# Initializing structural expectation maximization for learning Bayes net structures

I am using bnlearn in R to learn Bayesian network structures. It has a structural.em method for learning with missing data that works pretty much like any other EM - you specify a starting network, it computes a complete dataset (E-step) and then learns a network to maximize likelihood of model given the complete data (M-step) and so on. I am wondering what initialization I should choose for it. If I choose an empty graph, I get a consistent error about the evidence having zero probability (I'm guessing the model has a very low likelihood). Any ideas?