Variable Speed Limit Control Based on Deep Reinforcement Learning: A Possible Implementation

Today’s urban motorways cannot fulfill their purpose to simultaneously serve transit and local urban traffic with a high Level of Service. In the case of urban motorway infrastructure, the traditional “build only” approach is not always possible due to the lack of space. This study is focused on the Variable Speed Limit Control (VSLC) as one of the traffic control methods applicable for any type of motorway and Q-learning as one commonly used approach for designing learning based VSLC algorithms. The drawback of this methodology is the representation and exploration of the large state-action space as it is the case in its application for VSLC. This study introduces a Deep Q-Network to approximate the Q-function and presents a novel learning approach for the VSLC application with possibility to track vehicles on the microscopic level. The proposed reward function steers the learning towards the improvement of reward and prevention of oscillation among consecutive speed limits.