A General DRL-Based Framework Using Mode-Selection Tangent Time Projection for Mixed On-Ramp Merging

Safety and efficiency are essential factors in the on-ramp merging scenario. In a mixed-autonomy traffic, different queuing methods can greatly affect both safety and efficiency of the whole network, effect of which has been neglected by the model-free method. In this paper, we propose a general DRL-based framework called Mode-Selection Tangent Time Projection(MS-Ttp). By comparing to the classical spatial queuing method, the Tangent Time Projection method considers the instantaneous velocity of vehicles and their distances to the merging point, and makes a preliminary ranking. The Mode-Selection Adaptive Layer forms different modes through several combination of vehicles to optimize the time headway. Extensive experiments are performed using SUMO simulation platform, and the proposed framework has a better performance in different flow rate ratios and different penetration rates than the state-of-the-art baselines. Comparing with the IDM model, the proposed MS-TTP framework optimizes the average velocity by 74.80% and the on-ramp waiting time by 25.13%, greatly alleviate the congestion in the on-ramp merging scenario.