inessential worms (Re: [gnugo-devel] tuning patch)

[Arend Bayer]

Gunnar wrote:

> Arend wrote:
> > > - never give cutstone bonus for an inessential string
> >
> > Is this doing anything.
>
> It solves test case gunnar:26.
>
> > If I read the last loop in make_worms correctly, a (potential)
> > cutting stone can never be inessential.
>
> I haven't checked whether it's involved here, but there's a step to
> revise worm essentiality towards the end of make_dragons() too.

Ah, I forgot about that.

> > I stumbled across this in a case where it pretty much didn't make sense,
> > so I am not sure whether this policy is right.
>
> Please tell us more.

On reflection, I think I was expecting s.th. from the inessentiality
classification that it cannot provide.
It's maybe not a big problem, either.

The problem was with vie:14 and the stone at F4. If you run an
influence/territory evaluation with the white dragon G6 assumed to be
dead, then the F4 stone should not be considered territory for white.
That would work if we classified it as inessential. However, if
that dragon is assumed to be alive (e.g. when we are assuming it to
be white's turn), then it should be considered territory for white,
hence we can _not_ classify it as inessential.

Running gq -l golois/Goemate990902-3.sgf -t -w -d0x100000 reproduces
the problem: The move at G7 doesn't get a territory change at F4.
This won't be a huge problem in practice, because it's only a small
inaccuracy for owl move reasons.

I was hitting this in a different context. I am experimenting with a patch
that reworks the escape estimation and escape move generation inside
owl. For that I run an escape influence computation with the owl
dragon set to dead. The escape estimation somehow got confused by the
escape potential for G6 at F4 :-)

I will post a patch later.

Arend