Re: FW: [Bug-gnubg] Update: Checkerplay vs cube decision errors

[Joachim Matussek]

address@hidden schrieb am 19.05.04 20:14:23:
> 
> Hi,
> 
> Apart from noticing the misspelling of my name I have more comments:
> 

Sorry if I misspelled your name. But I read my item several times. Did not find 
it. 

> Since chequer errors  and cube errors really have  different "units" the
> ratio of the coefficients a2 and b is meaningless.
> 
> A meaningful measure is:
> 
> M = a2(N)*EPM / b(N)*EPC
> 
> M represents ratio  of rating points lost due to  chequer error adn cube
> error. Clearly it is player dependent. For me it is usually 10 in actual
> games I play.

This was not what I was talking about.

> Now in  the article this is  rewritten in the  form of Eq. 5.  The table
> below Eq. 5 plots something undefined but I guess it is
> 
> (a2(N)/UFM(N))/(b(N)/UFC)
> 
> with UFM  # unforced  moves and  UFC # unforced  cubes as  obtained from
> rollout.

Sorry about that. I did not explain these expressions well.

Checker error coefficient = a2(N)/(No. of unforced moves)
Cube error coefficient = b(N)/(No. of close or actual cube decisions)

Checker error/Cube error ratio = (a2(N)/(No. of unforced moves)/(b(N)/(No. of 
close or actual cube decisions) 

I am not talking about unforced cube decisions because this would mean _all_ 
cube decision. I am talking of close or actual cube decisions as defined by 
GNUBG.

> This is the  ratio of the coefficients of a bilinear  fit to total error
> rates of chequer and  cube play. It is not clear to  me this is really a
> better measure than a2/b. We still have the fact that cube decisions are
> really of a different kind  than chequer play decisions even though both
> can be expressed in equity losses.
> I don't understand the ssecond paragraph of the "Results". If the cutoff
> for  actual cube  decision is  decreased  to .05  the #  of actual  cube
> decision goes  down and the coefficient  b(N) will also go  down (it was
> measured assuming a fixed cutoff) so nothing changes.

Should be self-explaing now. It is not b(N) which goes down. It is b(N)/(No. of 
close or actual cube decisions) which goes down.
 
> In  the 3d  paragraph I  read  "... doubtful  to clain  cube error  were
> less..".   I  don't see  any  arguments  that  support that  claim.
> 
> The  opposite statement  "cube errors  are less  important  than chequer
> errors" has no rigorous meaning  either.  It just seems in practice that
> the rating  points you lose  due to cube  errors are much less  than the
> loss due to chequer play.

Some players concluded from reading your article that cube error were less 
important. You didn´t claim that.

> 
> Kees
> --

I hope this discussion will stay friendly. I am just a backgammon player who 
wants to understand this game.

Joachim Matussek
_____________________________________________________________________
Endlich SMS mit Bildern versenden! Das Bild selbst ist dabei gratis,
Sie bezahlen lediglich den Versand. http://freemail.web.de/?mc=021195