Quoting [EMAIL PROTECTED]:
Thanks, although if you're referring to
Modification of UCT with Patterns in Monte Carlo Go [Section 3.3]
http://hal.inria.fr/docs/00/12/15/16/PDF/RR-6062.pdf
I was aware of it already... the way I understood is was that
in some fixed(?) amount of simulations moves were restricted
to the local area and the rest of the simulations were done as
usually while _the objective was all the time winning the game_.
What I'm am curious about is if Monte Carlo Go/UCT has been
used for solving life and death/capturing/connection problems,
in which the objective is to determine whether the proposed
problem has a solution or not?
My old program Viking4 included a life and death analysis which was
depth first
and played sequences from the root until the status of the stones were certain
to be dead or with two eyes. This was not random since it tried moves given
priorities defined with patterns. But it did use win loss updating of
all nodes
stored in a tree very similar to what UCT does. It also used a homebrewed
equation how select move similar to what UCT does. For life and death analysis
this system was potentially quite strong, but the heavy pattern matching it
relied upon made it too slow to be of practical use in a playing program. I
also think that replacing my homebrewed stuff with UCT may make it more
efficient.
I have also been thinking of what you are thinking, and I see no reason why
MC/UCT cannot be extend to goal directed search. The question is what you are
going to use the information given by search. Is it useful or is normal MC/UCT
the most effective way to solve the problem of playing the best move in
position with multiple overlapping and conflicting local goals?
-Magnus
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
