[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [computer-go] has neural network been tried on this?
Arthur,
cleaning up my mailbox I noticed that in the ensuing discussions about the
(f)utility of NNs, nobody seems to have answered your original question:
> I've seen contributions about neural networks being trained to play go.
> With a little success, but only a little, I gather. I wonder has anybody
> demonstrated the ability of NNs to learn a task of the same sort of
> complexity as the following:
>
> 1)
> Imagine a single white block in contact with the bottom edge, fringed by
> black stones at its interface to other parts of the board. Does that white
> block have two or more eyes? (Not "could it make" but "does it now have")?
>
> ...
In the late 80s there was some work on learning topological properties
such as connectedness, holes, etc. with NNs. I don't remember the
specifics, but the upshot was that while NNs can learn these (*), they
don't learn them easily. The reason is that topological properties
are invariants that hold across an infinite variety of shapes (e.g.,
all shapes with two holes), while NNs tend to look at particular shapes.
(Technically, it's a question of separability in the input space.)
There has been some work on special NN architectures for learning
invariants, but since it's easy to precompute topological features such
as blocks, eyes, liberties, etc., you're much better off doing that,
then giving them as inputs to your NN, a la NeuroGo.
Regards,
- nic
(*) NNs can in principle learn *anything* - they're universal approximators.
The operative question is always whether they can learn a given task with
the given resources (hidden units, time, training data).
--
Dr. Nicol N. Schraudolph http://n.schraudolph.org/
Steinwiesstr. 32 mobile: +41-76-585-3877
CH-8032 Zurich, Switzerland tel: -1-251-3661
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://computer-go.org/mailman/listinfo/computer-go