Artificial Intelligence Verification Based on Operational Design Domain (ODD) Characterizations Utilizing Subset Simulation

For AI/ML verification, the definition of the Operational Design Domain (ODD) of the considered AI/ML constituent is crucial. Related to the ODD, the European Union Aviation Safety Agency (EASA) states: “The ODD defines the set of operating parameters, together with the range and distribution within which the AI/ML constituent is designed to operate, and as such, will only operate nominally when the parameters described within the ODD are satisfied”. Within this paper, an ODD characterization utilizing statistical distributions for their (marginal) ODD-parameters and the dependence structures among them (utilizing the concept of vine copulas) is proposed. In addition, this paper reflects the utilization of subset simulation, which is an advanced sampling process based on the basic Monte Carlo sampling that is tailored to the evaluation of specific areas of interest (typically failure regions of algorithms) with a fewer number of required samples. As discussed during the development of ED-324 / ARP6983” Process Standardd for Development and Certification / Approval of Aeronautical Safety-Related Products Implementing AI”, the concept of subset simulation can support AI/ML verification activities, which is outlined in this paper. A case study to illustrate the proposed methodologies is carried out and described. The contributions of this paper support the developments towards solving the challenges induced by the high dimensionality of the ODD and the complexity of the AI/ML model.