GreenLight: Green Traffic Signal Control using Attention-based Reinforcement Learning on Fog Computing Network

Traffic congestion is a pervasive problem causing severe environmental and economic issues. In recent years, traffic signal control using reinforcement learning (RL) has come a long way. Most existing studies focus on using distributed agents with data exchange among neighbors, which, however, increases network complexity and usage and still suffers from the lack of broader coordination. Meanwhile, the attention mechanism has achieved tremendous success, and advances in vehicle-to-infrastructure (V2I) communications have enabled real-time collections of granular data. However, integrating these technologies into traffic signal control remains under-explored. Therefore, we present GreenLight, a forward-thinking and eco-friendly traffic signal control framework that can be applied to V2I-equipped fog computing environments. For a large urban area, traffic signals are divided into clusters, each coordinated by a fog node with an RL agent. Intra-cluster indexed self-attention is applied to extract context-aware features that the fog-residing RL agent utilizes to determine the proper signal control command. Results of simulation experiments using both synthetic and real-world scenarios show that the presented framework yields lower waiting time, emissions, and fuel consumption compared to baseline methods, indicating its potential for next-generation transportation systems.