|
In Belief propagation the message that are passed between two nodes of a cluster graph have a specific interpretation. The message from one node to other specifies the node's belief over the sep-set/separator variables (variables that are common between the two connected nodes). This is easy to understand if the separator variables are discrete. 1) What is the interpretation when these variables are continuous ? 2) How is the messages in the continuous case handled programmatic-ally (for instance in the discrete case i.e. when the separator variables are discrete ,one can consider messages as vectors, which are exchanged between nodes) Thanks in advance |
|
I don't think there's a single continuous representation of the messages that is always sufficient, independently of the form of your distribution. If all your factors are gaussian, however, gaussian belief propagation can be used, and in that case the messages look like parametrizations of normal distributions. |
