AW: [gnugo-devel] Symmetry corrections

[Portela Fernand]

Arend wrote:
> I think I prefer the other one with fewer passes.

As you wish. Just revert the comparisons.

> While Gunnar's choice makes sense to me (the position in the database
> is _not_ arbitrary, and you could take that into account when fine-
> tuning pattern valuations),

Hmm, aren't the pattern databases simple arrays ? Let's assume D201 and
D1647 (just an example) have the same value, which one should get
prefered ? If it's the first one, then the ascending order (the one I
chose) is the right one, since D201 will have a "lower" address in
memory. Else, I'm missing sth...

> the move comparison is clearly arbitrary. You would get it the other
> way round in the same position just rotated around the center.

I'm aware of it.

> If we have such a situation, I'd rather prefer to have a FAIL in the
> regression tests so that we can keep tuning for it, than silently
> assuming we have solved the problem.

On the other side, a "lucky" pass doesn't impair Gnu Go's playing
performance. This patch won't prevent _any_ accidental pass/fail from
happening while doing tunings, it will only reduce (a lot I think) the
noise. Soon or later, some tuning will break the "lucky" passes and
care of it will then be taken (or maybe not). In the mean time, it
just won't disturb. Or am I again missing something ?

This said, I just stated my point of view and I'm not argueing. Just
do as you wish.

/nando