Traffic Signal Control Approach Combining Qmix and Self-Attention Mechanism

In recent years, to alleviate the immense pressure on urban traffic networks caused by the continuous rise in vehicle ownership in cities, numerous studies have attempted to apply multi-agent reinforcement learning algorithms to control and coordinate urban network traffic signals. This paper introduces a novel method termed Double Self-Attention Qmix (DSA-Qmix), which integrates the Qmix algorithm with a self-attention mechanism. By leveraging the intrinsic ability of Qmix to consider both local and global information, combined with the self-attention mechanism, the DSA-Qmix enhances the representation capability of the global value function and the observation space of agents. This enables it to effectively adapt to large-scale urban traffic network signal control scenarios. The experimental results demonstrate that the proposed DSA-Qmix method performs well under various road conditions and traffic volume changes, outperforming other baseline methods in both synthetic and real-world network simulations.