[Pachi] Pachi became strongest Go program on desktop systems. Are there any prospects?

Petr Baudis pasky at ucw.cz
Fri Aug 8 17:53:30 CEST 2014


  Hi!

  (I'm keeping pachi@ in Cc.)

On Fri, Aug 08, 2014 at 06:28:06PM +0400, Andrew P. wrote:
> Of course, these results aren’t very significant statistically, but how You can see at that forum earlier Pachi lost to the strongest at that day desktop Go program- MoGo more often, and now benefits against the MoGo more often.
> In that games You sparring newer Pachi versions with older versions or with other programs too (and may be in this case with such strongest on mobile platforms programs as Zenith and other)?
> What time and hardware settings You use in that games?

  I actually never pit Pachi against MoGo. :-)  That didn't occur to me
since I thought Pachi is already far stronger than Mogo; it surprises me
that apparently on lower-end (2 core) configurations it actually is
pretty close, that's an interesting datapoint.

  I typically test on 15x15 with 500s S.D.  This is a compromise with
regards to many factors that I gradually reached.  My primary sparring
partner during testing is GNUGo with massive reverse komi (say, 50, on
15x15) because GNUGo consumes almost no CPU during play; for more
serious testing, I also use Fuego.

  I regard Fuego as Pachi's most serious competitor since it's the other
strongest open source program.  Testing against commercial programs or
mobile programs isn't really technically feasible; I'd need a Linux
version, I can't compile it myself, some uneasy feelings about looking
at their tricks might arise and my results wouldn't be reproducible
by others.

> For analysis while You get fixes and strength enhancements it may be helpful to see played games from here: ..snip.. (most of them played with 2 minutes per move, 2 cores and 8 GB of memory for each program).
> You can compare how stronger became Pachi, for example, by sparring its versions with MoGo4.86 (./mogo --19 --time 120 --pondering 1 --nbThreads 2 --collectorLimitTreeSize 800000 --loadOB totalOB.ob). For example also, in these games Pachi compiled with: DOUBLE=1 in Makefile; spatial_hash_bits 30  in patternsp.h and used Patterns: gogod-handikgspachi-iter; was launched with: ./pachi -f book.dat -t 120 threads=2,max_tree_size=4096 (may be there are more strong launch variant for this amount of time, cores and memory?).

  Hmm, any particular reason for setting DOUBLE=1 ?  I think that will
just slow things down, it should make sense only for huge numbers of
playouts.

> Can You collaborate with GoGUI developers and together with them to address to developers of this Top Ten Distributions: http://distrowatch.com/dwres.php?resource=major (for example, in used by me OpenSUSE distribution there are 2 programs and some desks of chess- why there may not be 1 program and desk of Go)?

  Well, I'm afraid that's simply out of my scope of work; I'd rather
work on solving algorithmic Computer Go challenges.  By my previous
email I tried to imply that *anyone* could start working on this, even
without serious programming knowledge.  Maybe including you?  :)

  Kind regards,

				Petr Baudis


More information about the Pachi mailing list