Re: Pattern learning: update

[Henrik Rydberg <rydberg@xxxxxxxxxxxxxx>]

Antii,

That seems like a very nice algorithm! Both due to its
simplicity, and because of the connection to Kohonen's
work. For the record, I must say you Finnish guys have
made substantial contributions to AI in general of the
last two decades. I was very impressed by Kohonens
self-organizing maps some years back.

Now, the particular choice of metric: I interpet the operation
Corr[A,B] as the product of all the elements of A with the
elements of B. Is that what you mean? It's hard to say from
the definition with only diagonal elements shown. OK, so I
interpret the Corr() operation as a CONTAIN operation on your
logical structure, since any real position will have only one
of the values black,white,empty,limbo set at any time, and
hence the comparison (product in you language) of a real point
with your database of K patterns can only be zero or one,
depending on whether the pattern *contains* the particular
point or not. That seem's logical; if that's correctly
understood, it should mean the following:

In a simulated-annealing fashion, do the following:
* You construct an initial pattern from a real game.
* You pick the database-pattern that contains "the most"
  of your initial pattern.
* You make the database-pattern "more similar" to the initial
  pattern by a mixing process.

Is this correctly understood?

Henrik

-- 
Henrik Rydberg (http://fy.chalmers.se/~rydberg),
Department of Applied Physics, Chalmers University of Technology.