dflemin3
commented
5 years ago Also, I should consider this: MCMC can also track blobs, the 2nd returned parameter(s) in emcee's LnLike functions. approxposterior should include the functionality to handle them, either by saving them or ignoring them, depending on the user's needs.
Right now, no. Approxposterior can only derive posterior distributions for forward model inputs, i.e. initial conditions. It would be awesome to be able to derive posterior distributions for model outputs, like final water content for an exoplanet subject to hydrodynamic atmospheric escape. This can't be done using the BAPE or AGP algorithms since the GP predicts the likelihood of a given set of initial conditions by training on model input-output pairs, where the output is transformed to a likelihood by comparing to observations, perhaps using a Chi^2-like metic.
What I could do is allow the user to specify some ML algorithm to predict the output of interest, like a random forest model, that trains on the forward model input-output pairs, using cross-validation to optimize hyperparameters. I assume that this approach won't generalize well since the training sets will be small by design and the forward models are typically rather complicated. This would probably result in appreciable systematic error, but it's probably worth experimenting with for a bit to see if it's feasible.