What Makes It Testable? Conceptual Model for Safety Quantification

Edward Schwalb, Bernhard Bieder, Daniel Wiesenhütter

Jun 12, 2017 (modified: Jun 12, 2017) ICML 2017 MLAV Submission readers: everyone
  • Abstract: We propose a simple conceptual scenario based model to formalize and quantify safety validation. We introduce a failover model, whereby overall failure rates are reduced exponentially in the number of stages added, but the validation effort needed only increases linearly in the number of stages. We address validation of rare situations using synthetic data. We introduce multi-scenario and multi-model training and testing to quantify operational safety under poor conditions, failure and emergencies. Finally, we discuss regression introduced by a learning process, and hypothesize that safety-regression can be avoided in a broad class of system revisions.
  • Keywords: Autonomous Vehicles, Safety, Validation, Scenarios, Failover Model