Re: Learning patterns

[Antti Huima <huima@xxxxxx>]

On Tue, 8 Jun 1999, Dave Dyer wrote:

> Very interesting!
> 
> I'd like to see more information about the structure of the
> "representatives" you fed the learning program, and some references to
> the underlying QV algorithm.

The learning program is not fed with representatives, but by raw moves
from the expert games. The representatives are the return values of the
algorithm.

I don't know a canonical reference for VQ, but I studied it from the
following book, which is well-written and contains much interesting
information on self-organizing maps. They are in a sense a generalization
of VQ:

  Teuvo Kohonen: _Self-Organizing Maps_.
  Springer Series in Information Sciences, 30.
  ISBN 3-540-58600-8.
  Springer-Verlag, 1995.

And Henrik wrote:

> That's a nice piece of work! How should I interpret the large
> dont-dont-care areas - do you use very large patterns as well?

The algorithm operates internally using square matrices, but the don't
care regions are found by the algorithm itself. The problem is that if the
fraction of don't care intersections grows high, the `calibration' phase
does not work because the pattern matcher I use `explodes' when there are
too many don't care intersections. It is not a major proble, though, as
the `exactness' of the patterns is also a parameterizable value [and the
calibrating pattern matcher can be naturally enhanced].

> What strikes me though is that to me, it seems that your choice of
> metric is the important thing. Would you care to share that with us,
> or did I miss something?

I try to describe the algorithm in more detail on the same WWW page, as
you need to know also the updating part. I'll notify the list when the
description is there.

> Actually, a good metric should correspond to putting in good features
> from the beginning, right?. With good features, it should not matter
> what method you use to optimize your multi-dimensional space - NN or
> whatever.

Actually perhaps not, as there is nothing special in the beginning
situation. All the matrices are initialized identically and no specific go
knowledge is applied.

-- 
Antti Huima
SSH Communications Security Oy