Recently, as traffic congestion in urban areas has increased, there are many studies on reinforcement learning for traffic signal control that enable efficient traffic control in a complex environment that is difficult for humans to control. However, ...
Recently, as traffic congestion in urban areas has increased, there are many studies on reinforcement learning for traffic signal control that enable efficient traffic control in a complex environment that is difficult for humans to control. However, most existing reinforcement learning for traffic signal control studies are implemented based on simulation, and there are few cases where they are applied to the real world. In addition, existing reinforcement learning for traffic signal control methods use a step method that controls the signal every short step. However, a step method is inefficient in oversaturated traffic conditions with large differences between movements because the signal cannot be controlled based on the overall situation of the movements.
Therefore, this study focuses on transferring simulation-based reinforcement learning for traffic signal control to reality and develops an reinforcement learning for traffic signal control method that can respond to oversaturated traffic conditions. The action space is designed so that the agent derives an optimal signal set for every cycle length by understanding the traffic situation of all movements. During each cycle length performing signal optimization, the proposed model finds the optimal signal in the iterative strategy search process. We developed a kinematic wave-based mesoscopic model for a fast and accurate strategy search process. Based on the collected traffic information, the kinematic wave-based meso model estimates the traffic information of the entire link and obtains the status and reward.
The proposed reinforcement learning for traffic signal control method has been verified for field applicability through demonstration in the real world at a congested intersection in Seoul, Korea. As a result, the average queue length at intersection was improved by up to 11.4%.