Re: [gnugo-devel] about patchs for regress file 13x13b:8

[Evan Berggren Daniel]

On Fri, 4 Apr 2003, kevin yong wrote:

> Top moves:
> 1. F10 16.44
> 2. F11 16.44
> 3. K1  2.19
> ..
> Checking safety of a black move at F10
> Move generation likes F10 with value 16.44
> genmove() recommends F10 with value 16.44
>
> my appinion is that: it is basically works, but may
> need little more detailed tuning to make F10 has more
> value than F11. Factualy, the territory gain of F10 is
> only 1 point more than F11, without considering F10
> can also connect G9 etc, because which are beyond the
> scope of these 2 patterns.

I agree with this assessment.  I think there is a problem with these
patterns being j patterns at all, but I don't have a better suggestion.  I
imagine that if you run this with my value_moves patch from a little while
ago, you will find that gnugo gets the answer right.  If not, well, that
would be interesting to hear about too.

>
> I did read patterns.texi file several times, but I
> still don?t think I have fully understand how the
> whole pattern spec works yet. Eg. Pattern ED67, I
> don?t know what constraints we should put on and how
> to specify them.

The basic way it works is that the constraint is a C boolean expression,
with some custom preprocessing run on it.  Once the pattern is matched
against the board, other things are checked via the constraint.  eg,
reading constraints, whether the stones in question are alive or dead,
whether the stones are running in the right direction.

For ED67, the constraint needs to be such that blocking there is safe, and
is defending something useful, either territory on the edge or a
connection.  That's what the attack tactical constraints are intended to
do, though I don't think they do the best job of it.

Evan Daniel