PairUpLight: A Multi-agent Reinforcement Learning Approach for Coordinated Multi-intersection Traffic Signal Control

The management of heavy traffic demands has been significantly improved by employing synchronized traffic signal control at multiple intersections. Multi-agent Reinforcement Learning (MARL) techniques have been widely utilized to achieve this coordination. However, these approaches predominantly depend on manually crafted features from adjacent intersections, which impedes their generalization to new scenarios. Furthermore, while displaying high accuracy for specific traffic flow patterns, these methods often lack the necessary robustness for other patterns. In this study, our objective is to develop an effective signal timing plan by directly learning the minimal required communication between intersections from traffic data. We introduce a novel, comprehensive approach that combines multi-agent reinforcement learning with a learned communication mechanism. Our model incorporates a coordinated actor network and a centralized critic network to address the challenges of non-stationarity. We conducted extensive experiments comparing our model with other commonly used non-RL and benchmark MARL techniques. The evaluation results show that our proposed model, which relies only on local sensory input and a single message from neighboring intersections, excels in managing various traffic flow patterns. Furthermore, our model outperforms competing approaches in terms of robustness, resilience, and overall performance.