Seth Axen

Are there benchmarks showing that this is indeed a performance improvement in R? I tried something similar for PSIS.jl and found that in Julia at least it was always slower....

> You're sorting from the tail length to the end of the array (accidentally sorting everything _but_ the tail). Good catch! Yes, I'm seeing the same performance improvement you're seeing...

Why prefer ESS of 0 over NaN?

NA seems better, since IIUC, it indicates the value is unknown or unavailable.

SlicStan does this for Stan, and I imagine it may not be that hard for Soss to support automatic marginalization for the same classes of models. See https://arxiv.org/pdf/2010.11887.pdf.

I suspect many users of PPLs are unaware that if they marginalize out their discrete parameters `z` to sample their continuous parameters `θ`, they can still recover samples of `z`...

Something like this procedure would I think do it. I believe it only requires that `z` have finite support; then the marginals always take the form of a mixture model,...

> Thanks @sethaxen , the details here are really helpful. FWIW, the details are mine. I haven't fully processed the SlicStan paper yet; it's possible they do something...slicker. > I...

Thanks @baggepinnen, the fix resolved that issue! I'm now seeing the error you posted: ```julia julia> using Soss, Distributions [ Info: Precompiling Soss [8ce77f84-9b61-11e8-39ff-d17a774bf41c] julia> m = @model begin λ...

> Yeah, this is really tricky. Long-term, I think most MCMC should be iterable. Yes I don't currently see much of an issue with this in Julia. One can always...