Re: [gnugo-devel] Thickness ideas

[Evan Berggren Daniel]

On Wed, 13 Aug 2003, [ISO-8859-1] St?phane Nicolet wrote:

>
>
> On Jul 5, Arend wrote :
>
> > Btw, if you want to do more experiments with the influence code, one
> > thing I never got around to is the following:
> > you probably know the code that reduces the permeability around stones
> > (I think by a factor of 0.25 on direct neighbours and 0.5 on diagonal
> > neighbours). I think this should depend on the number of liberties a
> > stone has. A single stone with only 3 liberties cannot act freely in
> > response to a hane, similarly a two stone group with only four
> > liberties
> > (hane at the head of two stones), etc.
>
>
> Here is a first try to convince the influence code (in influence.c)
> to generate more thickness :
>
>
> -- Change A implements Arend's idea by turning off the permeability
>    reduction around weak stones. I have also lowered the reduction
>    on diagonal neighbours of strong stones to 20% (instead of 35%
>    previously) because I had noticed that when creating moyos
>    GNU Go often played a dubious diagonal 2-2 jump in the following
>    situations :
>
>    ..............
>    ..............
>    ...........XX.
>    .......X.X....    Black A would be better one point above (at B)
>    .....B..O.....    if Black wants to build a moyo at the top.
>    .....A.O.O....
>    ..............

Interesting.  Have you experimented at all with just changing the
permeability reduction around weak stones, instead of removing it
entirely?  Intuitively, that seems more correct, but there's nothing
particulard to back that up.

> - Change B further lowers the influence strength radiating from
>    stones which are under attack (short of liberty or surrounded,
>    in the sense that they don't have lots of second order liberties).
>    This is intented to have the influence code help isolated stones
>    or worms by adding a stone to the group to get the full influence
>    of the worm back.

Also sounds interesting.  Could be quite good, actually.  You might want
to consider looking at the group weakness and / or data from the owl code
as well here.

> My feeling is that change (A) is quite safe, while change (B) has
> a more drastic effect on the style of play, but need further tuning
> to get a proper effect.

Probably so.  With some tuning, hopefully drastic can become something
good :)

> I don't post the modifications as a patch because I have not tested
> them intensively, and I won't be able to test them further because
> I'm leaving on vacations for 3 weeks. For instance, I just don't have
> the time, before my train, to run the regression for change (A) and
> (B) independantly.

Patch form is good anyway.  Some of us like to play with these things when
we should be doing other stuff ;)

> At the moment, regression results for changes (A)+(B) are 42 PASS
> and 39 FAILs, but that's not the point, of course, because I see
> that as a work in progress. I thought, however, that this may be
> of interest for  the list, so feel free to use/test/modify as you
> will :-)

Definitely of interest.  I hope it works out well.  Enjoy your vacation :)

Evan Daniel