Include stateline capabilities in Stan?

[Michael Green]

There's a distributed MCMC software called stateline (https://github.com/NICTA/stateline) which implements random walk Metropolis-Hastings with parallel tempering. It claims to improve chain mixing and to provide an adaptive proposal distribution to speed up convergence, and allows the user to factorise their likelihoods (eg. over sensors or data). While that may be great I would love to have to opportunity of using the NUTS sampling algorithm in that framework. 

The biggest benefit would be to distribute several hundreds of chains easily through a server running on a high powered cluster. Think, 100's of users submitting chains to this server architecture running chains for their simulations. That would really kick some serious ass in my optics. 

Anyone has any experience with this or heard of any plans to join Stan and Statline attempts? To me it seems like a match made in heaven. 

Best,

Michael

[Bob Carpenter]

Multiple chains is trivial. Each can run in its own process.

Parallelizing the density calculation is challenging when
you need derivatives. We haven't put any thought into it other
than to realize it's hard (not impossible, just hard).

We would be interested in doing some kind of parallel warmup that
could exploit cross-chain information, at least when doing the
kind of global mass matrix adaptation we do now.

- Bob

> --
> You received this message because you are subscribed to the Google Groups "Stan users mailing list" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to stan-users+...@googlegroups.com.
> To post to this group, send email to stan-...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.