Discrete Variables

[rlouf]

Discrete Variables

We need to settle on an API design before the first release as it could affect the internals of inference quite a bit. There are several ways to go about sampling discrete variables:

• Gibbs Sampling
• discontinuous HMC
• find others

Of all these algorithms, Gibbs is probably the most challenging to implement. What would it look like?

kernel = Gibbs(
	x = HMC(100),
	z = HMC(100),
	y = RWMH(BinaryProposal())
)

Or to avoid repetitiions

kernel = Gibbs({
	HMC(100): ['x', 'z'],
	RWMH(BinaryProposal()): 'y'
})

Using the kernels as keys is slightly weird, but we do not have a choice since lists are non-hashable.