[Pachi] Multiprocessor

Petr Baudis pasky at ucw.cz
Thu Jan 5 00:17:32 CET 2012


On Fri, Dec 30, 2011 at 06:01:34PM +0100, Jean-m. a. wrote:
> Do you think it should be possible to implement an other function to
> pachi where it give you an estimate of its strenght according the
> hardware configuration and the time than you give him.  Or at least if
> not an absolute ranking a ranking compared to pachi with only ten
> second and one proc.

  If you run Pachi now, it prints a lot of output about its move search.
One of the last lines is:

	genmove in 16.52s (4852 games/s, 1213 games/s/thread)

  The important number is the "4852 games/s", this is the value that
essentially defines playing strength if you limit Pachi's moves by time.

  Short version, cookbook recipe:

  (i) Start Pachi with empty board and make it play the first move.
(Give it enough time, too!) Get the games/s number. You may want
to do this a couple of times and take an average.

  (ii) Compute this equation:

	13.3 - log(<measured games/s number>)/log(2)

(you can e.g. enter this as Google Search). Round it to get a rough
estimation of KGS kyu rank of your Pachi. (If you have a very fast
computer, around 0 is 1 dan, etc.)

  For example, on one computer, my Pachi prints:

	genmove in 12.83s (1316 games/s, 1316 games/s/thread)

Google says that

	13.3 - (log(1 316) / log(2)) = 2.93805623

therefore my Pachi could be slightly better than 3 kyu.


  Here is a short table of values for PachiIchi, PachiNi and PachiShi
that played many thousands of games in The Ranking Room of KGS:

	games/s  KGS rank
	1350	 3k (very good 3k)
	2450     2k (good 2k)
	4800     1k (average 1k)

  As you can see in the recently published Pachi paper, there is a
general strong exponential relation between # of games and Elo rating;
the increase is roughly 100 Elo per doubling, though for much weaker
Pachi version, this relation is less clear.

  It would be possible to make this more user-friendly, of course.  This
might make for a nice introduction project for developers who would like
to contribute to Pachi. :-)

  Patches are always welcome!

  Happy playing,

				Petr "Pasky" Baudis
	The goal of Computer Science is to build something that will
	last at least until we've finished building it.

More information about the Pachi mailing list