Exhaustive Hamiltonian Monte Carlo

[sammosummo]

I'm just bringing Exhaustive Hamiltonian Monte Carlo (XMC), as detailed here, to attention of the PyMC3 devs. Seems a fair bit more efficient than NUTS, and could be a good addition to PyMC3!

[twiecki]

That looks really neat, thanks for sharing. I wonder if they are using that in STAN.

[fonnesbeck]

Apparently they are but have not yet exposed it.

[twiecki]

stan-dev/stan#1859 is the PR.

[twiecki]

Took a look at the paper, doesn't seem like the numerical experiments suggest that it's really better than NUTS (only in select cases). Thus, I'm not sure we should jump on this right now as the upside does not seem all that big. But maybe I misread that, @betanalpha?

[betanalpha]

XHMC introduces an additional tuning parameter and an automated
procedure for setting that parameter is still an active topic of research.

On Nov 6, 2016, at 5:46 AM, Thomas Wiecki notifications@github.com wrote:

Took a look at the paper, doesn't seem like the numerical experiments suggest that it's really better than NUTS (only in select cases). Thus, I'm not sure we should jump on this right now as the upside does not seem all that big. But maybe I misread that, @betanalpha?

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or mute the thread.