Reliability sensitivity analysis with value of information

Reliability assessment and robust design provide decision support for engineering systems under uncertainty. Sensitivity analysis is a key element of such assessments. In many instances, it is relevant to understand if and how further efforts should be undertaken to reduce uncertainty, e.g. by collecting additional data or by improving models. The value of information concept provides a decision-theoretic measure to quantify the benefits of such uncertainty reduction, and has therefore been proposed as a basis for sensitivity analysis. In particular, the expected value of partial perfect information (EVPPI) and the expected value of sample information (EVSI) have been proposed as sensitivity measures (Pjörn 1997; Felli & Hazen 1998; Oakley 2009). This talk focuses on the definition and computation of EVPPI for reliability assessment of systems represented through physics-based models, in contrast to (Borgonove & Cillo 2017; Fauriat & Zio 2018), which consider Boolean system models. Because the value of information can only be calculated in the context of a decision, we propose and investigate general decision contexts that are representative for a wide range of applications. On this basis, we derive expressions for the EVPPI. Thereafter, we discuss three different computational strategies that can be employed to evaluate the EVPPI for systems that are represented by numerical models. Based on the results of a reliability analysis, these strategies work without additional runs of the (potentially costly) numerical model. Concepts and implementation are demonstrated on engineering problems. Finally, we draw some conclusions on the relationship between the EVPPI and common sensitivity measures in reliability assessments.