Re: [gnugo-devel] another semeai patch

[SP Lee]

> Paul wrote:
> > I didn't look into this, but i suspect that the failes are caused by
> > increased number of nodes.  That is, semeai code most likely runs
out
> > of nodes in those tests now, which is not something good for
accurate
> > reading.
>
> I would guess so too.
>
> > The logic of the patch is like this: "when semeai code doesn't know
> > what to do, and there is a dead eyeshape (but not two or more
certain
> > eyes), try to fill it".  Changing "==" to ">=" effectively removes
the
> > "no two certain eyes" cutoff and will of course waste nodes in
certain
> > positions.
> >
> > > It seems to me that this breakage is acceptable, but we
> > > might want to take a look at whether something can be
> > > done about strategy4.
> >
> > So, maybe we should rather try to make `probable_eyes_b' contain
more
> > accurate results (and not waste semeai nodes).
>
> Fully agreed. The "==" logic is correct and I'd much prefer to get the
> eye count right.
>
> > I will regress the change i proposed: ignoring lunches for dragon B
> > which are part of dragon A (like we do for dragon A's lunches which
> > are part of B).
>
> If that doesn't work out I have a reading patch (not yet regressed)
> which solves the tactical reading problem that's at the root of it
> all. This is the reading mistake:
>
> loadsgf games/nngs/beedee-gnugo-3.5.3-200401140035.sgf 280
> trymove B S17
> trymove W S16
> trymove B L12
> trymove W P10
> trymove B K9
> trymove W S10
> defend M10
>
> /Gunnar
>

I have made a patch in owl.c to ignore all "eat lunch" move during
semeai. It solves at least semeai:58. I have run the regression tests,
and I'm not sure whether the passes and fails are due to this patch or
other patches in cvs. I believe there should be not so many changes due
to this patch.

Btw, I have checked nngs:820 and found that D14 is absolutely not a bad
move. The black dragon can't escape even white doesn't play J13 or L9.
So even if the white loses the semeai, the points gained is rewarding
for white.


SP Lee

$ cvs diff -u owl.c
Index: owl.c
===================================================================
RCS file: /cvsroot/gnugo/gnugo/engine/owl.c,v
retrieving revision 1.187
diff -u -r1.187 owl.c
--- owl.c       7 Jan 2004 10:00:20 -0000       1.187
+++ owl.c       14 Jan 2004 22:15:10 -0000
@@ -1255,6 +1255,9 @@
   int k;

   for (k = 0; k < MAX_MOVES-1; k++) {
+  /* Eating lunch decreases own liberties during semeai */
+    if (owl_moves[k].name == "eat lunch")
+      continue;
     move = owl_moves[k].pos;
     if (move == NO_MOVE)
       break;