Machine Learning-Based AP Selection in User-Centric Cell-free Multiple-Antenna Networks

TL;DR

The paper tackles the challenge of real-time AP selection in user-centric cell-free massive MIMO by proposing learning-based schemes trained on datasets derived from LSF-based and Boosted sum-rate clustering. It develops both centralized (network-wide) and distributed (per-UE) deep neural networks to map clustering-derived features to AP activation, achieving sum-rate performance close to the heuristic baselines while greatly reducing computation time. The approach demonstrates substantial runtime gains, with the distributed model offering scalable, mobility-friendly operation and the BSR-based learning closely approximating the ideal CF performance. These findings suggest that learning-based AP selection can enhance the scalability, adaptability, and practicality of UCCF networks in dynamic wireless environments.

Abstract

User-centric cell-free (UCCF) massive multiple-input multiple-output (MIMO) systems are considered a viable solution to realize the advantages offered by cell-free (CF) networks, including reduced interference and consistent quality of service while maintaining manageable complexity. In this paper, we propose novel learning-based access point (AP) selection schemes tailored for UCCF massive MIMO systems. The learning model exploits the dataset generated from two distinct AP selection schemes, based on large-scale fading (LSF) coefficients and the sum-rate coefficients, respectively. The proposed learning-based AP selection schemes could be implemented centralized or distributed, with the aim of performing AP selection efficiently. We evaluate our model's performance against CF and two heuristic clustering schemes for UCCF networks. The results demonstrate that the learning-based approach achieves a comparable sum-rate performance to that of competing techniques for UCCF networks, while significantly reducing computational complexity.

Figures (4)

• Figure 1: UCCF clustered network.
• Figure 2: UCCF clustered network with distributed DNNs to manage AP selection.
• Figure 3: Sum-rate comparison of learning-based AP selection methods trained by LSF coefficients for $K=32$ and $M=64$.
• Figure 4: Sum-rate comparison of learning-based AP selection methods trained by BSR coefficients for $K=32$ and $M=64$.