Closed jlustigy closed 4 years ago
I like this idea and we'll continue to discuss it offline, but for now, I'm going to more closely look at #61 to expand, or rather implement, multiprocessing for approxposterior. Feel free to reopen if you want to take a stab at this @jlustigy!
Currently, I am getting varied training sets in
approxposterior
runs with the exact same initial conditions. I suspect that this has to do with the GP optimization (see #41 ) with random seeding, and how thefindNextPoint
method depends on the GP, and therefore so too does the training set thatapproxposterior
generates.However, this presents something of an opportunity to use
approxposterior
in parallel. Suppose we initializeN
identical approxposterior objects and run them all in parallel. Simultaneously, we train a GP surrogate on the ensemble training set and run MCMC. Theapproxposterior
"walkers" (if you will) never waste time deriving the approximate posterior using MCMC, and will constantly be helping to generate a master training set.I'm really just spitballing here. In general, it would be great to take advantage of multiple cores to build an optimal training set for the surrogate model.