Cooperative Lane-Changing Decision-Making Method for Connected and Automated Vehicles in Highway Merging Areas Based on Multi-Agent Deep Deterministic Policy Gradient

Abstract

Mandatory merging at expressway lane reductions often involves intensive inter-vehicle interactions, triggering lateral conflicts and longitudinal fluctuations that lead to traffic instability. This paper proposes a cooperative lane-changing decision-making framework for connected and automated vehicles (CAVs) using the multi-agent deep deterministic policy gradient (MADDPG) algorithm. The method integrates ego-vehicle states with the relative motion data of surrounding vehicles to establish a collaborative perception state space. Within the action space, the framework jointly optimizes lane-changing and car-following decisions of the merging vehicle while coordinating the longitudinal control of the target-lane following vehicle. A novel multi-objective reward function is introduced, balancing collaboration, safety, efficiency, and passenger comfort. Using a Python-SUMO co-simulation platform, the paper evaluates the proposed method against multi-agent proximal policy optimization and deep Q-network under free-flow and synchronized-flow conditions. Results indicate that, under free-flow conditions, the proposed method reduces collision risk by up to 1.27%, shortens lane-change decision time by up to 2.51 s, and increases the lane-change execution rate by up to 4.59%. Under synchronized-flow conditions, the method reduces collision risk by up to 1.16%, shortens lane-change decision time by up to 2.13 s, and increases the lane-change execution rate by up to 5.32%. Furthermore, the approach significantly enhances the level of coordination among CAVs. These findings demonstrate that the proposed method effectively balances safety, efficiency, and stability across varying traffic conditions, providing a practical technical solution for vehicle-infrastructure coordination and autonomous driving decisions in bottleneck zones.