Re: Re : Re : Re : [Bug-gnubg] Odd/even effect: post I made on bgonline

[Joseph Heled]

Yes, O being closed out has no move decisions to make, and is far from double or re-doubling if he takes - so such positions were not part of the training.

On Wed, Aug 26, 2009 at 7:45 PM, Massimiliano Maini <address@hidden> wrote:

>
>De : Joseph Heled <address@hidden>
>

>That is why getting a strong 0-ply is of the utmost importance.
>I am not sure what is the state of other bots.

We do agree on this, getting a strong 0ply NN is a must.
Actually that's why I was inquiring to know if the training
methodology of gnubg NN could be responsible of this kind of
behavior.

This whole odd/even effect thing seems to be much more visible
in gnubg than in other bots and it's the source of (possibly wrong)
beliefs like "gnubg 3ply cube is worse/better than 2ply cube"
or "2ply cube is worse/better than 3ply cube for closed out
positions" etc.

In the particular position (which is an extreme case):

   GNU Backgammon  Position ID: DwAAsN22AYgAAA
                   Match ID   : MAHgAAAAAAAA
   +-1--2--3--4--5--6-------7--8--9-10-11-12-+  O: gnubg
   | O  O  O  O  O  O |   |                  |  0 points
   | O  O  O  O  O  O |   |                  |  On roll
   |    O             |   |                  |
   |                  |   |                  |
   |                  |   |                  |
   |                  |BAR|                  |^ 7 point match (Cube: 1)
 X |                  |   |                  |
 XX | X                |   |                  |
 XX | X                |   |                  |
 XX | X                | X |                  |
 XX | X              O | X |       O          |  0 points
   +24-23-22-21-20-19------18-17-16-15-14-13-+  X: MaX

O on roll:
Position ID: DwAAsN22AYgAAA
Match ID: MAHgAAAAAAAA
Evaluator:  Crashed
       Win     W(g)    W(bg)   L(g)    L(bg)   Equity    Cubeful
static:  76.8     0.0     0.0     0.1     0.0    +0.534    +0.785
 1 ply:  88.0     0.0     0.0     0.1     0.0    +0.759    +0.923
 2 ply:  79.7     0.0     0.0     0.1     0.0    +0.593    +0.999
 3 ply:  87.7     0.0     0.0     0.0     0.0    +0.754    +1.000
 4 ply:  82.0     0.0     0.0     0.0     0.0    +0.639    +1.000

BGBlitz has:

O on move:   (Plies are starting with 1, so it's gnubg 0/1/2-ply)

1-ply  Green        83.0  0.0   0.0

2-ply  Green        81.9  0.0   0.0

3-ply  Green        84.4  0.0   0.0


So BGBlitz too has some swinging between odd and even plies, but
the amplitude is 2-4%, not 12%. Snowie seems to behave like BGBlitz,
just 2-3% swing.

MaX.