The recent explosion of high dimensionality in datasets for several domains has posed a serious challenge to existing Bayesian network structure learning algorithms. Local search methods represent a solution in such spaces but suffer with small datasets. MMHC (Max-Min Hill-Climbing) is one of these local search algorithms where a first phase aims at identifying a possible skeleton by using some statistical association measurements and a second phase performs a greedy search restricted by this skeleton. We propose to replace the first phase, imprecise when the number of data remains relatively very small, by an application of "Perturb and Combine" framework we have already studied in density estimation by using mixtures of bagged trees.