RE: [Bug-gnubg] Evaluation inconsistency

[Christopher D. Yep]

At 02:25 PM 9/16/2003 -0300, Albert Silver wrote:

> Naturally it is no problem to show the position, but as I said, it is
> completely unimportant whether my decision was wrong or right.
>
> GNU Backgammon  Position ID: 7gCAIOY2JgFCAA
>                  Match ID   : UQmgAkAAEAAA
>  +13-14-15-16-17-18------19-20-21-22-23-24-+     O: MartinKrainer
>  |                  |   |    X     O  O  X | OO  4 points
>  |                  |   |          O  O    | OO
>  |                  |   |          O  O    | O
>  |                  |   |                  | O
>  |                  |   |                  | O
> v|                  |BAR|                  |     21 point match
>  |                  |   |                  |
>  |                  |   |                  |
>  |                  |   |          X       |
>  |             X    |   |    X  X  X     X |     On roll
>  | X     X     X  O |   |    X  X  X  O  X |     2 points
>  +12-11-10--9--8--7-------6--5--4--3--2--1-+     X: SubtleOne (Cube: 2)
>
> What is important is the difference in the size of errors from (-0.007)
> to (-0.115) when *everything* in the percentages is improved:
>
>
> Cube analysis
> 3-ply cubeless equity  +0.345 (Money:  +0.343)
>    72.5%   0.0%   0.0% -  27.5%   9.7%   0.9%
>    ^^^^^                          ^^^^
> 3. Double, take         +0.671  ( -0.007)
>                                 ^^^^^^^^^
>
>
> Rollout cubeless equity  +0.369 (Money:  +0.362)
> 3. Double, take         +0.747  ( -0.115)
>                                 ^^^^^^^^^
> Player SubtleOne owns 2-cube:
>    73.0%   0.0%   0.0% -  27.0%   8.9%   0.9% CL  +0.369 CF  +0.862
>    ^^^^^                          ^^^^
> Player MartinKrainer owns 4-cube:
>    74.0%   0.0%   0.0% -  26.0%   8.8%   0.9% CL  +0.847 CF  +0.747
>    ^^^^^                          ^^^^

It probably just means that rollouts show a significantly different value 
to cube ownership than gnubg can initially see with a limited 3-ply 
horizon.  This probably happens frequently, especially in non-typical 
positions.  In the above position the cubeless equity stayed about the same 
(increasing from .345 to .369).  The ND equity changed from .679 to .862 
while the DT equity changed from .671 to .747.  The .679 value increases 
(with a rollout) because X will have many highly efficient chances to 
double later if he doesn't double this turn.  This is often the case in 
containment positions when X has hit a late shot (he can slowly close in on 
the drop/take point as he slowly extends his prime/board).  That the .671 
value would also increase (with a rollout) is not as crystal-clear to 
me.  Apparently O has less efficient redoubling opportunities than 
usual.  For example if he escapes one checker and suddenly hits a few shots 
in the outfield with his remaining back man he can quickly go from not good 
enough to too good in one roll.  Gnubg can't see all this in a 3-ply 
evaluation, but can "see" it in a rollout.

So, it's not surprising to me that both .679 and .671 increase (to .862 and 
.747 respectively) while the cubeless equity stays approximately the 
same.  The reason why this seems surprising is because one value increased 
about .18 while the other value increased by only about .08

Chris