Notice that this may have a large impact on the cube skill rate, but it
is the only thing that really makes sense."
Christian asked for comments at the time before committing, but those
working on GnuBG at the time (Including myself) didn't catch the potential
issue with the Luck adjusted Values.
It appears a couple of us are thinking that we'd like to recompute a2(N) and
b(N) against GnuBg and hen validate the data and fits. Your site has the
scripts you originally used and descriptions. Some of us were to redo your
simulations to produce the required bilinear fits on Cubes and Checker
errors.
Would you be interested in helping us out with this, or provide any caveats
or comments or issues you know of that those of us working on this might
find of benefit if we were to do it ourselves.
The GnuBG team would be appreciative for any assistance and comments you may
have on this issue.