Re: [Bug-gnubg] Predefined move filters

[Joseph Heled]

I cant see how such a tiny number (1296) can give any indication, unless 
the difference is huge, which I don't think is the case, and of course 
you can't use any form of variance reduction, since those depend again 
on which ply you believe in.

For 2 close players (like for a net and a contender) I see I need at 
least 50,000 money games to give a consistent answer - and that ONLY 
WHEN I run with "run games in pairs, alternating the the same dice 
sequence" option, which reduces a lot of the dice noise.

-Joseph


Gary Wong wrote:
> On Tue, Dec 31, 2002 at 12:31:51PM -0000, ershaw wrote:
>
> > I'd be happy to volunteer for this. How many trials do you want? I can throw
> > a 2Ghz PC at it, and probably give it up to a week.
>
>
> Thanks!  How about we just start with 1296 games, and we can add more
> after we see how that goes... if you want to use these settings:
>
>     set rollout trials 1296
>     set rollout truncation plies 0
>     set rollout varredn on
>     set rollout quasirandom on
>     set rollout cubeful off
>     set rollout initial on
>     set rollout seed 0
>     set rollout player 0 chequerplay plies 0
>     set rollout player 1 chequerplay plies 1
>     set rollout player 1 movefilter 1 0 8 8 0.2
>     new session
>     new game
>     set board 4HPwATDgc/ABMA
>     rollout
>
> then I'll do another 1296 with player 0 and 1 swapped.  I presume
> we're both using the 0.12b weights?
>
>
> > By your last para, I assume it rolls out with player 0 on roll each
> > time, so you have to swap over to get player 1 on roll.
>
>
> Exactly (or vice versa).
>
> Thanks very much,
> Gary.