I’m not sure that’s exactly right. I’m sure there are people who understand this better than me, but I think that PyMC get (in the model specification) a list of the parameters necessary to calculate the joint likelihood contributed by each node. It then runs over the list of nodes and passes each one the appropriate parameters and adds the log likelihood into the total. That is, it isn’t calculating a Markov blanket internally (if it makes sense to talk about a Markov blanket in this sense, it is being told he dependencies of each node) and it doesn’t use any such representation.
I’m curious if this is right or not.
Opher