Integration of the dominant node identification and separation model into conditional probability distribution elicitation process in Bayesian networks.

摘要

The number of conditional probability distributions (CPDs) of a variable in Bayesian networks grows exponentially with the number of its parents. In an all-binary-variable Bayesian network, the number of CPDs of a child node with n parent nodes is equal to 2n.; Conventionally, when no empirical data is available, all CPDs must be assessed by subject matter experts using a full-CPT method. Although it fully captures expert judgments, this method is extremely time-consuming and occasionally confusing, a fact which could lead to inconsistency in the CPDs and to inaccuracy in the Marginal Probability Distributions (MPDs) of the child node. Such inconsistency and inaccuracy could diminish the performance of both risk and systems failure analyses in systems engineering.; A number of approximate models, including canonical and causal independent models, were developed to improve the speed and quality of probability elicitation. By assuming either mutually independent or linear, additive relationships among; causal factors, these models require only n CPDs, instead of 2n, to be assessed by the experts. When used in appropriate situations, these models produce exceptionally accurate results, while requiring minimal time and resources. However, in many circumstances both the CPDs and MPDs generated by these models do not align with expert judgments, which result in an unsatisfactory output.; This dissertation presents a model to identify a dominant factor in families of Bayesian networks. A dominant factor is a variable that has a maximal impact on the child node yet a minimum of interactions with other parent nodes in the same family. Additionally, the developed model separates the identified node from the rest of the parents by using a modified Heckerman's Temporal Causal Independence Model. When implemented in suitable scenarios, the model, which is applicable to Bayesian networks in all problem domains, produces more accurate output, which is better aligned with expert judgments, than such of canonical models, while significantly reducing the number of CPDs assessed by experts.; Case study and simulation analyses were performed to validate and verify the model, as well as to identify the appropriate scenarios. MPDs generated by the model were tested against output from the canonical models.