Re: [gnugo-devel] owl tuning

[Evan Berggren Daniel]

> > Did you run regression?
>
> I actually don't remember.  I certainly ran some of them.  I'll rerun them
> now.

Seems I hadn't run a full set, and results are interesting.

3 PASS, 3 FAIL.

strategy:34   FAIL

Accidental.  The patch produces equivalent results to increasing the owl
node count.  There are other problems involved here, which I think can be
addressed by some other new patterns.  In either case, gnugo thinks n15 is
an attack on the w dragon.

trevorb:960   FAIL

I haven't analyzed this in detail, but Gnu Go thinks D5 defends D4 and
cuts apart the B dragon.  Not a good result, I think.

13x13:73      PASS

Properly solved.  This was the goal in writing the patch.

strategy4:190 PASS

Properly solved.  E14 is found as an owl defense of the B19 dragon.

strategy5:223 PASS

Properly solved.  Q7 is not counted as an attack on R8 now.

strategy5:232 FAIL

C12 is no longer counted as critical, which is a good thing.  However,
this means the move at D11 is undervalued.  After B D11, Gnu Go finds B14
as W's next move, both before and after the patch.  I think this is an
improvement in the owl reading, and I don't think Gnu Go really
understands the full value of the move either way.

Thanks

Evan Daniel