Let us say I have a simple network defined in the following image:
(I have this confusion of what to choose, Bayesian or Markov, because my nodes may have cycles, but they would have directions as well. I am thinking of markov networks as generalization of Bayesian. This may be a different question altogether, but it should not matter much for our question at hand)
This graph is something I would call distribution X.
I am interested in
posterior P(y | X, e, h),
where (y belongs_to Y), and Y is a random variable representing the output,
for the prior, we can assume Y is Normally distributed (at least for the time being)
e is a set of external parameters.
This is some information outside of the distribution which has strong correlations to some y belongs_to Y.
h is a history of selection.
How do I plan to update the distribution:
Use MCMC to sample from X.
- For each proposal of the sample, accept or reject it based on some (step?) function/ method f(ā¦). The parameters for this method will be e, proposal (this will be some node in the network), h (history), C (some contextual information). I want to have a custom evaluation of the proposal, rather than having some logp of previous sample determine the acceptance or rejection.
- If proposal is accepted, then the probability values in the original distribution X (mentioned above) would change. (for example, if node A was proposed and accepted, the probability of moving from Z to A would increase. In next iteration, if B is proposed and accepted, the probability of moving from A to B would increase (hence some other probabilities might decrease a bit), so on and so forth)
- Over the period of time, after a burn-in time, My original distribution X would have shifted to a posterior distribution, to which I can simply pass the (e) and get the appropriate node.
Now my specific questions are: Well, how do I go about each step? I understand MCMC at theorotical level, but do not know the APIs. How do I access the proposal? Where do I write a custom acceptance/ rejection code? How do I send the proposal to this location? I do not mind updating the probabilities of the distribution myself, but would like to know whether there is an automatic way? Once I have the iterations performed, how to use model with external parameters āeā ?
I thank you for your patience because English is no my fluent language. So expressing the problem may be little longer than required.