Re: nando_3_9.4b (was: RE: [gnugo-devel] nando_3_9.4)

[bump]

Nando wrote (about nando_3_9.4b):

> Btw, a short note FYI : I've submitted my code as soon as Arend asked for
> it, but I'm still working on it. Currently, I'm testing some "agressive"
> settings and I got an interesting side-effect. Globally, it now moves things
> in the regressions a bit more than the version I submitted : 14 passes and 9
> fails. But more important, I've been quite surprised to get this :
> 
> ---- File [safety.tst]
> 2 PASSED
> 3 PASSED
> 5 PASSED
> 
> Tuned this way, the engine seems a lot more worried about the safety of the
> groups (some of the fails are related to desparate attempts to save dead
> ones). I still need to work a lot on this (evaluate the cost in owl nodes
> for instance, check every breakage, etc), but I think it's worth it,
> specially if it helps the engine on this specific problem.

Safety is a big problem for GNU Go.

The fails you mention from attempting to save a dead group
must consist of situations where there are two mistakes:

(1) The owl code thinks the group is not dead;
(2) GNU Go actually thinks the group is safe.

So if you correct problem (2) without correcting problem (1)
you will get a fail as GNU tries to save the dead stones.
But you are only exposing an existing problem with the owl code.

If this is the case then the patch is good, even though
it breaks a few regressions.

How about sending us the version that gives these unexpected PASS?

Dan