A note on learning Bayesian networks

When dealing with Bayesian network learning, most of the authors concentrate their attention mainly to the problem of proper structure (acyclic directed graph) selection. In contrast to this, the present paper studies the possibility to fully exploit the information connected with conditional probability distributions estimates. It is shown that the commonly accepted approach does not comply with the maximum information principle. The paper concludes with a proposal of a new approach, which utilizes this information better.