Heuristics for optimizing ν-SVM? Do you know any good heuristics for finding optimal value of ν in case of ν-SVM classification? In this particular problem I have a radial basis kernel, if it helps.
 A: For optimisation, you don't need to perform a grid search; a Nelder-Mead simplex(fminsearch in MATLAB) approach is just as effective and generally much faster, especially if you have a lot of hyper-parameters to tune.  Alternatively you can use gradient descent optimisation - if your implementation doesn't provide gradient information, you can always estimate it by finite differences (as fminunc in MATLAB does).
The Span bound is a good criterion to optimise, as it is fast, but good old cross-validation is hard to beat (but use a continuous statistic such as the squared hinge loss).
HTH
n.b. nu needs to lie in [0,1], however this is not a problem, just re-parameterise as theta = logit(nu), and then optimise theta instead of nu.  You can then use more or less any numerical optimisation technique you like, e.g. Nelder-Mead simplex, gradient descent, local search, genetic algorithms...
A: There are no general heuristics, you should make a grid search, especially since the value of nu must be between 0-1.
