The advances of high-throughout technologies have produced huge amount of data regarding gene expressions or protein activities under various experimental conditions. The reverse-engineering of regulatory networks using these datasets is one of the top important research topics in computational biology. Although substantial efforts have been contributed to design effective inference methods, there are still a number of significant challenges to deal with the weak correlations between the observation data and the dependence of network structure on the order of variables in the systems. To address these issues, this work proposes a novel statistical approach to infer the structure of regulatory networks. Instead of using one single variable order, we generate a number of variable orders and then obtain different networks based on these orders. The weight of each edge for connecting genes/proteins is determined by the statistical measures based on the generated networks using different variable orders. Our proposed algorithm is evaluated by using the golden standard networks in Dream challenges and a cell signalling transduction pathway by using experimental data. Inference results suggest that our proposed algorithm is an effective approach for the reverse-engineering of regulatory networks with better accuracy.