[Computer-go] More data for DNN's?

Justin .Gilmer jmgilmer at gmail.com
Mon Jul 27 20:21:54 PDT 2015


The Maddison et al, and Clark et. al. papers,
http://arxiv.org/pdf/1412.6564v2.pdf
http://arxiv.org/abs/1412.3409,
both suffer from a lack of very high skilled games, and thus result to
using (possibly blitz) games from lower ranked players on KGS. Has there
been any effort to gather data from the Asian Go servers? Presumably
there's even more high level games available there. I don't see any mention
of this on
http://senseis.xmp.net/?GoDatabases.

Is anyone aware of other data sets available?

Also has anyone tried fooling around with training their own DNN's? I'm
curious if anyone has compared the weight tying method of Clark et al, with
the more brute force method of Maddison et al. where they generate
additional data by applying random rotations and reflections of the games.
-Justin
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://computer-go.org/pipermail/computer-go/attachments/20150727/f1c61de2/attachment.html>


More information about the Computer-go mailing list