Préservation de la performance des graphes dans un environnement perturbé de manière exogène (Ex.Apprentissage par Renforcement sur les graphes pour l'Optimisation de la mobilité urbaine)

Preserving Graph Performance In An Exogenously Perturbed Environment (Ex. Graph-Based Reinforcement Learning Optimization to Optimize Urban Mobility)

• Auteur / Autrice : Zihao Guo
• Direction : Andrea Araldo, Mounim A. El yacoubi
• Type : Projet de thèse
• Discipline(s) : Informatique, données, IA
• Date : Inscription en doctorat le 02/10/2023
• Etablissement(s) : Institut polytechnique de Paris
• Ecole(s) doctorale(s) : École doctorale de l'Institut polytechnique de Paris
• Partenaire(s) de recherche : Laboratoire : SAMOVAR - Services répartis, Architectures, Modélisation, Validation, Administration des Réseaux
• Partenaire(s) de recherche : Equipe de recherche : METHODES

Mots clés

Mots clés libres

Apprentissage par Renforcement

Optimisation

Théorie des graphes

Reinforcement learning

Optimisation

Graph theory

Résumé

Voir résumé en anglais

'This PhD position is available only to students eligible to the China Scholarship Council' (https://www.ip-paris.fr/en/education/international/cooperation-china) ''STARTING DATE IS FLEXIBLE''' ### MOTIVATION Accessibility[1] is a graph-theoretical metric that measures how well a certain place (aka node) is connected to the surrounding urban area (aka the rest of the transportation graph). A place enjoys high accessibility when it can be easily and rapidly reached from any other place in the area. Accessibility is a key indicator for a town or a district, as it determines quality of life, job attractiveness, value of properties. Current public transportation in large urban conurbations suffers from an evident inequity in the distribution of accessibility. Accessibility is high in city centers and poorer in the outskirts, resulting in poor Quality of Service (QoS) (large waiting and walking times to reach the nearest stop). This incentivizes travelers to use private cars,[8] with severe environmental and societal impact. The root cause of poor QoS in the outskirts is that public transportation is mainly provided by Fixed Route Transit (FRT), with a pre-determined schedule; this is inefficient, by construction, in the outskirts. ## RESEARCH ACTIVITY The PhD will be structured in 3 phases. In Phase 1, we will construct a multilayer graph[11] to model real cities. Since we only rely on open data (transit schedules in GTFS format, Open Street Maps, population densities), our procedure will be automatable and reproducible in several cities around the world. In Phase 2, we optimize a city, i.e., we take the aforementioned decisions, by formulating a graph optimization problem on the model constructed in Phase 1. The objective is to reduce the accessibility gap between city centers and outskirts. We apply Graph Embedding and Deep Reinforcement Learning,[3,6] which have been shown to be able to learn how to solve graph related problems on (small) networks and then apply this knowledge to other (bigger) ones. We can exploit this feature to transfer knowledge from a city to another. Finally, in Phase 3, we validate the optimized multimodal transit in a high detail fully econometric agent-based simulation (SimMobility[1]) in the Parisian Region. ## TEAM Supervisors: Assoc. Prof. Andrea Araldo and Prof. Mounim El Yacoubi (Institut Polytechnique de Paris – Télécom SudParis) External collaborators: Prof. Ben-Akiva (Massachusetts Inst. of Technology), Prof. Antonious (Technical Univ. of Munich), Assoc. Prof. Seshadri (Tech. Univ. of Denmark). ## CONTACT [e-mail] ## REQUIRED SKILLS The candidate must be proficient in Python (and, optionally, in C++). She/he must also be able to perform some basic server administration and work with remote servers (ssh). She/he must be proficient with github or similar platforms. A previous experience in deep learning or reinforcement learning is a big plus. ## REFERENCES [1] Biazzo, I., Monechi, B., & Loreto, V. (2019). General scores for accessibility and inequality measures in urban areas. Royal Society Open Science, 6(8). [2] https://data.iledefrance.fr/explore/dataset/horaires-prevus-sur-les-lignes-de-transport-en-commun-dile-de-france-gtfs/information/ [3] Facebook data for good: Population Density Maps, https://dataforgood.fb.com/tools/population-density-maps/ [4] D. S. Johnson, J. K. Lenstra, and A. Rinnoiy Kan. “The complexity of the Network Design Problem”. In: Networks (1978). [5] A. Barreto, D. Silver, et al. “Successor features for transfer in reinforcement learning”. In: NIPS (2017). [6] H. Dai et al. “Learning combinatorial optimization algorithms over graphs”. In: NIPS (2017). [7] A. Mittal, A. Dhawan, S. Manchanda, S. Medya, S. Ranu, and A. Singh. “Learning heuristics over large graphs via deep reinforcement learning”. In: arXiv (2019). [8] Welch, T.F., Mishra, S.: A measure of equity for public transit connectivity. Journal of Transport Geography 33, 29–41 (2013) [9] Quadrifoglio, L., Li, X.: A methodology to derive the critical demand density for designing and operating feeder transit services. Transportation Research Part B: Methodological 43(10), 922–935 (2009) [10] Steiner, K., Irnich, S.: Strategic planning for integrated mobility-on-demand and urban public bus networks. Transp. Science 54(6) (2020) [11] F. Asgari, A. Sultan, H. Xiong, V. Gauthier, and M. A. El-Yacoubi. “CT-Mapper: Mapping sparse multimodal cellular trajectories using a multilayer transportation network”. In: Computer Communications 95 (2016), pp. 69–81. [12] Calabrò, G., Araldo, A., Oh, S., Seshadri, R., Inturri, G., & Ben-Akiva, M. (2021, January). Integrating fixed and demand-responsive transportation for flexible transit network design. In TRB 2021