Open ParadaCarleton opened 3 years ago
(To be clear, this would not involve making Turing a dependency for ParetoSmooth. ParetoSmooth should work just fine without Turing, and always will, since I want to be sure it can be used with Stan, Gen, and Soss.)
As a first step I'll be moving ParetoSmooth under the TuringLang org, which should also help with ensuring the package can be maintained even if I find myself unable to continue working on it.
It would be nice to connect ParetoSmooth to the Turing ecosystem of importance samplers/variational inference, so that PSIS can be used as a step in Sequential Monte Carlo or Variational Inference. For SMC, PSIS could replace the typical importance sampling step; for variational inference, the variational inference distribution should be used as a proposal distribution for PSIS, allowing for improvements to VI (PSIS is consistent, unlike VI) and providing a diagnostic to judge convergence of the VI algorithm.