Problem

Traditional methods for online task offloading in wireless-powered mobile edge computing (MEC) networks face significant challenges. Devices with limited battery life and computational power struggle to handle intensive tasks. The joint optimization of binary offloading decisions (local vs. edge) and continuous resource allocation (wireless energy transfer and task transmission time) leads to complex mixed-integer programming problems. Existing heuristic algorithms often suffer from poor adaptability, slow convergence, and performance loss, making them unsuitable for real-time decision-making in highly dynamic channel environments, especially in large-scale networks.

Method

The QAROO (Quantum Attention-based Reinforcement learning for Online Offloading) framework is proposed to address these challenges. It utilizes a binary offloading strategy to co-optimize computing and energy resources in dynamic wireless channels. The problem is modeled as a Markov Decision Process, where an access point (AP) acts as an agent learning to maximize the weighted sum computation rate. QAROO introduces three key innovations: 1) It replaces traditional Deep Neural Networks (DNNs) with Recurrent Neural Networks (RNNs) to improve temporal modeling of channel states, capturing dynamic dependencies over time. 2) An Uncertainty-Guided Quantization (UGQ) method is developed to enhance exploration efficiency in the large discrete action space. UGQ generates diverse candidate actions by prioritizing adjustments to decisions with high uncertainty (probabilities near 0.5) and using random perturbations and forced bit flips to maintain diversity. 3) A hybrid Quantum Neural Network (QNN) combined with an Attention Mechanism is used for robust feature representation. The QNN encodes classical channel states into high-dimensional quantum features, while a multi-head self-attention module dynamically weighs these features to focus on relevant information, thereby improving decision quality.

Results

Experimental evaluations demonstrate that the QAROO framework significantly outperforms comparative methods, such as the classical DROO framework. RNN-based models exhibit faster convergence, achieving near-optimal normalized computation rates (exceeding 0.99) much quicker than DNN-based counterparts. They also show superior stability with lower variance in normalized rates and half the loss function value, indicating better generalization in dynamic environments. The Uncertainty-Guided Quantization (UGQ) method proves particularly effective in large-scale scenarios (20-30 devices), where it enables more comprehensive exploration of the exponentially growing action space, leading to higher average normalized rates compared to traditional Order-Preserving (OP) quantization. Specifically, RNN+UGQ achieved the highest average normalized rate of 0.998401 in the 30-device scenario. The integration of the quantum-attention hybrid architecture also shows improved performance, especially in high-dimensional feature refinement, by effectively capturing inter-channel correlations.

Implications

The QAROO framework offers a robust, efficient, and stable solution for online task offloading, enabling sustainable and resource-efficient applications in dynamic and large-scale Internet of Things (IoT) environments and wireless-powered mobile edge computing networks. By effectively addressing the co-optimization of computing and energy resources, QAROO provides a foundation for next-generation edge intelligence. The innovative use of RNNs for temporal modeling, UGQ for diverse exploration, and QNN+Attention for enhanced feature representation contributes significantly to overcoming the limitations of traditional offloading approaches. Future research will explore extending this framework to multi-cell MEC networks, supporting partial task offloading, handling mobility scenarios, and investigating deployment on actual quantum hardware for further energy efficiency optimization.