Re: computer-go: Neural Nets: suggesting and evaluating

["Nicol N. Schraudolph" <compgo@xxxxxxxxxxxxxxx>]

> I've mentioned it to people before, and it has always seemed obvious to me 
> as well. However training times increase dramatically because of time spent 
> doing tactical reading, and *then* training times increase dramatically 
> because you are presenting something more complicated to the net and it is 
> harder to learn it.

I seriously doubt your second point.  Mapping a good tactical reading into
(say) a positon value is clearly much easier than doing it from a raw board
position, so training should be faster if anything.  If you train from
recorded games, you can do the tactical analysis for each position just
once and store the results, which should make this quite affordable.

It is generally true that machine learning can only do so much, so it
always pays to do as much preprocessing as you can.  This is after all
what we do - in fact it is not that computers play go worse than other
games, it is just that humans play it better, because go ties into all
the preceptual preprocessing machinery we have in our heads.

Best,

- Nici.

-- 
    Dr. Nicol N. Schraudolph                 http://n.schraudolph.org/
    Steinwiesstr. 32                         mobile:  +41-76-585-3877
    CH-8032 Zurich, Switzerland                 tel:      -1-251-3661