Thesis defence of Tatiana BABICHEVA
« Machine Learning for the distributed and dynamic management of a fleet of autonomous taxis and shuttles »
10 mars 2021 – 9h30
Online : Zoom
Follow this event :
Connection link
Jury :
- Mme Leïla KLOUL, MCF (HDR), Université de Versailles Saint-Quentin-en-Yvelines, FRANCE – Directeur de these
- M. Dominique BARTH, Professeur, Université de Versailles SainDt-Quentin-en-Yvelines, FRANCE – Co-encadrant de these
- M. Alain QUILLIOT, Professeur des Universités, Université Clemont-Ferrand II, FRANCE – Rapporteur
- M. Akka ZEMMARI, Maître de conférences (HDR), Université de Bordeaux, FRANCE – Rapporteur
- M. Wilco BURGHOUT, Directeur de recherche, KTH Royal Institute of Technology, SUÈDE – Co-encadrant de these
- M. Jakob PUCHINGER , Professor, CentraleSupélec, FRANCE – Examinateur
- M. René MANDIAU, Professeur des universités, Université Polytechnique Hauts-de-France , FRANCE – Examinateur
- M. S. M. Hassan Mahdavi, PhD, Vedecom, Mobilab – Invité
Résume
In this thesis are investigated methods to manage shared electric autonomous taxi urban systems under an online context in which customer demands occur over time, and where vehicles are available for ride-sharing and require electric recharging management.
We propose heuristics based on problem decomposition which include road network repartition and highlighting of subproblems such as charging management, empty vehicle redistribution and dynamic ride-sharing.
The set of new methods for empty vehicle redistribution is proposed, such as proactive, meaning to take into account both current demand and anticipated future demand, in contrast to reactive methods, which act based on current demand only.
We provide reinforcement learning in different levels depending on granularity of the system.
We propose station-based RL model for small networks and zone-based RL model, where the agents are zones of the city obtained by partitioning, for huge ones. The complete information optimisation is provided in order to analyse the system performance a-posteriori in offline context.
The evaluation of the performance of proposed methods is provided in a set of road networks of different nature and size.
The proposed method provides promising results outperforming the other tested methods and the real data on the taxi system performance in terms of number of satisfied passengers under fixed fleet size.