bug-gnubg
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [Bug-gnubg] Re: Strange FIBS ratings


From: Albert Silver
Subject: RE: [Bug-gnubg] Re: Strange FIBS ratings
Date: Tue, 9 Sep 2003 23:39:43 -0300

Hi Kees,

> Never mind, I converted the files myself.

First of all, I'm astonished. I sent the files almost as soon as you
reiterated the request for the original .mat files, didn't you get them?
I can send them again of course, though it is evidently pointless now.
Still, this worries me as I've had other problems getting e-mails to
their detsination.

> Full results are attached for you. One file is with your analysis,
other
> with  a  fresh  0-ply  analysis   only.   Note  there  is  no
noticable
> diff. between 0ply and your 2ply estimates.

Yes, it is very close, so this is good to know.

> As you can see it is close to what you reported, so at least we agree:
> 
> 1862 +- 23 (actual 1700).
> 
> I have no definite explanation for why it is so far off your 1700
rating,

At least you understand the origin of my concern now.

> but possibilities are:
> 
> 1) The estimation is flawed; human play can't be modeled by noise.

Possible, and this is what Zare has suggested. I do wonder how off it
would be though even if this were the case.

> 2) You have made a recent quantum leap in playing strength

Possibly, but if I have, I must admit it's hard for me to judge. 

> 3) Luck (the variance of 23 is not reliable as estimated from the
actual
> variations in R over a small set of 27).

I sent 33 matches, unless I misunderstand the '27' mentioned. Ah ok. I
just opened the file and see that you averaged results where more than
one match was played against the same opponent.

> How to explain  it works so precisely for myself?   Maybe because I
only
> have been playing for a few  years and have learned mostly from bots
and
> therefore my "style" is bottish and better modeled by noise than
yours?

Also possible, but my number one teacher has really been GNU/Snowie.
Still, I have taken classes with Tom Suzanski at GG, and he has a
distinctly human approach discussing game plans according to score, etc.
I know this has changed the way I think, but I'm certain GNU is still
the biggest instructor overall. And the Gammonline articles and forum. 

> What  is needed  now is  more analysis  of games  by players  with
known
> ratings for validation/refutation. I am willing  to do the work if I
get
> the data to work on.

I'll play more and send you my matches as they come. In any case, thanks
for looking into this.

                                                Albert

> 
> Kees






reply via email to

[Prev in Thread] Current Thread [Next in Thread]