Re: [gnugo-devel] Latest news from the Olympiad

[Gunnar Farnebäck]

I wrote:
> > Yesterday's unfinished game was decided by the Jury to the favour of
> > Indigo,
> 
> Annoying, but mostly because GNU Go misread.

See the end of regression/games/olympiad2004/19x19/gnu-ind.sgf.

The problem is in the lower left corner:

12 O O O O O X X . O O X
11 O X X X O O . X X X O
10 X X X X O . X X O O O
 9 . X O X O O O X O X X
 8 O X O O O X O O O X X
 7 . X X X X X O . O O X
 6 . . X . . X O O X X X
 5 . X O X O O X X X X X
 4 X O O + . . X O X O O
 3 X O . O O . X O X O O
 2 X O O . X . X O X X X
 1 . O . X . . . . . . X
   A B C D E F G H J K L

Black playing first can of course win the semeai trivially by making
two eyes with A7, A6, or A5. I think black E4 would also win the
semeai, although not quite so easily.

The problem is that white playing first is not entirely dead. If black
tries to keep white dead, the combination of white A1 and A7, followed
by D4, leads to a ko for the life and death of both. Black can
sacrifice the A4 string to gain certain life but in that case white
also lives, so it would be a failure.

Also interesting is the variation W A6, B A5, W A7, leaving black with
only one eye, but in this case black eventually wins the semeai. Still
we would want the owl code to see this attack, so that the semeai code
will get to analyze it too. The problem here, however, is that we have
a case of "eyeshape with cutting point",

12 O O O O O X X . O O X
11 O X X X O O . X X X O
10 X X X X O . X X O O O
 9 . X O X O O O X O X X
 8 O X O O O X O O O X X
 7 O X X X X X O . O O X
 6 O . X . . X O O X X X
 5 X X O X O O X X X X X
 4 X O O + . . X O X O O
 3 X O . O O . X O X O O
 2 X O O . X . X O X X X
 1 . O . X . . . . . . X
   A B C D E F G H J K L

which have troubled us a long time. GNU Go thinks black is locally
alive in seki here, which of course is wrong. Paul, should we try to
revive your old patch for this problem?

Another thing to investigate here is that there seems to be some kind
of persistent cache problem. When the position is loaded GNU Go wants
to play black A5 to eliminate a combination attack related to the
semeai, but that doesn't happen when replaying the full game.

/Gunnar