A two-level collaborative multi-agent reinforcement learning framework for variable speed limit control of freeway tunnel groups

Compared to freeway mainline sections, freeway tunnels present a higher risk of collisions and pose greater challenges for rescue operations, owing to the distinctive features of tunnel lighting and structures. Moreover, the safety demands for tunnel groups, which consist of multiple tunnels, are more pronounced. Recent studies have demonstrated the promising application of reinforcement learning in implementing variable speed limit (VSL) control on freeways to enhance safety and efficiency. However, existing approaches cannot adequately capture the interaction of traffic flows across different tunnels in group scenarios, nor can they efficiently train policies under highly complex environments. This study proposes a VSL strategy based on a two-level collaborative multi-agent reinforcement learning (TCMARL) framework to improve the safety and efficiency of freeway tunnel groups. The framework realizes coordinated control at two levels: (1) capturing potential spatial dependencies among tunnel agents through a spatial graph model, thereby achieving synergy at the information input level; and (2) constructing a mixing network that fuses features of different tunnel agents and introducing an importance weight vector to optimize the global Q-value output, achieving synergy at the action output level. To improve training efficiency and policy generalization, a model-based reinforcement learning mechanism is further incorporated to generate short-horizon virtual rollouts. Simulation experiments are conducted in four tunnel-group scenarios, including two-, three-, four-, and five-tunnel configurations. Comparative results show that the proposed framework achieves faster convergence and better safety and efficiency performance than benchmark methods. Additional ablation and sensitivity analyses confirm the complementary roles of the two-level collaboration design and the model-based planning mechanism, while out-of-training testing under unseen demand profiles demonstrates robust transferability. Beyond algorithmic performance, the findings provide policy relevant evidence for corridor level speed management in smart freeway tunnel systems, informing practical guidelines on coordinated VSL deployment and operational governance.