[Computer-go] Teaching Deep Convolutional Neural Networks to Play Go

Álvaro Begué alvaro.begue at gmail.com
Mon Dec 15 13:49:57 PST 2014


On Mon, Dec 15, 2014 at 4:39 PM, Dave Dyer <ddyer at real-me.net> wrote:
>
> Combining simple metrics such as 3x3 neighborhood, position on the board,
> and proximity to previous play, you can easily get to an average score
> of 85%, without producing noticeably good play, at least without a search
> to back it up.
>


Their results don't use that metric exactly, but they quote an average rank
of 5.21 for the pro move, which probably translates to somewhere around 97%
in your metric (using 180 as a typical number of available moves).


Álvaro.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://computer-go.org/pipermail/computer-go/attachments/20141215/e7e6ed8e/attachment.html>


More information about the Computer-go mailing list