UnifSrv: AP Selection for Achieving Uniformly Good Performance of CF-MIMO in Realistic Urban Networks

Under the ideal assumption of uniform propagation, cell-free massive MIMO (CF-mMIMO) provides uniformly high throughput over the network by effectively surrounding each user with its serving access point (AP) set. However, in realistic non-uniform urban propagation environments, it is difficult to consistently select good limited serving AP sets, resulting in significantly degraded throughput, reintroducing “edge-effect” for the worst-served users. To restore the uniformly good performance of scalable CF-mMIMO in realistic urban networks, we formulate a novel multi-objective optimization problem to jointly achieve high throughput by maximizing the sum data rate, uniform throughput by maximizing Jain’s fairness index of the throughput per user, and scalability by minimizing the serving AP set size. We then propose the UnifSrv AP selection algorithms to solve this optimization problem, consisting of a deep reinforcement learning (DRL)-based algorithm UnifSrv-DRL and a heuristic algorithm UnifSrv-heu. We conduct a comprehensive performance evaluation of scalable CF-mMIMO under realistic urban network distributions, propagation, and mobility patterns, showing that the prior benchmark AP selection schemes fail to provide uniformly high throughput in practice. By contrast, UnifSrv at least doubles the throughput compared to prior benchmarks, or achieves comparable throughput but with half of the serving AP set size. Importantly, our heuristic algorithm achieves equivalent throughput to our DRL one, but with orders of magnitude lower complexity. We thus for the first time propose an AP selection algorithm that achieves uniformly good CF-mMIMO performance in realistic urban networks with low complexity.

💡 Research Summary

This paper addresses a critical gap in the deployment of cell‑free massive MIMO (CF‑mMIMO) systems: while theoretical analyses often assume a uniform Poisson point process (PPP) distribution of access points (APs) and a simple distance‑based path‑loss model, real‑world urban environments exhibit highly non‑uniform AP placements and severe building‑induced shadowing. Under such realistic conditions, existing AP‑selection schemes either select an excessively large set of APs to maintain channel quality (sacrificing scalability) or limit the AP set size at the cost of dramatically reduced throughput, thereby re‑introducing the classic “cell‑edge” effect.

To restore the promised uniformly high throughput of CF‑mMIMO, the authors formulate a novel multi‑objective optimization problem that simultaneously (i) maximizes the sum data rate, (ii) maximizes Jain’s fairness index of per‑user throughput, and (iii) minimizes the total number of AP‑UE connections. The formulation includes explicit constraints on the maximum number of serving APs per user and the maximum number of users served per AP, reflecting the scalability requirements of a practical O‑RAN‑based distributed architecture.

Two solution approaches are proposed. The first, UnifSrv‑DRL, transforms the three objectives into a weighted scalar reward and solves the problem with a Deep Q‑Network. The state comprises the current cooperation matrix and channel statistics; actions correspond to adding or removing a specific AP from a user’s serving set; the reward captures improvements in sum‑rate, fairness, and connection reduction. Experience replay and ε‑greedy exploration are employed to obtain a stable policy.

Recognizing the high training cost of deep reinforcement learning, the authors design a low‑complexity heuristic, UnifSrv‑heu. Starting from the strongest AP for each user, the algorithm iteratively expands the serving set only when Jain’s fairness falls below a predefined threshold. The threshold is dynamically lowered, allowing additional APs to be added in a controlled manner. This method requires only sorting of channel gains and simple threshold checks, yielding orders‑of‑magnitude lower computational complexity while achieving virtually identical performance to the DRL solution.

The performance evaluation uses site‑specific ray‑tracing channel models for two distinct cities—Seoul (medium‑density) and Frankfurt (high‑density)—with realistic AP placements (edges or corners of buildings) and mobility patterns generated by the V‑Walk model for 50 moving users. A PPP‑based baseline with random waypoint mobility is also considered for comparison.

Results show that both UnifSrv‑DRL and UnifSrv‑heu dramatically outperform seven representative benchmark schemes (including pure UE‑centric, clustered UE‑centric, constrained UE‑centric, and proportional‑fairness DRL). The proposed algorithms at least double the average sum‑rate compared to the best existing method, or achieve comparable sum‑rate with roughly half the average number of AP‑UE connections. The 5‑percentile user throughput improves substantially, and Jain’s fairness index remains above 0.95, confirming that the worst‑served users benefit without sacrificing overall efficiency. Importantly, the heuristic’s runtime is two to three orders of magnitude lower than the DRL approach, making it suitable for real‑time network operation.

In addition to delivering a practical AP‑selection framework, the paper contributes a comprehensive study of CF‑mMIMO performance under realistic urban propagation, highlighting the inadequacy of prior assumptions and establishing a benchmark for future work. The authors also discuss integration with O‑RAN’s distributed CPU hierarchy, potential extensions to multi‑band (sub‑6 GHz/mmWave) scenarios, and the possibility of online policy adaptation or transfer learning to further reduce training overhead.

Overall, the work convincingly demonstrates that jointly optimizing throughput, fairness, and scalability via the proposed UnifSrv algorithms enables CF‑mMIMO to achieve uniformly high performance in realistic urban networks, thereby bridging the gap between theory and practical deployment.