[Computer-go] Representing Komi for neural network

[Álvaro Begué]

On Fri, Mar 20, 2015 at 8:24 PM, Hugh Perkins <hughperkins at gmail.com> wrote:

> On 1/12/15, Álvaro Begué <alvaro.begue at gmail.com> wrote:
> > A CNN that starts with a board and returns a single number will typically
> > have a few fully-connected layers at the end. You could make the komi an
> > extra input in the first one of those layers, or perhaps in each of them.
>
> That's an interesting idea.  But then, the komi wont really
> participate in the hierarchical representation we are hoping that the
> network will build, that I suppose we are hoping is the key to
> obtaining human-comparable results?
>

I don't see why komi needs to participate in the hierarchical
representation at all. The representation is supposed to learn higher-level
notions like good shape, life and death, territory... The effect of komi
can easily be incorporated into the mix at a later stage, since it has no
bearing on what's good shape, what's alive or dead or what constitutes
territory.



On Fri, Mar 20, 2015 at 8:24 PM, Hugh Perkins <hughperkins at gmail.com> wrote:

> On 1/12/15, Álvaro Begué <alvaro.begue at gmail.com> wrote:
> > A CNN that starts with a board and returns a single number will typically
> > have a few fully-connected layers at the end. You could make the komi an
> > extra input in the first one of those layers, or perhaps in each of them.
>
> That's an interesting idea.  But then, the komi wont really
> participate in the hierarchical representation we are hoping that the
> network will build, that I suppose we are hoping is the key to
> obtaining human-comparable results?
>
> But on the other hand, in the general case, where we want to give a
> variety of inputs to the computer, eg a map, and an x/y position, has
> anyone come up with a clean, effective way of combining these inputs
> into the net?  I dont recall seeing any such attempt/paper?
> - if we feed the map into a conv net, and the x/y pos into the fc
> layers, it seems like the x/y pos wont really participate in any
> hierarchical representation?
> - if we have 100 conv input planes for each possible value of x, and
> another 100 for each possible value of y, seems like overkill ... ?
> - feeding reals into neural nets, which have layered activation
> functions, empirically doesnt work well, and logically doesnt sound
> like it should work that well
> - contemplating just feeding them in as visual representations of the
> number, printed each on a single plane :-D
>
> Are there some papers/research/approaches in the area of combining
> non-image inputs into convnets, in such a way that the non-image
> inputs participate in the hierarchical structure, and at the same
> without creating hundreds of input planes, for each single natural
> input, which planes might contain only 5-10 bits of actual
> information?
> _______________________________________________
> Computer-go mailing list
> Computer-go at computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://computer-go.org/pipermail/computer-go/attachments/20150320/0c61b494/attachment.html>