/Leveraging Deep Learning and Multi-agent Reinforcement Learning for Traffic Forecasting and Intelligent Traffic Light Optimization

Leveraging Deep Learning and Multi-agent Reinforcement Learning for Traffic Forecasting and Intelligent Traffic Light Optimization

PhD - Antwerpen | More than two weeks ago

Harnessing the Power of Deep Learning for Next-Generation Urban Traffic Management

Apply

Traffic state predictions can be used for proactively managing the likely traffic in the near future to avoid congestion and unsafe situations as well as lower emissions. For effective traffic management and governance, accurate traffic predictions are essential as they give access to valuable information such as the propagation of congestion. In this PhD, we will examine how these predictions can be used to adapt timings on smart traffic lights. Next to motorized traffic, cyclists and pedestrians will be able to move more smoothly at intelligent intersections. Furthermore, a smoother flow of traffic and a reduction in unnecessary stops and accelerations of cars and trucks will have a positive impact on the environment, aiding the general plans of sustainable urban mobility with the aim to give a higher benefit to society. Data-driven techniques will be investigated and combined with multi-agent reinforcement learning to learn the relation between roads helping the traffic lights to decide action. This results in alternative management strategies based on various scenarios such as extending the green signal phase for specific traffic flows that are causing or may be impacted by future traffic congestion. These approaches will be assessed/benchmarked in terms of accuracy, reliability of the outcomes and the computational load (uncertainty analysis/sensitivity analysis of the model scalability) and combined in a data and resource efficient hybrid methodology for multi-modal forecasting in large-scale road networks.

Required background: Engineering Technology, Engineering Science, Computer Science or equivalent

Type of work: 70% modeling/simulation, 20% experimental, 10% literature

Supervisor: Siegfried Mercelis

Co-supervisor: Joris Finck

Daily advisor: Ynte Vanderhoydonc, Eridona Selita

The reference code for this position is 2024-157. Mention this reference code on your application form.

Apply