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

[Joseph Heled]

Look guys, any rollout is suspect - for moves or cubes.

Under "normal" circumstances you hope that errors happen infrequently enough or far ahead in the future so as not to affect the main signal from averaging over the huge number of possible sequences of rolls.

When gnubg (say) misunderstands a position and thinks it is a cube when the no-cube is correct, a rollout may be tainted since it will make the incorrect cube on many continuations. same argument for moves. 

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

has anyone done an extensive test for comparing 2ply rollouts with 0ply rollouts and 1ply rollouts, for random positions. I think this will be very interesting, and may provide some basis of discussion for/against using 0ply rollouts. It is very easy to generate millions of positions where 0ply plays incorrectly, and in some very badly. You have to start with positions from games played against people, though.

-Joseph

On Tue, Aug 25, 2009 at 11:33 PM, Massimiliano Maini <address@hidden> wrote:

>De : Joseph Heled <address@hidden>

>À : Massimiliano Maini <address@hidden>
>Cc : "address@hidden" <address@hidden>

>Envoyé le : Mardi, 25 Août 2009, 12h13mn 13s
>Objet : Re: Re : [Bug-gnubg] Odd/even effect: post I made on bgonline

>
>The
net was trained "to evaluate relative equities"  for moves by
>adding
pairs of positions for the different moves, but also on some
>absolute
positions to get the cube decisions right.
>So, it may be that the accuracy is less in positions far from a
>possible cube - but it will surprise me.

What do you mean with "pairs of positions" ? The same board setup with
player 1 / player 2 on roll ?

Have you got any other explanation for the behavior described here :

http://www.bgonline.org/forums/webbbs_config.pl?noframes;read=46778

>Unless I am mistaken GNUbg still is the strongest 0ply player
around
>(at least it was so at the time),

I'm not sure anybody has done a 0ply comparison of the bots recently
mainly because nobody is using 0ply anymore, not even for rollouts.
But I won't be surprised if gnubg is still the strongest, actually I
do hope so.

However, the fact gnubg may misplay that badly the cube decisions as the
above may make users doubt about its real strength, e.g. would you trust
a cubeful rollout where the position above can be reached ?


MaX.