I have used the Bayes Net Toolbox to build a small network, which consists of 3 nodes and is shown below.

enter image description here

Node 1 is a Bernoulli random variable, node 2 is a Gaussian random variable and node 3 is a softmax random variable with 3 possible values. The data is incomplete, so I use the EM algorithm to estimate the parameters. But the log-likelihood decreases from the beginning and stops with the last two iterations being equal.

Does anyone know the possible solution to this situation?


If you included a prior distribution, then an EM algorithm would converge to the mode of the posterior distribution. In such a case, the log likelihood may decrease, but the combined term of (log likelihood + log prior) would increase at every iteration.


Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.