computer-go: SVM go

[Joshua Berryman <joshuab@xxxxxxxxxxxx>]

I've just finished a quick summer project to test SVMs for go. I used
gnugo to evaluate the scores of randomly generated game states, then
trained an SVM to guess the scores of previously unseen states, hence an
evaluation function, hence play.

I used an RBF kernel, representing the board as a vector of 1s -1s and 0s.

After about 20k states, the RAM of the workstation started to fill up in
the training stage: evaluation was still fast. Play was poor, it never
beat gnugo but it would often make good openings, and sometimes hold one
or two corners to the end.

The curious might want to try Edinburgh's online dissertation archive in a
few days for the full write up; I'm leaving the college now so tata.


Joshua Berryman

PS the point was that SVMs can give a compact representation of the value
hypersurface with arbitrary accuracy (in the case of the RBF kernel).
It didn't work because the training time+space is polynomial in the number
of examples, and gets pretty big. Could be parallelised (is that a word?)
or dealt with incrementally to avoid the RAM explosion.