Research Topics

Advancing Demand Modeling

Leveraging existing and new data sources, this research direction aims at developing new demand modeling methodology (both microscopic, e.g. activity-based models; or macroscopic, such as aggregated demand models) by taking advantage of core methodological domains:

  • Probabilistic graphical models
  • Bayesian non-parametric models
  • Discrete choice models
  • Deep neural networks
  • Deep generative models

Prediction under Stress Scenarios

The transport system as we know it today is a dynamic one. While there is plenty of research and practice approaches for habitual behaviour (e.g. commuting), there is much less in stress scenarios (e.g. special events, incidents, breakdowns, roadworks). More importantly, it is in such stress scenarios that we need good predictions!… This research direction dedicates to problems such as predicting demand in special events, safety levels in vehicle interactions or forecasting changes in speed, capacity or clearance time in traffic incidents.

Complex Simulation

What will be the impact of new smart mobility systems, such as automated mobility and mobility-as-a-service? How will individuals change their travel and activity patterns? How will these changes impact the design and operations of these innovative services? We develop simulation model and software architectures to tackle these questions and the future challenges of tomorrow focusing on four key features:

  • Agent-based simulation
  • Demand-supply interactions
  • Multi-modal dynamic traffic assignment
  • Microscopic traffic simulation

Mobility Management and Operations

New technologies, business models, data sources and user preferences are increasingly emerging in today’s mobility market. Developing new solutions is at the core of our research philosophy and this research line aims precisely at contributing to this mobility market with new technological and service designs, deployments and operations for an efficient and innovative mobility system, such as:

  • Deep reinforcement learning for traffic control
  • Dynamic rescheduling of fleet operations
  • Intelligent truck platooning
  • Real-time incentives and tradable permits in mobility systems

Robustness under Uncertainty

More often than not, we have noisy (or missing) data, or simply too much complexity to handle at the moment. This leads to well-known problems such as heteroscedasticity (non-constant variance) or homogeneity (correlation with the error term), and a bad modeling will certainly lead to bad predictions. In this research, we focus on two general approaches:

  • Heteroscedastic Models – models that can deal with non-constant variance in the error term
  • Distribution Exploration Models – instead of only predicting the mean, we focus on other properties if the predictive distribution. For example, when we predict two quantiles (e.g. 5% and 95%), we can create a prediction interval (this will say “the model predicts that the value will lie between X and Y with Z% probability) or even approximate the full distribution!