Re: [gnugo-devel] Patch: removal of value()s from patterns.db

[Arend Bayer]

On Sat, 26 Oct 2002, Gunnar Farneback wrote:

(...)
> but the main reason for the small regression
> delta is probably that too few of these patterns have any
> corresponding test cases.

What I would dream of to make this kind of tuning more efficient by some
orders of magnitude is a good integration of GNU Go into Kombilo, the
Go database program by Ulrich Goertz.

Kombilo can find all geometrical matches of a pattern in ten thousands of
games within a few seconds. If then the pattern constraint could be checked
via GTP, it wouldn't be so hard to let automatically go GNU Go over all
the matches and reduce the list to those situations where the pattern
actually would get used. Then it would be easy to get a good impression
of the benefits and drawbacks of a pattern; this would take many more
situations into account than the regressions can, and wouldn't have the
systematic bias of the regression tests (which often fail to give examples
of new mistakes introduced by new tuning).

I don't think it would be much more work than hacking GTP into kombilo.


Arend