QoE-Driven Multi-Task Offloading for Semantic-Aware Edge Computing Systems

Mobile edge computing (MEC) provides low-latency offloading solutions for computationally intensive tasks, effectively improving the computing efficiency and battery life of mobile devices. However, for data-intensive tasks or scenarios with limited uplink bandwidth, network congestion might occur due to massive simultaneous offloading nodes, increasing transmission latency and affecting task performance. In this paper, we propose a semantic-aware multi-modal task offloading framework to address the challenges posed by limited uplink bandwidth. By introducing a semantic extraction factor, we balance the relationship among transmission latency, computation energy consumption, and task performance. To measure the offloading performance of multi-modal tasks, we design a unified and fair quality of experience (QoE) metric that includes execution latency, energy consumption, and task performance. Lastly, we formulate the optimization problem as a Markov decision process (MDP) and exploit the multi-agent proximal policy optimization (MAPPO) reinforcement learning algorithm to jointly optimize the semantic extraction factor, communication resources, and computing resources to maximize overall QoE. Experimental results show that the proposed method achieves a reduction in execution latency and energy consumption of 18.1% and 12.9%, respectively compared with the semantic-unaware approach. Moreover, the proposed approach can be easily extended to models with different user preferences.

💡 Research Summary

The paper tackles the growing challenge of data‑intensive, multi‑modal applications in 6G‑era mobile edge computing (MEC) where uplink bandwidth is often the bottleneck. By integrating semantic communication (SemCom) with MEC, the authors introduce a “semantic extraction factor” that controls how much task‑relevant information is retained before offloading. A higher factor preserves more semantics (better task performance) but increases the amount of data to be transmitted; a lower factor reduces transmission load at the cost of possible performance degradation.

To evaluate offloading decisions holistically, the authors propose a unified Quality‑of‑Experience (QoE) metric that aggregates three components: execution latency, computation energy consumption, and task performance (e.g., accuracy). Each component is normalized using a logistic function, and user‑specific preference weights allow the metric to adapt to latency‑sensitive, energy‑saving, or performance‑driven users. This design resolves the lack of a common evaluation scale across heterogeneous tasks such as image classification, text classification, and visual question answering (VQA).

The optimization problem is cast as a Markov Decision Process (MDP). The state captures per‑user task size, channel conditions, and current resource allocations. The action space comprises (i) the semantic extraction level, (ii) uplink bandwidth and power allocation, and (iii) edge‑server CPU share. The immediate reward equals the QoE value, and the goal is to maximize the long‑term expected reward. Because the problem involves both continuous (power, bandwidth) and discrete (semantic level) decisions, a multi‑agent Proximal Policy Optimization (MAPPO) algorithm is employed. Each user acts as an independent agent while a centralized critic evaluates the joint reward, enabling coordinated resource sharing without excessive communication overhead.

Experiments are conducted on three representative modalities: (1) image classification using a ResNet‑based semantic encoder, (2) text classification with a Transformer‑based encoder, and (3) VQA where both visual and textual features are jointly compressed. Baselines include a semantic‑unaware MEC scheme and a scheme that applies semantic compression but uses static resource allocation. Results show an average QoE improvement of 12.68 % over the semantic‑unaware baseline. Specifically, VQA—being the most data‑heavy—achieves a >30 % reduction in transmitted bits while keeping accuracy loss below 2 %. Overall execution latency drops by 18 % and energy consumption by 13 % across all tasks. By adjusting the preference weights, the framework can prioritize latency, energy, or performance, each yielding 20‑25 % gains in the targeted metric.

Key contributions are: (1) the first MEC offloading framework that jointly optimizes semantic compression, communication, and computation; (2) a fair, logistic‑based QoE metric that unifies heterogeneous task evaluations; (3) a MAPPO‑based multi‑agent solution capable of handling hybrid action spaces; and (4) demonstration of user‑customizable QoE optimization.

Limitations include reliance on pre‑trained semantic encoders (which may not generalize to all domains) and the need for periodic policy retraining when channel dynamics change rapidly. Future work is suggested on lightweight on‑device semantic encoders, meta‑reinforcement learning for fast policy adaptation, and distributed MAPPO across multiple edge servers for collaborative offloading.