[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [computer-go] Designing faster, better influence functions



Gunnar wrote


To begin with multiplication doesn't correspond to convolution in the
discrete case but to circular convolution, so it's absolutely
necessary to pad the board, possibly with reflected stones. The amount
of padding needed depends on type of padding and the size of the
convolution kernel. Second FFT can only be done on sizes which are
composite numbers, the smaller the prime factors the better.

This second point isn't strictly true, there are fast ways of coping with
smallish primes.  I find references to a Winograd Fourier transform.  The
fact that the data is real should give a factor of 2 in time.  Using a
reflected 38 X 38 board with following symmetry

pq
bd

where 'b' is the original board would mean the data is symmetrical (sort of),
which might give another factor of 2.

However, I share your skepticism about the usefulness of a linear influence
function.
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/