Provide a resimulation Metropolis TransitionKernel

[axch]

Idea for usage:

kernel = tfp.mcmc.ResimulationMetropolis(
  target_log_prob_fn=...,
  proposal_dist=tfp.some.distribution,
  seed=124)

[davmre]

Do you have a reference describing what this is? Does 'resimulation' mean the same thing as an independence sampler (i.e., always propose from the given distribution ignoring the current state of the chain)?

[axch]

Yes. At least, I meant constructing valid transition kernels from independence proposals. I don't know how standard or non-standard a term that is.

[SiegeLordEx]

One somewhat hacky way of implementing this is by creating a new next_state_fn for RandomWalkMetropolis that ignores the current state. Usage would look like:

kernel = tfp.mcmc.RandomWalkMetropolis(
  target_log_prob_fn=...,
  new_state_fn=resimulation_fn(tfd.Distribution))

EDIT: Nevermind, RandomWalkMetropolis assumes symmetry around the current state, but perhaps that's a defect of RandomWalkMetropolis.

[axch]

Symmetry is a useful assumption, because it lets you use proposals whose densities you can't evaluate -- as long as they are symmetric. I think it's reasonable for these two to be different classes.

We could also add the class of asymmetric, dependent proposals with tractable densities. That looks like a TransitionKernel made out of a function that accepts the current state and emits a Distribution representing the (conditional) proposal. One would have invoke this function backwards to compute the corresponding term in the acceptance ratio.