hello all,
I am searching literatures on parameter learning for BNs with completely
unobservable nodes. A simple example is H->X->Y->Z. We are given the
BN structure and a collection of training data {<h,z>}. Since X and Y
are completely unobservable, in the data set none of X and Y 's values
are provided. The goal is to learn the CPTs of the BN.
I am interested in comments or papers in two directions. First, it is
known that EM and gradient descent approaches can solve the
problem. Are there papers theoretically or empirically comparing the
two approaches? I noted that Salakhutdinov's ICML93 showed a
relationship between the two approaches. Second, it is also known that
both approaches find local optimality. Are there any proposals
attempting to escape from the local maxima ?
I will compile the responses and send back to this list.
Many thanks,
Weihong
