Gian-Carlo Pascutto wrote:
Remi Coulom has done some work in this area:
http://remi.coulom.free.fr/QLR/
It sounds very interesting (v-optimal sampling). But I don't understand
it enough to implement it. Your idea sounds simpler, but the enumeration
would be a problem, for parameters with wide ranges where we don't know
where to start.
I am working on a tool for automatic parameter optimization. Right now
it is in a very experimental state. You can find it there:
http://remi.coulom.free.fr/QLR/QLR.tar.bz2
Right now only BAST is a practical method to optimize parameters with my
tool. It is UCT over a recursive binary partitioning of the parameter
space. It does not work well for high-dimensional optimization. Even for
1-dimensional problems, I expect QLR will outperform it significantly.
If you have access to Springer publications, that paper by Ken Chen
describes another method:
http://www.springerlink.com/content/759660u175x64663/
It combines UCB ideas with a complex genetic algorithm. It seems very
complicated for my taste.
It is probably more convenient to wait a little that I improve my
parameter-optimization system, and you'll have a ready-made tool.
My tool also allows to easily implement new parameter-optimization
algorithms, test them on artificial problems, before applying them to a
Go program.
Right now, only binary outcomes are supported. I will add win/loss/draw
results later.
Rémi
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
