RE: [computer-go] Neural Nets

["Mark Boon" <tesuji@xxxxxxxxx>]

> -----Original Message-----
> From: Imran Ghory [mailto:imran@xxxxxxxxxxxxxxxxx]
> Sent: Saturday, December 04, 2004 13:13
> To: tesujisoftware@xxxxxxxxxxxxxxxxx; computer-go
> Subject: RE: [computer-go] Neural Nets
>
>
> On Sat, 4 Dec 2004, Mark Boon wrote:
> > >
> > > http://www.cs.bris.ac.uk/Publications/pub_info.jsp?id=2000100
> > >
> > > Particularly section 9 might interest you.
> > >
> >
> > I get 'Connection refused' when I click that. I'm really interested, as
> > feature detection to produce a score is probably what you'd
> need for aji.
> > Not an absolute score of course, but more like a score adjustment. I'm
> > afraid that neural nets are too slow though, any information on that?
>
> It looks like the server is just down at the moment, you can grab a copy
> from the citeseer cache (at http://citeseer.ist.psu.edu/673480.html) if
> it's still down later.
>
> My paper covers more of the applications of TD learning to board games,
> for what your interested in I'd recommend you have a look at some of
> donald beal's papers. I've probably got one of his papers on my computer
> somewhere, if you want it let me know and I'll email it to you directly.
>
> As for speed it really depends on how you're using it, if you want speed
> you might be able to get away with using a one layer neural network as all
> that is is essentially an array of weightings.
>
> So for example if you had say 100 feature detectors which populate an
> array results[100], you'ld then have an array (of floats) weigh[100]
> which will be populated by a neural net training algorithm. When you want
> to use it all you'ld have to calculate is the sum of results[i]*weight[i]
> for all i. It's common to then pass this value through a normalizing
> function like (1/(1+e^-x)) but that's not neccesary.

I see... That may be too simplistic.

The thing with aji is that the occurence of one feature would not
necessarily influence the score at all. Two would maybe mean a small
adjustment, where three could mean a considerable adjustment. All depending
on which features are involved and their relative location. Similar with
weak groups.

I'm thinking more in the lines of determining an area of interest, look for
certain features, and adjust the score based on their combined occurence in
the area and their relative position or distance. Basically it becomes a
kind of pattern-matcher where the features are the input, instead of the
board-data, and you'd want to allow for imperfect matches to still influence
the score.

I'll try the link again later...


_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/