Re: [gnugo-devel] 3.5.3 vs 3.5.4 (owl attack patch)

[SP Lee]

> max-d wrote:
> >
> > A- i wrote
> >
> > > i had done such a match with board-size 13,i got
> > >
> > > gnugo 3_5_4 82
> > > gnugo_3.4 88
> >
> > B- SP Lee pointed  out some 3-5-4's mistakes
> >
> > C- I had run a lot of games which showed that 3-5-3 did better than
3.4
> > vs Handtalk
> >
> > I tried to verify things with an other match (size 13 level 10)
> >
> > = gnugo353 91
> > = gnugo_3_5_4 67
> >
> > almost 57%
> > Seems that something was broken in gnugo_3_5_4
>
> List of patches that can be theoritically responsible for something
> bad between 3.5.3 and 3.5.4:
>
> paul_5_4.2a. A simple followup to paul_5_4.2 below.
> arend_5_4.3. Break-in code now uses new cache. (very unlikely)
> splee_5_4.1b. Revision of estimate_lunch_eye_value(). Supercedes
> splee_5_4.1a. Revision of splee_5_4.1. Semeai tweak.
> paul_5_4.2. Semeai patch.
> paul_5_4.1. Seki code improvements. (very unlikely)
> gunnar_5_4.7. Tuning. (large patch)
> martin_5_4.1b. Revision of martin_5_4.1a. Revision of martin_5_4.1.
> Owl tuning. (might waste node, don't know counters)
> evan_5_4.2a. Revision of evan_5_4.2. Tuning. (hardly)
> gunnar_5_4.1. Tuning. (very good breakage -- unlikely)
> arend_5_4.1. Breakin fix. (very unlikely)
>
> Not too many candidates.  Anyone wants to check?  The main problem
> is to determine on what to check.
>
> Paul
>
>

Again for the twogtp game I referred to before in
http://mail.gnu.org/archive/html/gnugo-devel/2004-02/msg00060.html. On
move 38 gnugo was checking the status of black group q5. Gnugo 3.4 is
able to find an attack move on q6, but gnugo 3.5.4 just finds the group
is alive. There is an owl attack pattern, A1124a, revised in 3.5.4,
which is related to this owl analysis. Unfortunately, there are not so
many owl attack patterns which can be matched in this situation, if
A1124a can't be matched. The move sequence should be considered W:Q6
B:S5 W:R6(A1124a, without W:R6 B escapes easily).

After the following patch gnugo is able to find the attack move q6.


SP Lee

Index: owl_attackpats.db
===================================================================
RCS file: /cvsroot/gnugo/gnugo/patterns/owl_attackpats.db,v
retrieving revision 1.107
diff -u -r1.107 owl_attackpats.db
--- owl_attackpats.db 4 Feb 2004 16:07:37 -0000 1.107
+++ owl_attackpats.db 24 Feb 2004 18:55:14 -0000
@@ -4566,10 +4566,11 @@
 # tm New Pattern (3.1.22) (see nngs:770 decide-dragon K4)
 #   see also D1140
 # gf Revised. Might still be too general. (3.5.4)
+# spl revised (3.5.5)

 ?OY
 x*.
-xx.
+xxx

 :8,-,value(70)