[gnugo-devel] Owl (performance) tuning

[Nando]

[although quite different, this patch should supercede nando_3_13.3b]

First some statistics about the changed patterns :

Before: 

hits    good    bad     total   perf
VA9     1287    26548   27835   4.62%
VA44    1902    69612   71514   2.66%
A107    53      1786    1839    2.88%
(these values might be a little outdated, I got them from a build a couple
patches before 3.3.13)

After:

hits    good    bad     total   perf
VA9     1307    17470   18777   6.96%
VA44    1577    35910   37487   4.21%
A107    53      1312    1365    3.88%


A small comment (which can be generalized to other cases) about why VA44
performs so badly : let's imagine the position described by VA44 has
actually a 50% chance to work and let's suppose we don't use the optics
code. Then VA44 would of course get these 50% good hits. But the optics
code probably gets almost all of the "good" cases and probably generates
very few misses (if any). As a result, VA44 matches very often against
positions where the move doesn't work.



Performance on regression tests:

* reading nodes decrease :      2.1%
* owl nodes decrease :          6.8%


Breakage : 0 PASS and 3 FAILs

A bit deceptive, but I wasn't looking for PASSes and I think that the patch
is still valuable if we consider that :

owl:46          FAIL

Well, I tried my best but I couldn't find the refutation of E1. If a dan
player could tell me, I would try again to correct the problem.

viking:1        FAIL

Here, I think that the PASS is a lucky one, based on the uncertainty of an
attack on P17. IMO, whether O15 has a chance to kill or not shouldn't matter,
specially since I believe that it actually doesn't kill, if the defender
doesn't tenuki. Also, E4 looks a bit overvalued to me.

But it's a very serious problem, since O15 is almost mandatory in such
positions. Unfortunately, JH431 doesn't match because of the additional
stone at Q14. And I guess, we don't have the right tools to generate
following from sgf files.

  ------+
  ?.....|
  ?.....|
  ?.....|
  ?.X...|
  ...X..|
  .*XO..|
  ..O...|
  ......|
  
  :8,sFU


strategy:55     FAIL

If we observe that CVS also plays C14 if the owl node limit is increased
to 1200, then we must conclude that this FAIL is accidental. The performance
gain just makes the problem visible.


/nando

- new autohelper : owl_eye_size()
- owl tuning

nando_3_14.1
Description: Text document