Abstract
In the aftermath of natural disasters, cellular infrastructure is often compromised, leaving survivors and first responders without critical connectivity. This paper presents UnVAIL (UAV-Assisted Imitation Learning), a novel framework that utilizes Unmanned Aerial Vehicles (UAVs) as mobile communication relays. By employing behavioral cloning to replicate human-expert scheduling logic, UnVAIL optimizes data buffer management and energy consumption. Experimental results demonstrate that the system achieves 97.52% accuracy in mimicking expert decisions with sub-millisecond latency, making it suitable for real-time deployment on low-power hardware.
Introduction
Natural disasters such as wildfires or floods frequently disable ground-based cellular infrastructure. In these scenarios, UAVs can serve as "flying base stations" to restore connectivity. However, managing these mobile relays involves complex trade-offs between ground user equipment (UE) buffer management, UAV energy constraints, and the computational overhead of real-time optimization.
Traditional mathematical optimization methods are often too slow for onboard processing, while standard Reinforcement Learning (RL) suffers from long training times that are impractical during active emergencies. UnVAIL addresses these challenges by using Imitation Learning (IL) to provide a fast, model-free control mechanism.
Methodology: The UnVAIL Framework
The core of UnVAIL is a Behavioral Cloning (BC) approach that trains a Deep Neural Network (DNN) to mimic a high-compute expert simulator.
The Expert and the Learner
The expert simulator uses human-like logic to optimize an Energy/Delay Throughput (EDT) utility function, deciding which user to serve based on queue lengths and spatial proximity. The learner is a DNN with three hidden layers (80, 160, and 80 neurons) that maps the current status of all user buffers to the optimal next action.
Action Space and Mobility
The UAV operates in a circular orbit divided into 36 sectors (10° each). It can move clockwise, counter-clockwise, or hover. Unlike systems with fixed hover times, UnVAIL implements dynamic service times, adapting its stay based on real-time demand to maximize energy efficiency.
Experimental Setup and Results
The researchers developed a custom UAV-Relay Simulator in Python to generate 28 hours of expert interaction data (100,000 seconds), covering the delivery of approximately 550,000 packets.
Performance Metrics
- Accuracy: UnVAIL achieved a 97.52% accuracy in replicating expert scheduling decisions on a dedicated test set.
- Efficiency: The model's performance in terms of packet delivery rates and energy consumption was nearly identical to the expert's.
- Latency: Decisions are made in less than one millisecond, enabling real-time control on hardware like the Nvidia Jetson Nano.
Discussion and Future Work
While UnVAIL shows high fidelity to expert behavior, it faces challenges with distribution shift. When packet arrival rates were altered from the training distribution, accuracy dropped to 79.4%. This indicates that while behavioral cloning is highly efficient, it may require additional techniques like Dataset Aggregation (DAgger) or hybrid RL approaches to improve robustness in unseen environments. Future research will also explore multi-UAV coordination to scale coverage in larger disaster zones.
Summary updated