Potentially interesting SMC algorithms

[devmotion]

I had a short conversation with Fredrik Lindsten about SMC algorithms that might be interesting to implement. Apart from PGAS, his suggestions were:

• Our divide and conquer SMC is something that I think would be very well suited for PPL. I have mentioned this to a lot of people (including Lawrence, David Broman, ...) but no one has taken the bait yet (https://arxiv.org/abs/1406.4993).
• Some type of "Twisted SMC" would also be very beneficial for PPL, but that's more of a research topic to figure out how to construct the sampler itself so that it can be implemented in a generic way.
• Last but not least, Christian's VSMC (https://arxiv.org/abs/1705.11140).

[donhausk]

Thanks a lot for these remarks!

[cscherrer]

Anglican has a nice list here:
https://www.robots.ox.ac.uk/~fwood/anglican/literature/index.html

IPMCMC seems particularly nice:
https://www.robots.ox.ac.uk/~fwood/anglican/assets/pdf/rainforth-icml-2016.pdf

[devmotion]

IPMCMC is implemented in https://www.github.com/TuringLang/AdvancedPS.jl/tree/master/src%2Fcontrib%2FAdvancedSMCExtensions.jl, but currently broken. It just has to be updated to the new interface which hopefully is quite straightforward.

Another SMC algorithm you guys might find interesting, which gives an asymptotic speedup from O(n^-1) to O(n^-2) by using quasi-Monte Carlo.

[ParadaCarleton]

Another interesting algorithm -- perhaps we could try integrating ParetoSmooth.jl, and letting users replace the importance-sampling step with PSIS? I've never seen it done before, but it definitely seems like an interesting/worthwhile experiment.