Joint Resource Management for Energy-Efficient UAV-Assisted SWIPT-MEC: A Deep Reinforcement Learning Approach

Yue Chen, Hui Kang, Jiahui Li, Geng Sun, Boxiong Wang, Jiacheng Wang, Cong Liang, Shuang Liang, Dusit Niyato

TL;DR

This paper introduces a novel framework for 6G IoT networks where a UAV acts as both a flying power station and a mobile edge server. By utilizing a specialized Deep Reinforcement Learning (DRL) algorithm called SAC-SK—which integrates Kolmogorov-Arnold Networks (KAN) and Simple Recurrent Units (SRU)—the authors successfully balance UAV energy consumption with terminal battery longevity and fairness across remote IoT devices.

What Problem Does This Paper Address?

In the push toward 6G, many Internet of Things (IoT) devices will be deployed in remote or disaster-stricken areas where ground-based infrastructure is non-existent. These devices face two major hurdles: limited battery life and insufficient processing power for complex tasks.

While Simultaneous Wireless Information and Power Transfer (SWIPT) and Mobile Edge Computing (MEC) are promising solutions, combining them on a mobile platform like an Unmanned Aerial Vehicle (UAV) is incredibly difficult. The system must juggle a "trilemma":

  1. UAV Energy: Flying and transmitting data consumes the UAV's limited onboard battery.
  2. Terminal Sustainability: IoT devices need to harvest enough energy from the UAV to stay alive.
  3. Task Deadlines: Computational tasks must be processed (locally or offloaded) within strict time windows.

Traditional optimization methods struggle with the "hybrid" nature of this problem (deciding where to fly is continuous, while deciding which task to offload is discrete) and the non-linear realities of wireless charging.

Methods

The authors propose SAC-SK, an enhanced version of the Soft Actor-Critic (SAC) algorithm. Their approach features three key innovations:

  • Action Simplification Mechanism: To handle the complex hybrid action space, the authors use a greedy strategy. They theoretically prove that as a UAV approaches a terminal, energy costs for the terminal decrease while harvesting efficiency increases. By simplifying offloading decisions (directing communication to the nearest terminal with tasks), they reduce the problem to a more manageable continuous action space.
  • Simple Recurrent Units (SRU): Unlike standard LSTMs or GRUs, SRUs are lightweight and highly parallelizable. They are used here to capture long-term temporal dependencies in task arrivals and channel conditions without the usual computational bottleneck.
  • Modified Kolmogorov-Arnold Networks (KAN): Moving away from traditional Multi-Layer Perceptrons (MLPs), the authors use KANs. Instead of fixed weights on edges, KANs use learnable functions on the "edges" (univariate B-splines). This allows for much better mathematical fitting of the complex, non-linear relationships inherent in wireless power transfer and flight dynamics.

Datasets

The paper does not use a pre-existing external dataset. Instead, it employs a rigorous simulation environment based on established 6G channel models:

  • Terminal Distribution: Generated using Poisson disk sampling to create realistic network topologies.
  • Models: The simulation incorporates a probabilistic Line-of-Sight (LoS) air-to-ground channel model and a non-linear Logistic-function-based Energy Harvesting (EH) model, which is more accurate than the linear models often used in simpler studies.
  • Scenarios: Evaluation is performed across three distinct spatial distributions: Uniform, Clustered, and Asymmetric.

Key Insights and Ideas

  • The "Flying Battery" Fairness: The authors don't just aim for high average energy; they use Jain’s Fairness Index to ensure the UAV doesn't just "babysit" the closest nodes while letting distant nodes die.
  • Directional Antennas are Crucial: By modeling the UAV with a directional antenna, the system achieves higher gain and better energy beamforming, which significantly boosts the efficiency of the SWIPT component.
  • KAN > MLP: A major "aha moment" is the effectiveness of Kolmogorov-Arnold Networks in this domain. Because the physics of RF energy harvesting and propulsion power are highly non-linear, the flexible functional mapping of KANs outperforms the rigid weight-based mapping of standard neural networks.
  • Generalization: The SAC-SK model showed a remarkable ability to adapt to "unseen" terminal distributions (like highly clustered or asymmetric layouts) without needing to be retrained from scratch.

Key Results

The proposed SAC-SK algorithm was compared against standard DRL baselines (SAC, DDPG, TD3, and PPO):

  • Terminal Energy: Achieved a 47.86% improvement in average retained energy for IoT terminals compared to standard SAC.
  • Fairness: Improved charging fairness by 65.15%, ensuring a more equitable distribution of power across the network.
  • System Efficiency: Reduced total system energy consumption by approximately 109.25 Joules relative to the baseline SAC.
  • Convergence: The integration of SRU and KAN allowed the model to reach a stable, high-reward policy faster and with less variance than traditional MLP-based architectures.

Limitations and Future Work

The authors acknowledge several areas for expansion:

  • Static Terminals: Currently, the IoT devices are stationary. Real-world applications might involve mobile sensors (e.g., on livestock or vehicles).
  • Single UAV: The model focuses on one UAV. Scaling this to a coordinated swarm (multi-UAV) would introduce new challenges in interference management and collision avoidance.
  • Interference: The current model does not fully account for potential signal interference in extremely dense deployments.
  • Future Directions: The authors suggest integrating Digital Twin technology and Blockchain for more secure and transparent resource management in future iterations.

Summary updated

Abstract

The integration of simultaneous wireless information and power transfer (SWIPT) technology in 6G Internet of Things (IoT) networks faces significant challenges in remote areas and disaster scenarios where ground infrastructure is unavailable. This article proposes a novel autonomous aerial vehicle (UAV)-assisted mobile edge computing (MEC) system enhanced by directional antennas to provide both computational resources and energy support for ground IoT terminals. However, such systems require multiple tradeoff policies to balance UAV energy consumption, terminal battery levels, and computational resource allocation under various constraints, including limited UAV battery capacity, nonlinear energy harvesting characteristics, and dynamic task arrivals. To address these challenges comprehensively, we formulate a biobjective optimization problem that simultaneously considers system energy efficiency and terminal battery sustainability. We then reformulate this nonconvex problem with a hybrid solution space as a Markov decision process (MDP) and propose an improved soft actor-critic (SAC) algorithm with an action simplification mechanism to enhance its convergence and generalization capabilities. Simulation results have demonstrated that our proposed approach outperforms various baselines in different scenarios, achieving efficient energy management while maintaining high computational performance. Furthermore, our method shows strong generalization ability across different scenarios, particularly in complex environments, validating the effectiveness of our designed boundary penalty and charging reward mechanisms.
BibTeX
@article{chen2025joint,
  author = {Yue Chen and Hui Kang and Jiahui Li and Geng Sun and Boxiong Wang and Jiacheng Wang and Cong Liang and Shuang Liang and Dusit Niyato},
  title = {{Joint Resource Management for Energy-Efficient UAV-Assisted SWIPT-MEC: A Deep Reinforcement Learning Approach}},
  journal = {IEEE Internet of Things Journal},
  year = {2025},
  url = {https://openreview.net/forum?id=GHVCSUqhDW},
}
Figure