[Pachi] Regression ?
pasky at ucw.cz
Sun Apr 10 03:37:44 CEST 2016
On Fri, Apr 08, 2016 at 11:20:17PM +0200, lemonsqueeze wrote:
> I'm pretty sure about the tests on my side, i couldn't believe the outcome
> at first so i went through everything i could think of and set it up again.
> But it'd be good if someone can replicate just to make sure i'm not going
> crazy =)
> All the 19x19 tests in this thread were with 180k sims, 16 threads,
> DCNN code not even compiled in (orig Makefile with DCNN=1 commented out) so
> board size shouldn't matter at all. If it's indeed a threading issue not
> sure if 2 threads will be enough to make it happen...
You are right, I can't reproduce it with two threads either:
. 385 0.317 0.024 19--45.5-1-gnugo10-pm-32e02-_300-playout=moggy,threads=2
. 437 0.32 0.022 19--45.5-1-gnugo10-pm-febab-_300-playout=moggy,threads=2
I will try with four, but that's as high I can go for now...
sthalik's struct uct_node layout changes could have had some unforeseen
threading consequences wrt. lockless stats updates, but that should have
happenned *before* the regression you reported.
> On 04/08/2016 06:37 PM, Petr Baudis wrote:
> >Well, to be fair, your DCNN changes are only enabled for 19x19.
> >I've started some tests too, my results:
> >/ 78 0.538 0.056 15--45.5-1-gnugo10-pm-32e02-_500-playout=moggy
> >/ 96 0.5 0.051 15--45.5-1-gnugo10-pm-febab-_500-playout=moggy
> >. 222 0.432 0.033 19--45.5-1-gnugo10-pm-32e02-_600-playout=moggy
> >. 208 0.361 0.033 19--45.5-1-gnugo10-pm-febab-_600-playout=moggy
> >(single-thread, against gnugo with -45.5 komi; 15x15 with -t _500, 19x19
> >with -t _600; these are my standard autoplay settings)
> >So nothing really definite came out of that. Someone really well versed
> >with statistics could probably think of paired test that'd tell us at
> >this point, but that's not me without a few hours pouring over textbooks
> >and the internet. :) I'll switch 19x19 to threads=2 as you suggest.
> >It's also possible your original results are a statistical fluke, though
> >not terribly likely. :-) (I feel silly asking again - but are you sure
> >your binaries couldn't be mixed up with some that use caffe?)
If you have good ideas, good data and fast computers,
you can do almost anything. -- Geoffrey Hinton
More information about the Pachi