Optimizing Intersection Navigation With Multishot Update Mechanism: A Real-Time Distributed Reinforcement Learning for Ground and Aerial Moving Objects

초록

As the Internet of Things (IoT) reshapes modern transportation, there is a growing need for an intelligent navigation strategy that addresses the challenges of various moving objects (MOs) like ground vehicle, indoor robots, and unmanned aerial vehicles (UAVs), while also overcoming recent global positioning system (GPS) vulnerabilities. These vulnerabilities include signal unavailability in obstructed environments, geopolitical or institutional dependencies of GPS services, and risks of centralized system failures, such as single-node disruptions. To address these issues, we introduce UDMSU-distributed reinforcement learning (DRL), a real-time user-demand-based DRL algorithm for intersection navigation, incorporating a multishot update mechanism. Our decentralized, GPS-independent approach meets real-time user preferences, such as time, energy, and safety-while ensuring private navigation by restricting data exchange to localized road segment interactions with the intersection agent. The proposed multishot Q-table update (MSQU) mechanism enhances data efficiency by reusing Q-values across multiple Q-table updates, thereby accelerating learning convergence. Extensive experiments show that our method effectively adapts to dynamic environments with diverse user demands, while reducing computational complexity and memory usage compared to other methods.

키워드