Urban Traffic Signal Control at the Edge: An Ontology-Enhanced Deep Reinforcement Learning Approach

Static or traditional rule-based urban traffic signal control approaches are inadequate to handle ever-changing and stochastic urban traffic. Recent works in Intelligent Traffic Signal Control (ITSC) have extensively explored Deep Reinforcement Learning (DRL) to enhance traffic signal adjustment based on real-time traffic. However, DRL-based ITSC is data-driven, and with current communication technologies, especially in dense urban traffic networks, it is unrealistic to assume perfect data transmission. Partial detection of vehicles, noisy sensor data, and unexpected traffic events should be addressed promptly. Ontology-based Intelligent Traffic Signal Control (OITSC) was proposed to address these challenges and validated under synthetic grid maps. In this paper, we have improved the original OITSC to handle more complicated urban scenarios using DRL. A road network in Dublin city center is selected to simulate realistic urban traffic scenarios. The reduction in the average vehicle waiting time demonstrates the efficiency of OITSC in improving urban traffic under partially observable and noisy traffic information, with the inclusion of prioritized vehicles in the face of unexpected road events. We have also provided an edge-computing deployment strategy, which is expected to provide robust and fast responses for complex urban traffic, especially for unexpected cases.