The "failed attempt 1" model using a Bayesian network seemed to utterly fail to converge using PyMC. I don't have any experience with Bayesian inference, so I'm not sure whether this is a model flaw, an model implementation flaw, a problem too hard to PyMC to handle, or user error (e.g., insufficient thinning). We're applying Bayesian rule with an exponentially small posterior probability, but I would naively expect that to be standard operating procedure for a large Bayesian network.
The "failed attempt 1" model using a Bayesian network seemed to utterly fail to converge using PyMC. I don't have any experience with Bayesian inference, so I'm not sure whether this is a model flaw, an model implementation flaw, a problem too hard to PyMC to handle, or user error (e.g., insufficient thinning). We're applying Bayesian rule with an exponentially small posterior probability, but I would naively expect that to be standard operating procedure for a large Bayesian network.