Wrong parameter distributions from DREAM

[emitra17]

DREAM has different final distributions than Parallel Tempering and MCMC. PT and MCMC match the result from Harmon et al 2017, but that's unsurprising given that the paper used an earlier version of the same code.

Good evaluation to confirm who's right would be to try to reproduce the distributions in Kozer 2013.

[emitra17]

Confirmed that PT/MCMC are correct and DREAM is not, by running a trivial example where we know the correct distribution.
DREAM produces a distribution that is too narrow.
Also, for box-constrained variables, DREAM undersamples near the boundary.

[emitra17]

@ryants This has been sitting here for a while. Should we just de-support DREAM?

[ryants]

Probably... :( I keep meaning to implement a simpler version (the regular DE-MC without the fancier stuff DREAM should have).

It's always something we can add in later (and we still have decent bayesian MCMC functionality without it)

Regarding de-support, I suppose we just omit the DREAM option and parameters from the main docs. No need to discard the code, and I'm sure there's a way to tag the class as non-functional so that the API docs indicate it is a work-in-progress/stagnation

[emitra17]

Sounds like a good plan re: hiding it from the main docs. I'll look into how to tag it for the API docs.

[ryants]

I'd like to keep this issue open for posterity. I'll retag it accordingly