AIM5LA: A Latency-Aware Deep Reinforcement Learning-Based Autonomous Intersection Management System for 5G...

Area of Science:

• Intelligent Transportation Systems
• Artificial Intelligence
• Networked Systems

Background:

• Autonomous vehicles (AVs) require advanced intersection management for safety and efficiency.
• Existing Autonomous Intersection Management (AIM) systems often neglect communication network latency.
• Decentralized cooperative systems offer a path to collective intelligence for AVs.

Purpose of the Study:

• To introduce AIM5LA, a novel latency-aware AIM system utilizing deep reinforcement learning for 5G networks.
• To address the impact of communication latency on the decentralized control of AVs in AIM.
• To develop a robust and resilient multi-agent control policy for AVs at urban intersections.

Main Methods:

• Implementation of a Multi-Agent Deep Reinforcement Learning (MADRL) approach within the AIM5LA framework.
• Integration of 5G communication network latency awareness, including historical data and future predictions.
• Development of a decentralized cooperative system for controlling AVs at urban intersections.

Main Results:

• AIM5LA demonstrated significant safety improvements, eliminating collisions entirely (average reduction from 27 to 0).
• The system achieved comparable performance in travel time and intersection waiting time metrics.
• AIM5LA guarantees collision-free operation, outperforming other AIM systems.
• Compared to traffic light systems, AIM5LA reduced waiting time by over 99% and time loss by over 95%.

Conclusions:

• AIM5LA offers a robust and resilient solution for autonomous intersection management by incorporating 5G network latency.
• The proposed system significantly enhances AV safety and traffic efficiency in urban environments.
• Latency-aware AIM is crucial for the successful deployment of fully autonomous vehicle systems.