The aim of this paper was to present results of algorithms implementation for learning in bayesian networks from incomplete data, based on Expectation Maximization algorithm (EM). Both structure and probabilities learning problems were focused, but, differently of previous works published, details of implementation and test results are described, considering distance between distributions and topological differences. Algorithms were developed using UnBBayes System, benchmarks ALARM and ASIA were used for testing, being obtained successfully underlying networks.