New package: SimulationBasedInference v0.1.0

[JuliaRegistrator]

• Registering package: SimulationBasedInference
• Repository: https://github.com/bgroenks96/SimulationBasedInference.jl
• Created by: @bgroenks96
• Version: v0.1.0
• Commit: 7da58d9bf9e4b7fff4982e4a55bea1ec3611f1d8
• Reviewed by: @bgroenks96
• Reference: bgroenks96/SimulationBasedInference.jl@7da58d9#commitcomment-141036787
• Description: A flexible toolkit for simulation based inference in Julia

[github-actions]

Your new package pull request met all of the guidelines for auto-merging and is scheduled to be merged when the mandatory waiting period (3 days) has elapsed.

Since you are registering a new package, please make sure that you have read the package naming guidelines: https://pkgdocs.julialang.org/v1/creating-packages/#Package-naming-guidelines

---

If you want to prevent this pull request from being auto-merged, simply leave a comment. If you want to post a comment without blocking auto-merging, you must include the text [noblock] in your comment. You can edit blocking comments, adding [noblock] to them in order to unblock auto-merging.

[gdalle]

[noblock]
Hi @bgroenks96 and congrats on the new package! Can I ask how it aims to differentiate itself from other similar initiatives, most notably Turing and its MCMC ecosystem?

[bgroenks96]

[noblock]

Hi @gdalle,

Thanks for your interest.

I am a long time user of Turing, and I love it's simplicity and ease-of-use for many statistical modeling use cases.

However, after several years of use, I must say that Turing is woefully inadequate for use-cases where the forward model is computationally expensive. The interface is not designed in a way that makes sense for such models. For example, it lacks the ability to save and recover deterministic outputs of the model; one must call generated_quantities which requires re-running the model which is extremely wasteful in cases where doing so is non-trivial. Furthermore, it is sorely lacking in non-gradient based sampling methods, especially when it comes to parallel sampling (i.e. ensemble or particle-based algorithms, not multi-chain). SimulationBasedInference aims to fill this gap.

The closest existing package to this one is probably DataAssim but this is focused on classical data assimilation methods, i.e. filtering problems applied to model state, rather than simulation-based Bayesian inference more broadly.

I will add a README section soon to highlight these points.

EDIT: I should also mention that SimulationBasedInference actually includes an extension module for Turing as well as DynamicHMC and Gen (the latter is still WIP) which allows the prior model to be expressed using its PPL and integrates with its sampling algorithms to cover cases where Turing's samplers would be applicable.

[gdalle]

[noblock]
Thanks for the explanation! Just wanted to make sure that the package creator didn't miss the elephant in the room, which you obviously didn't :)