Open semester projects and master theses:
Although carpooling has been long celebrated as a promising solution to reducing vehicular traffic and emissions during peak hours, it is still limited within families and for long-distance trips. For instance, EPFL launched a carpooling platform to encourage its staff and students to carpool in their daily commute. However, limited trips are posted and very few users remain active. The inconvenient trip matching comes to be a major obstacle. In this project, we will build an idealized carpooling system for daily commute trips that features flexible role assignment and centralized trip matching.
During this project, you are expected to:
- Formulate and solve the carpooling matching as optimization problems
- Analyze EPFL mobility survey data to build profiles of potential carpooling users
- Build simulation to evaluate the proposed carpooling system
Contact: Prof. Kenan Zhang ([email protected])
Mobility-as-a-service (MaaS) aims to provide seamless multi-modal mobility options to traveler and has great potentials to improve sustainability and accessibility of urban transportation systems. One critical factor in MaaS system design, however often overlooked, is the transfer cost, e.g., extra walk and wait due to mode transfers. This project aims to estimate the transfer cost perceived by travelers using data-driven methods. Specifically, we will connect a traveler trajectory dataset to transit database, identify mode transfers, and develop a network model to estimate the transfer cost.
During this project, you are expected to
- Learn to work with large spatiotemporal dataset
- Identify mode transfer from traveler trajectories
- Develop a multi-modal transit network with transfers
- Estimate and analyze the perceived transfer cost
Contact: Dr. Rui Yao ([email protected])
On-demand meal delivery service (e.g., UberEat) has developed rapidly in recent years, particularly during COVID lockdowns. Similar to ride-hailing platforms, on-demand meal delivery platforms promptly collect orders and assign them to couriers for pickup and delivery. Differently, some platforms take the bundling strategy that groups several orders with close pickup and delivery locations into bundles. Accordingly, each courier service trip consists of multiple pickups and deliveries. Typically, solving the order bundling requires extensive evaluations on candidate bundles and each evaluation involves solving a Pickup-and-Delivery-Route-Planning (PDRP) problem. This imposes great computational challenge for real-time operations. In this project, we will explore a data-driven solution to the order bundling problem. Specifically, we will investigate key features that affect the bundle performance and then build clustering algorithms, a class of unsupervised machine learning models, using these features.
During this project, you are expected to
- Learn the agent-based on-demand meal delivery simulation framework
- Identify the influential features of order bundling using simulation and historical data
- Develop clustering algorithms and implement them in the simulation
- Evaluate the performance of proposed bundling algorithms and compare them with benchmark algorithms
Contact: Prof. Kenan Zhang ([email protected])
Besides the imbalance between supply (e.g., road capacity) and demand (e.g., vehicular flows), traffic congestion is largely due to the selfish routing behaviors of private vehicles. Theoretically, if all vehicles choose their routes to minimize their own travel time, the system converges to a state called user equilibrium (UE). On the other hand, if all vehicles are centrally controlled and routed to minimize total travel time in the network, the system ends up at the system optimum (SO) state. Our previous study shows that we can actually approach SO by strategically controlling a small fraction of vehicles. Moreover, the controlled vehicles are found traveling between a limited number of origin-destination (OD) pairs. Motivated by this observation, this project aims to further study these critical OD pairs, explore their distribution in general networks, and design metrics to evaluate the “critical” level of an OD pair.
During this project, you are expected to
- Learn ORCS and implement its solution algorithm
- Conduct numerical experiments to investigate critical OD pairs in different networks
- Generalize the similarity of recognized OD pairs
- Propose features and metrics to evaluate the critical level of OD pairs
Contact: Dr. Hossein R. Farahani ([email protected])
Ongoing projects:
Traffic congestion has long been a critical issue in large cities. As the well-known solution, congestion pricing is however arguably unfair as it tends to favor wealthier travelers. To address such an equity issue, we proposed a mechanism named CARMA that makes travelers to bid for access to scarce transport resources with non-tradable mobility credits, in anticipation of other travelers’ decisions as well as their own future trips. In our preliminary study, we numerically showed the redistribution of credits plays a key role in the system efficiency and found its strong connection to classic road pricing schemes. This project aims to further investigate credit redistribution as a design problem and deepen the understanding of its impact on system performance.
During this project, you are expected to
- Learn the base model of CARMA and reproduce numerical results
- Review the literature on congestion pricing with the bottleneck model
- Explore different credit redistribution schemes and evaluate their performances
Supervisor: Kenan Zhang
Students: Yasser Tahiri, Mohamed Mansour
“Braess Paradox” is a well-known phenomenon in traffic networks, where adding one road link to the network instead worsens traffic congestion. This discouraging consequence is rooted in the selfish routing behavior of private vehicles. For example, a newly built highway may attract a lot of people and finally intensify congestion on upstream and downstream roads. In this project, we will identify the existence of Braess links in both stylized and general road networks, explore their common properties, and develop heuristics to identify them without extensive computations.
During this project, you are expected to
- Learn the static traffic assignment model and implement its solution algorithms
- Conduct numerical experiments to recognize Braess links in different networks
- Generalize the similarity among recognized Braess links
- Develop heuristics to identify Braess links
Supervisor: Hossein Farahani
Student: Julien Ars
Previous semester projects:
This project analyzed the PubliBike network at EPFL and surrounding municipalities, including demand pattern and user satisfaction, using EPFL mobility survey and operational data. It also benchmarked current service against alternative solutions to identify key challenges and opportunities in promoting bikesharing. The project was led by the EPFL Mobility Office
Supervisor: Luca Fontana, Luca Pellandini, Kenan Zhang
Student: Evangelia Gkola, Gregoire Ecuyer
This project studied the carpooling behaviors of EPFL staffs and students using the mobility survey and data of an EPFL carpooling platform. It also studied the implication of carpooling on EPFL’s parking and CO2 emission reduction strategies. The project was led by the EPFL Mobility Office.
Supervisor: Luca Fontana, Alexia Couturier, Kenan Zhang
Student: Martin Simon, Marine Jacob
On-demand mobility covers a wide range of services including taxis, e-hailing (e.g., Uber), ride-pooling, and bike-sharing. Although different in detailed operations, these services share some fundamental characteristics regarding the movements of vehicles, their interactions with users, as well as the spatiotemporal distribution of demand and supply. This project investigated an open-source simulation of on-demand mobility and tested it potential for generalizing different service modes and market conditions.
Supervisor: Kenan Zhang
Student: Mathis Magnin
The perturbed utility model (PUM) is a discrete choice model that represents an individual’s decision as a vector of choice probabilities and describes the utility of each alternative as the sum of systematic utility and a convex perturbation function of the choice probability vector. PUM has been shown to generalize a wide range of discrete choice models (e.g., MNL) meanwhile enjoying a higher computational efficiency in model inference. However, a challenge is the specification and estimation of the perturbation function. In this project, we explored the potential of learning the perturbation function using neural networks (NN).
Supervisor: Rui Yao
Student: Jingren Tang
Current navigation apps (e.g., Google Maps) usually recommend walking paths based on distance and elevation, while ignoring how much pollution the pedestrian may be exposed to along the walk. This project studied the pollution-aware path planning problem for pedestrians in the urban environment using the spatiotemporal environmental data collected by Sparrow.
Supervisor: Rui Yao
Student: Anne-Valérie Preto