I can't help going back to somethings said in the thread
"Some thought on the fascinating and elusive luck".
https://groups.google.com/g/rec.games.backgammon/c/eIKGjPQXISU/m/rzpNtdqUAAAJ
On August 28, 2022 at 6:41:02 AM UTC-6, Axel Reichert wrote:
I found some old "mutant" session results with expert
checker play mimicking GNU Backgammon and random
cube play (double, hold, beaver, take, pass). After 1000
games the mutant was trailing with 17394 points versus
gnubg's 114822.
As I had indicated before, that is 15% win rate for the
"mutant" with "zero cube skill".
I wanted to know the "cube error rate" for the mutant
which would be easy to figure out, if Axel shared his
game files, by just letting XG or Gnubg analyse them.
Since he will neither share the game files or help with
calculating it, I'm trying to do it myself by deriving it
from the "overall/average error rates" by XG or Gnubg.
With combined checker and cube error rate, 15%-85%
win rates require between 415 to 575 ELO difference
between the players. For example, to win 15% against
a bot of "supernatural" level, a player would need to be
between "advanced" and "intermediate" level. See:
http://www.columbusbg.org/Ratings/
http://www.bgonline.org/forums/webbbs_config.pl?read=53409
https://www.gnu.org/software/gnubg/manual/html_node/Overall-rating.html
In Axel's experiment against "expert" (checker) level,
the "random player" would have to be of about 1,500
ELO, i.e. "casual player" which is higher than "beginner"
and "distracted/awful".
If the same holds true also for cube error alone, I think
it would be quite telling about the so-called "cube skill".
I'm assuming that if I can eliminate the checker skill
from the equation, I should be able deduct the cube
skill from the bot's overall/average error rate number.
I looked but could find any documentation on how the
bots use the checker and cube error rates to come up
with an overall/average error rate. Does anyone know
how XG or Gnubg figures it out?
XG also calculates "expected win percent" based on
the overall error rate. I couldn't find anything about how
it calculates that either. I would appreciate any direct
answers or links to relevant documentations.
As you may have guessed, what I'm after is to show
that if there was as much cube skill as hyped, random
cube player could not achieve a 15% win rate...
MK
--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)