Adaptive traffic light control based on reinforcement learning under different stages of autonomy

With the development of big data, artificial intelligence and vehicle road coordination, China intelligent transportation system is changing to autonomous transportation system. During the development of autonomous transportation, the traffic application scenarios will also change, which will affect the control efficiency of the same control strategy. In order to show the impact of the traffic control strategy on traffic in different stages of autonomous traffic, we use edge computing to simulate different stages of autonomous traffic, by establishing three scenarios of user-end computing, roadside-end computing and cloud computing, which respectively correspond to the three autonomous traffic stages of auxiliary autonomy, high autonomy and full autonomy. With the extensive application of reinforcement learning in the traffic field, deep Q-Learning network(DQN) has become a effective solution of traffic signal adaptive control. We use DQN to adaptively control the traffic lights, and through comparison of different scenarios, we show the differences of the same control strategy under different stages of autonomy. Through the experiment, we can conclude that the expected reward of user-end computing is 6.7% and 19.8% higher than roadside computing and cloud computing. Meanwhile, it shows that under the condition of full autonomy, the sensitivity and robustness of the control strategy are improved, effectively improving the traffic efficiency.