The scalability study has a serious but correctable flaw in it. It turns out that due to a typo, level 8 and 9 for leela light are in fact level 10 and 11. In other words, there is a 2 level gap. Woops! This of course means the "real" level 8 and 9 are missing.
I have already applied a fix which involves hacking up the processing scripts to remap the mislabeled data. The correction will appear on the web page and graphs as soon as the processing scripts recycle. Unfortunately, this probably means the line is not as parallel as it was starting to look. - Don Heikki Levanto wrote: > On Mon, Apr 14, 2008 at 12:25:15PM -0400, Don Dailey wrote: > >> Gian-Carlo provided a light version of Leela, and we are extending the >> scalability study in order to answer the question, "Do light play-outs >> scale similarly to heavy play-outs?" In other words, can we expect >> to see a parallel line on the graph? >> >> http://cgos.boardspace.net/study/13/index.html >> > > > Interesting. > > By now the corner of Mogo's curve around 4/1600 must be pretty well > established. It looks like Leela has a slight corner around 6/1800. I still > stand by my suspicion that this point happens near the playing strength of > the raw playouts. So I predict that a Leela with very weak playouts will show > a slight turn of the curve somewhat lower, possibly around 1200-1300 elo. > > Would be fun to know how well these programs do with just MC playouts, > without any tree search. With some decent large number of playouts... > > > -H > > > _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
