Deep Mixture of Experts Network for Resource Optimization in Aerial-Terrestrial CF-mMIMO Systems under URLLC

TL;DR

This paper introduces a deep learning framework combining channel prediction and mixture of experts for resource optimization in aerial-terrestrial CF-mMIMO systems supporting URLLC, enhancing efficiency and reliability.

Contribution

It proposes a novel hybrid aerial-terrestrial CF-mMIMO network with a Transformer-based channel predictor and a deep MoE for uplink resource optimization under URLLC constraints.

Findings

The channel prediction network improves CSI accuracy for high-mobility UEs.

The deep MoE network effectively allocates power based on heterogeneous UE requirements.

Numerical results show significant performance gains in communication reliability and efficiency.

Abstract

As a critical component of sixth-generation (6G) wireless networks, ultra-reliable and low-latency communication (URLLC) is expected to support real-time and reliable information exchange in low-altitude environments. However, achieving URLLC often incurs significant resource overhead, including increased bandwidth consumption, higher transmit power, and denser access point (AP) deployment, which pose significant challenges to both spectral efficiency (SE) and energy efficiency (EE). Besides, existing iterative optimization algorithms are computationally intensive and struggle to meet the latency requirements of URLLC. To address these challenges, we propose a hybrid aerial-terrestrial cell-free massive MIMO (CF-mMIMO) network to support diverse services, along with a channel prediction network and a deep mixture of experts (MoE) network for uplink optimization. First, we design a channel prediction network (CP-Net) to mitigate channel aging caused by high-mobility user equipment (UE). CP-Net employs three Transformer-based sub-networks for aged channel state information (CSI) prediction, while a channel quality-aware loss function is introduced to improve the prediction accuracy of weak links. Based on the predicted CSI, we develop a deep MoE network (MoE-Net) for power allocation comprising three expert models targeting different objectives. Then, we introduce a weighted gating network (WT-Net) to learn an efficient adaptive combination of expert outputs. The proposed framework better captures heterogeneous UE requirements and improves communication performance under URLLC constraints. Numerical results demonstrate the effectiveness of the proposed method.