Re: [Bug-gnubg] Possible cube evaluation problem

[Philippe Michel]

On Mon, Mar 19, 2018 at 11:30:28AM +0100, Terje Pedersen wrote:
> Hi!
> 
> One of my users was upset that a position which XG evaluated as
> Double/Take was evaluated to a No Double/Take by Gnu. The score 2A-5A
> may make the cube action a bit tricky.
> 
> XGID=-AAbB-EBBB-----a--acbbbb--:0:0:1:00:3:0:0:5:10
> 
> Analyzed in 4-ply
> Player Winning Chances:   74.79% (G:2.12% B:0.01%)
> Opponent Winning Chances: 25.21% (G:1.40% B:0.01%)
> 
> Cubeless Equities: No Double=+0.5233, Double=+1.6234
> 
> Cubeful Equities:
>        No double:     +0.8108 (-0.0040)
>        Double/Take:   +0.8149
>        Double/Pass:   +1.0000 (+0.1851)
> 
> Best Cube action: Double / Take
> 
> eXtreme Gammon Version: 2.19.211.pre-release, MET: Kazaross XG2
> 
> Gnu:
> 
> tu0iAAxlvm0AAA:cAmgAAAAGAAE
> 
> Cube analysis
> 2-ply cubeless equity  +0.506 (Money:  +0.487)
>   0.739 0.021 0.000 - 0.261 0.012 0.000
> Cubeful equities:
> 1. No double            +0.912
> 2. Double, pass         +1.000  ( +0.088)
> 3. Double, take         +0.742  ( -0.169)
> Proper cube action: No double, take (65.7%)

It looks like there is something wrong with the even-ply cubeful 
evaluations at the score:

1-ply; ND 0.77 DT 0.76
2-ply: ND 0.91 DT 0.74
3-ply: ND 0.81 DT 0.80
4-ply: ND 0.86 DT 0.75
rollout: ND 0.84 DT 0.86

The equities are a bit off and the cube action slightly wrong, but the 
results for 3-ply and even 1-ply are reasonable. The numbers at even 
plies lead to the same small error on actual play but are really 
inadequate for an analysis.

For money play ND vs. DT numbers seem consistent at all plies :
2-ply: ND 0.80 DT 0.79
3-ply: ND 0.84 DT 0.82
4-ply: ND 0.82 DT 0.81