Multi-Agent Deep Reinforcement Learning for Regional Traffic Signal Control Based on Dynamic Weight Decomposition

Conventional traffic signal control methodologies are deficient in adapting to rapid traffic flow variations and capturing the complex dynamic interactions between intersections within regional road networks. In order to address this specific issue, the present study proposed the Qatten (Q-value Attention Network)-TSC algorithm. The algorithm was constructed on the basis of the dynamic weighted value decomposition principle and was built upon the multi-agent QMIX (Q-value Mixed Network) framework. The model employed a multi-head attention mechanism to effectively fuse individual agent Q-values with global states and individual features to compute global Q-values. Furthermore, the model incorporated multidimensional state information to comprehensively characterize complex traffic networks. Extensive experiments were conducted on small- and large-scale SUMO simulation platforms based on the real road network of Yangzhou. The experimental results demonstrated that in comparison to VDN and QMIX, Qatten-TSC attained average reward increments of 26.4% and 3.46%, correspondingly, in small-scale road networks, and 34.12% and 12.81%, correspondingly, in large-scale road networks. Furthermore, in large-scale scenarios, the average time loss was reduced by 19.28% and 7.15%, respectively, while the average speed increased by 3.00% and 0.87%, respectively. In addition, the baseline algorithm (Qatten) is unstable and poor-performing. The dynamic weighting mechanism is robust and effective, even as the road network complexity increases.