Validation and Verification Challenges in a Machine Learning Algorithm for Connected Vehicles - Design Science Research of Developing a Most Probable Path Algorithm

Abstract

Machine Learning (ML) software in connected and automated vehicles puts new demands on safety regulators and industry standards to keep up with the explosive evolution of technology in the automotive domain. This thesis reports a practical example of developing an ML-based algorithm that predicts the most probable path for an arbitrary vehicle, without knowing the destination. This work is done in collaboration with Carmenta Automotive AB as an industry partner, a company that is aiming to increase situational awareness for vehicles on the roads. The thesis methodology follows an iterative design science research (DSR) approach, developing an artifact consisting of an ML model connected to the company’s system. The literature highlights the challenges of validating and verifying (V&V) an ML component, as there are currently no applicable standards for ML software in the automotive domain. This DSR attempts to showcase V&V activities on ML models trained with different data characteristics to assess whether the challenges surrounding V&V can be mitigated when validating the data-driven most probable path algorithm.