Dynamic Speed Control in Highway Bottlenecks Using Deep Reinforcement Learning

To enhance the traffic efficiency and safety of vehicles in highway bottleneck areas, this paper constructs a typical scenario of three-lane merging into two - lane on the SUMO simulation platform and proposes a dynamic speed-limit control method based on the Deep Q Network (DQN). The traffic state vector is composed of macroscopic quantities such as the average cross-section occupancy and average traffic speed. The action space is a discrete set of speed - limit values. The reward function comprehensively considers the total system time, total travel time, speed ratio, and queue length, aiming to actively adjust the vehicle speed, suppress shock waves, and prevent bottleneck spread before congestion occurs. The method adopts experience replay and target networks to improve training stability, and enables online interaction between the agent and SUMO through the traci interface to achieve closed - loop control of state collection, action issuance, and reward feedback. Compared with the fixed speed - limit and uncontrolled baselines, the experimental results show that the proposed method can significantly reduce the total delay and average number of stops, and effectively reduce energy consumption and pollution. It can maintain good robustness under different traffic flow intensities and CAV penetration rates.