convergence for variational inference

[Alp Kucukelbir]

hi folks,

sorry for the silence.

i'm about to try out / implement code to monitor convergence for ADVI.

attached is a one-page pdf with my status. any/all thoughts are welcome!

i have a question: is there existing code for running window computations in stan? if not, any suggestions?

cheers,
alp

[Marcus Brubaker]

I have no idea about the running window computations but my hunch is no. 

Anyway, looks pretty reasonable to me for convergence options.  I suspect the running window will be critical and you'll probably want a looser tolerance than what BFGS uses, but I suspect the specifics of what values to use will become clear once it's implemented and you've tested it on a few problems.

Cheers,

Marcus

--
You received this message because you are subscribed to the Google Groups "stan development mailing list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Alp Kucukelbir]

thanks marcus.

agreed. a looser tolerance would make sense.

cheers,
alp

> To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.

> For more options, visit https://groups.google.com/d/optout.
>

> --
> You received this message because you are subscribed to a topic in the Google Groups "stan development mailing list" group.
> To unsubscribe from this topic, visit https://groups.google.com/d/topic/stan-dev/dqdDzMSI_UE/unsubscribe.
> To unsubscribe from this group and all its topics, send an email to stan-dev+u...@googlegroups.com.