基于多智能体的大规模路口交通信号灯协同控制研究
Research on Multi-Agent Based Cooperative Control of Traffic Lights at Large-Scale Intersections

随着城市化建设加快，车辆数量增加，道路负载变大，交通拥堵问题日益严重，目前大多数的控制方法仍局限于孤立的路口，整体路网拥挤程度仍存在较大的优化空间。因此，本文以交通信号灯控制为研究对象，深度强化学习算法为基础，针对基于多智能体的交通信号灯协同控制问题展开研究。首先将信号灯协同控制问题抽象为马尔可夫决策过程，基于Python开发平台搭建智能体交互环境，设计大规模路口信号灯决策下的DQN算法并进行调试运行，结果表明算法训练出了有效的红绿灯控制策略，并在各个路口都能够学习到公平策略，能够提高路网整体通行效率。最后通过与传统的信号固定配时方案进行对比实验，验证算法具有良好的优化性能。通过调整算法的超参数对训练结果进行对比分析，研究不同超参数对网络训练的影响，以及超参数对项目研究的重要性。
With the continuous acceleration of urbanization construction, the number of vehicles continues to increase, the road load gradually increases, and the problem of urban traffic congestion is becoming increasingly serious. Currently, most control methods are still limited to isolated intersections, and there is still significant room for optimization of the overall road network congestion level. Therefore, this paper takes the traffic light control as the research object, and based on the Deep reinforcement learning algorithm, carries out relevant research on the multi-agent based traffic light cooperative control problem. Firstly, the problem of signal light collaborative control is abstracted as a Markov decision process. Based on the Python development platform, an intelligent learning and interaction environment is built, and the DQN algorithm for large-scale intersection signal light decision-making is designed and debugged. The results show that the algorithm has trained effective red and green light control strategies, and fairness strategies can be learned at each intersection, which can improve the overall traffic efficiency of the road network. Finally, through comparative experiments with traditional signal fixed timing schemes, the algorithm was verified to have good optimization performance. Compare and analyze the training results by adjusting the hyperparameters of the algorithm, study the impact of different hyperparameters on network training, and the importance of hyperparameters for project research.