Re: [Bug-gnubg] Error rate of predefined settings?

[boomslang]

Hi Timothy,

What it did was letting a bot play on FIBS against human players. I started with 0 noise for about 10000 matches, then increased the noise to 0.007 for another 10k matches, increased again etc. For each noise setting I could estimate its actual strength on FIBS based on the match outcomes and the accompanying opponents ratings, i.e. estimates were made using logistic regression, not by analysing the matches with GNU backgammon.

For 0 ply, zero noise ('expert') it estimated about 2035; for 0.015 noise it estimated 1825 etc.

What I meant with my remark is this: in my opinion, an intermediate player will hoover around 1650 on FIBS, so a better noise level for GNU's 'intermediate' will be 0.012.  Similar, a beginner will hoover around 1400 on FIBS, so 0.034 would be a better setting for 'beginner'. A 1900 player on FIBS is IMO an expert, so 0.010 noise is more suitable for the 'expert' preset, etc.

Bottom line is, GNU backgammon noise levels for the presets are too high, particularly for intermediate/casual player/beginner.

On Sunday, 2 March 2014, 20:17, Timothy Y. Chow <address@hidden> wrote:

On Sun, 2 Mar 2014, boomslang wrote:
> I cannot give you Snowie error rates for the several settings, but I can
> give you an estimated FIBS rating for various noise settings.
[...]
> Looking at the table, I would say that the noise settings for
> intermediate, casual player and beginner are set too high.

Thanks, boomslang!  If I understand your comment, you're saying that if we
were to record a bunch of matches and analyze them using, say, world-class
settings, then the so-called "Intermediate" player would be assessed as
being worse than "Intermediate" (and similarly for "Casual Player" and
"Beginner")?  But "Expert" and "Advanced" would (usually) get assessed as
"Expert" and "Advanced" respectively?

Tim

_______________________________________________
Bug-gnubg mailing list
address@hidden
https://lists.gnu.org/mailman/listinfo/bug-gnubg