Multi-mode Light: Learning Special Collaboration Patterns for Traffic Signal Control

To alleviate traffic congestion, it is a trend to apply reinforcement learning (RL) to traffic signal control in multi-intersection road networks. However, existing researches generally combine a basic RL framework Ape-X DQN with the graph convolutional network (GCN), to aggregate the neighborhood information, lacking unique collaboration exploration at each intersection with shared parameters. This paper proposes a multi-mode Light model that learns the general collaboration patterns in a road network with the graph attention network and trains simple Multilayer Perceptron for each intersection to capture each intersection's unique collaboration pattern. The experiment results demonstrate that our model improves average by 27.19% compared with the state-of-the-art transportation method MaxPressure and average by 4.57% compared with the state-of-the-art reinforcement learning method Colight.