So if I understand correctly, at each stage of the modeling, we would have a population of particles, and after the first iteration (first set of observations), each element of the population would be a trained PyMC3 model (i.e., a MultiTrace), and we would evolve the particles by posterior sampling?
I’m not sure how that would work, since PyMC3 will only sample tuples of latent values that are already seen in the trace (if I understand how PyMC3 does posterior sampling correctly). That seems like it would be very likely to fail if used repeatedly. So likely I’m not fully following the concept.