RE: [Bug-gnubg] FW: Analysis Question

[Christopher Yep]

When gnubg evaluates a play it throws away the existing evaluation (and/or 
rollout) for that play (ignoring the cache).  This seems desirable, 
otherwise the .sgf file would be large.

For example, consider the pre-defined "world class" move filter.  Gnubg 
analyzes up to 8 plays within .16.  For a given position, suppose that 
gnubg analyzes 4 plays (A, B, C, D) at 2-ply.  Now, suppose that you want 
to redo the analysis using supremo.  You can't do it without first 
re-analyzing A, B, C, and D at 0-ply (the supremo move filter says to 
analyze up to 16 plays within .32, using 0-ply evaluations to determine 
whether a play is within .32 of the best 0-ply play).  These 0-ply 
evaluations may still be in the cache (I don't know precisely what 
evaluations are stored in the cache), but they won't be there if you save 
the .sgf file and open it again in a new gnubg for example (plays A, B, C, 
and D will only have 2-ply evaluations in the .sgf file).

Snowie doesn't throw away lower-ply evaluations (or rollout 
information).  It keeps all the information in the file, allowing the file 
to grow large.  For example, if you select 20 plays, evaluate all of them 
at 3-ply, then re-evaluate them at 1-ply, then re-evaluate them all at 
3-ply again, these last 3-ply evaluations will all display 
instantly.  Similarly, if you roll them all out, you can cycle between 
1-ply, 3-ply, and rollout displays instantly (Snowie doesn't have to take 
time to re-evaluate or re-rollout).

Chris


At 04:52 AM 9/3/2009, Michael Depreli wrote:
> I checked to see how Snowie 3 handles the "outside the new filter issue".
> After reanalysis it drops the moves down to the level it would have been 
> had the larger filter been
> run in the first place. At least that way it's consistent.
> How come if gnubg has the analysis already stored from the narrower 
> filters it simply just can't analyse
> any NEW moves that fall within the wider filters? It doesn't need to 
> reanalyze the others.
>
>
>
>
> > Date: Thu, 3 Sep 2009 00:14:10 +0200
> > Subject: Re: [Bug-gnubg] FW: Analysis Question
> > From: address@hidden
> > To: address@hidden
> > CC: address@hidden; address@hidden
> >
> > Wouldn't work. Suppose that you have a half made analysis and then
> > restart it. Since the move filter is not stored you would have to go
> > through the match and re-analyse at plies N-1 to make sure that all
> > moves are within the current filter. Time lost. And suppose that you
> > some moves are now outside the new filter. Should they then be
> > dropped?
> >
> > I guess that with a lot of jumping through hoops we could make this
> > work and half way consistent, but is it really worth it? What we do
> > now is at least consistent.
> >
> > Christian.
> >
> > On Wed, Sep 2, 2009 at 11:22 PM, Michael Petch<address@hidden> 
> wrote:
> > >
> > > I've been asked this by others. Personally I'm not sure I would 
> change how
> > > it operates. If one knows this, its just as easy to clear the 
> analysis and
> > > redo it. Just an opinion
> > >
> > > ------ Forwarded Message
> > > From: "address@hidden" <address@hidden>
> > > Date: Wed, 2 Sep 2009 12:39:12 +0000
> > > To: Michael Petch <address@hidden>
> > > Subject: Analysis Question
> > >
> > > Hi Michael
> > > I part analyzed a match then stopped it as I wanted to increase the move
> > > filters.
> > > Saved settings and re-ran the analysis.
> > > Gnubg only analyzes at the larger filters from where the previous 
> analysis
> > > finished.
> > > Is there a reason for this?
> > > Obviously it would be good if it added analysis for the extra moves 
> included
> > > with the wider filters.
> > >