revisit convergence criteria

[pavanramkumar]

current convergence criterion based on norm(grad) is not robust to dataset size. we need to make it a function of:

• rate of change in loss,
• norm(grad)/norm(beta) and
• n_features

See e.g.:

• https://docs.scipy.org/doc/scipy/reference/optimize.minimize-lbfgsb.html
• https://cran.r-project.org/web/packages/mize/vignettes/convergence.html

[jasmainak]

I think this is finished in #327