Decreasing rhat for Black-Box Likelihood Model

I’ve sampled hundreds of thousands of draws with metropolis and not gotten convergence. If it’s a hard problem to sample it’s just going to be rough going with metropolis, no way around it.

Can you explain anything more about what you’re trying to do? With no other information I guess I could suggest trying to put more (very much more) informative priors than just uniform, but it’s hard to give totally unconditional advice.

1 Like