[Computer-go] Pachi Authors

Petr Baudis pasky at ucw.cz
Mon Apr 18 10:55:38 PDT 2011


On Mon, Apr 18, 2011 at 05:49:11PM +0200, Jean-loup Gailly wrote:
> > One more question on the recent Pachi wins against pro players
> > in handicap games: Did you use dynamic komi; and if so, in which way?
> pachi supports several dynamic komi options, including a rather
> sophisticated
> "adaptive" method. But in distributed mode I found that the linear method
> works best. Black gives itself initially a disadvantage of 7 points per
> handicap
> stone (so 49 points in the h7 game) and this disadvantage drops linearly to
> 0
> after 200 moves. This way all machines in distributed mode compute
> the same dynamic komi.

  I'll just expand on this that the reason adaptive komi does not work
well in distributed mode is probably since there is no implementation
for synchronization of komi among nodes - so they could share results
that may well be obtained with different komi values.

  We decided that implementing infrastructure for komi synchronization
is not worth the trouble since while I measured adaptive dynkomi to be a
tangible improvement (in single-node mode), the inter-node synchronization
might be a bit messy and it is probably more economical to spend effort on
other things.

				Petr "Pasky" Baudis
UNIX is user friendly, it's just picky about who its friends are.

More information about the Computer-go mailing list