Re: [gnugo-devel] influence and territory values

[David G Doshay]

On 14, May 2006, at 12:41 PM, Gunnar Farnebäck wrote:

> David wrote:
> > > > Following from F6, option I is us G4, them C6, us D3, and option  
> > > > L is
> > > > us D3, them C6, us G4.
> > > >
> > > > For option I: InfW=-7.00, TerrW=-3.00, InfB=-29.00, TerrB=-3.00
> > > > For option L: InfW=-11.00, TerrW=-3.00, InfB=-31.00, TerrB=-3.00
> > > >
> > > > Why would the same board evaluate differently?
> > >
> > > No idea. How are InfW and InfB computed?
> >
> > These are just short names for:
> >
> >      influence4black = influence_score(INITIAL_INFLUENCE( BLACK ));
> >      territory4black = influence_evaluate_position( BLACK, NULL,
> > &status );
> >      influence4white = influence_score(INITIAL_INFLUENCE( WHITE ));
> >      territory4white = influence_evaluate_position( WHITE, NULL,
> > &status );
> >
> > calculated directly from the functions in engine/influence.c
>
> If I understand things correctly, the GTP commands
>
> boardsize 9
> clear_board
> play black f6
> play white g4
> play black c6
> play white d3
> initial_influence black territory_value
> captures black
> captures white
>
> should show some difference compared to

I do not see why they should, but they did.
> boardsize 9
> clear_board
> play black f6
> play white d3
> play black c6
> play white g4
> initial_influence black territory_value
> captures black
> captures white

Note that the different returned values were for Influence, not
territory values, but yes. Those two sequences were sent to 2
different CPUs, and the values returned were different.

> but I get identical results. Plausibly there could be some badly
> initialized or reset data structure

This is a plausible reason, but it seems unlikely. The nodes are
unlikely to get out of sync in that way ... but I will check again.

> involved but then debugging would
> require doing the same intermediate processing.

We have not re-run these sequences since the tournament, and
perhaps we should.

> In other words, in
> order to get any further with this I would need complete and
> preferrably minimal code to reproduce the problem.
>
> /Gunnar

Complete SlugGo code is easy ... minimal is harder. There is a
bunch of code there that is required for the cluster communication.

I have done some restructuring since making the code available
for download. Early in the week I will tar up the new bundle and
link to it instead. After I put up that link I will email again with
some thoughts on how to simplify figuring out what SlugGo is
doing.

Building SlugGo also requires installation of lam-mpi.

Thanks.


Cheers,
David