Re: computer-go: Learning from existing games

[Markus Enzenberger <compgo@xxxxxxxxxxxxxxxxxxxxx>]

On Friday 17 January 2003 06:21, schraudo@xxxxxxxxxxxxxxxxx wrote:

> I always thought your way of dynamically wiring up the network was
> very cool, and saw it as a major reason for your good results.  So it
> comes as quite a surprise to me that you're abandoning that technique
> - would you mind telling us why the change of heart?  Was it just too
> much hassle to code, or did you find something fundamentally wrong
> with it?

the new architecture is even more complicated.
The old network had relations between points as hardcoded features,
like an additional input. The new idea was, that the most important 
relation between two points on the board is if white or black can 
probably build a connection between them. If you assume local tactical 
subgames are more or less independent of each other, this relation is 
transitive.

So there should be two advantages: the network can learn relations
(i.e. connectivity patterns) and the propagation in the last step
that uses a pathfinding algorithm would save a lot of layers that
would be necessary for propagation of correlations along the board.

So far, the new architecture is better than the old one with the same 
input features, but only slightly. It is slower and less stable during 
the training, but I think, it has more potential for the future.

- Markus