Re: [gnugo-devel] Adapt semeai test suite STS-RV for GNU Go.

[Gunnar Farnebäck]

Emanuele Cisbani wrote:
> Work has begun.
>
> I posted the first set of 26 tests "ported" to GNU Go,
> and the problem suite folder to put in games directory.
>
> http://trac.gnugo.org/gnugo/ticket/41
>
> Summary: 17/26 passes. 0 unexpected passes, 9 unexpected failures
>
> Only one failure is a misunderstanding of semeai problem
> (GNU Go fail to read a simple seki solution).

The failure is a very simple semeai with no eyes, two common liberties
and three (white) vs two (black) outer liberties. This is a typical
case where the lack of semeai understanding in the tactical reading
fools up the analysis. Since black has only four liberties a tactical
attack is found, but not the tactical defense which requires filling
in an outer liberty of the adjacent five liberty string.

Fixing this in the tactical reading is not very difficult but doing it
efficiently is. Maybe it's possible to solve it in the dragon
amalgamation instead by making it more suspicious about tactically
dead four liberty strings.

Actually it's quite easy to disable amalgamation over four-liberty
strings altogether, see the appended patch. It does solve this semeai
mistake as well as testcases semeai:48 and semeai:147. Probably it
does bad things elsewhere though.

/Gunnar

Index: patterns/conn.db
===================================================================
--- patterns/conn.db    (revision 2407)
+++ patterns/conn.db    (working copy)
@@ -629,7 +629,8 @@
 Bc
 dA

-;(attack(A) || attack(B)) && !disconnect_helper(c,d)
+;((attack(A) && lib(A) < 4) || (attack(B) && lib(B) < 4))
+;&& !disconnect_helper(c,d)


 Pattern CC506