Re: [gnugo-devel] endgame tuning

[Evan Daniel]

On Fri, 10 Sep 2004 03:02:12 +0200, Gunnar Farnebäck
<address@hidden> wrote:
> > In general, I think that if all the move does is pull a stone out of
> > atari, the shape considerations are relatively irrelevant compared to
> > the value of the groups directly involved.  If it is a tactical
> > defense and there are other liberties, then shape may well still be
> > relevant.  Of course, I could be wrong about all that, but that was
> > the logic behind the change.
>
> I would say that giving the shape bonus is questionable in all cases
> for this pattern but pulling out a string in atari often deprives the
> opponent of easy eye space. Actually a tactically captured string with
> remaining liberties might even have a fair bit of aji, so it could be
> argued that it's more important to draw out a string in atari than one
> with liberties. I guess the bottom line is that I can accept reducing
> or removing the shape bonus overall but that I don't see a strong case
> for differentiating with respect to liberties.

I tried reducing the shape value , and that produced a new FAIL at
trevord:760, which looks basically real to me.  There are other issues
with the test (it's a basic example of a counterproductive kikashi
before the correct move), which long term clearly need to be solved in
other ways, but I think short term the pattern is more likely to
prevent such moves in its current form than it is to cause problems
like in endgame:910.

A new version of the patch is at
http://www4.ncsu.edu/~ebdaniel/endgame.patch.  I would propose that
the CC81 shape reduction should not be merged, though others are
certainly welcome to disagree.

I probably won't get to figuring out influence tuning any time soon to
do eg CB343 properly, so I'll defer to other's judgement as to whether
it's worth merging as is.

Complete delta:

endgame:850 PASS
endgame:910 PASS
endgame:920 PASS
trevord:760 FAIL
kgs:70 PASS
olympiad2004:113 PASS
olympiad2004:121 PASS

Evan Daniel