Traffic Signal Control Based on Reinforcement Learning with Graph Convolutional Neural Nets

Traffic signal control can mitigate traffic congestion and reduce travel time. A model-free reinforcement learning (RL) approach is a powerful framework for learning a responsive traffic control policy for short-term traffic demand changes without prior environmental knowledge. Previous RL approaches could handle high-dimensional feature space using a standard neural network, e.g., a convolutional neural network; however, to control traffic on a road network with multiple intersections, the geometric features between roads had to be created manually. Rather than using manually crafted geometric features, we developed an RL-based traffic signal control method that employs a graph convolutional neural network (GCNN). GCNNs can automatically extract features considering the traffic features between distant roads by stacking multiple neural network layers. We numerically evaluated the proposed method in a six-intersection environment. The results demonstrate that the proposed method can find comparable policies twice as fast as the conventional RL method with a neural network and can adapt to more extensive traffic demand changes.