Re: [Bug-gnubg] Odd rollout results of classic reference position

[Joern Thyssen]

On Mon, May 19, 2003 at 09:26:37AM +0100, Ian Shaw wrote
> Øystein wrote:
> 
> >I think this position may be misplayed in the rollouts. If it keeps the cube 
> >for a roll, X may be able to get a incorrect take from O. I guess that 
> >what's happening in the rollouts.
> 
> 
> > Can you try to play this position a couple of times against GNU and see if 
> > it does something funny in subsequent cube actions? (I guess chequer play 
> > will be really simple in this position.)
> 
> 
> You are correct, Øystein.  After No Double, X 62 6/0 6/4, Dance, O
> takes if X doubles. A rollout shows this to be a 0.03 pass. There are
> others, no doubt.

Does 2-ply get this cube decision right? If yes, then you really need to
do the original cubeful rollout with 2-ply cube decisions. If no, do
4-ply the this cube decision right? If yes, then you really need to the
the original cuebful rollout with 4-ply cube decisions? If no, ....

> I have also learnt that a rollout isn't necessarily more accurate than
> an evaluation. 0 through 4-ply all get this problem right.

Cubeless rollouts "only" depends on the chequer play decisions being
correct. 

Cubeful rollouts also depends on cube decisions being correct. This is a
much more severe condition, as the neural nets are often good giving
relative equities, but have problems with absolute equities.

Jørn