Re: [Bug-gnubg] bug in 3 ply equities

[Joseph Heled]

Øystein Johansen wrote:
> Hugh Sconyers wrote:
>
>  GNU Backgammon  Position ID: AAAgYNs2BKAAAA
>                  Match ID   : cAklAAAAAAAA
>  +13-14-15-16-17-18------19-20-21-22-23-24-+     O: gnubg
>  | X                |   |             X  X | OOO 0 points
>  |                  |   |                  | OOO
>  |                  |   |                  | OOO
>  |                  |   |                  | OOO
>  |                  |   |                  | OO
> v|                  |BAR|                  |     1 point match (Cube: 1)
>  |                  |   |                  |
>  |                  |   |                  |
>  |                  |   |                  |
>  |          X  X  X |   | X  X  X          |     Rolled 21
>  |          X  X  X |   | X  X  X  O       |     0 points
>  +12-11-10--9--8--7-------6--5--4--3--2--1-+     X: user
>
> Like this?
>
> In this situation odd-ply and even-ply does match each other very well. 
> The esponation is simply that the crashed net is not very well trained 
> in these position type. The crashed net still needs a bit more training. 
> To me it lloks like the net understands the pos when it's vied from O 
> side, but overestimates X winning chances when it's viewing the position 
> from X's viewpoint.

I agree we need a better net for containment cases, a sub-class of 
crashed. The bad news is I am still unable to train such a beast, 
perhaps due to chicken-and-egg problem, perhaps because of other 
problems as well. I need fresh ideas, which might come over time, or if 
someone else joins me in working on the nets.

However, positions such the above are not a big concern at this stage. I 
know it is an eyesore to see such evaluations. I know your confidence 
might be shaken each time you see it, but my main concern is playing 
strength. GNUbg checkers play is reasonable here, and will not make 
wrong cube decisions at most scores. Later, when we get play problems 
out of the way, we can aim higher and see if we can get more absolute 
equities right.

If you expect the net to get it right every time, you are in for a big 
disappointment. If you care about absolute equities, you must perform a 
rollout (which can be wrong as well due to imperfect play). A bot is far 
from perfect, it is just a strong player.

-Joseph

> Dear Hugh,
> If you have reference positions like this, where you know the right(ish) 
> outcome, the project would be very happy add this to the training database.
>
> I can't say that the crashed net is the "worst" trained network, but 
> it's simplest to find positions gnubg doesn't handle correctly in the 
> crashed position class.
>
> -Øystein
>
>
>
>
>
> _______________________________________________
> Bug-gnubg mailing list
> address@hidden
> http://mail.gnu.org/mailman/listinfo/bug-gnubg