This paper introduces CityTransfer, a framework for effective transfer learning in multi-agent traffic signal control across heterogeneous city networks. It integrates a behavioral encoder, a meta-learner, and an adaptive coordination mechanism to achieve superior zero-shot transfer performance and stable control across varying intersection geometries and traffic patterns.
Key findings
CityTransfer reduces adaptation time by 65% compared to learning from scratch.
The framework maintains stable control across different intersection geometries and traffic patterns.
CityTransfer demonstrates superior transfer performance on synthetic and real-world networks.
Limitations & open questions
The framework's effectiveness may be limited in cities with highly unique traffic patterns not seen during training.
The generalization capabilities across very diverse urban environments require further validation.