Re: [Bug-gnubg] Errors in evaluation of races with backgammons

[Joseph Heled]

Hi,

This position just falls under the special code I wrote ages ago.
There might be errors (bugs) there, but please remember that the
"exact" is *not*   a '8 vs 2 1 0 0 3 cubeless bearoff', since clearing
checkers from the back is not the same as bearoff - you *can* move any
checker you want. I think that was the point behind the special code
....

-Joseph

On 27 February 2012 12:23, Philippe Michel <address@hidden> wrote:
> It looks like there are some bugs in the evaluation of backgammon chances in
> races, in raceBGprob() in eval.c . For instance, this position :
>
>    +24-23-22-21-20-19------18-17-16-15-14-13-+  O: gnubg (Cube: 2)
>    | O  O     O  O  O |   |    O             |  0 points
>    |                O |   |                  |
>    |                O |   |                  |
>    |                  |   |                  |
>    |                  |   |                  |
>    |                  |BAR|                  |v
>  X | 8                |   |                  |
>  X | X                |   |                  |
>  X | X  O             |   |                  |
>  XX | X  O           O |   |                  |  On roll
>  XX | X  O        O  O |   | O                |  0 points
>    +-1--2--3--4--5--6-------7--8--9-10-11-12-+  X: pm
>
> evaluates as :
>
> Position ID:    pU4AWhz/AAAAAA Match ID:       QQkAAAAAAAAA
>
> Evaluator:      Race
>
>        Win     W(g)    W(bg)   L(g)    L(bg)   Equity    Cubeful static:
> 1.000   1.000   0.678   0.000   0.000    +2.678    +2.678
>
> The backgammon ratio is wrong by about 0.1.
>
> It is supposed to be computed exactly (not by a net) as the winning chance
> of a 8 vs 2 1 0 0 3 cubeless bearoff, either by not-particularly-obvious
> internal functions or with the bearoff databases.
>
> In the example above, it is the internal functions that are used and there
> may be a bug in them or, perhaps more likely, they are not used properly.
>
> If I twist the code in eval.c to only use the databases, I get :
>
>        Win     W(g)    W(bg)   L(g)    L(bg)   Equity    Cubeful static:
> 1.000   1.000   0.784   0.000   0.000    +2.784    +2.784
>  1 ply: 1.000   1.000   0.784   0.000   0.000    +2.784    +2.784
>  2 ply: 1.000   1.000   0.744   0.000   0.000    +2.744    +2.744
>  3 ply: 1.000   1.000   0.744   0.000   0.000    +2.744    +2.744
>  4 ply: 1.000   1.000   0.709   0.000   0.000    +2.709    +2.709
>
> Now the bg rate is accurate for 0 and 1 ply, but not for deeper searches.
> The diminished backgammons surely don't come from misplays by X so there
> must be another problem here, either in shorter races or in the collating
> process to get the result from higher plies.
>
> _______________________________________________
> Bug-gnubg mailing list
> address@hidden
> https://lists.gnu.org/mailman/listinfo/bug-gnubg