Deep Reinforcement Learning to Minimize Traffic Congestion with Emergency Facilitation

Abstract

In the domain of intelligent traffic light control, which real-time traffic data to consider has a huge impact on the efficiency and performance of the traffic light control system. The rewards and state representations used in previous studies can mislead the agent in some cases. This paper examines those problems and proposes a solution using the standard deviation of the vehicle waiting time. Existing studies have not yet provided emergency facilitation. This paper proposes a method that provides emergency facilitation. The proposed method is self-evaluated with another version of the proposed method under both synthetic and real-world data, and it has proven that consideration of standard deviation has a significant impact on performance. The proposed method is also evaluated with a statistical method and a fixed time and has outperformed both of them. Buy considering vehicle type it was able to approximate the emergency vehicle waiting time to zero which was initialy at 20s when starting the training. With the help of standard deviation of waiting time, It was able to approximate the regular vehicle waiting time to 21s which was initialy at 60s when starting the training. The proposed method was able to record 21.588s of average waiting time of regular vehicles at the testing phase outperforming against methods.