Relative risk analysis using Bayesian networks and evidential reasoning

Bayesian networks provide a unified and consistent framework for analysing and expressing risks and thus, have been widely applied to safety and reliability studies. Yet, most of them focus on using the advances of Bayesian theorem and posterior probabilities to risk prediction and diagnosis (forward and backward inference) and assume that the risk related prior probabilities could be easily obtained from subjective expert judgements if the associated objective historical failure statistics is incomplete or unavailable, although in many circumstances this is not the realistic case. This paper, therefore, discusses and deals with some of the practical challenges of implementing Bayesian reasoning in relative risk analysis (from a Bayesian view), which is corresponding with those positivism risk analysis from a classical perspective, including risk Tanking using the sensitivity analysis of Bayesian networks. It emphasizes the introduction of a novel "Noisier or" approach on the basis of an evidential reasoning algorithm for obtaining the Bayesian prior probability distributions conditioned on multi-state parents. Consequently, analysts can assign subjective probabilities with single condition and synthesise them using the evidential reasoning algorithm (and its attached computing software-IDS) without adopting the somewhat mathematically sophisticated procedure of specifying prior distributions with multiple ones. An example related to the terrorism threats in container supply chains is presented to illustrate the proposed ideas.