Learning Bayesian networks by floating search methods

In this work, two novel sequential algorithms for learning Bayesian networks are proposed. The presented sequential search methods are an adaptation of a pair of algorithms proposed to feature subset selection: Sequential Forward Floating Selection and Sequential Backward Floating Selection. As far as we know, these algorithms have never been used for learning Bayesian networks. An empirical comparison among the results of the proposed algorithms and the results of two sequential algorithm (the classical B-algorithm and its extension, the B3 algorithm) is carried out over four databases from literature. The results show promising results for the floating approach to the learning Bayesian network problem