Re: [gnugo-devel] atari_atari and restricted_defend1

[Inge Wallin]

Gunnar wrote:
> 3) Redesign. The restricted_defend1() was a simple way out for
> atari_atari() but doesn't really scale for more general combination
> sequences. It should be possible to find all relevant defense moves by
> considering direct liberties, captures of opponent neighbors, and
> possibly some more moves which are best found through pattern
> matching.

Yes.

> Similarly I would recommend pattern matching instead of the general
> attack_threats() function to find the attack moves. That should give
> better selectivity (i.e. avoid including really silly attack threats).

Of course, you are right again.  But if I had started coding on
pattern matching code we wouldn't have anything right now.  Besides, I
was thinking of adding pattern matching to attack_threats, and to let
it take a parameter that tells it which ways to try (liberties,
second-order liberties, patterns, etc).

The only really silly attack threat that I can think of is a
self-atari, and those could be avoided by testing for it explicitely.
Do you have other examples?  Remember that we are talking about
combination attacks here and most of the time a few stones have to be
sacrificed before the real attack appears.

Dan wrote:
> I'm running the regressions now. It doesn't break anything
> in atari_atari.tst, and it rejects the bogus move at move
> 146 in viking2.sgf while finding the correct one.

Did you try setting aa_threat_depth to something bigger than 1?

        -Inge