Deep reinforcement learning for multi-Objective latency minimization in 5G heterogeneous vehicular networks

Modern intelligent transportation systems rely fundamentally on vehicular networks capable of ultra-responsive communication, a capability brought into reach by 5G and beyond technologies for applications such as autonomous driving. A persistent obstacle within these fast-changing, mixed-technology networks is the dual and often competing requirement to slash both data transmission delays and power usage, a multi-faceted optimization that existing solutions, focused on singular objectives, fail to address adequately. Confronting this trade-off, we present SONG-PO-DRL, a new hybrid framework that fuses a swarm-optimized non-dominated sorting genetic algorithm with deep reinforcement learning to drive adaptive, multi-criteria decision-making. Evaluation with seven algorithms as baselines, using simulation implemented with OMNET++, Simu5G, Veins, and SUMO, shows that this approach surpasses current state-of-the-art methods in critical areas such as latency and competitively; energy consumption, cluster head longevity, congestion levels, and processing overhead. By providing a foundational design strategy, this framework enables the development of higher-performance, more stable, and scalable vehicular networks, thereby contributing directly to the evolution of dependable and energy-conscious future transportation infrastructures.