ARD and RVM regressions: Convergence

[sophiegif]

Hi,
thank you for the code, it was very easy to use and seems to work fine!

I just have a small question: I am using the regression tools, and it seems that in your first example (sinc) the algorithm stops way before the actual convergence, because the default number of iterations is 300. By forcing more iterations I obtain:
RVR(coef0=1, copy_X=True, degree=3, fit_intercept=True, gamma=1, kernel='rbf',
kernel_params=None, n_iter=10000, tol=0.001, verbose=True)
gives : Iteration: 1537, number of features in the model: 9
Algorithm converged !

Is it to be expect to have that much number of iterations? If so, do you think the approximation at 100 or 300 iterations is valid?

Thank you!!

[kroscek]

Note also that VBGMMARD is so slow for [311000,930] data dimension, even though me using MKL libraries and using minibatchkmeans for computing cluster center to be fed into mean initialization for GMM.