Digital Twin-Enabled Reinforcement Learning for Fault-Resilient Urban Traffic Signal Control

The trend of digitalization and smart cities is enabling new Traffic Signal Control methods, with urban transportation Digital Twins acting as a key asset, enabling data-based approaches for reducing urban traffic congestion. We propose a Digital-Twin-enabled Deep-Q-Network Traffic Signal Control scheme, trained to improve the signal control efficiency of a single intersection in spite of possible sensor failures. The controller is trained using a Digital Twin of the western half of Södermalm, Stockholm, with rewards mirroring the Max-pressure signal control. We demonstrate in simulations that our proposed control significantly outperforms the Maxpressure benchmark in cases involving sensor failures, while achieving comparable performance without them.