DNLight: Learning Efficient Evaluation for Traffic Signal Control

Traffic signal control (TSC) is an effective way to alleviate traffic congestion. Since traditional methods cannot adapt to complex and changeable dynamic traffic flows, reinforcement learning (RL) methods have attracted widespread attention from researchers. However, due to the lack of efficient policy evaluation modules, the evolution process faces the challenges of long learning time and slow convergence speed. To achieve efficient policy evaluation and speed up the evolution process, we propose a model, called as DNLight, which uses advanced traffic state design and dueling network to facilitate evaluation. Specifically, for a target intersection in the network, DNLight can not only express the phase need of queuing vehicles and running vehicles more accurately, but also evaluate the advantage of the strategies and promote the evolution of advantageous strategies. We conduct experiments on real-world datasets, and the results show that our proposed model has better performance (an average improvement of 3.65%) and stronger stability compared to the best baseline method.