Ifsttar PhD subject


French version

Detailed form :

Title : Offline and online Machine learning approaches to calibrate multimodal macroscopic fundamental diagrams

Main host Laboratory - Referent Advisor COSYS - LICIT  -  LECLERCQ Ludovic      tél. : +33 472142429 
Director of the main host Laboratory LECLERCQ Ludovic  -  
Laboratory 2 - Referent Advisor COSYS - GRETTIA  -  ZARGAYOUNA Mahdi  -    -  tél. : +33 181668698
PhD Speciality Génie Civil / Ingénierie du trafic
Axis of the performance contract 1 - COP2017 - Efficient transport and safe travel
Main location Bron
Doctoral affiliation ENTPE
Planned PhD supervisor LECLERCQ Ludovic  -  Université Gustave Eiffel  -  COSYS - LICIT-ECO7
Planned PhD co-supervisor ZARGAYOUNA Mahdi  -  Université Gustave Eiffel  -  COSYS - GRETTIA
Planned financing Contrat doctoral  - Université Gustave Eiffel


Context and motivations

In recent years, we have seen profound changes in urban mobility, with both an abundance of supply due to the appearance of new mobility services (car-sharing, transportation on demand) and the development of new practices (bicycles, shared scooters, etc.). With multimodal planers and access to real-time information, passengers are increasingly able to manage their multimodal trips in an optimal way. Two key elements in travelers' decisions are their multimodal choice model and the estimation of their corresponding travel times. These changes therefore raise questions about the way transportation systems are characterized. Within the framework of the RESIFLEX project, GRETTIA and LICIT, two laboratories of the Gustave Eiffel University have decided to combine their expertise to develop an open source multi-agent simulator meeting this vision, called OpenMobiFlex.


The objective of this thesis is to support the specification, design and implementation of OpenMobiFlex, representing the entire urban system as a multi-agent system. This modeling will allow to take into account the decisions at the individual scale of the actors of the system. The defined system will be multimodal, describing the whole transport offer, from soft modes to motorized modes, individual or collective. The main feature of this platform is to integrate simple physical models to characterize travel time functions while intrinsically taking into account the interactions between the different modes. Intelligence is moved from model construction in the form of a physical equation to static learning of travel time functions from historical data from heterogeneous sources. These functions then allow the reproduction of the travel conditions encountered on the chosen travel modes and feed the route choice model.
One of the main results expected from the thesis proposed here is to define and implement the methodological framework allowing the calibration of travel time functions from historical data. The analysis will be based on multimodal data collected by GRETTIA within the framework of the ClaireSITI platform on several cities in France (Toulouse, Rennes, Seine Saint-Denis, ...) and on data collected by LICIT on the metropolis of Lyon.


The contributions of this thesis will be based on machine learning and multi-agent modeling and simulation. On the one hand, the use of multi-agent modeling allows a differentiated representation of the human actors in the system (different profiles, different preferences, different decision mechanisms, etc.) and to observe heterogeneous and realistic collective behaviors. Multi-agent simulation is today a mature field with proven and efficient platforms, some of which integrate tools for managing geographic data required for transport applications. On the other hand, we will rely on advanced learning methods to merge heterogeneous data from different sources/modes of transportation in order to characterize the correlations between travel times of different modes of travel. This work will initially be carried out on historical data and may be complemented by a real-time approach allowing to directly assimilate the collected data and thus adjust the calibration of the different parameters.
It will then be possible to use the platform's full predictive potential to test both current evolution scenarios and breakthrough scenarios (integration of new mobility services, new travel management policies, major incidents, etc.).

Existing infrastructures and tools

OpenMobiFlex is based on the significant advances obtained by the LICIT laboratory within the framework of the ERC MAGnUM on the issue of multimodal and multi-scale dynamic simulation and on the long-standing experience of the GRETTIA laboratory on multi-agent simulation (SIM4T platform).

Profile of the candidate
We are looking for motivated and talented candidates who have experience in machine learning. Knowledge of traffic models and/or multi-agent simulation platforms will be highly appreciated. The candidate must have very good English language skills (spoken and written).

Keywords : Machine learning, Agent-based systems, Urban mobility, Simulation, multimodal travel times
List of topics
Applications closed