(mis)Using Bayesian networks for dynamic risk assessment

Abstract

Abstract Bayesian network (BN) and dynamic Bayesian network (DBN) are extensively utilized in modeling and risk assessment of dynamic systems and events. Considering the tremendous work devoted to BN and DBN in the field of system safety and reliability assessment during the past decade, in this chapter we tried not to bore the readers by replicating the commonplace applications of BN and DBN but rather illustrate some misapplications of BN and DBN to dynamic risk assessment. Particularly, it will be discussed that conventional BN is not an appropriate technique for dynamic risk assessment, and more importantly, its probability updating feature is quite often misused for this purpose. On the other hand, probability adapting (also known as sequential learning), which has received less attention than probability updating, is demonstrated to be more relevant to dynamic risk assessment when it comes to the application of conventional BN. Furthermore, DBN is illustrated to be an intrinsically appropriate technique for dynamic risk assessment thought its modeling features could easily be misused by negligent modelers. Some risk importance measures and how they can be calculated using BN and DBN are also discussed. It will be illustrated that Birnbaum importance measure, compared to others, is more sensitive to probability changes, making it a more suitable metric to dynamic risk assessment.