This project will explore applications of Reinforcement Learning (RL) for a real-world application in the railway domain.
To address increasing demand, railway providers aim to increase traffic density on the existing network. However, dense traffic can result in delayed trains and infrastructure disruptions, which can impact planned trips in a large part of the network and are prone to cascading effects. Therefore, real-time rescheduling of trains, in case of delays and disruptions, is a vital.
To foster research on RL-based solutions for the VRSP in the rail domain, the flatland-association developed flatland-rl, an open-source digital simulator. Alongside the environment, they provide a starter kit, benchmark scenarios, and baseline agents. Several flatland-rl challenges have already been hosted at top AI venues (e.g.,NeurIPS).
The core contribution of this thesis will be to advance new methodologies to improve collaboration and coordination among independent RL agents, and will be evaluated in flatland-rl.
The topic is suitable for an MSc-level student and prior experience with RL is essential.