Application of value of information theory in adaptive metamodeling for reliability assessment

Abstract

The present paper discusses the application principles of value of information theory in adaptive metamodeling for reliability analysis. Metamodeling for reliability purposes has become particularly relevant in recent years. The usage of metamodels allows surrogating the, costly to evaluate, performance functions of engineering structures. Adaptive Kriging procedures are examples of the successful application of metamodel- ing in reliability analysis. Efficient adaptive Kriging involves the usage of some notion of improvement in what ultimately is an unsupervised decision making scheme that selects points to enrich the model. Therefore, the decision to select a point to enrich the experimental design should consider the utility of each candidate in the expectation of improvement of the metamodeling accuracy. Within this context, a comprehensive discussion on the application of value of information for reliability metamodeling is presented. Since the candidate points and surrogate are jointly built in a virtually costless model, it is possible to know the virtual outcome of the enrich- ment decisions. In many circumstances, points in the experimental design may provide redundant information. Furthermore, a priori knowledge on the performance function may be applied to weight the expected outcome of exploration and exploitation. Value of information considerations adds value to reliability metamodeling that uses adaptive methods, and is of interest for efficient design and optimization of complex structures, such as bridge structures.