I'm currently trying to wrap my head around evidence propagation in bayesian network (simple tree propagation) but I'm having trouble finding information about the process.

As an example, let's take this graph from a set of slides.

enter image description here

The initialization of A and B is already done, C and D are now being initialized by B sending $\pi$-messages to them.

A $\pi$-message is computed like this

$\forall b \in \Omega_B : \pi_{B \rightarrow C}(b) = \frac{P'(b)}{\lambda_{C\rightarrow B }(b)}$

with reference to the $\lambda$-message that we get from

$\forall b \in \Omega_B : \lambda_{C\rightarrow B }(b)= \displaystyle\sum_{\forall c \in \Omega_C}^{} P(c|b)* \lambda (c) $

in this step it's still simple since the $\lambda$-value of B is 1, so the $\pi$-message is passed as shown.

Here the resulting P' and $\pi$-value differs from the other nodes tho, how did that happen? In general, how is the $\pi$-value computed based on an incoming $\pi$-message?

Now we receive evidence about C , telling us that $c_2$ happens. Because of this, the values for $c_2$ are set to 1.

enter image description here

How do we get the values of the $\lambda$-message that are shown here? How are they used to adjust the $\lambda$-value of the node above it?

  • $\begingroup$ Have you read any of Judea Pearl's books or papers? Steffen Lauritzen has also made some good references. $\endgroup$ – Chill2Macht May 1 '16 at 16:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.