Re: [Bug-gnubg] Strange bearoff analysis

[Jim Segrave]

On Thu 12 Jun 2003 (19:12 +0000), Joern Thyssen wrote:
> On Thu, Jun 12, 2003 at 08:21:40PM +0200, Jim Segrave wrote
> > 
> > Using gnubg (development tree) from June 4th, still with 0.13b weights.
> > I was analysing a match I played against it with 2 ply more or less
> > Supremo settings when this came up:
> > 
> > (jes) set matchid UQnrAFAAKAAA
> > (jes) set board OwAAwCYAAAAAAA
> >     GNU Backgammon  Position ID: OwAAwCYAAAAAAA
> >                     Match ID   : UQnrAFAAKAAA
> >     +24-23-22-21-20-19------18-17-16-15-14-13-+  O: gnubg
> >  OO | O  O             |   |                  |  5 points
> >  OO | O  O             |   |                  |  
> >  OO |    O             |   |                  |  
> >  OO |                  |   |                  |  
> >  OO |                  |   |                  |  
> >     |                  |BAR|                  |v 7 point match
> >  XX |                  |   |                  |  
> >  XX |                  |   |                  |  
> >  XX |                  |   |                  |  
> >  XX | X  X             |   |                  |  Rolled 62
> >  XX | X  X     X       |   |                  |  5 points
> >     +-1--2--3--4--5--6-------7--8--9-10-11-12-+  X: jes (Cube: 2)
> > 
> > 
> > 
> > Analysis and hints both say (this is the hint output, but the analysis
> > is identical in it's values)
> > 
> >     1. Cubeful 0-ply    4/off                        Eq.:  +0.906           
> >    
> >        0.953 0.000 0.000 - 0.047 0.000 0.000
> >         0-ply cubeful [expert]
> >     2. Cubeful 0-ply    4/off 2/off                  Eq.:  -0.925 ( -1.831)
> >        0.037 0.000 0.000 - 0.963 0.000 0.000
> >         0-ply cubeful [expert]
> > 
> > Now I simply cannot see how taking two men off is a -1.831 blunder. In
> > fact, there's one case where taking 2 men off matters - gnubg rolls
> > doubles and I follow with double 1's. If I have two pieces on both the 1
> > and 2 points, I will lose, if I have only 3 pieces, I will still win.
> 
> I can't reproduce your results. I get:
> 
> (jth) set evaluation chequerplay evaluation plies 0
> `eval' and `hint' chequerplay will use 0 ply evaluation.
> (jth) hint
>     1. Cubeful 0-ply    4/off 2/off                  Eq.:  +0.516
>        0.758 0.000 0.000 - 0.242 0.000 0.000
>         0-ply cubeful [expert]
>     2. Cubeful 0-ply    4/off                        Eq.:  +0.500 ( -0.016)
>        0.750 0.000 0.000 - 0.250 0.000 0.000
>         0-ply cubeful [expert]
> 
> In both cases O has roughly 1/6*5/6 + 5/6*5/6*1/6 = 25% (approx), so
> this analysis looks right.
> 
> Can you try setting up the two resulting positions with gnubg on roll,
> and post the output from "eval", please?

Hmm. On my home machine, I get the above sensible results. The odd
ones come from a copy installed at work. I'm wondering if my bearoff
database or weights file is corrupt in some way. I had a copy built in
March which was installed in /usr/local/bin, where I first saw the
problem. I also had a recent copy which gave the same strange
results. The only thing they'd have in common are the databases in
/usr/local/share. I'll try to look at that tomorrow after work, but I
begin to think there's something very odd there.


-- 
Jim Segrave           address@hidden