Toward a Methodology for the Verification and Validation of AI-Based Systems

Verification and validation (V&V) is the cornerstone of safety in the automotive industry. The V&V process ensures that every component in a vehicle functions according to its specifications. Automated driving functionality poses considerable challenges to the V&V process, especially when data-driven AI components are present in the system. The aim of this work is to outline a methodology for V&V of AI-based systems. The backbone of this methodology is bridging the semantic gap between the symbolic level at which the operational design domain and requirements are typically specified, and the sub-symbolic, statistical level at which data-driven AI components function. This is accomplished by combining a probabilistic model of the operational design domain and an FMEA of AI with a fitness-for-purpose model of the system itself. The fitness-for-purpose model allows for reasoning about the behavior of the system in its environment, which we argue is essential to determine whether the system meets its requirements. While this work only provides an outline of such a methodology, we point out future research directions toward a full methodology for the V&V of AI-based systems.