I guess neural networks is fine for learning pattern priorities for
example. There are probably just simpler and faster methods for doing
that.
Anyway a good project would be learning 3x3 patterns for MC heavy
playouts with a large number of extra features such as exact liberty
counts, distance from edge and corner etc. I guess the best use of a
neural network would be to create a mapping from a huge feature space
to a move ordering, where more straightforward pattern matching would
fail because the number of patterns are too high. The problem is then
of course that the learning of the neural network might never converge
to something useful... But I cannot judge that.
A second idea would be to focus on some basic parameters in the search
part of an MC program for example how to best mix AMAF, Winrates, and
prior pattern values. These parameters tend to be static in programs
but maybe they should be dynamic depending on stage of game,
evaluation and other things.
There is a program called NeuroGo which is quite strong. Read the papers at
http://www.cs.ualberta.ca/~emarkus/neurogo/
and get some inspiration from there.
-Magnus
Quoting Petr Baudis <pa...@ucw.cz>:
Hi!
Is there some "high-level reason" hypothesised about why there are
no successful programs using neural networks in Go?
I'd also like to ask if someone has a research tip for some
interesting Go sub-problem that could make for a nice beginner neural
networks project.
Thanks,
--
Petr "Pasky" Baudis
A lot of people have my books on their bookshelves.
That's the problem, they need to read them. -- Don Knuth
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
--
Magnus Persson
Berlin, Germany
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
