Deep reinforcement learning-based dynamic green wave control strategy

The complexity of the existing urban road network is high, and the cost of coordinated control of the entire road network with green waves is high, green wave sections are now widely used in cities. Therefore, aiming at the limitations of traditional green wave control methods, this paper proposes a dynamic green wave coordinated control model based on DQN, which aims to optimize the traffic efficiency of green wave sections in urban road networks. Deep reinforcement learning technology is used. In the model, the intelligent body (intersection) realizes real-time perception and response to the dynamic traffic environment by learning the experience of traffic condition and historical decision making. By optimizing the phase difference and public signal period, the model is able to carry out intelligent green wave coordinated control of urban arterial roads under different time periods, traffic flows and bus demand conditions. Finally, simulations are created using SUMO to validate the offline coordinated control scheme, genetic algorithm and dynamic model proposed in this paper. The simulation results show that the dynamic control model coordinated with the green wave reduces the average vehicle deceleration by 27.8% and 8.3%, the lane occupancy by 50.8% and 44.8%, and the total number of starts and stops by 54.3% and 49% reduced upside direction compared to the two schemes. In the downstream direction, there was a reduction in average delay by 23.6% and 6.6% respectively, lane occupancy by 27.9% and 2.3%, and total starts and stops by 38.7% and 17.5%.