Distributed MoE-based Uplink Detection for Cell-Free Communication Systems

Technical Summary: Distributed MoE-based Uplink Detection for Cell-Free Communication Systems

Problem Statement
Cell-free Massive MIMO is a pivotal technology for beyond-5G networks, utilizing distributed access points (APs) to jointly serve user equipments (UEs) and mitigate inter-cell interference. However, effective uplink signal detection in these systems faces a fundamental trade-off. Conventional linear detection methods, such as Minimum Mean Square Error (MMSE) and Zero Forcing (ZF), offer a practical balance between performance and fronthaul load but are inherently limited by linear processing constraints. Conversely, fully centralized processing, which forwards all received signals to a Central Processing Unit (CPU), achieves optimal performance but incurs prohibitive fronthaul overhead and latency. Existing distributed schemes, such as Large-Scale Fading Decoding (LSFD), attempt to bridge this gap by using statistical channel information for weighted combining; however, their reliance on expectation-maximization processes and statistical channel assumptions often leads to performance degradation under complex, deterministic channel conditions.

Methodology: DMoE-DetNet
To address these limitations, the authors propose DMoE-DetNet (Distributed Mixture of Experts Detection Network), a deep learning-based framework that leverages a division of labor between instantaneous and statistical channel information. The architecture operates as follows:

1. Local Expert Networks (APs): Each AP acts as a "local expert." Instead of simple linear combining, each AP employs a Convolutional Neural Network (CNN) to perform non-linear feature extraction. These networks utilize Instantaneous Channel State Information (ICSI) and received signals to generate preliminary local detection results and fused feature representations.
2. Central Processing Unit (CPU) - Feature Fusion: The CPU receives the local detection results and feature vectors from all APs. A Transformer-based encoder is employed to capture complex spatio-temporal dependencies among users, refining the global feature representation.
3. Intelligent Gating Network: A gating network at the CPU dynamically weights the contributions from different APs. Unlike conventional LSFD which relies on static statistical expectations, this gating network utilizes Statistical Channel State Information (SCSI) (e.g., spatial correlation matrices) to compute adaptive fusion weights. This allows the system to prioritize APs with favorable long-term channel conditions.
4. Final Detection: The weighted fusion of expert outputs is passed through a linear detector to output symbol probabilities.

Crucially, the trained framework operates using only the detected signals from APs and SCSI (which can be acquired via prior knowledge maps), avoiding the need to transmit raw received signals or full ICSI to the CPU.

Key Contributions

• Novel Architecture: The introduction of a distributed Mixture of Experts (MoE) framework specifically tailored for cell-free uplink detection, distinguishing between non-linear local processing (ICSI-driven) and adaptive global fusion (SCSI-driven).
• Non-Linear Processing: Moving beyond the linear constraints of MMSE and LSFD by utilizing CNNs for local feature extraction and Transformers for global dependency modeling.
• Adaptive Gating: Replacing static, expectation-based combining weights with a learnable gating network that dynamically adjusts to channel statistics, offering a more robust solution for deterministic channel environments.
• Efficiency: The design maintains low fronthaul overhead by transmitting only processed local estimates and statistical parameters rather than raw signal data.

Simulation Results
Extensive simulations were conducted comparing DMoE-DetNet against fully centralized processing, local processing with averaging, and local processing with LSFD combining.

• Symbol Error Rate (SER): DMoE-DetNet significantly outperforms conventional linear distributed methods (LSFD and averaging) across various Signal-to-Noise Ratio (SNR) levels.
• Performance Gap: While fully centralized processing remains the theoretical upper bound, DMoE-DetNet closely approaches this performance limit while requiring substantially less fronthaul interaction.
• Robustness: The proposed method demonstrates superior performance in heterogeneous channel conditions and varying numbers of UEs and APs, validating its ability to learn effective non-linear strategies that static statistical methods cannot capture.

Significance and Claims
The paper claims that DMoE-DetNet represents a significant advancement in AI-enabled communication systems. By successfully integrating the Mixture of Experts paradigm with cell-free architectures, the framework overcomes the performance ceiling of linear distributed detectors without incurring the overhead of fully centralized processing. The authors assert that this approach showcases the potential of deep learning to realize the full potential of distributed processing in beyond-5G networks, offering a practical path toward near-optimal detection with manageable implementation complexity.