Re: [gnugo-devel] nando_3.9.4c ?

[bump]

> Ok, I'm seeing things more clearly now. 
> 
> I'm almost positive that the passes in safety.tst (btw, it is ineffective on
> the new tests Arend just added) are actually only caused by the "thing" I
> added later, and has most probably nothing to do with the work on GAIN/LOSS
> codes.

In addition to the tests in safety.tst let me mention two
safety examples from other tests. There is century2002:50.
Then there is also nngs1:8, a test which currently passes with
the CVS version. It just barely passes, though, as I'll
explain.

It should be clear to anyone that the W dragon at L8 
is in need of defense.

The following seemingly unrelated pattern fixes century2002:270.

Pattern A422
# db New pattern (3.3.9)

X..?         try to cut
X.*X
??O?

:8,-,value(61)

bde?
bc*a
??O?

;(owl_escape_value(a) > 0 || owl_escape_value(b) > 0)
;&& xplay_attack_either(*,c,d,e,c,e)

After this pattern, nngs1:8 fails. Adding an ATTACK
pattern has paradoxically changed the status of the
group from WEAK to WEAKLY_ALIVE. It accomplishes this
by stopping the owl code from running out of nodes.

The attack on the L8 dragon is severe, but if you
look at the variations generated, it seems that
GNU Go thinks it reads out life. It is willing to
sacrifice parts of the dragon, which is one reason
that what Nando is doing may prove important.

The addition of this pattern also breaks nngs:1700
and passes nngs3:420, century2002:130 and century:2002:270
for a net gain of one test. I haven't investigated the
failure of nngs1700.

Dan