Re: [Bug-gnubg] Bug in analysis - position enclosed - more details

[Ned Cross]

Hi,

 

New comments embedded below (NC).

 

Subject: Re: [Bug-gnubg] Bug in analysis - position enclosed - more details

> Ned Cross wrote:
> >
> > Hi,
> >     I don't know the definition of "crashed" position.  I used the term
> > because "evaluate" prior to the roll gave the output below.
> >
>
> Right. I see that I forgot to check in a fix for this. I am posting it now.
>
> The contact net shows 85.8%   % 14.2% after 6/1 5/1

 

(NC)   Good to know that perhaps GNU will evaluate positions like this better in the future, since the "contact" net is probably better trained to handle them.  I will rerun the analysis as soon as a new build is available.

>
> > I'm wondering if there are two separate issues here:
> >
> > 1) That GNU doesn't understand this position.
> >
> > 2) That the reported 2-ply win %  for the candidate move ("after the roll")
> > does NOT equal the reported 2-ply win %  for the position resulting from the
> > candidate move ("before the next roll"). Right or wrong, shouldn't those two
> > numbers be the same?
> >
>
> Should be. Perhaps one is cubelss and the other cubefull? Are you sure the
> parameters are the same?

 

(NC) The output in my original post was generated from a GNU 2-ply analysis (cubeful checker play) with 2-ply cube. The analysis parameters were the same.  After digging deeper, I have found the cause of the discrepancy.  The value for 2-ply analysis AFTER the move is the same as the value for 1-ply analysis BEFORE the next move.  (Similarly, a 3-ply move analysis of the 6/1* 5/1 move predicts the same 82.2% wins as the 2-ply cube analysis before the next roll).  This does not seem correct to me.  Shouldn't 2-ply = 2-ply if there is no move in between?

>
> > Ned