“Modelling patterns of individual mobility”: Mehdi Katranji defends his thesis

VEDECOM is pleased to announce that Mehdi Katranji will be defending his computing thesis entitled “Deep learning in individual mobility”.

His thesis defence will take place on 16 December 2019 at the mobiLAB.

JURY

Mr Alexandre Caminada, Université Nice Sophia Antipolis, Thesis Supervisor

Mrs Latifa Oukhellou, IFSTTAR, Referee

Mr Marc Barthelemy, CEA, Referee

Mr Fouad Hadj Selem, VEDECOM, Thesis Co-Supervisor

Mr Laurent Moalic, UHA, Thesis Co-Supervisor

Mr Frédéric Precioso, Université Nice Sophia Antipolis, Examiner

ABSTRACT

Understanding mobility is a major issue for the authorities responsible for organising mobility and urban planning. Our thesis focuses on “individual mobility” – a term we employ in the absence of any formal definition of human mobility. In our introduction, we outline the applications used to enhance our understanding of human mobility, alongside the relevant stakeholders.

This will be followed by an account on the state of the art relating to different transport models. Transport studies are not in a readily usable format for mobility stakeholders seeking to implement mobility solutions or policies. Transport models convert the initial data to deliver information in a usable and workable format. This is then used to determine the prerequisites for creating a learning model: understanding of available dataset typologies, strengths and weaknesses. We will also give an overview of the four-step transport model, in use since 1970, before discussing how methodologies have developed over recent years.

We will then present our own models for individual mobility. These automatic learning models allow us to gain a clearer, more comprehensive overview of individual mobility, without further investigation. The commonality between these different models is that they focus on the individual, in contrast to traditional methods based on locality. We build on the principle that individual people make decisions based on their perception of the local environment.

The final chapter of our paper, our main theoretical contribution, seeks to improve the robustness and performance of these models. In doing so, we study the deep learning methodologies of restricted Boltzmann machines. Following an account on the state of the art relating to this family of models, we explore strategies for their viability in the applications environment.