Moves diagnostics

[Peter Pauli]

Hi all,

I am using emcee in a hadron physics analysis where I try to estimate (depending on the problem) some 10-30 parameters (real and imaginary parts of complex numbers), some of them fairly correlated.

The likelihood is quite multimodal and might exhibit some real mathematical ambiguities.

I see that I have low acceptance, very long auto-correlation times and that one parameter in particular gets stuck easily.

I am currently experimenting with different moves and also wrote a few moves myself using some domain specific knowledge to explore the likelihood efficiently. Now I am wondering if there are good ways to diagnose the combinations of steps which I use.

E.g. if I use 3 different moves I would like to know if all three have low acceptance rates or if it is one in particular. Is there a way to do this efficiently? Are there recommendations on how to pick a good set of moves?

Thanks a lot,

Peter

[Dan Foreman-Mackey]

Hi Paul,

The main metric to use is the required computation per effective sample, but you can't really compute that until you've run a long chain, making the development cycle long. I don't really know of general tips for speeding this up. When I've done this I normally try to use some combination of physical insight and intuition to find something that works well enough. emcee doesn't currently track per-move stats (e.g. acceptance fraction), but I've always thought that would be a useful feature! I don't think there are any open issues or PRs on the GitHub repo, but perhaps it would be good to brainstorm some ideas.

Dan

--

You received this message because you are subscribed to the Google Groups "emcee users" group.

To unsubscribe from this group and stop receiving emails from it, send an email to emcee-users...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/emcee-users/026167ee-c721-4ae3-967a-1a049387233bn%40googlegroups.com.