Re: A high dimensional example to compare rotational NUTS with default stan

[Bob Carpenter]

We need to understand the full mass matrix adaptation for NUTS
even after we add RM-HMC, because it'll be MUCH faster to run.
We only need to do one estimation and matrix inversion rather than
inverting matrices and calculating third-order derivatives on each
sub-leapfrog step!

Any of the free parameters should be added as command-line options.
For default settings, I think it makes most sense to consider
fully standardized inputs, but we should also talk about this
at the meeting.

- Bob


On 1/2/13 12:08 AM, Yuanjun Gao wrote:
> Dear all,
>
> Happy new year!
>
> I tried to use a correlated multivariate normal distribution to compare the performance of rotational NUTS(both preset the matrix or estimate in the warm-up phase) with default stan. Simulation results shows that default stan does suffer for high correlated parameters while the non-diagonal matrix does provide better results and scales well(even for dimension like 200).
>
> Optimal procedure for estimating a covariance matrix may need more discussion since there are several parameters(like shrinkage parameter and time point for doing the estimation) that are set quite arbitrarily. I am not sure if this is still worth researching since it seems that RMHMC maybe the focus for next semester and hopefully it will tackle those issues.
>
> The report is attached, comments are very welcomed. Thank you very much.
>
>
>
>
>
> Best,
> Yuanjun
>

[Andrew Gelman]

Just to confirm, our next mtg will be Tues 8 Jan, 12:30pm in my stat ofc, right?
A

> --
>

[Bob Carpenter]

Yes. And we'll set up a google+ hangout.

Our last meeting was today in your office,
though only Michael B and I were there :-)
On the other hand, it was a very high productivity
to time meeting (despite our digression into
British TV) because we now have a spec for
an interface of Stan models and auto-dif that
I can implement and will let Michael do RM-HMC :-)

- Bob

[Matt Hoffman]

Hi gang,

Are we meeting?

Matt

> --
>
>