How to use regularisation in FirstOrderMinimizer

[Dmitry Novikov]

Hi everyone,

I have a class that implements FirstOrderMinimizer and needs to apply L1 regularisation. In my understanding the regularisation stuff should go to FirstOrderMinimizer::adjust. But then not the adjusted (regularised) value is used in convergence check but the non-adjusted one (see FirstOrderMinimizer::infiniteIterations):

    val newCInfo = convergenceCheck.update(x, grad, value, state, state.convergenceInfo)

This could've been overcome if I've used FirstOrderMinimizer::adjustFunction for regularisation but this would be out touch with e.g. AdaptiveGradientDescent where the adjust() is used.

Could you please elaborate on how to use adjust/adjustFunction?