search

CDCgov / Rt-without-renewal

https://cdcgov.github.io/Rt-without-renewal/
Apache License 2.0
17 stars 3 forks source link

Test using pidgeons.jl inference methods with EpiAware models #457

Open SamuelBrand1 opened 6 days ago

SamuelBrand1 commented 6 days ago

The Pigeons.jl package exposes parallel tempering for Turing models. PT is considered an effective inference methodology for posterior distributions with complicated geometries, and the Pigeons.jl package looks very interesting.

Identifiability/complex posterior geometry problems abound in stats epidemiology so it would be good to look at pigeons.

SamuelBrand1 commented 5 days ago

I think a good test case would be: https://cdcgov.github.io/Rt-without-renewal/dev/showcase/replications/mishra-2020/

As mentioned in the vignette, there is non-identifiability between the neg bin observation dispersion and the noise of the AR(2) process for $\log R_t$. It would be interesting to see how Pigeons.jl deals with that.

seabbs commented 1 day ago

I thought the most/more interesting element was being able to run across many cores at once?

SamuelBrand1 commented 1 day ago

I don't think that the multi-core aspect is as exciting per se, its more that by having a lot of cores you can create a "communication" bridge between an easy to sample distribution and a hard to sample distribution via a large number of intermediate distributions.

However, its still MH... just from theoretical considerations I expect it to be outperformed by NUTS on a few cores for not-so-hard distributions. Happy to see some counter evidence though.

SamuelBrand1 commented 1 day ago

Full disclosure: I'm talking about parallel tempering as I learnt about it back in the day. Pigeons also has AutoMala https://proceedings.mlr.press/v238/biron-lattes24a.html I need to get to the bottom of whether that is an alternate sampler or blended into their meta-algo.