Bayesian Network Structure Learning and Its Applications

This paper discusses the purposes and methods of Bayesian network structure learning, then proposes a new algorithm for this task. Based on a fully connected potential graph, we enter the expert knowledge and prior knowledge in order to reduce the query space of the structures. By using CI (conditional independence) tests, it can be pruned a fully connected potential graph to a best PG, which is expected to approximate the undirected version of the underlying directed graph. The experimental results of fault diagnosis in automobile are provided to illustrate the feasibility and efficiency of the new algorithm.