Networked Traffic Signal Control (NTSC) has become an essential component in Intelligent Transportation Systems (ITS). To satisfy both scale and coordination challenging requirements for large-scale networked traffic signal control, this paper proposes a Communication-Qmix (CQmix) approach based on Qmix and Long Short-Term Memory (LSTM) communication module correspondingly. Firstly, we apply Qmix as the foundation for balancing large-scale and effective optimization, benefifiting from its centralized-training and decentralized-execution mechanism. Then a communication module based on LSTM is implemented for effective global coordination. Further, random origin-destination demands (ODs) about different maximum traffific flflows and occurrence times are performed to adapt the actual traffific flflow pattern in practice. We conduct experiments on both synthetic and complicated real Zhongguancun road networks, and the proposed CQmix demonstrates its superiority over the baseline methods.

Networked Traffic Signal Control (NTSC) has become an essential component in Intelligent Transportation Systems (ITS). To satisfy both scale and coordination challenging requirements for large-scale networked traffic signal control, this paper proposes a Communication-Qmix (CQmix) approach based on Qmix and Long Short-Term Memory (LSTM) communication module correspondingly. Firstly, we apply Qmix as the foundation for balancing large-scale and effective optimization, benefifiting from its centralized-training and decentralized-execution mechanism. Then a communication module based on LSTM is implemented for effective global coordination. Further, random origin-destination demands (ODs) about different maximum traffific flflows and occurrence times are performed to adapt the actual traffific flflow pattern in practice. We conduct experiments on both synthetic and complicated real Zhongguancun road networks, and the proposed CQmix demonstrates its superiority over the baseline methods.