[Computer-go] Neural Net move prediction

Brian Sheppard sheppardco at aol.com
Thu Feb 4 13:19:47 PST 2016

The method is not likely to work, since the goal of NN training s to reduce the residual error to a random function of the NN inputs. If the NN succeeds at this, then there will be no signal to train against. If the NN fails, then it could be because the NN is not large enough, or because there were aspects of training that were set up incorrectly.


What this comes down to: my experience is that you would be better off making a single network that is twice as large.


This feedback applies strictly to the exact method being proposed. It is very likely that there are other ways to use multiple networks that would be significant improvements over using a single network for the whole space.


From: Computer-go [mailto:computer-go-bounces at computer-go.org] On Behalf Of Huazuo Gao
Sent: Thursday, February 4, 2016 7:51 AM
To: computer-go at computer-go.org
Subject: Re: [Computer-go] Neural Net move prediction


Sounds like some kind of boosting, I suppose?


On Thu, Feb 4, 2016 at 7:52 PM Marc Landgraf <mahrgell87 at gmail.com <mailto:mahrgell87 at gmail.com> > wrote:


lately a friend and me wondered about the following idea.

Let's assume you have a reasonably strong move prediction DCNN. What
happens if you now train a second net on the same database.
When training the first net, you tried to maximize the judgement value
of the expert move. But for the second net you now try to maximize the
maximum of the judgement of both nets. This means, that the second net
does not profit from finding moves the first net can easily find, but
instead will try to fill in the weaknesses of the first net.
In practical application the easy static usage would be to first
expand the top2 candidates of the first net, then mix in the top
candidate of the second net, then again the next 2 candidates from the
first net, etc.

What do you guys think about that?

Cheers, Marc
Computer-go mailing list
Computer-go at computer-go.org <mailto:Computer-go at computer-go.org> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://computer-go.org/pipermail/computer-go/attachments/20160204/bff48fe5/attachment.html>

More information about the Computer-go mailing list