Answers are grounded in this paper and cite their sources.
Cross-architecture empirical benchmark reveals factors influencing SLA-violation prediction in O-RAN environments.
Federated learning (FL) on O-RAN edge gNBs is a natural candidate for slice-level SLA-violation forecasting when raw per-user telemetry is undesirable or infeasible to pool centrally (FL itself does not provide formal privacy guarantees; see Section 8 L16). Yet the deployment design space — sequence backbone, federation algorithm, client-heterogeneity partition, commodity edge hardware — has been studied one axis at a time, leaving the joint behaviour under realistic O-RAN traffic characterised only anecdotally. This paper presents the first cross-architecture empirical FL benchmark on the public Colosseum/ColO-RAN testbed: 3 backbones (LSTM, Mamba, Spiking-SSM) × 5 algorithms (FedAvg, FedProx, FedAdam, SCAFFOLD, FedDyn) × 6 partitions (natural-by-BS + Dirichlet α∈{0.05..5.00}) × 10 seeds = 900 cells, with NVML idle-baseline-subtracted training-energy measurement on RTX 4080 and a 15-cell controlled random-split ablation on 4× V100-SXM2-32GB to probe the leading mechanism hypothesis. Four findings: (1) the natural base-station partition uniformly outperforms every Dirichlet α across all 3 architectures and 5 algorithms in this setup, contradicting the standard Dirichlet-stress trend; (2) server-side algorithmic gains are small — FedAdam improves over FedAvg by at most +0.016 AUC in the Dirichlet sweep, and a 5-seed FedSWA probe at LSTM × natural-by-BS exceeds FedAdam by only +0.0004 AUC, suggesting no large LookAhead-style gain in this representative cell; (3) selective-state-space backbones interact catastrophically with control-variate algorithms at moderate skew (Mamba × SCAFFOLD × α∈{0.10, 0.50}); (4) architecture choice traverses 10× in commodity-GPU energy and an order of magnitude larger AUC range than algorithm choice. We provide a reference benchmark with paired-bootstrap CI95 over 270 paired distributions, NVML-instrumented per-cell energy, and a preregistered statistical protocol, released under Apache-2.0 with Croissant metadata and a Zenodo DOI on acceptance.
No takes yet. Share an insight, caveat, or question.
Hsiu-Chi Tsai (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: