Approximate Bayesian network formulation for the rapid loss assessment of real-world infrastructure systems

This paper proposes to learn an approximate Bayesian network (BN) model from Monte-Carlo simulations of an infrastructure system exposed to seismic hazard. Exploiting preliminary physical simulations has the twofold benefit of building a drastically simplified BN and of predicting complex system performance metrics. While the approximate BN cannot yield exact probabilities for predictive analyses, its use in backward analyses based on evidenced variables yields promising results as a decision support tool for post-earthquake rapid response. Only a reduced set of infrastructure components, whose importance is ranked through a random forest algorithm, is selected to predict the performance of the system. Further, owing to the higher importance of evidenced nodes, the ranking method is enhanced with a recursive evidence-driven BN-building algorithm, which iteratively inserts evidenced components into the subset identified by the random forest algorithm. This approach is applied to a French road network, where only 5 to 10 components out of 58 are kept to estimate the distribution of system performance metrics that are based on traffic flow. Sensitivity studies on the number of selected components, the number of off-line simulation runs and the discretization of variables reveal that the reduced BN applied to this specific example generates trustworthy estimates.