Physically Constrained Federated Additive Models for O-RAN SLA-Risk Prediction

Aubida A. Al-Hameed; Maryam Hafeez; Mohammed M. H. Qazzaz; Syed A. Zaidi

REVIEW 3 major objections 3 minor 22 references

Physically Constrained Federated Additive Models for O-RAN SLA-Risk Prediction

T0 review · 3 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A federated additive model that hard-codes the known physical direction of each radio KPI can predict per-slice SLA violations in O-RAN without learning any physically inverted effect, at a cost of only 0.04 to 0.07 AUC.

desk verdict Core constraint-preservation trick is correct; the 0.04 AUC cost claim is confounded with model capacity, and the 1.00 consistency jump needs a mechanism. read the letter →

arxiv 2607.21665 v1 pith:WSIEVR7M submitted 2026-07-23 cs.LG cs.SYeess.SY

Aubida A. Al-Hameed , Mohammed M. H. Qazzaz , Maryam Hafeez , Syed A. Zaidi This is my paper · ORCID

classification cs.LGcs.SYeess.SY

keywords federatedlearningneuraladditivemodelsmonotonicconstraintsO-RANSLApredictionsplineincrementsinterpretabilityradioKPIs

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tackles a real O-RAN problem: predicting per-slice SLA violations without pooling sensitive tenant KPI data and without losing auditability. It demonstrates that standard neural additive models, although interpretable, learn KPI effects that contradict wireless physics—for instance, better channel quality appearing to raise risk—and that federated averaging makes this worse. The proposed Monotone FedNAM represents physically unambiguous KPIs as monotone splines whose signed-increment parameterization guarantees that the constraint survives FedAvg aggregation. On a seven-base-station testbed, it eliminates all monotonicity violations, raises constrained-shape consistency from 0.71 to 1.00, generalizes to an unseen scheduling policy, and reduces uplink traffic by 65%, at a cost of 0.04 to 0.07 AUC.

What carries the argument

The load-bearing mechanism is the monotone linear spline with a signed-increment parameterization on a shared 16-knot grid. Each constrained feature's shape is a cumulative sum of strictly positive or negative increments, making the function monotone by construction. Because FedAvg is a sample-weighted average, averaging these increment vectors preserves the sign, so the global shape inherits the constraint. An anchoring step subtracts each client's shape value at a public reference point, resolving the additive model's offset ambiguity and making client shapes comparable before aggregation.

What would settle it

Train Monotone FedNAM on a dataset with genuine tenant SLA contracts and compare predicted risk against actual violations; if the accuracy gap relative to unconstrained models grows well beyond 0.04–0.07 AUC, or if monotonicity violations reappear on features whose true effect is opposite to the assumed direction, the central claim fails. A cheaper check: inject a synthetic feature whose true effect is non-monotone but label it monotone and observe whether the constraint forces a large, detectable accuracy drop.

Watch

Extended reading notes

Core claim

The central claim is that monotonicity constraints can be preserved under federated averaging by construction rather than by hope. If all clients use the same spline knots and a signed-increment parameterization, then FedAvg of the increment vectors preserves the sign, so each aggregated feature shape is monotone in the intended direction. This converts a physical prior into a hard guarantee. On the ColO-RAN dataset, Monotone FedNAM removes all monotonicity violations (average violation rate falls from 0.652 for unconstrained FedNAM to 0.000), raises constrained-shape consistency from 0.712 to 1.000, transfers to an unseen scheduler, and reduces client-to-server uplink traffic by 65%, with a

Load-bearing premise

The evaluation depends on hand-set SLA-violation proxies (e.g., eMBB rate below 2 Mbps) standing in for real tenant contracts; if these proxies do not reflect actual SLAs, the model is learning to predict a synthetic target and the reported performance does not transfer.

Editorial extensions

If this is right

Deployed as an O-RAN Near-RT RIC xApp, the model gives operators per-slice SLA-risk scores with per-KPI additive explanations that are physically consistent, so an auditor can verify each effect.
The 65% uplink reduction makes frequent federated retraining practical: each constrained feature is transmitted as 16 spline increments instead of hundreds of MLP weights.
The constraint manifests force an explicit, documented decision about which KPI directions are physically unambiguous, turning an implicit modelling choice into an auditable engineering artifact.
The method's accuracy cost is bounded (0.04–0.07 AUC) on this testbed, so physical validity does not require sacrificing predictive performance.
The leave-one-scheduler-out result suggests the governed shapes capture transferable wireless relationships, not just dataset idiosyncrasies.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The spline-increment-plus-FedAvg trick is a general recipe for any federated additive model with known monotone directions; it could be applied to credit scoring, medical risk, or other domains where domain validity must survive aggregation.
The paper's label proxies are a major caveat: if real tenant SLA contracts differ from the threshold rules, the absolute AUC numbers may not transfer. A natural next test is a deployment with signed per-tenant SLAs.
The monotone assumptions themselves may be too rigid in rare regimes (e.g., under strong interference, higher CQI could coincide with higher risk); testing the method under adversarial or extreme radio conditions would reveal where conservative constraints become harmful.
Because the rApp and xApp split is only simulated here, a real-time implementation on an O-RAN controller would settle whether the 2.5-second inference cadence and model size meet near-real-time latency requirements.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Read this one for the constraint-preservation trick. The core idea—represent constrained features as linear splines with softplus increments and FedAvg the signed increment vectors—is correct, and Proposition 1 is a clean one-line proof. The paper also anchors each shape at a public reference, which neatly handles the federated identifiability problem. That combination is genuinely new relative to FedNAMs, and the O-RAN rApp/xApp framing is a nice touch. The paper does a lot of things right: the problem is well motivated, the testbed is real, the limitations are mostly stated, and the leakage check (removing delivered rate) is the kind of sanity test that should be standard.

The soft spots are empirical. The headline tradeoff claim is not fully supported. The 0.042 AUC drop relative to unconstrained FedNAM is confounded with model capacity: constrained features are 16-knot splines with 2,563 total parameters, while FedNAM uses MLP subnetworks with 7,315. There is no unconstrained spline baseline with the same knots and the same parameterization but with increments allowed to be negative. Without that, you cannot attribute the drop to the monotonicity prior rather than to the fixed linear shape. The same confound affects the uplink savings: those come from the compact spline representation, not from the sign constraint. The abstract's 'cost of physical consistency' overstates what the experiment shows.

A second issue: the constrained-shape consistency jump from 0.712 to 1.000 is not explained. Monotonicity alone does not force per-client shapes to be perfectly correlated—two monotone functions can have different slopes. The paper should either define the metric differently or show why all clients converge to identical shapes. The exact 1.000 looks suspicious.

Minor: the SLA labels are hand-set threshold proxies for all three slices, not just URLLC; the Discussion only flags the URLLC one. The thresholds are plausible, but the evaluation remains on a synthetic target from an emulated testbed with static users, as acknowledged.

Who should read it: anyone working on interpretable federated learning, constrained shape functions, or O-RAN service assurance. It is a solid applied contribution, not a breakthrough. It deserves a serious referee. I would send it out, but I would ask for an unconstrained spline baseline and a closer look at the consistency metric before the tradeoff claims are accepted.

Referee Report

3 major / 3 minor

Summary. The paper proposes Monotone FedNAM, a federated additive model for per-slice SLA-violation risk prediction in O-RAN. Continuous KPIs with an unambiguous physical direction are represented as 16-knot linear splines with softplus (strictly positive) increments, so each client's shape is monotone by construction; a short proof (Proposition 1) shows that FedAvg of these signed increment vectors preserves monotonicity. The model is positioned as a Non-RT RIC rApp that produces a governed model deployable as a Near-RT RIC xApp. Experiments on the ColO-RAN testbed compare against local/centralized NAMs, FedAvg-MLP, and an unconstrained FedNAM. The paper reports that Monotone FedNAM eliminates all monotonicity violations, raises constrained-shape consistency from 0.71 to 1.00, transfers to an unseen scheduler, reduces uplink traffic by 65%, and costs 0.042 AUC relative to unconstrained FedNAM and 0.072 relative to FedAvg-MLP.

Significance. If the empirical claims are supported, the paper is a useful contribution to interpretable and physically constrained federated learning for O-RAN. The central design idea is sound: representing constrained features as signed spline increments whose FedAvg preserves the sign is a simple, auditable mechanism with a correct proof. The authors also provide a concrete deployment path (rApp to xApp) and report model sizes and communication costs, which is the right level of concreteness for a systems-oriented ML paper. The evaluation on a public O-RAN testbed is a strength, as is the honest acknowledgment of the proxy-label limitation in Section V. However, two load-bearing empirical points need strengthening: the 1.000 constrained-shape consistency figure is not entailed by monotonicity alone, and the reported AUC cost is confounded with a capacity reduction because the constrained model uses far fewer parameters than the unconstrained baseline.

major comments (3)

[§III-C and Table IV] The reported jump in constrained-shape consistency to 1.000 is not explained by the monotonicity constraint. The metric is the mean pairwise Pearson correlation between per-client local shape functions after local training. Clients train on different non-IID data, so their softplus increments should differ; monotone functions need not have pairwise correlation 1 (e.g., two increasing piecewise-linear functions with different slopes/spacings). Either the metric is actually measuring something else (e.g., sign agreement or violation rate), in which case 1.000 is trivially by construction and should be described as such, or the authors need to provide the mechanism and per-feature correlation values. This is load-bearing because the abstract and Section IV-c use this number to claim that client shapes are aligned.
[§IV-a and Table III] The claim that enforcing physical consistency 'costs 0.042 AUC relative to unconstrained FedNAM' is confounded with model capacity. Monotone FedNAM replaces the constrained MLP subnetworks with 16-knot splines, reducing the parameter count from 7,315 to 2,563. The AUC drop could be due entirely to the lower-capacity spline representation, not to the sign constraint. The authors should add an unconstrained spline baseline (same knot grid, same 16 increment parameters per feature, but with increments allowed to be negative) to isolate the effect of monotonicity. Without this baseline, the 'at a cost of 0.042 AUC' statement is unsupported.
[§III-B and §V] The SLA-violation labels are hand-set proxy thresholds (eMBB delivered rate < 2 Mbps, MTC < 0.015 Mbps, URLLC buffer > 28.8 B or error > 5%) standing in for tenant SLA contracts. The paper acknowledges a 'proxy URLLC label,' but the issue applies to all three slices and to every accuracy number in Tables III–IV, as well as to the scheduler-transfer result. Since these thresholds are not derived from real tenant contracts, the model predicts a synthetic target and the 'proactive service assurance' framing is stronger than the evidence supports. Please add a sensitivity study across plausible contract definitions, or explicitly re-scope the conclusions to proxy-based risk prediction.

minor comments (3)

[Section I vs. Section III-B] The Introduction states that 'SLA labels follow per-slice service profiles rather than ad hoc thresholds,' but Section III-B defines thresholds that are themselves ad hoc (half of offered load, a Little's-law buffer estimate). Reconcile the wording.
[Equation (9)] It is not fully clear whether the intercept β0 and the unconstrained subnetwork parameters are included in the same ψ_j notation and averaged with the same sample weights. A sentence clarifying the exact grouping would help reproducibility.
[Section IV-d] The leakage-robustness ablation is described only in prose ('AUC falls from 0.943 to 0.927...'); consider adding it to a table for completeness.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the zero-violation guarantee is an explicit design property, not a hidden reduction.

full rationale

The claimed derivation chain is self-contained. Proposition 1 is an independent algebraic fact: a weighted average of non-negative spline increments is non-negative, so FedAvg preserves monotonicity under the stated assumptions (shared knots, signed-increment parameterization, client-side sign satisfaction). The monotonicity directions are imposed from wireless domain knowledge (Table I) rather than fitted from labels, so the resulting zero violation rate is a designed guarantee, and the paper says so explicitly ('removes these violations by construction', Section IV-b). The accuracy figures, the AUC trade-off, the scheduler-transfer result, and the uplink-traffic reduction are empirical or arithmetical and are not forced by the model's definition. The only near-tautological element is that a model parameterized with softplus increments reports no monotonicity violations; because this is transparent and not used as evidence for the transferability/accuracy claims, it does not constitute circular reasoning. There is no load-bearing self-citation chain, no fitted parameter renamed as a prediction, and no imported uniqueness theorem. A separate experimental-design concern—the absence of an unconstrained-spline baseline, so the 0.042 AUC cost may reflect model capacity rather than the monotonicity constraint—is a correctness/confounding issue, not circularity.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim depends on hand-chosen design decisions (Q=16 knots, SLA label thresholds) and on domain assumptions about the physical direction of each KPI and the adequacy of additive structure. No new physical entities are introduced. The FedAvg preservation proof uses only the algebra of convex combinations.

free parameters (3)

Spline grid size Q = 16
Chosen in Section II-C to balance shape resolution against update size; every constrained feature is transmitted as 16 increments per round.
SLA label proxy thresholds = eMBB 2 Mbps; MTC 0.015 Mbps; URLLC 28.8 B / 5% error
Hand-set in Section III-B to define positive labels; a ±20% sensitivity check is reported, but the thresholds remain the definition of the target.
Training hyperparameters = E=3, batch size 8192, 30 rounds, Adam (learning rate not reported)
Section III-C specifies these; they affect the empirical results but not the by-construction guarantee.

assumptions (6)

standard math Weighted average of non-negative increments is non-negative
Proposition 1 rests on this algebraic fact; correct and trivial.
domain assumption Additive model class is adequate for SLA-risk prediction
Equation (2) assumes the joint KPI effect is a sum of per-feature functions; interactions are ignored.
domain assumption Assigned monotonic directions in Table I are physically correct
If, for example, better CQI does not lower risk in some regime, the constraint encodes an error.
domain assumption SLA labels can be proxied by threshold rules
Section III-B defines labels from delivered-rate and buffer thresholds; real SLA contracts may differ.
domain assumption ColO-RAN testbed data are representative of operational O-RAN
Seven base stations, static users, one emulated scenario; the paper acknowledges this limitation.
standard math Anchoring at a_j=0 fixes federated identifiability without changing predictions
Section II-D: subtracting the value at a public reference and folding it into the intercept leaves predictions unchanged.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Physically Constrained Federated Additive Models for O-RAN SLA-Risk Prediction." pith.science (2026). https://pith.science/paper/WSIEVR7M

@misc{pith2026260721665,
  author       = {Pith},
  title        = {Pith review of: Physically Constrained Federated Additive Models for O-RAN SLA-Risk Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WSIEVR7M}},
  note         = {Machine review of arXiv:2607.21665}
}

read the original abstract

Proactive service assurance in O-RAN requires predicting per-slice SLA violations before they occur. The prediction model must be auditable by operators and must train across base stations without pooling per-slice KPIs, which are commercially sensitive because slices are leased to individual tenants. Neural additive models (NAMs) offer auditability because each KPI contributes through a visible shape function. However, visibility alone does not guarantee physical validity. On the ColO-RAN testbed dataset, unconstrained NAMs learn effects that contradict wireless physics, for example predicting higher risk when channel quality improves. This failure appears under both local and centralized training, and non-IID federated averaging worsens it. We present Monotone FedNAM, a federated additive model in which KPIs with unambiguous physical direction are represented as monotone splines whose constraints survive FedAvg aggregation by construction, while contestable KPIs remain unconstrained. The model trains and operates as a Non-RT RIC rApp and is compact enough for deployment as a Near-RT RIC xApp. Monotone FedNAM eliminates all monotonicity violations, raises constrained shape consistency from 0.71 to 1.00, generalizes to an unseen scheduling policy, and reduces uplink traffic by 65%, at a cost of 0.04 to 0.07 AUC. These results show that physically constrained federated additive models can support auditable SLA risk inference for multi-tenant O-RAN service assurance

Figures

Figures reproduced from arXiv: 2607.21665 by the authors.

**Figure 2.** Unconstrained FedNAM learns an inverted channel-quality effect; [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 1 linked inside Pith

[1]

Under- standing O-RAN: Architecture, interfaces, algorithms, security, and re- search challenges,

M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Under- standing O-RAN: Architecture, interfaces, algorithms, security, and re- search challenges,”IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 1376–1411, 2023

2023
[2]

Taking 5G RAN analytics and control to a new level,

X. Foukas, B. Radunovic, M. Balkwill, and Z. Lai, “Taking 5G RAN analytics and control to a new level,” inProc. ACM MobiCom, 2023

2023
[3]

From network sharing to multi-tenancy: The 5g network slice broker,

K. Samdanis, X. Costa-Perez, and V . Sciancalepore, “From network sharing to multi-tenancy: The 5g network slice broker,”IEEE Com- munications Magazine, vol. 54, no. 7, pp. 32–39, 2016

2016
[4]

Generic network slice template,

GSMA, “Generic network slice template,” GSM Asso- ciation, Official Document NG.116, Version 10.0, 2024. [Online]. Available: https://www.gsma.com/newsroom/gsma resources/ ng-116-generic-network-slice-template-v8-0/

2024
[5]

Communication-efficient learning of deep networks from decentralized data,

B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inArtificial intelligence and statistics. Pmlr, 2017, pp. 1273– 1282

2017
[6]

Advances and open problems in federated learning,

P. Kairouz, H. B. McMahanet al., “Advances and open problems in federated learning,”Foundations and Trends in Machine Learning, vol. 14, no. 1–2, pp. 1–210, 2021

2021
[7]

Federated deep reinforce- ment learning for resource allocation in O-RAN slicing,

H. Zhang, H. Zhou, and M. Erol-Kantarci, “Federated deep reinforce- ment learning for resource allocation in O-RAN slicing,” inProc. IEEE Global Communications Conference (GLOBECOM), 2022

2022
[8]

Federated meta- learning for traffic steering in O-RAN,

H. Erdol, X. Wang, P. Li, J. D. Thomas, R. Piechocki, G. Oikonomou, R. Inacio, A. Ahmad, K. Briggs, and S. Kapoor, “Federated meta- learning for traffic steering in O-RAN,” inProc. IEEE Vehicular Technology Conference (VTC2022-Fall), 2022

2022

Show all 22 references

[9]

Peer-to- peer federated learning based anomaly detection for open radio access networks,

D. Attanayaka, P. Porambage, M. Liyanage, and M. Ylianttila, “Peer-to- peer federated learning based anomaly detection for open radio access networks,” inProc. IEEE Int. Conf. Communications (ICC), 2023

2023
[10]

DRL-based latency-aware network slicing in O-RAN with time-varying SLAs,

R. Raftopoulos, S. D’Oro, T. Melodia, and G. Schembra, “DRL-based latency-aware network slicing in O-RAN with time-varying SLAs,” in Proc. IEEE Int. Conf. Computing, Networking and Communications (ICNC), 2024

2024
[11]

Explainable AI in 6G O-RAN: A tutorial and survey on architecture, use cases, challenges, and future research,

B. Brik, H. Chergui, L. Zanzi, F. Devoti, A. Ksentini, M. S. Siddiqui, X. Costa-P ´erez, and C. Verikoukis, “Explainable AI in 6G O-RAN: A tutorial and survey on architecture, use cases, challenges, and future research,”IEEE Communications Surveys & Tutorials, 2024

2024
[12]

Explainable and robust artificial intelligence for trustworthy resource management in 6G networks,

N. Khan, S. Coleri, A. Abdallah, A. Celik, and A. M. Eltawil, “Explainable and robust artificial intelligence for trustworthy resource management in 6G networks,”IEEE Communications Magazine, vol. 62, no. 4, pp. 50–56, 2024

2024
[13]

EXPLORA: AI/ML explainability for the open RAN,

C. Fiandrino, L. Bonati, S. D’Oro, M. Polese, T. Melodia, and J. Widmer, “EXPLORA: AI/ML explainability for the open RAN,”Proceedings of the ACM on Networking (CoNEXT), vol. 1, pp. 1–26, 2023

2023
[14]

Explanation- guided deep reinforcement learning for trustworthy 6G RAN slicing,

F. Rezazadeh, H. Chergui, and J. Mangues-Bafalluy, “Explanation- guided deep reinforcement learning for trustworthy 6G RAN slicing,” inProc. IEEE Int. Conf. Communications Workshops (ICC Workshops), 2023

2023
[15]

Neural additive models: interpretable machine learning with neural nets,

R. Agarwal, L. Melnick, N. Frosst, X. Zhang, B. Lengerich, R. Caruana, and G. E. Hinton, “Neural additive models: interpretable machine learning with neural nets,” inProceedings of the 35th International Conference on Neural Information Processing Systems, ser. NIPS ’21. Red H...

2021
[16]

Generalized additive models,

T. J. Hastie, “Generalized additive models,”Statistical models in S, pp. 249–307, 2017

2017
[17]

Monotonic networks,

J. Sill, “Monotonic networks,” inAdvances in Neural Information Processing Systems (NeurIPS), 1997

1997
[18]

Certified monotonic neural networks,

X. Liu, X. Han, N. Zhang, and Q. Liu, “Certified monotonic neural networks,” inAdvances in Neural Information Processing Systems (NeurIPS), 2020

2020
[19]

Constrained monotonic neural networks,

D. Runje and S. M. Shankaranarayana, “Constrained monotonic neural networks,” inInternational Conference on Machine Learning. PMLR, 2023, pp. 29 338–29 353

2023
[20]

Fednams: Performing in- terpretability analysis in federated learning context,

A. Nanda, S. B. Balija, and D. Sahoo, “Fednams: Performing in- terpretability analysis in federated learning context,”arXiv preprint arXiv:2506.17466, 2025

2025 arXiv
[21]

ColO- RAN: Developing machine learning-based xApps for open RAN closed- loop control on programmable experimental platforms,

M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “ColO- RAN: Developing machine learning-based xApps for open RAN closed- loop control on programmable experimental platforms,”IEEE Transac- tions on Mobile Computing, vol. 22, no. 10, pp. 5787–5800, 2023

2023
[22]

Goodfellow, Y

I. Goodfellow, Y . Bengio, A. Courville, and Y . Bengio,Deep learning. MIT press Cambridge, 2016, vol. 1, no. 2

2016

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Under- standing O-RAN: Architecture, interfaces, algorithms, security, and re- search challenges,

M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Under- standing O-RAN: Architecture, interfaces, algorithms, security, and re- search challenges,”IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 1376–1411, 2023

2023

[2] [2]

Taking 5G RAN analytics and control to a new level,

X. Foukas, B. Radunovic, M. Balkwill, and Z. Lai, “Taking 5G RAN analytics and control to a new level,” inProc. ACM MobiCom, 2023

2023

[3] [3]

From network sharing to multi-tenancy: The 5g network slice broker,

K. Samdanis, X. Costa-Perez, and V . Sciancalepore, “From network sharing to multi-tenancy: The 5g network slice broker,”IEEE Com- munications Magazine, vol. 54, no. 7, pp. 32–39, 2016

2016

[4] [4]

Generic network slice template,

GSMA, “Generic network slice template,” GSM Asso- ciation, Official Document NG.116, Version 10.0, 2024. [Online]. Available: https://www.gsma.com/newsroom/gsma resources/ ng-116-generic-network-slice-template-v8-0/

2024

[5] [5]

Communication-efficient learning of deep networks from decentralized data,

B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inArtificial intelligence and statistics. Pmlr, 2017, pp. 1273– 1282

2017

[6] [6]

Advances and open problems in federated learning,

P. Kairouz, H. B. McMahanet al., “Advances and open problems in federated learning,”Foundations and Trends in Machine Learning, vol. 14, no. 1–2, pp. 1–210, 2021

2021

[7] [7]

Federated deep reinforce- ment learning for resource allocation in O-RAN slicing,

H. Zhang, H. Zhou, and M. Erol-Kantarci, “Federated deep reinforce- ment learning for resource allocation in O-RAN slicing,” inProc. IEEE Global Communications Conference (GLOBECOM), 2022

2022

[8] [8]

Federated meta- learning for traffic steering in O-RAN,

H. Erdol, X. Wang, P. Li, J. D. Thomas, R. Piechocki, G. Oikonomou, R. Inacio, A. Ahmad, K. Briggs, and S. Kapoor, “Federated meta- learning for traffic steering in O-RAN,” inProc. IEEE Vehicular Technology Conference (VTC2022-Fall), 2022

2022