Hamiltonian Monte Carlo in PROC MCMC SAS/STAT 14.1

[Garnett]

What experiences have folks had with ALG = HMC(stepsize=<> nsteps=<>) in PROC MCMC? In my experience, I can only get it to work with trivial problems that are just as easily run using metropolis-hastings. For example, I am fitting a hierarchical logistic regression model, and using HMC yields posteriors that are completely immobile unless the step size is extremely small (e.g. stepsize = 0.0001). Even then the posterior samples are highly localized, as one might expect. (I get convergence using the default sampler, but it takes a while and I'd rather speed things up.) Any thoughts or general advice on setting the tuning parameters for Hamiltonian Monte Carlo? 

[Garnett]

I ran ten chains with uniform(0,0.0001) randomly select stepsize and nsteps random uniform(1,100). The only chains that moved at all had step sizes less than 0.002. These chains explored the posterior, but barely. Convergence diagnostics were far from acceptable after 10,000 iterations. Still no luck using HMC for complex models. Default metropolis sampling works much better for these problem, and, for those problems where I can get HMC to work, the default sampler works just fine.

Please share any experiences that you have with HMC!