Human-like Merging Control of Intelligent Connected Vehicles on the Acceleration Lane

To develop a merging control algorithm for intelligent connected vehicles (ICVs) on freeway acceleration lanes interacting with human-driven vehicles (HDVs) on the mainline, we propose a merging control model (DQN-RF). This model integrates the deep Q-network (DQN) algorithm and the random forest (RF) algorithm. First, a roadside data acquisition platform was established to collect the naturalistic merging behavior data of HDVs at a typical merging zone with an acceleration lane on the G70 freeway in China. Second, a human-like merging decision model using RF was built using historical merging environmental contextual data and the merging urgency of the merging vehicle on the acceleration lane as input. We constructed a simulated merging scenario featuring an acceleration lane on the freeway using the simulation of urban mobility (SUMO) platform. Utilizing the Python language, we developed a testing script environment for the deep reinforcement learning algorithm. Additionally, we introduced a longitudinal acceleration control algorithm based on DQN. Finally, the DQN-RF merging control model, which embedded the RF merging decision algorithm into the DQN longitudinal acceleration control algorithm, was established to embrace merging decision control and longitudinal acceleration control in a comprehensive framework. The default lane-changing control algorithm in SUMO, known as "LC2013," was also combined with the proposed DQN algorithm to serve as a baseline model. The simulation results show that, with the same acceleration action value space [ - 1, 2] m • s~z , compared to the DQN-LC2013 model, the DQNRF model receives a higher total reward value. The average accelerations of the ICV for the DQN-RF and DQN-LC2013 models are 0. 55 and 0. 09 m • s~z , respectively. Furthermore, the average speeds are 21. 4 and 19. 7 m • s_ 1 , respectively. There are no stop-and-wait phenomena observed when the DQN-RF model is adopted, while there are seven stop-and-wait events in 100 turns of simulation when the DQN-LC2013 model is adopted. The proposed DQN-RF merging control model can realize human-like merging decisions and improve the merging efficiency and success rate of the ICV. The DQN-RF model can be used for merging decision control and longitudinal acceleration control of the ICVs on the freeway acceleration lane. © 2024 Chang'an University. All rights reserved.