Re: [gnugo-devel] 100 game series

[Gunnar Farneback]

Dan wrote:
> One game, match078.sgf is misscored by W. The error
> is reproducible since --score last reports a 7.5
> win by W. There is actually a 7.5 point win by B.
> I'll investigate this later. The sgf file is
> attached.

The aftermath scoring gets it right, as expected. Interesting
difference in playing style, btw. White managed to build a center moyo
in a way that GNU Go has never managed before. Not that all of the
moyo building moves were good or well-timed, but I think it was
encouraging anyway.

Particularly annoying was the tendency to make "two space diagonal
jumps", leaving a big gap in the position. The worst example of this
is P6 but also K14 and N10 are similar. I guess one way of solving it
is to make the pattern Intrusion21 more powerful, but I leave it to
Arend to try to tune this.

Move 66 (when N10 was played) involves a serious semeai mistake. The
traces say

| Considering semeai between N18 and M13
| Changing matcher_status of N18 from CRITICAL to ALIVE.
| Changing safety of N18 from CRITICAL to ALIVE.
| Changing matcher_status of M13 from CRITICAL to DEAD.
| Changing safety of M13 from CRITICAL to ALIVE.

Notice the inconsistency in matcher_status and safety change for M13.
That is fixed by the appended patch. Still it is clearly incorrect to
change M13 to DEAD since it trivially can live by connecting at M14,
so there's some more bug here. Maybe Dan can comment on this.

> Score-dependent behavior consists of playing more cautiously
> when ahead. This hopefully has the effect of increasing the
> win rate though decreasing the margin. I'm not sure I believe
> that it's more meaningful to turn this off.

I'm also doubtful.

> About the uncertain owl reasons, in the game
> Teto-gnugo-3.1.25-200202190228.sgf available at
> http://www.lysator.liu.se/~gunnar/gnugo/nngs/ , at move
> 110, B finds an uncertain defense reason at F17 for the
> upper left corner. It thinks the group is alive but the
> node count is exhausted. B thinks it's behind by a few
> points (less than 10) so the uncertain defense reason is
> ignored. As a result the group gets killed. Can anything
> be done about this?

Even better would have been to connect at H15 in move 104, especially
since that would have cut white at the same time.

> One thing I thought about was that it might be good for
> twogtp to have an option of playing one or three game
> kadoban. Thus if you run the option --kadoban 3 the
> handicap would change after someone wins 3 games in a
> row. This way you'd find out the right handicap.

Sounds like a good idea.

/Gunnar

Index: engine/semeai.c
===================================================================
RCS file: /cvsroot/gnugo/gnugo/engine/semeai.c,v
retrieving revision 1.35
diff -u -r1.35 semeai.c
--- engine/semeai.c     11 Feb 2002 04:07:12 -0000      1.35
+++ engine/semeai.c     20 Feb 2002 17:53:50 -0000
@@ -240,7 +240,7 @@
          if (b_matcher_status == ALIVE_IN_SEKI)
            b_matcher_status = ALIVE;
          update_status(apos, a_matcher_status, a_status);
-         update_status(bpos, b_matcher_status, a_status);
+         update_status(bpos, b_matcher_status, b_status);
        }
       }
     }