I’m also not familiar with the literature on the subject but I would like to ask the people here that work on Sequential Monte Carlo (SMC) their input (for example @aloctavodia, @junpenglao). The basic idea of SMC is to approximate the distribution with an ensemble of particles in parameter space. The positions of the particles are updated through a sequence of steps that rely on the unnormalized posterior density.
- Could a trace sampled with NUTS be used as a hot start for the SMC ensemble?
- Could the update steps take advantage of HMC for the ensemble of particles?
- If both of these were true, could this be used as a form of “update” of the posterior given new information, as what is done by particle filters?