Life and death problem

[sanxiyn]

GNU Go seems to misread an attached rather simple life and death problem. White at B1 kills.

[gunnar]

I was just about to add this as a testcase when I discovered that ld_owl:326 (top left),

is essentially the same. So let's see what can be done to solve it instead. Doing

Sulitelma 511% ../interface/gnugo -l games/life_and_death/mixed4.sgf --decide-dragon B11 -o vars.sgf
finished examine_position
B11 cannot be attacked (7 variations)
B11 can be defended at B12 (2 variations)

and checking the variations in vars.sgf turns up (variation 2):

vital move at B12 (variation 2, hash 77449e0e3d5186d5)
owl_attack B11: 0 (2 or more secure eyes)
(next variation: 3)

This analysis is clearly wrong since C12 is not the solid eye it's supposed to be, despite black not having a safe move at D13.

This is where owl_vital_apats.db comes in handy. Its very purpose is to point out where the eye analysis has been too optimistic. Let's test a pattern like

xXOo    destroy eye
X.XO
OX.*
----

:8,-,value(45)

xXOo
XaXO
OX.*
----

;owl_mineye(a)==1

Seems to work,

Sulitelma 513% ../interface/gnugo -l games/life_and_death/mixed4.sgf --decide-dragon B11 -o vars2.sgf
finished examine_position
B11 can be attacked at B13 (4 variations)
B11 can be defended at B12 (2 variations)

This pattern is added in the attached patch. The effect on the rest of the regressions to follow in a later comment.

[gunnar]

owl tuning

[gunnar]

Only a very small impact on node counts for the rest of the regressions:

ld_owl:326      PASS 1 B13 [1 B13]
1 PASS
Total nodes: 1681186140 3315287 12604741 (+0.014% +0.067% +0.0025%)