Wireless Communications Meets Artificial Intelligence: An Illustration by Autonomous Vehicles on Manhattan Streets

摘要

Interactions of multiple smart agents serve a fundamental aspect of Internet of Things (IoT), or known as social IoT. Such smart agents are equipped with sophisticated machine learning for mobile operation, such as autonomous vehicles and robots. Although wireless networking is intuitively important in such scenarios, there lacks investigations to provide a holistic and in-depth understanding on wireless networked multi-agent systems. In this paper, we disruptively use reinforcement learning to model each agent of artificial intelligence, and explore the interplay between wireless communications and multi-agent systems. Autonomous vehicles navigating over Manhattan streets serve the illustrating system. The first new finding is the need to modify reinforcement learning of policy exchange due to getting information from other agents through wireless communication. The advantage of applying wireless communication is clearly observed. We also demonstrate the impacts of communication errors to result in penalty in system performance. Multi-agent systems equipped with direct vehicle- to-vehicle communication and vehicle-to-infrastructure communication are compared to initially conclude favorable using infrastructure of small cells. Finally, we explore multiple access communication over multi-agent systems by employing real- time ALOHA. Different from traditional thinking on reliable delivery of packets using re-transmit after collisions, real-time ALOHA discards re-transmission mechanism to ensure in-time contributions from wireless communication on the learning algorithm of a multi-agent system with satisfactory performance.