SPEAKER: Maja Aaslyng Dall, PhD fellow at ITU
ABSTRACT: Bayesian probabilistic programs are good at hiding errors and bugs, which in turn has led to examples of published papers, where the conclusion based on the model output has later been shown to be invalid. Testing can help mitigate this. Unfortunately, testing can be difficult, especially if you do not have a background in computer science, which is one of the reasons why automatic testing can be useful. The goal for my thesis is to introduce automatic testing focusing on which properties we can test. Which of those can be tested in a model independent way? And finally, to develop a DSL to implement the tests. My own spin on this project will hopefully include either: 1. Quantifications of how well the priors are explored, and/or 2. Looking into how Symbolic Execution can be explored with a branch and bound approach using SAT-Metropolis.