MMOS+ Ordering Search Method for Bayesian Network Structure Learning and Its Application

Abstract

To address the problem of a reduced efficiency due to an increase of the search space, it has been proposed that priors could be added as constraints to the OS+ algorithm, which are Parent and children (PC) sets of each node obtained using the Max-min parent and children (MMPC) algorithm. Experimental results indicate that compared to other competitive methods, the proposed algorithm yields better solutions while maintaining high efficiency. Bayesian network (BN) sensitivity analysis is also proposed, which allows the network structure to be determined via a proposed ordering search method. We performed sensitivity analysis to determine the accuracy of the airborne avionics system, for which a simulation model is constructed to generate data samples, and the main effect of each error index is obtained using different sensitivity analysis methods. Experimental results indicate that the proposed BN method produces more accurate results when there is insufficient sample data, and this method can elucidate causal relationships that are present in the data.