Defense of Yikai WANG (automatic control department)

Advanced driver assistance systems (ADAS) have undergone considerable development in recent years. However, ADAS enabling autonomous driving are usually based on fixed rules, which on the one hand limits their ability to adapt to differences in driver behavior, and on the other hand does not allow them to dynamically adjust their own behavior according to the situation and preferences of those drivers.

This lack of flexibility can lead to a "style misalignment" between the driver and his system, causing decision-making conflicts that can lead to regaining control, which can affect driving safety, as well as degrading trust and acceptability of the system.

In this context, the CoCoVéIA project, funded by the Agence Nationale de la Recherche (N° ANR-19-CE22-0009-01) and led by LAMIH UMR CNRS 8201, was born.

The previous project, CoCoVeA, had already developed a Level 2 controller based on a shared driving mechanism. On this basis, CoCoVéIA introduces a self-learning mechanism for the driver's driving preferences, exploiting the decisional conflicts between the driver and his ADAS as the main source of data.

CoCoVéIA is the first project of its kind to introduce a self-learning mechanism for the driver's driving preferences.

The work in this thesis focuses on the design, development and validation of a driving preference model, based on artificial intelligence technologies, and enabling the control of a vehicle in autonomous mode and the gradual adaptation of this model to the driver's preferences.

In terms of methodology, the research began with pre-experimentation with drivers to identify the parameters linked to driving preferences, i.e. driver behavior in terms of action when faced with a particular situation to manage.

The data analysis made it possible to define the situation parameters to be taken into account as inputs to the preference model, and the behavior parameters to be taken into account as outputs from this model. Various neural network learning algorithms were tested with the datasets obtained during the pre-experiments.

Neuronal network learning algorithms were tested with the datasets obtained during the pre-experiments.

From the selected algorithms, a network is then trained with a large dataset to obtain an initial model enabling vehicle control.

After validation of this model's behavior during online use, a small volume of new data representing new preferences is used to adjust this model, enabling new driving preferences to be learned.

Experiments conducted on a driving simulator have validated the effectiveness of the proposed approach. The results show that the combination of a Transformer-type neural network architecture and the fine-tuning technique enables new driving preferences to be learned efficiently from a limited volume of data, thus ensuring rapid adaptation of the system to individual driving preferences.

The results of these experiments are presented below.

Rapporteurs:

• Mrs Lydie NOUVELIERE, Professeur des Universités, Université de Saint-Étienne
• Mrs Naïma AITOUFROUKH-MAMMAR, Maître de Conférences HDR, Université d'Évry Val d'Essonne

Examiners:

• Mr Jimmy LAUBER, Professeur des Universités, Université Polytechnique Hauts-de-France
• Mr Jean-Christophe POPIEUL, Professeur des Universités, Université Polytechnique Hauts-de-France
• Mr Bernard RIERA, Professeur des Universités, Université de Reims

Thesis supervisor:

• Mr Serge DEBERNARD, Professeur des Universités, Université Polytechnique Hauts-de-France
  .