Re: [gnugo-devel] breakdown of GnuGo 3.2 in Goban

[Arend Bayer]

> > In one game, it misread a ladder, quite far into the ladder, while taking
> > about 3 minutes per move (at level 10) on the above machine.
(...)

> Of course the tactical reading code is very stable and would
> not misread a ladder. However the owl code thinks this is
> an owl defense. I haven't seen this bug before.

Well, it sort of misread the ladder. It doesn't think it's an attack on
G16, but while defending F18, it keeps playing the ladder until the
attacker gets pushed beyond a reading depth limit in attack2(), where
the attacker gives up because a boundary string is still in atari.

The patch below solves this. However, it will in turn cause a problem
in cases where the ladder succeeds. This is hopefully the smaller
problem, though.

Arend


Index: engine/reading.c
===================================================================
RCS file: /cvsroot/gnugo/gnugo/engine/reading.c,v
retrieving revision 1.90
diff -u -p -r1.90 reading.c
--- engine/reading.c    9 Dec 2002 23:04:16 -0000       1.90
+++ engine/reading.c    15 Dec 2002 08:29:11 -0000
@@ -1375,7 +1375,8 @@ defend2(int str, int *move, int komaster
     }
   }

-  break_chain2_defense_moves(str, &moves);
+  if (stackp < depth)
+    break_chain2_defense_moves(str, &moves);

   if (stackp <= backfill_depth) {
     special_rescue5_moves(str, libs, &moves);