Metropolis: ideal balance between slow-mixing and high rejection rate?

10⁵ samples in 8 chains with metropolis results in this far-from-converged traces (again with constant ad hoc proposal distributions):
metropolisFail10e5Figure_1

I got DEMetropolis running a few trials and it looks promising. I will compare it with Metropolis with enabled tuning.

This whole thing reveals how much I don’t know, so I’m back to reading Gelman’s book and also the DEMetropolis paper.

2 Likes