[Bug-gnubg] Re: gnubg settings

[Massimiliano . Maini]

>Sorry to bug you again,

No problem. I'm forwarding this to the gnubg mailing list since the topic
is of general interest and since somebody else may answer as well (and
better
than me, for sure).

>but I use Supremo, but don't know about pruning
>neural nets.  With Supremo, I am 3-ply for everything.  As I am just using
>it to analyze matches, and don't mind how long it takes, am I using the
best
>settings?  Some at the Zone say chequers should be 2-ply, while others say
>that cube should be 2-ply.

First of all, supremo is 2-ply in gnubg. Confusion may arise since gnubg's
2-ply is Snowie's 3-ply ...

Best setting is probably 7-ply with move filter huge and no pruning nets,
but this would take close to forever for a single game.
I don't think anybody in the world has ever gone beyond 3-ply (Snowie's
4-ply).
And for rollouts, 2-ply seems to be more than enough (0-ply is ultra-fast
and
in some situations, give accurate results).

My personal advice, use supremo for both chequer and cube play/analysis.

[Pruning nets : that's something added a while ago to speed-up evaluation
Should be almost twice as fast playing at almost the same level. It's an
option(I think it's ON by default) in the panel where you chose the level,
like expert world class, supremo etc]

>Also, are the default thresholds for what is
>doubtful, bad, and very bad good?  I know that some people at Zone insist
>that these must be altered as well.

Thresholds can be altered, but I think more or less anybody uses the
default
ones. However, this isn't very important: the thing that matters is the
equity
loss ... -0.050 is -0.050, up to you to consider it as good, doubtful, bad
or
very bad.

>Is there a good sticky document on GNUBG to address setting the program to
>do the most exhasustive analysis?

Not I've heard of, maybe it's on the todo list ...

>Sorry to be such a pest, but I will at least pass on any information to
>players at the Zone (at least the ones who take the game more seriously by
>analyzing their matches).

No problem, but post to the mailing list (address@hidden) to reach the
maximum number of people that may be interested/have something to say.


MaX.