I wonder if CGOS-like tournaments could be distributed?
Instead of an author's single machine playing 1000 games over a long period of
time, programs would be distributed over hundreds or thousands of computers,
which would agree pairwise to play each other, enabling the testing of many
variations of a program. A little tweaking, and we could have a sort of
co-evolutionary system which searches for optimum parameter settings for that
killer version of a program.
We did something like this on a small scale at the Congress; having more
computers than programs, we started multiple instances and manually scheduled
simultaneous games; program x.1 played a.1 while x.2 played b.1, etc. An
automatic scheduler would start instances as needed, and set up multiple
matches between closely-ranked programs; it might use something like the
McMahon method, or a Swiss tournament, or whatever is determined to be an
optimal method for generating solid ratings.
To be fair, we'd have to discard results from computers which drop off the
network, and re-play that match.
Terry McIntyre <[EMAIL PROTECTED]>
We must stop dressing up the slaughter of foreigners as a great national cause.
-- Sheldon Richman
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
