MADRL for CAVs at Tomorrow’s Intersections

Full title—Multi-Agent Deep Reinforcement Learning to Manage Connected Autonomous Vehicles at Tomorrow’s Intersections

In this paper, we present a new advanced approach to AIM based on end-to-end Multi-Agent Deep Reinforcement Learning (MADRL) and trained via Curriculum through Self-Play, called advanced Reinforced AIM (adv.RAIM). adv.RAIM can autonomously learn the complex dynamics of real-life traffic, and then provide a new path for building smarter AIM capable of proactively controlling CAV speed at intersections.

The results show impressive improvements when compared to traffic light control techniques, as well as better performance than other recently proposed AIMs, highlighting the advantages of using MADRL.

Full Article: IEEE Transactions on Vehicular Technology, Early Access, April 2022