Task offloading for multi-UAV asset edge computing with deep reinforcement learning

Samah A. Zakaryia

doi:10.1007/s10586-025-05382-1

Return

Task offloading for multi-UAV asset edge computing with deep reinforcement learning

2025-07-31

DOI：10.1007/s10586-025-05382-1

Original

Original request for help

Save

Abstract

En 中文

Mobile Edge Computing (MEC) addresses the demands of computation-intensive network services by providing processing capabilities at the network’s edge, reducing service latency. Unmanned Aerial Vehicles (UAVs) enhance MEC due to their flexibility, broad coverage, and reliable wireless communication. This paper investigates task offloading in a UAV-assisted MEC system with multiple collaborating UAVs, optimizing task allocation in binary offloading mode to enhance system efficiency. System performance is evaluated in terms of energy consumption and task delay, while the proposed framework jointly optimizes UAV trajectory design, binary offloading, computation resource allocation, and communication resource management to achieve better resource utilization. To overcome the limitations of traditional heuristic-based approaches, we propose an innovative Deep Reinforcement Learning (DRL)-based task offloading framework and conduct a comparative analysis of Deep Deterministic Policy Gradient with Distance to Task Location and Capability Matching (DDPG with DTLCM), Deep Q-Network (DQN), and Q-Learning. Our approach leverages DTLCM to enhance decision-making, allowing UAVs to assign tasks more intelligently by factoring in both their proximity to the task and their computational capabilities. Experimental results show that DDPG with DTLCM consistently outperforms both DQN and Q-Learning, delivering more stable task allocation and better adaptability in dynamic scenarios. While DQN is more stable than Q-Learning in high-dimensional state spaces, it still suffers from discretization constraints. In contrast, DDPG with DTLCM benefits from a continuous action space, enabling more flexible and fine-tuned decision-making. These insights underscore the strength of policy-gradient DRL methods, particularly when augmented with context-aware mechanisms like DTLCM, in advancing UAV-assisted MEC task scheduling and resource optimization.

Keywords:

Deep reinforcement learning

Unmanned aerial vehicle (UAV)

Task offloading

Mobile edge computing

Multi-UAV

DTLCM