RHMC... is it self-adapting like NUTS?


#1

Hi!

I have a very simple model in mind to try out RHMC… but before diving into this I was wondering if the RHMC implementation in Stan includes the NUTS trick?

So is the RHMC implementation as user-friendly as NUTS or more like plain vanilla HMC?

I found that even with experience with NUTS it is hard to figure out stable sampling parameters for vanilla HMC (at least for me).

Sebastian


#2

Yes. See Michael’s paper on GeoNUTS (generalizes no-U-turn to Riemannian):

Also see the one on SoftAbs (stable metric that can be computed):

The ODE integrators haven’t been generalized for higher-order derivatives. And I think there’s still a tuning parameter in there somewhere—I don’t know how well adaptation has been tuned.