I wander if it is foreseen to implement the following HMC improvement over NUTS developped by Hoffman et al?
Thanks
I wander if it is foreseen to implement the following HMC improvement over NUTS developped by Hoffman et al?
Thanks
Anybody interested?
Does this improvement on NUTS/HMC is not usefull for Numpyro? I’ve seen BlackJax implementing this new algo…
Just curious, would ChEES-HMC be better than NUTS?
Yes. As Hoffman states in the “Discussion” part the aticle
We have proposed ChEES-HMC, an adaptive Hamilto-
nian Monte Carlo variant that can automatically tune
its step size and trajectory-length parameters. ChEES-
HMC is competitive with the widely used NUTS algo-
rithm in terms of effective samples per gradient evalua-
tion, but it really shines when running many parallel
chains on a GPU, since its lack of control-flow over-
head lets it fully exploit these accelerators’ massive
floating-point computational resource