Abstract
Deep Reinforcement Learning (DRL) has become a widely adopted approach for adaptive traffic signal control, owing to its flexibility and effectiveness. However, DRL faces dimensionality and convergence challenges in large-scale multi-agent settings. Additionally, simplified representations of traffic conditions limit adaptability under mixed environments with pedestrians and weather factors. Moreover, in Vehicular Ad-Hoc Networks (VANETs) scenarios, low Vehicle-to-Infrastructure (V2I) penetration exacerbates the complexity of traffic observation and coordination. In this paper, we propose a DRL-based robust adaptive traffic signal control with a low V2I penetration rate under mixed environment (DRATSC). Specifically, DRATSC builds upon the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) framework, modeling intersections as autonomous agents and leveraging V2I to obtain real-time data. An attention mechanism is employed to dynamically focus on neighboring intersections, optimizing traffic signal coordination. The method overcomes low V2I penetration through scenario reconstruction and models mixed traffic environments by incorporating weather impacts, pedestrian interactions, and diverse action spaces. Extensive experiments conducted in the Simulation of Urban Mobility (SUMO) platform show that the proposed method improves overall performance by 40% over conventional approaches and outperforms IPDALight across most scenarios.
Funding Information
  • National Natural Science Foundation of China (62402304)
  • Shanghai Sailing Program (24YF2713100)
  • National Key Research and Development Program of China (2023YFB2504800)

This publication has 35 references indexed in Scilit: