Learning Bayesian Network Structures from Small Datasets using Simulated Annealing and Bayesian Score

This paper proposes a new technique for learning the structure of Bayesian networks from data. The algorithm is based on the search and score approach. Simulated annealing is used as search method, and Bayesian score as a measure of goodness. This algorithm is not new; however, we are proposing two new important steps. First, we exploit a classical resampling strategy to restrict the selection of parents of a given node during the search phase. This step avoids significant computation in similar approaches. Second, a refining step to prune erroneously added arcs is considered at the end phase of the algorithm. These ideas were tested with the well-known ALARM network. We found an improvement for small datasets on the number of correct and wrong arcs discovered.