Models with problematic seeds

[Daniel Lee]

The last Jenkins model build failure was caused by a bad seed. I'll
post the problematic models with seeds as we come across them. They
might indicate a misspecified model, a bug, or some other issue.

We're having trouble with this model:
models/bugs_examples/vol2/beetles/beetles_probit
with seed:
1073571327

the sampled beta value is problematic.

[Daniel Lee]

We're having issues with the dogs example too:

models/bugs_examples/vol1/dogs
seed = 1121078627 (randomly generated)

[Jiqiang Guo]

That is good to know.  I hope this is the reason for a sporadic segfault I got in RStan when
running the dogs example though not necessarily.  I am trying to figure what causes the segfault for days. 

Jiqiang 

[Bob Carpenter]

On 6/30/12 10:06 AM, Daniel Lee wrote:
> On Friday, June 29, 2012 11:55:25 AM UTC-4, Daniel Lee wrote:
>
> The last Jenkins model build failure was caused by a bad seed. I'll
> post the problematic models with seeds as we come across them. They
> might indicate a misspecified model, a bug, or some other issue.
>
> We're having trouble with this model:
> models/bugs_examples/vol2/beetles/beetles_probit
> with seed:
> 1073571327
>
> the sampled beta value is problematic.

What I'm seeing is that with that seed, the beta value
stabilizes and then never moves.

I can only imagine its mass was adapted down to a small value.

I'll add instrumentation to our csv output for the
mass matrix.

At the same time, I'll add output for the adapted
step size. Then, if step size variation is turned on,
I'll also output the actual step size used for each
iteration.

- Bob

[Daniel Lee]

Here's another:
models/bugs_examples/vol2/orange/orange
seed = 1135909277 (randomly generated)
param 27, has 0 variance. That's theta_mu[3] (1 indexed)

[Daniel Lee]

vol1/dogs, seed=1985516974

By the way, I am seeing a seg fault with some of the models when I
shift parameters to transformed parameters.

[Bob Carpenter]

On 7/10/12 12:42 AM, Daniel Lee wrote:
> vol1/dogs, seed=1985516974
>
> By the way, I am seeing a seg fault with some of the models

Stan should catch problems before segfaults, so there's a problem
even if the models aren't right. Can you recreate a seg fault?
If so, let's track this down tomorrow.

If Stan segfaults, I believe it'll crash R around it.

> when I
> shift parameters to transformed parameters.

I don't see how you could shift parameters to transformed
params. If one parameter was a transform of the other,
the original sampler shouldn't be able to move. If not,
you shouldn't be able to shift to transformed.

I'd guess the problem is in initialization of transformed params.

- Bob