Re: [Bug-gnubg] Misreported JSD

[Jim Segrave]

On Tue 13 Mar 2007 (12:26 -0400), address@hidden wrote:
> Okay. FYI, in my case both settings are cubeful. The differences are in
> the number of trials and the ply levels.
> 
> Thanks.
> 
> Chase
> 
> 
> On Tue, 13 Mar 2007 09:32:40 +0100, "Jim Segrave" <address@hidden> said:
> > On Tue 13 Mar 2007 (00:49 +0100), Christian Anthon wrote:
> > > I get 1.830 when extending the rollout. Which version of gnubg are you
> > > using, and are you using threads as this might be a threading issue?
> > 
> > Yes, the threaded code had a bug which would show up if:
> > Settings->Rollouts choices differ from the rollout settings. The main
> > issue is if one is cubeless and the others are not.
> > 

That seems a bit strange. As far as extending rollouts goes, the only
thing kept as part of the rollout data is the number of trials already
done, not the number requested nor the conditions for stopping.

During one trial of rollout of any particular position, gnubg is using
the rollout settings which were done with the original rollout. At the
end of a trial, it makes a decision about move ranking and the current
displayed equities using your current rollout settings, not the ones
used during the actual process of doing the rollout of that
trial. This obviously causes some difficulty when one setting is
cubeful and one is cubeless. 

Hmm, were you doing a truncated rollout? In that case, it's possible
that it rolled out the correct number of moves using the rollout's
settings, then did the final evaluation after those moves using the
normal settings. I'd need to spend some time looking at the code to
figure out what it does there.

It might help if you were able to mail me the .sgf file with the
rollout you are trying to extend (if it's part of a match, please tell
me which move in which game). I'd like to explore what the effect is
of differences between the regular rollout settings and the settings
for the rollout being extended .

-- 
Jim Segrave           address@hidden