ADVI for time series data

[Dustin Tran]

Begin forwarded message:

From: Dustin Tran <dt...@g.harvard.edu>

Subject: Re: ADVI for time series data

Date: August 3, 2015 at 9:44:24 PM EDT

To: Jalaj Bhandari <jb3...@columbia.edu>, stan-...@googlegroups.com

Cc: Andrew Gelman <gel...@stat.columbia.edu>, Bob Carpenter <ca...@alias-i.com>, Alp Kucukelbir <a...@cs.columbia.edu>

Thanks for the update—adding to stan-users mailing list. There’s a known issue with bad numerical gradients and log probabilities, which can lead to nonsensical results. Have you tried the experiment on the https://github.com/stan-dev/stan/pull/1545 branch? (run `git checkout feature/issue-1544-include-gradient-and-elbo-exceptions`)

Dustin

On Aug 3, 2015, at 9:35 PM, Jalaj Bhandari <jb3...@columbia.edu> wrote:

Hi,

I as trying to compare ADVI estimates for time series data using a simple AR(1) process. PFA the R code for generating simulated data and the Stan code. 

Stan gives me perfect estimates while ADVI gives me the following error, I am wondering why...:

Gradient evaluation took 0.000846 seconds

1000 iterations under these settings should take 0.846 seconds.

Adjust your expectations accordingly!

  iter       ELBO   delta_ELBO_mean   delta_ELBO_med   notes 

Exception: Exception thrown at line 20:

stan::math::normal_log: Scale parameter is -0.327583, but must be > 0!

Diagnostic information: 

Dynamic exception type: std::domain_error

std::exception::what: Exception thrown at line 20:

stan::math::normal_log: Scale parameter is -0.327583, but must be > 0!

Thanks

Jalaj

<AR.stan><AR_data_dump.R>