Re: AW: [gnugo-devel] a patch from a newbie :)

[Arend Bayer]

On Tue, 10 Sep 2002, Portela Fernand wrote:

> Dan wrote:
>
> > Could you explain the meaning of HALFWIN and HALFLOSS and the
> > general intent of the patch?
>
> The meanings would be something like :
>
> * HALFWIN : attack failed ( dragon not killed ), but something
>             "substantial" can be captured.
> * HALFLOSS : defense successfull ( dragon alive ), but something
>              "substantial" has to be sacrificed.

Good!

> > I looked at blunder:12. Arend: is there any relevance of the recent
> > patch arend_3_7.2 to this problem? It is a case where part of the
> > dragon can be discarded so the dragon cannot be owl attacked. Since
> > the number of liberties is small, I would think D4 should be a
> > tactical attack if fourlib_depth is sufficient.
>
> blunder:12 is the typical case where the atari_atari code fails,
> mainly because after W D4 B D3, the only one attacked string has
> suddenly 5 liberties (I guess, this is the "reason" of the C6 liberty
> in the problem). At the same time, the OWL code "sees" the correct
> sequence, but fails to consider it a WIN, since it doesn't kill the
> dragon.

I think Dan was talking about the tactical reading code in reading.c
(atari_atari is the code in combination.c -- it is very good but by far
not as important as the reading code). You're right that black is
tactically alive by our definition after D4- D3 because it has five
liberties. So arend_3_7.2 can do nothing here. (And note that D6 could
have an aritrary number of outside liberties, so this is really not an
tactical reading problem.)

I think it is absolutely correct to address this as an owl problem.
If we want to do anything useful with connection handling we will have
to find such a way to value loosing tails.

I only looked at the patch briefly, will look at it in more detail later.
It seems to me that for the move valuation, you treat defending a dragon
that is attackable with code HALFWIN as it would be as valuable as
defending the whole dragon:
gnugo -l games/blunder10.sgf -L 2 -t
Top moves:
1. D4  39.03

Of course, D4 is only worth ~15 points here. This will cause problems
in other situations, where it is better to tenuki here to grab a 35
points play elsewhere -- iow, we need to add some code in the move
valuation to handle this.

Arend

P.S.: Btw, thanks a lot for changing all the "3 - acode" stuff -- that
should have been done for consistency a while ago. Probably we should
use the REVERSE_RESULT macro here all the time, which should be moved
from cache.h to a more prominent place in liberty.h.