Task offloading an‎d resource allocation in mobile network edge via multi-agent reinforcement learning

1/21/2026 12:00:00 AM

With the rapid emergence of novel applications an‎d the expansion of the Internet of Things (IoT), next-generation mobile networks are experiencing unprecedented growth in both traffic volume an‎d computational deman‎ds. Nevertheless, end-user devices, due to their low cost, small size, lightweight, an‎d portability, often suffer from limited storage capacity an‎d processing capabilities. Multi-access edge computing (MEC) has been proposed as a promising solution to address these limitations. A key challenge in MEC for mobile networks lies in determining effective data offloading policies for processing an‎d efficient resource allocation. This study aims to propose an approach to address this challenge. Specifically, the objective is to determine the optimal data offloading location in order to minimize energy consumption an‎d response time. Once the offloading location is identified, resource allocation becomes more straightforward. To this end, we adopt an actor–critic framework enhanced with an attention mechanism. This approach represents an implementation of centralized learning with distributed execution, an‎d it has been tailored to the specific characteristics of the problem under investigation to achieve improved results. After determining the offloading location, processor an‎d communication channel resources are allocated for data transmission. For processor resource allocation, a deterministic optimal relation is utilized, whereas for communication channel allocation, we propose a simple yet computationally efficient algorithm. Simulation results demonstrate that the proposed modifications to the actor–critic with attention mechanism (MAAC) progressively enhance system performance. Compared to the baseline actor–critic with attention, the proposed method achieves reductions in energy consumption an‎d latency, while also lowering the number of deadline violations by approximately 10% under equivalent conditions.

محاسبات لبه با دسترسي چندگانه , بارگذاري داده , يادگيري تقويتي , عملگر- منتقد

Multi-access edge computing , data offloading , reinforcement learning , actor–critic

Ali Reza Eskandari

Dr. Morteza Analoui