Is there a recommended sampler for a situation with black-box likelihood function for which no efficient derivative information can be extracted? In my case, PyMC3 always uses the Slice sampler for variables involved in the likelihood. I happen to have all continuous and bounded parameters, but some (most) are used in the likelihood. What other samplers could I try for comparison?
I’m a big fan of the DEMetropolisZ sampler. It doesn’t require gradients, is faster than Slice (in my experience), and requires only a few chains for efficient parameter estimation.
2 Likes