ABSTRACT
Managing the reliability of a complex system is a problem involving many different decisions. Typically, these decisions are made on different points of time, and thus are based on unequal sets of information. In the first part of this chapter we discuss the effects these decisions have on reliability evaluations. We also suggest a general methodology for optimizing decisions under such circumstances. To illustrate the ideas we consider a simple example where a multicomponent system is inspected regularly during its lifetime. At each inspection time a decision is made whether to repair failed components or just to leave the system in its present state.
In the second part of this chapter, combining the opinions of k experts about the reliabilities of n components of a binary system is considered, the case n = 2 being treated in detail. Our work in this area generalizes results in papers by Huseby (1986 , 1988) on the single component case. Since the experts often share data, he argues that their assessments will typically be dependent and that this difficulty cannot be handled without making judgments concerning the underlying sources of information and to what extent these are available to each of the experts. In the former paper the information available to the experts is modelled as a set of observations Y 1, ..., Ym. These observations are then reconstructed as far as possible from the information provided by the experts and used as a basis for the combined judgment. This is called the retrospective approach. In the latter paper, the uncertain quantity is modelled as a future observation, Y, from the same distribution as the Yi's. This is called the predictive approach. For 108 the case n > 1, where each expert is giving opinions about more than one component, additional dependencies between the reliabilities of the components come into play. This is for instance true if two or more components are of similar type, are sharing a common environment or are exposed to common cause failures. In this chapter the generalized predictive approach is considered. Our complete work in this area is presented in Natvig (1990).
