@article{M20926B0E,
title = "Offline Reinforcement Learning Based UAV Training for Disaster Network Recovery",
journal = "The Journal of Korean Institute of Communications and Information Sciences",
year = "2024",
issn = "1226-4717",
doi = "10.7840/kics.2024.49.1.1",
author = "Jeyeon Eo, Dongsu Lee, Minhae Kwon",
keywords = "Unmanned Aerial Vehicle, Reinforcement Learning, Offline Reinforcement Learning, Sparse Reward Environment",
abstract = "In disaster scenarios, ongoing research seeks to swiftly restore disrupted networks using unmanned aerial vehicles (UAVs) when conventional wired and wireless infrastructures falter. Current research predominantly relies on online reinforcement learning, wherein UAVs acquire behavioral policies through real-time interactions.
However, network restoration in three-dimensional spaces presents formidable challenges due to the scarcity of reward signals in low-probability success scenarios, rendering traditional reinforcement learning approaches less effective. To address these challenges, this paper proposes an approach that integrates Long short-term memory (LSTM) into offline reinforcement learning, utilizing a fixed pre-collected dataset to enable safe policy learning without direct real-world interaction. The LSTM's capability to assign rewards to action sequences contributing to success facilitates smoother policy development even within sparse reward environments. Empirical simulation experiments confirm the effectiveness of our method in enabling UAVs to efficiently recover partial network loss."
}