HiLight: Heterogeneous Traffic Signal Control for Automatic Drive Guidance Based on Multi-Agent Reinforcement Learning

Self-driving cars need to pre-sense traffic light status for better route planning. However, the heterogeneity and complexity of modern urban intersections (including crossroads and intersections) pose challenges to traffic signal control systems. Traditional traffic signal control methods based on multi-intelligence reinforcement learning are often based on four-way intersections, while ignoring the more complex traffic network conditions in the real world. To address this problem, this paper designs HiLight, a small-scale heterogeneous intersection scheduling system, which performs a finer-grained division of intelligences for heterogeneous traffic intersections, designs the rewards, states, and actions of the intelligences, extracts the information in the environment through an embedding layer, and uses the MADDPG algorithm to make decisions on actions. This system aims to develop intelligent signalling systems that can accurately understand and effectively manage diverse and heterogeneous traffic conditions, thus better adapting to various complex urban traffic environments.