Establishing Verification and Validation Objectives for Safety-Critical Bayesian Networks

The assurance of autonomous systems and the technologies that drive them is a major research challenge in the safety-critical systems engineering domain. The nature of many of these Machine Learning (ML) and Artificial Intelligence (AI) approaches raises a number of additional, technology-specific assurance concerns. One such approach is the Bayesian Network (BN) probabilistic modelling framework. Bayesian Networks and the family of modelling techniques they belong to form the basis of many AI applications. However, little research has been conducted into the assurance of BN-based systems for use in safety-critical applications. This paper explores some of the key distinctions between BN-based software-intensive systems and conventional software systems. It introduces a modelling framework that explicitly captures BN-based system-specific considerations and facilitates both the communication of assurance concerns between safety practitioners and system stakeholders, and the subsequent safety analysis of the system itself. It demonstrates how this approach can be used to develop specific verification and validation objectives for a BN-based system in a medical application.