Re: [Bug-gnubg] Wildly different evaluations by CRASHED network

[Joern Thyssen]

On Wed, Jul 03, 2002 at 12:35:46PM +0100, Ian Shaw wrote
>     GNU Backgammon  Position ID: AAAAt+3uAAAAAA
>                     Match ID   : cAkAAAAAAAAA
>     +24-23-22-21-20-19------18-17-16-15-14-13-+  O: GnuBg
> OOO |                  | O |                  |  0 points
> OOO |                  | O |                  |  
>  OO |                  | O |                  |  
>  OO |                  |   |                  |  
>  OO |                  |   |                  |  
>     |                  |BAR|                  |v (Cube: 1)
>     |                  |   |                  |  
>     |                  |   |                  |  
>     |          X  X  X |   |                  |  
>     | X  X  X  X  X  X |   |                  |  On roll
>     | X  X  X  X  X  X |   |                  |  0 points
>     +-1--2--3--4--5--6-------7--8--9-10-11-12-+  X: ian
> 
> static:       0.559   0.000   0.000   0.000   0.000   (+0.119  (+0.168))
>  1 ply:       0.932   0.000   0.000   0.000   0.000   (+0.863  (+1.000))
>  2 ply:       0.629   0.000   0.000   0.000   0.000   (+0.259  (+0.373))
>  3 ply:       0.931   0.000   0.000   0.000   0.000   (+0.863  (+1.000))
>  4 ply:       0.752   0.000   0.000   0.000   0.000   (+0.504  (+0.839))
> 
> 
> Double, take        : +0.839
> Double, pass        : +1.000   (+0.161)
> No double           : +0.803   (-0.036)
> 
> Correct cube action: Double, take
> 
>                               Win    W(g)   W(bg)  L(g)   L(bg) Equity
> Cubeful
>             Current position 0.9480 0.0000 0.0000 0.0000 0.0000 +0.8960
> n/a
>               Standard error 0.0137 0.0000 0.0000 0.0000 0.0000  0.0251
> n/a
> 2-ply 108 +vr cubeless
> 
> 
> Note the huge discrepancies between the plies, depending on which side of
> the board GnuBg is looking from. This is a humungous pass but GnuBg beavers
> even at 2-ply.

Yes, this is typical behaviour in positions where the neural net is not
good; you see some huge differences between even and odd ply. I think
before Josephs crashed network gnubg's evaluations would be even worse.

However, judging from the series 0.559, 0.629, 0.752,... you would
only have to do an 8-ply analysis before gnubg would evaluate correctly
:-)

Jørn

-- 
Joern Thyssen, PhD
Vendsysselgade 3, 3., DK-9000 Aalborg, Denmark
+45 9813 2791 (private) / +45 2077 2689 (mobile) / +45 9633 7036 (work)