Approximate Bayesian computation?

[juliohm]

Is it possible to perform ABC with Turing.jl?

[yebai]

At the moment, ABC is not directly supported yet. However, it should be easy to add ABC give that we have other simulation-based inference algorithms including IS, SMC etc. Please have a look at the importance sampling code below and see whether it can help you.

https://github.com/yebai/Turing.jl/blob/master/src/samplers/is.jl

[ChrisRackauckas]

Yeah, this would be nice to have. Does it require anything different about the likelihood? I am quite unfamiliar with it but hear it's good for the parameter estimation stuff we've been doing with DiffEqBayes.jl.

[juliohm]

As a follow up, a package came out with ABC: https://github.com/marcjwilliams1/ApproxBayes.jl

It would be nice to have it in Turing.jl, which is more general than just ABC.

[yebai]

@juliohm Thanks for the pointer. It looks interesting. Would be good if someone could give it a try to use ApproxBayes.jl as a backend of Turing.jl. I'm happy to help with along the way.

PS. The coroutine-based SMC implementation in Turing is the fastest SMC at the moment. I think coroutines (copying) from Turing might help ABC as well if the simulator involves complex stochastic control flows.

[yebai]

ABC should be implemented as a standalone package supporting the AbstractMCMC interface.

[fipelle]

@juliohm @yebai I am actually working on that and I hope to publish a short proof of concept soon (relevant discussion). Long story short, I am defining a faux distribution to handle summary statistics rather than the usual log-likelihood. It seems to be working fine with simple problems and ABC algorithms. However, I am still trying to figure out how to implement more advanced ABC algorithms on the basis of SMC (relevant issue). Any help on the latter would be very appreciated!