An Interacting Wasserstein Gradient Flow Strategy to Robust Bayesian Inference

Abstract

Model Updating is frequently used in Structural Health Monitoring to determine structures' operating conditions and whether maintenance is required. Data collected by sensors are used to update the values of some initially unknown physics-based model's parameters. Bayesian Inference techniques for model updating require the assumption of a prior distribution. This choice of prior may affect posterior predictions and subsequent decisions on maintenance requirements, specially under the typical case in engineering applications of little informative data. Therefore, understanding how the choice of prior may affect the posterior prediction is of great interest. In this paper, a Robust Bayesian Inference technique evaluates the optimal and worst-case prior in the vicinity of a chosen nominal prior, and their corresponding posteriors. This technique employs an interacting Wasserstein gradient flow formulation. Two numerical case studies are used to showcase the proposed algorithm: a double-banana-posterior and a double beam structure. Optimal and worst-case prior are modelled by specifying an ambiguity set containing any distribution at a statistical distance to the nominal prior, less or equal to the radius. Examples show how particles flow from an initial assumed Gaussian distribution to the optimal worst-case prior distribution that lies inside the defined ambiguity set, and the resulting particles from the approximation to the posterior. The resulting posteriors may be used to yield the lower and upper bounds on subsequent calculations used for decision-making. If the metric used for decision-making is not sensitive to the resulting posteriors, it may be assumed that decisions taken are robust to prior uncertainty.