Re: [gnugo-devel] games against 3.3.16

[Arend Bayer]

Heikki Levanto wrote:

> On Tue, Jun 17, 2003 at 11:24:05PM +0200, Arend Bayer wrote:
> >
> > Well, apart from classical corner joseki moves, I somewhat concur
> > that we should rethink how much we trust our patterns:
>
> Sounds reasonable.
>
> It should be possible (and I wish I had the time and energy) to check
> all regression failures, see which joseki patterns were involved in the
> move generation of the wrong move(s), and try to disable them one by one
> and see how much else breaks. This would take for ever to run, but could
> be done almost automatically. It could result in a list of interesting
> patterns, specifying what they do to regressions (pattern X breaks test
> NNN, but removing it would break tests MMM, KKK, and LLL)

The problem I see with this is that the regression tests might give
much too small coverage, i.e. there are certainly needed patterns which just
do not appear in any regression test.

I am looking forward to Kombilo v6. Uli mentioned that it should
contain a better separation between the UI and the search routines, that
allow to use the search routine in python scripts.

That should make it _very_ easy to write a script that (after you've
built a gnugo version with and without the pattern) gives you _all_
instances in some 1000s of games where the pattern would actually make
a difference for the move generation.
(I.e. a sort of kombilo-aided matcher_check.)

Arend