HELPING GNU GO SEE THE CONNECTION

Run gnugo -l tune.sgf --quiet -L40 -a -t to obtain GNU Go's move valuations. you can add -w to the options if you want more information. GNU Go values the moves in this position as follows:


1. J2 14.945239
2. B17 11.633148
3. E14 9.264934
4. E10 8.704933
5. F11 7.922349
6. E2 7.750444
7. K2 7.652647
8. Q12 7.132079
9. G11 6.922761
10. R12 6.803124

The move at J2 is not hugely overvalued. Rather the move at K2 is undervalued. Here is the breakdown of the 7.65 measly points found for K2:


K2: 7.000000 - block or expand territory
K2: 0.650000 - influence
Move generation values K2 to 7.650000

Reading the full report we find that the only pattern matched at K2 is CB241, which looks like this:


Pattern CB241

.. block off area
.*
OX

:8,XOb,NULL

..
ba
OX

;odefend_against(a,b)

At the very least, a pattern of type C should be matched here, since K2 establishes a connection between two dragons. Let's add one. Add the following pattern to patterns/patterns.db, then recompile GNU Go.


Pattern EC68
# db added (2.7.201)

O.XOo connect underneath
..*.o
.....
-----

:8,CXda

O.Abo
..*.o
.....
-----

; lib(A)<4 && lib(b)>2 && !attack(A)

The significance of the constraint is that if the X stone has only 3 liberties and the right O stone has at least three, then the pattern almost guarantees connection. There are rare cases where it could be cut, and these could be handled by a more elaborate constraint, but even in such a case the move at least threatens to connect, and is probably beneficial, so perhaps we don't need to bother with this. The last constraint excludes the case where the X stone can be tactically captured, since if that is true, there is a better move.

After this change, GNU Go finds the right move K2 with the following move reasons:


K2: 13.333333 - owl attack/defend for H3
K2: 7.000000 - block or expand territory
K2: 0.115500 - strategic effect on J4
K2: 0.710000 - strategic effect on M3
K2: 0.650000 - influence
Move generation values K2 to 21.808832

This simple change is thus sufficient to correct GNU Go's behavior at this move.

Continue to the Next page.


Return to GNU's home page.

Return to the GNU Go development page.

Please send comments on these web pages to webmasters@www.gnu.org, send other questions to gnu@gnu.org.

Copyright (C) 1999 Free Software Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111, USA

Verbatim copying and distribution of this entire article is permitted in any medium, provided this notice is preserved.

Updated: 1 Dec 2000 bump