Evaluation of NUTS — more comments on the paper by Hoffman and Gelman

Here is my second post on the paper by Matthew Hoffman and Andrew Gelman on “The No-U-Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo”, available from arxiv.org. In my first post, I discussed how well the two main innovations in this “NUTS'” method — ending a trajectory when a “U-Turn” is encountered, and adaptively setting the stepsize — can be expected to work. In this post, I will discuss the empirical evaluations in the NUTS paper, and report on an evaluation of my own, made possible by the authors having kindly made available the NUTS software, concluding that the paper’s claims for NUTS are somewhat overstated. The issues I discuss are also of more general interest for other evaluations of HMC.

No U-Turns for Hamiltonian Monte Carlo – comments on a paper by Hoffman and Gelman

Matthew Hoffman and Andrew Gelman recently posted a paper called “The No-U-Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo” on arxiv.org. It has been discussed on Andrew’s blog.

It’s a good paper, which addresses two big barriers to wider use of Hamiltonian Monte Carlo — the difficulties of tuning the trajectory length and tuning the stepsize to use when simulating a trajectory. The name “No-U-Turn Sampler” (NUTS) comes from their way of addressing the problem of tuning the trajectory length — repeatedly double the length of the current trajectory, until (simplifying a bit) there is a part of the trajectory that makes a “U-Turn”, heading back towards its starting point. This doubling method is clever, and (as I discuss below) one aspect of it seems useful even apart from any attempt to adaptively set the trajectory length. They also introduce a method of adapting the stepsize during the burn-in period, so as to achieve some desired acceptance probability.

However, I don’t think these are completely satisfactory ways of setting trajectory lengths and stepsizes. As I’ll discuss below, these problems are more complicated than they might at first appear.