A DRL-Based Multi-Modal Local Congestion Index Method for Mixed On-Ramp Merging

The safety and efficiency of entrance ramp merging are key factors in relieving traffic congestion. Congestion near the merging point is a propagation process, and in hybrid autonomy, connected and autonomous vehicles (CAVs) need to instantly perceive local congestion and devise reasonable strategies to alleviate ramp congestion, a factor often overlooked by reinforcement learning-based methods. In this paper, we propose a deep reinforcement learning (DRL)-based multi-modal local congestion index (MM-LCI) method, which consists of a local congestion index method and a multi-modal adaptive layer. The local congestion index method considers both vehicle observations and local traffic flow states to develop more effective driving strategies. The multi-modal adaptive layer optimizes queuing strategies by distinguishing different modes through combinations of CAVs and various types of vehicles. Extensive experiments conducted on the simulation of urban mobility (SUMO) platform demonstrate that the proposed method outperforms state-of-The-Art benchmark methods under various traffic flows and penetration rates. Compared with Flow, the MM-LCI model improves average speed by 27.86% and reduces average waiting time by 13.25%. No significant reduction in safety was observed, effectively alleviating congestion at the entrance ramp.