Stopping criterion of support vector machines

Question

I am writing my bachelor's thesis using support vector machines. Is it right, that the "epsilon"-parameter is the tolerance when optimization the error (including c)? So if the conststraint of max. margin and min. error is at least closer to its optimum than "epsilon" the optimization stops?

Answer 1

Since it seems you´re using LIBSVM, you should probably look at this technical paper by the LIBSVM authors (in particular section 3).

EDIT: yes, what you're saying is a reasonably accurate description of what the termination criteria of the LIBSVM algorithm is. When you do write your report, don't say some random dude on the internet told me, cite the original document describing the algorithm.

EDIT2: m and M are defined on page 1891 towards the end, they are basically the minimum and maximum components of the gradient of f times y.