Re: [gnugo-devel] gnugo 3.4 problems

[Gunnar Farnebäck]

John wrote:
> Hi, I hope this can help gnugo.
> move 118 ( reduce itself liberty, very strange )
> move 120 ( repeat this bad move again and again )

What's going on here is a semeai in the upper right corner:

M N O P Q R S T
. . . X . X . O 19
. . . X X O O O 18
. X X X O O . . 17
X O O O X X O O 16
. O . X . X O X 15
. O O X X X X . 14
. . O O O O X X 13
. . . X . O O O 12

At move 118 GNU Go (X) plays an own outer liberty at O15 and then
follows up in move 120 with the send-two-return-one move at S19. This
move is then repeated time and time again while white picks up most of
the rest of the board.

(Actually I can't quite reproduce this. I get the following results:

        118  120
3.4     O15  T17
3.6     S19  S19
3.7CVS  S19  S19

which in any case show a huge problem in current versions.)

This is not the first time we have run into the send-two-return-one
problem. See 
http://lists.gnu.org/archive/html/gnugo-devel/2003-06/msg00065.html
with comments and testcases reading:184-187.

That time the problem could be solved by some tactical reading
patches, see
http://lists.gnu.org/archive/html/gnugo-devel/2003-06/msg00157.html
with comments.

I haven't checked yet what exactly goes wrong in the analysis this
time but considering how devastating the problem is I'm inclined to
try some heavy artillery, namely an antisuji pattern. Can someone give
an example where the send-two-return-one move is good or some other
reason against adding the pattern?

/Gunnar