DACLight: traffic signal control with deep reinforcement learning for actual deployment

Reinforcement learning for optimizing traffic signal control is becoming increasingly popular, but the actual deployment of reinforcement learning algorithms poses certain challenges. The difficulties primarily lie in the following aspects:(1) The detection equipment installed at most intersections are mainly geomagnetic coils and video detectors and the detection range is limited. (2) Common academic traffic control algorithms usually allow random phase switching of signal phases. However, in practice, considering the habits of drivers and the existence of left-turn waiting lane, the application of random switching signal control mode may introduce potential risks. Therefore, it is necessary to use traffic signal control methods with fixed-phase sequence. In this paper, we propose the DACLight to address the deployment problems. Under the assumption of fixed phase sequences, DACLight utilizes the attention mechanism to make decisions on traffic signal control strategies based on queue lengths of each lane obtained from intersection detectors. Experimental results conducted in both the real-world dataset scenario in Zhengzhou and the public dataset scenario in Hangzhou demonstrate that the algorithm outperforms other baseline experiments.