Open dmcdougall opened 7 years ago
We already have (I believe) Gelman-Rubin
I called it "Brooks-Gelman" and yeah, it's in there, sort of. It's a pretty stupid implementation (I'm the one that did it O_O), honestly, and should be removed IMHO. Instead, I think a ConvergenceDiagnostics
or ConvergenceMonitor
(or whatever) base class would be the way to go, from which we could subclass implementations and then the user could attach it to the SIP (or sampler, or whatever level is appropriate) to monitor convergence as sampling progresses or just use it to post process convergence. Brooks-Gelman scales O(n_samples^2), so encapsulating it to have options like lagging and only recomputing every m samples would be good to do.
We already have (I believe) Gelman-Rubin. We could do with implementing these others, however: