Elicitation and Validation of Graphical Dependability Models

摘要

We discuss elicitation and validation of graphical dependency models of dependability assessment of complex, computer-based systems. Graphical (in)dcpendency models are network-graph representations of the assumed conditional dependences (statistical associations) of mul-tivariate probability distributions. These powerfully 'visual', yet mathematically formal, representations have been studied theoretically, and applied in varied contexts, mainly during the last 15 years. Here, we explore the application of recent Markov equivalence theory, of such graphical models, to elicitation and validation of dependability assessment expertise. We propose to represent experts' statements by the class of all Markov non-equivalent graphical models consistent with those statements. For any one of these models, we can produce alternative, but formally Markov equivalent, graphical representations. Comparing different graphical models highlights subsets of their underlying assumptions.