forum.alglib.net

Hi!

I've been looking all over the internet for a QP solver, and then I found ALGLIB.
I need to maximize the following function (I'm implementing an SVM):

sum(i from 1 to L) a_i - 0.5*a'*H*a

a being a vector of length L, a_i the elements of a and H an L*L symmetric matrix.

The constraints:

0 <= a_i <= C where C is a predetermined constant

and

sum(i from 1 to L) a_i*y_i = 0

y_i are the classes of each of the L training examples i have provided the SVM, and y_i takes on the values of either 1 och -1.

It is this last constraint that I am uncertian on how to implement using the minasa subpackage. Is it even possible?

It is impossible for now (unless you try to play with penalty functions). But it is possible that next release of ALGLIB will include optimizer for a combination of bound and linear constraints.