A dual-network DRL approach for traffic flow modelling and AV driving strategy in expressway merging areas under AV–HDV mixed traffic environment

Deployment of autonomous vehicles (AVs) is transforming expressway merging zones into complex mixed-traffic environments. Multi-agent reinforcement learning enables vehicle cooperation but is demanding in computational and infrastructural costs, while prevailing deep reinforcement learning (DRL)-based single-vehicle intelligence approaches rely on single network to generate both car-following and lane-changing decisions, introducing coupled rewards and reducing learning efficiency. This study presents a dual-network DRL approach for AV driving strategy in merging zones. Two dedicated networks independently optimize lane-changing and car-following decisions, supported by coupled and action-specific rewards. Evaluated on Jungong Road on-ramp merging section in Shanghai against four baselines, the DDRL strategy improves efficiency, safety, and comfort by up to 12.4%, 79.8%, and 5.3% over rule-based and single-vehicle intelligence baselines. When compared to multi-vehicle cooperation baselines, it requires only half the computation time with acceptable under-performance of less than 5%, featuring a robust balance between traffic operation performance, computational efficiency and deployment feasibility.