Regional Intelligent Traffic Signal Control System Based on Multi-agent Deep Reinforcement Learning

Urban traffic congestion is becoming a severe problem. A reinforcement learning-based intelligent traffic signal control system can alleviate traffic congestion problems. However, when reinforcement learning faces complex multi-agent decision problems, it is difficult for single-agent reinforcement learning algorithms to realize the complex relationships among multiple agents. Therefore, integrating reinforcement learning with multi-agent techniques in multiple ways is becoming inevitable. Based on reinforcement learning, multi-agents obtain optimal regional strategies through continuous interaction and improvement. This paper proposes a new approach to regional intelligent traffic signal control combining multi-agent techniques and deep reinforcement learning. The method effectively reduces the average trip waiting time (ATWT) and the total waiting queue length (TWQL) of vehicles in the region while performing better on the reward function compared to the deep Q network (DQN) algorithm. This approach improves traffic efficiency and increases resource utilization efficiency while reducing congestion time, road utilization, and resource waste.