Usage of regularisation in FirstOrderMinimizer

[Dmitry Novikov]

Hi everyone,

From what I see in e.g. OWLQN.scala it is supposed to use FirstOrderMinimizer::adjust to apply regularisation and not FirstOrderMinimizer::adjustFunction but then if I look to FirstOrderMinimizer::infiniteIterations I see that non-adjusted values are used in convergence check:

    val (value, grad) = calculateObjective(adjustedFun, x, state.history)

    val (adjValue,adjGrad) = adjust(x,grad,value)

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

Please elaborate on why is value used here instead of adjValue? It seems to me that regularised values should be used in convergence update.

By the way this may be a duplicate of another topic of mine, I've posted it earlier today but could not find it anywhere so decided to post again, sorry if any inconvenience was introduced.

Kind regards,

Dmitry

[David Hall]

Yeah I think that's a bug. Please open an issue on github and/or a PR.

Your post got caught in moderation. It sometimes takes me a day or two to approve posters for the first time.

--
You received this message because you are subscribed to the Google Groups "Scala Breeze" group.
To unsubscribe from this group and stop receiving emails from it, send an email to scala-breeze...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/scala-breeze/0fd65261-0e6d-40f2-acf1-453e8df2b849o%40googlegroups.com.