A Bayesian Approach to Adversarially Robust Life Testing

Dorina Weichert, Alexander Kister, Sebastian Houben, Gunar Ernis, Tim Wirtz

Keywords: Life tests, Bayesian Inference, Active Learning

Abstract: In materials science and engineering, the lifetime of materials and products is tested by costly manual characterization procedures that are standardized only in certain cases. In this paper, we investigate a modular Bayesian approach to lifetime testing that can reduce the number of experiments and, thus, the overall cost of experiments. The approach is based on the correct definition of the probability of the outcome of an experiment, e.g., its likelihood. Since this is usually unknown, we extend it to the adversarial setting, finding an experimental procedure that is robust to a given set of probabilities in the worst case. By simulations, we empirically show the advantages of this procedure over the state-of-the-art and the basic approach, potentially reducing the number of costly experiments.

Submission Number: 89