Re: [gnugo-devel] endgame tuning

[Arend Bayer]

Gunnar wrote:

> Evan wrote:
> > I also targeted Gifu03:202 a bit; GNU Go's answer had changed relative to
> > the game move (and improved) but the relative value of the correct move
> > and the game move was still wrong; now it's not.  Perhaps the test needs
> > revision.
> 
> I still think that H7 is the globally correct move and important to
> find, but I'm open for a second opinion from a stronger player.

I agree. The move at L7 may be sente, but the exchange L7 vs L8 is
a negative one. B1 is worth 5.5 points, and H7 seems worth significantly
more.

10 . . . O . X O . O + . O . X X X . O O 10     BLACK (X) has captured 0 stones
 9 . . . O . O . . . . . . . . O X O . . 9
 8 . . . . . . O . . O . . . . X . . O . 8
 7 . . O . X . . . . . . O . . . . . . . 7
 6 . . O X . . . . . . X O . . . . O . . 6
 5 . . O X . . . X . . . X X . . X O O . 5
 4 . . O O X . . . . + . . . . . + X O O 4
 3 . . . O X . . X . . . . X . . X . X X 3
 2 . O O X X . . . . . . . . . . . . X . 2
 1 . . X . . . . . . . . . . . . . . . . 1
   A B C D E F G H J K L M N O P Q R S T

The problem as I see it is the following:

We have a block for white influence at G5 and F6. (This can be seen with
view.pike by showing the "white permeability" for "white influence,
dragons known".) Which is logical, since when White tries to intrude 
with F6), it can be blocked at G5, and can be blocked at F6. But of
course Black cannot block at both places.

So maybe we need s.th. analogous to Gunnar's "vulnerabilities" in the
connection reading. We should note that the blocks at G5 and F6 try to
block influence coming from G6. And consequently remove one of the
blocks.

This solution could help against the problem of sometimes horribly
undervalueing opponent's intrusions into GNU Go's moyos.

How do we learn that the block at F5 is trying to block influence from
F6? Either we add this information to the patterns in barriers.db.
Or we check it _after_ spreading the white influence: We just check all
intersections with white influence, whether it has two neighbors with
blocks. If yes, we remove one of them, and respread the influence
from this intersection.

Arend


> Index: patterns/barriers.db
> ===================================================================
> RCS file: /cvsroot/gnugo/gnugo/patterns/barriers.db,v
> retrieving revision 1.62
> diff -u -r1.62 barriers.db
> --- patterns/barriers.db      24 Jan 2004 04:04:56 -0000      1.62
> +++ patterns/barriers.db      24 Jan 2004 09:03:55 -0000
> @@ -3389,4 +3389,90 @@
>  >non_xterritory(b);
>  
>  
> +Pattern Nonterritory56a
> +# gf New pattern. (3.5.4)
> +# See gifu03:202.
> +
> +oOo
> +...
> +...
> +..X
> +
> +:8,t
> +
> +oco
> +.b.
> +eaf
> +.dX
> +
> +;oplay_connect(a,b,a,c)
> +
> +>non_xterritory(d);
> +>non_xterritory(e);
> +>non_xterritory(f);
> +
> +
> +Pattern Nonterritory56b
> +# gf New pattern. (3.5.4)
> +# See gifu03:202.
> +
> +oOo
> +...
> +...
> +..X
> +
> +:8,t
> +
> +oco
> +.b.
> +ea.
> +fdX
> +
> +;oplay_connect(a,e,d,b,a,c) && oplay_connect(a,d,e,b,a,c)
> +
> +>non_xterritory(f);
> +
> +
> +Pattern Nonterritory57
> +# gf New pattern. (3.5.4)
> +# See gifu03:202.
> +
> +Ooo
> +...
> +...
> +..X
> +
> +:8,t
> +
> +boo
> +...
> +dae
> +.cX
> +
> +;!oplay_disconnect(a,a,b)
> +
> +>non_xterritory(c);
> +>non_xterritory(d);
> +>non_xterritory(e);
> +
> +
> +Pattern Nonterritory58
> +# gf New pattern. (3.5.4)
> +# See gifu03:205.
> +
> +O??
> +.XO
> +..?
> +
> +:8,t
> +
> +O??
> +aXb
> +?c?
> +
> +;!oplay_disconnect(a,a,b)
> +
> +>non_xterritory(c);
> +
> +