likely llc

I remember sitting in her office, with Eunjee Lee, PhD. We’re looking at a trace plot, this is a Metropolis-Hastings embedded in a Gibbs sampler, and there were multiple. We call it a “Metropolis-within-Gibbs.” So we’re looking at a trace plot. And she says zoom in, zoom in. So I zoom in. And she pauses, and stares at the trace plots for a a few seconds, and she says, “I remember having this problem.” Set your Metropolis-Hastings to 1 iteration (I had it at 2 or 3). She had used a slice-sampler, but I couldn’t get it to work, I asked someone at SAS later, who was a Bayesian, but perhaps it was due to initialization conditions. I just used the MH.

When I was in Finland, I was with Michael Riis Anderson, PhD, working on the Stan math library, trying to figure out what they call, “The Birthday Problem.” See BDA3, chapter 21, “Gaussian Process Models.” And I’m asking him questions, on our lunch break, and he says, while waving his hands as if to say, don’t feel bad, (may be I’m misquoting him), “…debugging probabilistic algorithms is like an art.”