Probabilistic Inference

[rafaelkaufmann]

Seriously -- WebPPL already has a JS inference implementation that is perfectly sufficient for most simple use cases, and it would make Squiggle >10x more useful.

[OAGr]

I agree inference would be useful, but it also could make things a lot more complicated if we're not careful.

WebPPL is neat, but is really optimized for inference-heavy situations.

With Squiggle, we've found that there's a lot of use to be had by regular risk analysis, without inference. It's a surprisingly uncommon complaint. Note that there is a clear history of powerful risk analysis tools that don't really focus on or do inference.

I'm nervous about how much inference setups could scale, if we tried applying inference on a set of functions that import a deep chain of other assumptions and functions.

Going forward, I do plan on eventually adding some ability to use inference in limited blocks/sections, but this might require some special syntax changes and stuff. This isn't trivial work.

If we added some functionality from WebPPL, this would introduce a great deal of complexity, if we weren't careful.

Right not I don't get the sense that funders in particular are excited about this, but we could continue to have it in mind and stuff.