Models and algorithms for the exploration of the space of scenarios: toward the validation of the autonomous vehicle. (Modèles et algorithmes pour l'exploration de l'espace des scénarios: vers la validation du véhicule autonome)

Marc Nabhan

Abstract: Autonomous vehicles represent highly complex systems, where multiple types of failures could occur leading to a wrong execution on the road. Therefore, each component should be thoroughly tested to anticipate potential failures and mitigate them. Simulation testing methods are used to complement real test driving in the validation process. The aim of this CIFRE PhD thesis is to feed an industrial project at Renault novel algorithms and methods to facilitate the validation of the command law requirements by performing Model-In-the-Loop (MIL) testing. The main contributions of this PhD thesis are threefold: detecting a maximum number of failures of the command law, detecting scenarios as close as possible to the border separating zones of failed and safe scenarios, and building explainable border models to identify the border as accurately as possible. The algorithms of the first two objectives use machine learning (Random Forest) and optimization (CMA-ES) techniques to abide with the industrial requirement of reducing the computing power needed, and three approaches are considered to build border models while comparing their performances and explainabilities: Neural Networks, Mixed-Integer Linear Programming (MILP), and Genetic Programming (GP) applied to symbolic regression.