RStan and inits in ADVI

[Bob Carpenter]

[moved from stan-users]

We should be able to set inits in the constrained space. Is
that what CmdStan does, or is that also limited to unconstrained?
The necessary transforms are all there --- they're being used
for HMC inits.

I just wanted to bring this up because it should go into the
underlying service commands that Daniel's coding.

- Bob

> On Apr 15, 2016, at 9:54 AM, Ben Goodrich <goodri...@gmail.com> wrote:
>
> On Friday, April 15, 2016 at 7:29:56 AM UTC-4, Sebastian Weber wrote:
> The vb function in rstan to run ADVI currently does not allow me to specify initials which renders ADVI useless for me in my setting (i.e. with meaningful initials which I can give in cmdstan my model just runs super fast). Is it planned to add this option in rstan 2.10? This is desperately needed.
>
> I think this already works. Except you can only set the initial mean vector to the multivariate normal distribution in the unconstrained space, not the variance-covariance matrix.
>
> Ben
>
>
> --
> You received this message because you are subscribed to the Google Groups "Stan users mailing list" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to stan-users+...@googlegroups.com.
> To post to this group, send email to stan-...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[Ben Goodrich]

On Friday, April 15, 2016 at 11:59:28 AM UTC-4, Bob Carpenter wrote:

[moved from stan-users]

We should be able to set inits in the constrained space.  Is
that what CmdStan does, or is that also limited to unconstrained?

The user can put initialize in the constrained space from R and it gets transformed to the constrained space. I wasn't clear in the original message but was just trying to say that we are missing the feature of being able to initialize the covariance matrix. This would be useful if you have a big dataset that you run NUTS on and then get incrementally more data that you would be willing to do fullrank ADVI on with informative multivariate normal priors in the unconstrained space.

Ben

[Alp Kucukelbir]

this will require change to the backend of ADVI. currently, there is no way to initialize the variance/covariance part of the variational approximation.