Data-Driven Cellular Mobility Management via Bayesian Optimization and Reinforcement Learning

Basem Shihada; David L\'opez-P\'erez; Giovanni Geraci; Mohamed Benzaghta; Sahar Ammar

REVIEW 3 major objections 5 minor 53 references

Data-Driven Cellular Mobility Management via Bayesian Optimization and Reinforcement Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Per-cell handover settings chosen by Bayesian optimization or reinforcement learning beat fixed standard benchmarks in a 30-cell London network.

desk verdict A sound, genuinely novel application of high-dimensional BO to per-cell handover optimization, held back by missing error bars and an unvalidated ray-tracing channel model, but still worth refereeing. read the letter →

arxiv 2505.21249 v1 pith:ULZA722O submitted 2025-05-27 cs.IT math.IT

Mohamed Benzaghta , Sahar Ammar , David López-Pérez , Basem Shihada , Giovanni Geraci This is my paper

classification cs.ITmath.IT

keywords mobilitymanagementhandoveroptimizationBayesianTuRBOdeepreinforcementlearningPPOtime-to-triggerradiolinkfailure

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

This paper claims that handover quality in a dense urban cellular network improves when the two standard handover parameters—A3-offset, how much stronger a neighbor cell must be before a switch, and time-to-trigger (TTT), how long that condition must persist—are chosen per cell from data instead of being fixed network-wide. It offers two data-driven ways to choose them: high-dimensional Bayesian optimization over all cells' settings, and a deep reinforcement learning agent that picks the serving cell in real time without any thresholds. In a ray-traced model of a 30-cell London deployment, both approaches reduce ping-pong handovers and radio link failures relative to the 3GPP set-1 and set-5 benchmark configurations, with per-cell Bayesian optimization cutting ping-pongs by as much as 73% at 60 km/h. The result matters because mobility management is a bottleneck in dense networks, and per-cell data-driven tuning promises better handover behavior without any hardware change.

What carries the argument

The central machinery is joint per-cell optimization of A3-offset and TTT, with high-dimensional Bayesian optimization implemented as Trust Region BO (TuRBO): multiple local Gaussian-process surrogate models, each confined to a hyperrectangle that grows or shrinks with success or failure, keep the 60-dimensional search tractable. The objective is a weighted sum of ping-pong rate and either handover-failure or radio link failure rate, and each evaluation is a system-level simulation. The DRL counterpart replaces threshold parameters entirely: a PPO agent observes a state made of the street ID, serving-cell ID, time-of-stay, candidate-cell RSRP, and a short history of serving cells, and takes a binary stay-or-hand-over action, with a reward that penalizes ping-pongs and radio link failures.

What would settle it

A field trial on the same ten London sites, applying the HD-BO-recommended per-cell A3-offset and TTT configuration to real users on the five studied streets and measuring ping-pong and radio link failure rates, would settle the claim; if the reported reductions over the two benchmark configurations (for example the 73% ping-pong reduction at 60 km/h) do not appear, the gains are an artifact of the simulation model rather than a property of the network.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the handover trade-off between ping-pongs and failures is not fixed by network-wide defaults; it can be optimized, cell by cell, from observed performance. The paper formulates two weighted objectives—one penalizing ping-pongs plus handover failures, the other penalizing ping-pongs plus radio link failures—with sixty decision variables for thirty cells. The high-dimensional Bayesian optimizer finds per-cell A3-offset and TTT settings that, in the London simulation, drive radio link failures to zero in the targeted scenarios and cut ping-pongs by up to 73% relative to the 3GPP set-1 benchmark. A PPO-based reinforcement learning agent that selects serving cells directly reaches comparable ping-pong and radio link failure performance, but needs 10–250 times more simulation iterations to converge; with transfer learning, its training cost drops by a factor of 2.5 when moving from ground users to aerial users.

Load-bearing premise

The load-bearing premise is that the ray-traced propagation model built from the London map, with concrete buildings, five reflections and one diffraction at 2 GHz, reproduces the real radio environment closely enough that the handover settings found in simulation would also reduce ping-pongs and link failures on the live network.

Editorial extensions

If this is right

Per-cell handover parameter tuning is viable at network scale: with 30 cells and 60 variables, the optimizer converges in tens to low hundreds of simulations.
A threshold-free learned handover policy can match tuned threshold-based settings, so networks may choose between interpretable parameter curves and direct cell-selection policies.
Transfer learning lets one speed-specific model adapt to another speed; for the DRL agent it cuts training iterations by 2.5× while keeping ping-pong and radio link failure rates at 0%.
Optimizing against radio link failures rather than handover failures is the safer objective: the paper reports it also eliminates handover failures, while the reverse leaves residual outages.
When measurements are costly, HD-BO is the more practical option because it needs far fewer simulator runs than DRL.

Reading between the lines

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

Editorial inference: The per-cell settings HD-BO finds could be mapped against street geometry and building layout; if optimal A3/TTT values cluster by line-of-sight exposure rather than by cell identity, the method could be generalized to sites without training data.
Editorial inference: The 2.5× DRL training reduction is demonstrated for one source-target pair, ground to aerial users at 30 km/h; the paper leaves open whether the same warm-start works across cities, altitudes, or speed gaps.
Editorial inference: Because the DRL reward contains only ping-pong and RLF penalties, adding a throughput or latency term may change the learned policy; whether ping-pong reduction trades away user throughput is not tested.
Editorial inference: The entire comparison depends on the ray-traced channel model; replacing it with measured RSRP traces from the same sites would convert the proposed loops into a closed-loop field trial and test the reported gains outside simulation.

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

The genuinely new thing here is pushing TuRBO to optimize 60 handover parameters (A3-offset and TTT per cell) in a realistic 30-cell London layout, and using transfer learning to reuse that optimization across speeds and between ground and aerial users. That is a real niche and the paper executes it clearly. The DRL component is less novel—PPO for handover decisions is well-trodden—but it serves as a useful baseline and the comparison is honest.

The paper does several things well. The system model is detailed: Sionna ray tracing, real site geometry from OpenStreetMap, explicit HO event A3 logic, L1/L3 filtering, T310, ping-pong definition. The benchmark comparison to 3GPP set-1/set-5 is sensible, and per-cell optimization shows a clear advantage over uniform thresholds. I also appreciate that the authors report an unsuccessful transfer-learning case (pedestrian speeds) rather than only the successes.

The soft spots are real but not fatal. The headline numbers are single point estimates with no error bars or multiple seeds; for a stochastic simulation, that is hard to evaluate. The HOF definition is simplified (SINR only at HO complete), which the authors admit and which weakens the PP-HOF study. The claim that RL-PPO needs '10 to 250 times' more iterations doesn't exactly match Table III (the ratios run from about 26 to 233), though that is a minor wording issue. There is no code or data release, so reproducibility is limited.

The biggest concern is the unvalidated Sionna RT propagation model. The paper says 'real-world cellular deployment scenario' but that is only the geometry; the channel itself is simulated with default material parameters and no calibration to drive-test or MDT measurements. The optimized per-cell parameters could simply be absorbing biases of this specific ray-tracing configuration. I don't think this undermines the algorithmic comparison—both HD-BO and DRL are being optimized against the same simulator—but it does weaken the claim that the specific parameter values would transfer to the real London network. A sensitivity analysis over reflection/diffraction settings or a sanity check against a simple empirical path-loss model would go a long way.

Who is this for? Researchers working on SON, mobility robustness, and data-driven network tuning. It deserves peer review; the idea is sound and the execution is mostly careful. Before publication, I'd want error bars, a note reconciling the iteration-ratio claim with Table III, and some discussion of the channel-model sensitivity.

Referee Report

3 major / 5 minor

Summary. The paper proposes two data-driven approaches for cellular mobility management: (i) high-dimensional Bayesian optimization (HD-BO) based on TuRBO to tune per-cell A3-offset and TTT parameters, and (ii) a deep reinforcement learning (DRL) approach based on PPO that directly selects the serving cell without predefined thresholds. The methods are evaluated in a site-specific 30-cell London deployment modeled via Sionna ray tracing, comparing against 3GPP set-1 and set-5 benchmarks for ground UEs at 3, 30, and 60 km/h and for aerial UEs at 150 m. The paper reports that both HD-BO and DRL outperform the 3GPP benchmarks in terms of ping-pongs and RLF/HOF, with HD-BO being more sample-efficient, and that transfer learning reduces DRL training time by a factor of 2.5 while preserving performance. The central claim is that data-driven per-cell optimization provides practical gains over uniform 3GPP configurations.

Significance. If the reported results are robust, the paper makes a useful contribution by demonstrating that HD-BO can scale to a 60-variable mobility optimization problem and by providing a direct comparison between parameter-based BO and parameter-free DRL on a realistic deployment geometry. The use of a real-world topology (London streets, actual site locations) and standard 3GPP HO models strengthens the application-oriented value. The transfer-learning results are also of practical interest. However, the paper's significance is partially undermined by the lack of statistical confidence (single point estimates, no error bars or multiple seeds) and by the absence of any validation of the ray-traced channel model against field measurements, which is critical for the claimed 'real-world' applicability.

major comments (3)

[Section II-A and all reported KPIs] The site-specific propagation model is the load-bearing foundation of every reported KPI, but it is never validated. Section II-A specifies Sionna RT with itu_concrete materials, 5 reflections, and 1 diffraction at 2 GHz, yet the paper provides no calibration against drive-test/MDT measurements from the 'production radio network' it references, no comparison to an empirical path-loss model, and no sensitivity analysis over ray-tracing parameters. Since HD-BO tunes 60 parameters directly to simulation outcomes (and PPO learns a policy from simulated rewards), any systematic bias in the simulated channel will be absorbed into the optimized A3/TTT values and the learned policy. The abstract's claim to validate the approach on a 'real-world cellular deployment scenario' and the conclusion's implication of practical deployability are therefore not supported by the evidence as presented. The authors should either add a calibration experiment against field measurements, provide a sensitivity analysis showing that conclusions are stable across ray-tracing configurations, or substantially temper the claims to be explicitly about the simulation model rather than the real network.
[Section IV-C and abstract] The statement that 'RL-PPO requires 10 to 250 times more iterations before convergence compared to HD-BO' is not consistent with Table III. For the weight configuration w_PP = 1, w_RLF = 9, the ratios are 320/140 ≈ 2.3, 300/75 = 4, and 270/45 = 6, all below 10. Only the w_PP = 9, w_RLF = 1 rows fall in the 10–250 range (25.6, 233, and 143, respectively). This claim appears in the abstract and in Section IV-C and must be corrected to reflect the dependence on the KPI weights, or Table III must be reconsidered.
[Section III-B and all results tables/figures] All KPI numbers (Tables I, II, IV; Figs. 2–9) are presented as single-point estimates without error bars, confidence intervals, or multiple independent seeds. The authors themselves acknowledge in Section III-B that 'repeated evaluations at the same query point x may yield different outcomes due to the intrinsic stochasticity of the environment.' Without repeated runs, it is impossible to assess whether the observed differences between HD-BO, RL-PPO, and the 3GPP baselines are statistically significant, or whether the claimed improvements could be within simulation noise.

minor comments (5)

[Eq. (12)] The policy objective in Eq. (12) has mismatched parentheses: the expression should be min(ρ_t A_ψ, clip(ρ_t, 1−τ, 1+τ) A_ψ) with a closing parenthesis after A_ψ, but the text writes 'min(ρtAπψ , clip(ρt, 1−τ, 1+τ )Aπψ ]' mixing parentheses and brackets. Please fix the notation.
[Section III-C, heading 'Minimizing HOF vs. RLF'] The heading 'Minimizing HOF vs. RLF' is confusing because the subsection actually compares two KPI studies (ping-pongs vs. HOF and ping-pongs vs. RLF), not 'HOF vs. RLF' directly. Consider renaming the heading to 'Comparing Ping-Pongs vs. HOF and Ping-Pongs vs. RLF'.
[Section III-E, Fig. 7] The convergence curves in Fig. 7 are single realizations and the claim that 'performance declines by only 3%' is not quantified or supported with error bars. Adding multiple seeds and reporting the mean and spread would strengthen this result.
[Section IV-C, Table III] The iteration counts in Table III are not clearly defined for each method. It should be stated what one iteration means for HD-BO (one objective evaluation) and for RL-PPO (one policy update or one environment rollout), as this is essential for interpreting the sample-efficiency comparison.
[Section I-C, contribution bullet] The bullet in Section I-C states that HD-BO reduces ping-pongs by 73% for GUEs at 60 km/h 'compared to 3GPP benchmarks.' It should specify which benchmark (set-1 or set-5) and clarify that this is for per-cell optimization, as the comparison is not unique across figures.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the HD-BO and DRL comparisons are benchmarked against external 3GPP set-1/set-5 configurations, and the only same-author dependency (UAV antenna tilts from [8]/[47]) is an environment setup, not a predicted output.

full rationale

The central claims, that per-cell HD-BO optimization of A3-offset/TTT and a PPO-based DRL policy outperform the 3GPP set-1 (TTT 480 ms, A3-offset 3 dB) and set-5 (TTT 40 ms, A3-offset -1 dB) benchmarks, are evaluated against external 3GPP TR 36.839 configurations under the same Sionna RT simulation. No reported KPI reduces by construction to a fitted input: the benchmark values (Tables I-II, Figs. 3 and 9) are independent parameter choices, not outputs of the proposed methods, and the paper even reports a case where HD-BO does not beat set-1 (Table I, PP-RLF ping-pongs 56.4% vs set-1 48.8%), showing the favorable numbers are not forced. The claim that set-1 and set-5 are feasible points in the search box (A3 in [-1,3] dB, TTT in [40,480] ms) only guarantees domination of the optimizer on the weighted training objective, not the per-KPI headline claims. The only same-author citation that materially shapes a result is the antenna-tilt configuration for Case Study 2: 'we first optimize the electrical antenna tilts of all BSs following the approach in [8]' (Section III-D, also citing [47]). This is an environment setup inherited from prior published work (IEEE Globecom 2023) with an independent objective (maximizing GUE and UAV rates); it is not a prediction derived from the present mobility KPIs, and it is not a definitional reduction, so it is at most a minor self-citation dependency. The transfer-learning claims (Fig. 7, Table IV) are empirical convergence statements with warm-started initial datasets, and the Fig. 7 min-max normalization is anchored to the 100%-target-condition result, not to the claim being tested. The genuine weaknesses lie outside circularity: (i) the Sionna RT model (Section II-A: itu_concrete materials, 5 reflections, 1 diffraction, 2 GHz) is not calibrated against operator drive-test or MDT measurements, so the abstract's 'we validate our approach using a real-world cellular deployment scenario' asserts deployment realism without validating the propagation model; and (ii) the KPIs are reported in-sample on the same simulation used for optimization, with no held-out deployment. Both are correctness and external-validity risks, not circular reductions.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

Everything in the paper rests on a simulated environment and on hand-chosen objective weights; these are listed below. No new physical entities or mediators are introduced, so invented_entities is empty.

free parameters (5)

Objective weights wPP and wRLF = wPP=9,wRLF=1 and wPP=1,wRLF=9
Chosen by hand to shape the ping-pong vs RLF tradeoff; no sensitivity analysis is provided, and the reported gains depend on these weights.
TuRBO hyperparameters = tau_succ=3, tau_fail=15, L_init=0.8, L_min=2^-7, L_max=1.6
Taken from the open-source TuRBO repository; these affect search behavior and convergence but are standard defaults.
PPO hyperparameters = lr=0.0001, gamma=0.95, entropy=0.002, clip=0.2, actor [256,128,256], critic [64,64,64]
Selected through extensive experimentation; training time and final policy quality depend on these choices.
History length n in DRL state = not specified
The state includes a history Hn_t of n previous serving BSs, but the value of n is never given.
Initial dataset size N0 = 60 for per-cell optimization (60 variables)
Follows the recommendation in [38]; transfer learning uses 100%, 50%, and 0% splits of target-scenario data.

assumptions (6)

domain assumption Sionna RT simulation with itu_concrete and up to 5 reflections and 1 diffraction accurately represents the real London deployment.
Section II-A. All results depend on the fidelity of this unvalidated propagation model.
ad hoc to paper HOF can be approximated as SINR below Qout at the moment the HO complete message is sent.
Section II-B. This simplification may not match 3GPP HOF definitions and affects KPI study #1.
domain assumption UE mobility along five selected streets at constant speeds is representative of real network traffic.
Section II-A and Case Study #1. The generalization to real networks depends on this selection.
domain assumption The weighted objective functions of ping-pongs vs HOF and ping-pongs vs RLF capture the operator-relevant tradeoff.
Section III-A. The weights are hand chosen and no sensitivity analysis is provided.
standard math Gaussian process and trust region heuristics in TuRBO efficiently model the nonconvex objective.
Section III-B; relies on [38]. Standard method assumptions for Bayesian optimization.
standard math PPO converges to a good policy for the MDP defined by the simulator.
Section IV-B; standard RL assumption, but no convergence guarantees for this stochastic environment.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Data-Driven Cellular Mobility Management via Bayesian Optimization and Reinforcement Learning." pith.science (2026). https://pith.science/paper/ULZA722O

@misc{pith2026250521249,
  author       = {Pith},
  title        = {Pith review of: Data-Driven Cellular Mobility Management via Bayesian Optimization and Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ULZA722O}},
  note         = {Machine review of arXiv:2505.21249}
}

read the original abstract

Mobility management in cellular networks faces increasing complexity due to network densification and heterogeneous user mobility characteristics. Traditional handover (HO) mechanisms, which rely on predefined parameters such as A3-offset and time-to-trigger (TTT), often fail to optimize mobility performance across varying speeds and deployment conditions. Fixed A3-offset and TTT configurations either delay HOs, increasing radio link failures (RLFs), or accelerate them, leading to excessive ping-pong effects. To address these challenges, we propose two data-driven mobility management approaches leveraging high-dimensional Bayesian optimization (HD-BO) and deep reinforcement learning (DRL). HD-BO optimizes HO parameters such as A3-offset and TTT, striking a desired trade-off between ping-pongs vs. RLF. DRL provides a non-parameter-based approach, allowing an agent to select serving cells based on real-time network conditions. We validate our approach using a real-world cellular deployment scenario, and employing Sionna ray tracing for site-specific channel propagation modeling. Results show that both HD-BO and DRL outperform 3GPP set-1 (TTT of 480 ms and A3-offset of 3 dB) and set-5 (TTT of 40 ms and A3-offset of -1 dB) benchmarks. We augment HD-BO with transfer learning so it can generalize across a range of user speeds. Applying the same transfer-learning strategy to the DRL method reduces its training time by a factor of 2.5 while preserving optimal HO performance, showing that it adapts efficiently to the mobility of aerial users such as UAVs. Simulations further reveal that HD-BO remains more sample-efficient than DRL, making it more suitable for scenarios with limited training data.

Figures

Figures reproduced from arXiv: 2505.21249 by the authors.

**Figure 2.** SINR for GUEs at 3 km/h: the 3GPP benchmark configurati [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** GUE ping-pongs and RLF performance at a certain speed [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 6.** Figure 6: SINR achieved by GUEs of different speeds when the net [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: Convergence of transfer learning applied on Case Stu [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

**Figure 9.** Figure 9: Ping-pongs and RLF performance for GUEs at different [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 49 canonical work pages

[1]

Capacity an d power con- sumption of multi-layer 6G networks using the upper mid-ban d,

D. Lopez-Perez, N. Piovesan, and G. Geraci, “Capacity an d power con- sumption of multi-layer 6G networks using the upper mid-ban d,” in Proc. IEEE ICC , 2025

work page 2025
[2]

The future of wireless broadband in the peak smartphone era : 6G, Wi-Fi 7, and Wi-Fi 8,

E. Oughton, G. Geraci, M. Polese, M. Ghosh, W. Webb, and D. Bubley, “The future of wireless broadband in the peak smartphone era : 6G, Wi-Fi 7, and Wi-Fi 8,” IEEE Wireless Commun. Mag. , 2025

work page 2025
[3]

Incorporating mobil ity prediction in handover procedure for frequent-handover mitigation in small-cell networks,

S. M. Shahid, J.-H. Na, and S. Kwon, “Incorporating mobil ity prediction in handover procedure for frequent-handover mitigation in small-cell networks,” IEEE Trans. Network Sci. and Eng. , 2024

work page 2024
[4]

User mobility evaluation for 5G small cell networks based on individual mobility model,

X. Ge, J. Y e, Y . Y ang, and Q. Li, “User mobility evaluation for 5G small cell networks based on individual mobility model,” IEEE J. on Sel. Areas Commun., 2016

work page 2016
[5]

Geometry-based analysis of op timal han- dover parameters for self-organizing networks,

M.-T. Nguyen and S. Kwon, “Geometry-based analysis of op timal han- dover parameters for self-organizing networks,” IEEE Trans. Wireless Commun., 2020

work page 2020
[6]

A comprehensive survey on mobility management in 5G heteroge neous networks: Architectures, challenges and solutions,

E. Gures, I. Shayea, A. Alhammadi, M. Ergen, and H. Mohama d, “A comprehensive survey on mobility management in 5G heteroge neous networks: Architectures, challenges and solutions,” IEEE Access , 2020

work page 2020
[7]

What wil l the future of UA V cellular communications be? A ﬂight from 5G to 6 G,

G. Geraci, A. Garcia-Rodriguez, M. M. Azari, A. Lozano, M . Mezzavilla, S. Chatzinotas, Y . Chen, S. Rangan, and M. Di Renzo, “What wil l the future of UA V cellular communications be? A ﬂight from 5G to 6 G,” IEEE Commun. Surveys Tuts. , vol. 24, no. 3, pp. 1304–1335, 2022

work page 2022
[8]

Designing cellular networks for UA V corridors via Bayesian optimizat ion,

M. Benzaghta, G. Geraci, D. López-Pérez, and A. V alcarce , “Designing cellular networks for UA V corridors via Bayesian optimizat ion,” in Proc. IEEE Globecom , 2023, pp. 4552–4557

work page 2023

Show all 53 references

[9]

Mobilit y management in 5G and beyond: A novel smart handover with adaptive time-t o-trigger and hysteresis margin,

R. Karmakar, G. Kaddoum, and S. Chattopadhyay, “Mobilit y management in 5G and beyond: A novel smart handover with adaptive time-t o-trigger and hysteresis margin,” IEEE Trans. Mobile Comput. , 2022

2022
[10]

Stochastic geometry analysis o f user mobility in RF/VLC hybrid networks,

R. Arshad and L. Lampe, “Stochastic geometry analysis o f user mobility in RF/VLC hybrid networks,” IEEE Trans. Wireless Commun. , 2021

2021
[11]

Performance evaluation of handover triggerin g condition estimation using mobility models in heterogeneous mobile n etworks,

A. A. Malik, M. A. Jamshed, A. Nauman, A. Iqbal, A. Shakee l, and R. Hussain, “Performance evaluation of handover triggerin g condition estimation using mobility models in heterogeneous mobile n etworks,” IET Networks, 2024

2024
[12]

Pr ivacy- preserving handover optimization using federated learnin g and LSTM networks,

W.-C. Chien, Y . Huang, B.-Y . Chang, and W.-Y . Hwang, “Pr ivacy- preserving handover optimization using federated learnin g and LSTM networks,” Sensors, 2024

2024
[13]

Multi-agent Q-learning for real-time load balancing user association and handover in mobile netw orks,

A. Alizadeh, B. Lim, and M. Vu, “Multi-agent Q-learning for real-time load balancing user association and handover in mobile netw orks,” IEEE Trans. Wireless Commun. , 2024

2024
[14]

Op- timal beam association for high mobility mmwave vehicular n etworks: Lightweight parallel reinforcement learning approach,

N. V . Huynh, D. N. Nguyen, D. T. Hoang, and E. Dutkiewicz, “Op- timal beam association for high mobility mmwave vehicular n etworks: Lightweight parallel reinforcement learning approach,” IEEE Trans. Com- mun., 2021

2021
[15]

Enabling proportionally-fair mobility management with reinforcem ent learning in 5G networks,

A. Prado, F. Stöckeler, F. Mehmeti, P . Krämer, and W. Kel lerer, “Enabling proportionally-fair mobility management with reinforcem ent learning in 5G networks,” IEEE J. on Sel. Areas in Commun. , 2023

2023
[16]

Intelligent han dover manage- ment enabled by O-RAN and deep reinforcement learning,

J. Dai, S. Mahboob, H. Wang, and L. Liu, “Intelligent han dover manage- ment enabled by O-RAN and deep reinforcement learning,” in Proc. IEEE VTC, 2024, pp. 1–6

2024
[17]

Efﬁcient dro ne mobility support using reinforcement learning,

Y . Chen, X. Lin, T. Khan, and M. Mozaffari, “Efﬁcient dro ne mobility support using reinforcement learning,” in Proc. IEEE WCNC , 2020

2020
[18]

A deep learni ng approach to efﬁcient drone mobility support,

Y . Chen, X. Lin, T. Khan, and M. Mozaffari, “A deep learni ng approach to efﬁcient drone mobility support,” in Proc. ACM MobiCom W orkshop on Drone Assisted Wireless Commun. for 5G and Beyond , 2020, pp. 67–72

2020
[19]

Mobil ity man- agement for cellular-connected UA Vs: Model-based versus l earning-based approaches for service availability,

I. A. Meer, M. Ozger, D. A. Schupke, and C. Cavdar, “Mobil ity man- agement for cellular-connected UA Vs: Model-based versus l earning-based approaches for service availability,” IEEE Trans. on Network and Ser . Man., 2024

2024
[20]

REQIBA: Regression and deep Q-learning for intelligent UA V cellula r user to base station association,

B. Galkin, E. Fonseca, R. Amer, L. A. Dasilva, and I. Dusp aric, “REQIBA: Regression and deep Q-learning for intelligent UA V cellula r user to base station association,” IEEE Trans. V ehicular Tech., 2022

2022
[21]

Taking the human out of the loop: A review of Bayesian optimi zation,

B. Shahriari, K. Swersky, Z. Wang, R. P . Adams, and N. De F reitas, “Taking the human out of the loop: A review of Bayesian optimi zation,” Proc. IEEE , vol. 104, no. 1, pp. 148–175, 2015

2015
[22]

Optimi zing coverage and capacity in cellular networks using machine le arning,

R. M. Dreifuerst, S. Daulton, Y . Qian, P . V arkey, M. Bala ndat, S. Kasturia, A. Tomar, A. Y azdan, V . Ponnampalam, and R. W. Heath, “Optimi zing coverage and capacity in cellular networks using machine le arning,” in Proc. IEEE ICASSP , 2021, pp. 8138–8142

2021
[23]

A differentiable thr oughput model for load-aware cellular network optimization through gradien t descent,

L. Eller, P . Svoboda, and M. Rupp, “A differentiable thr oughput model for load-aware cellular network optimization through gradien t descent,” IEEE Access, 2024

2024
[24]

Bayesian and multi-armed contextual meta-optimization for efﬁcien t wireless radio resource management,

Y . Zhang, O. Simeone, S. T. Jose, L. Maggi, and A. V alcarc e, “Bayesian and multi-armed contextual meta-optimization for efﬁcien t wireless radio resource management,” IEEE Trans. on Cognitive Communications and Networking, 2023. 15

2023
[25]

Bayesian optimiz ation for radio resource management: Open loop power control,

L. Maggi, A. V alcarce, and J. Hoydis, “Bayesian optimiz ation for radio resource management: Open loop power control,” IEEE J. Sel. Areas Commun., vol. 39, no. 7, pp. 1858–1871, 2021

2021
[26]

Cell-free data power control via scalable multi-objective Bayesian optimisation,

S. S. Tambovskiy, G. Fodor, and H. Tullberg, “Cell-free data power control via scalable multi-objective Bayesian optimisation,” in Proc. IEEE PIMRC, 2022, pp. 1–6

2022
[27]

Energy sav- ings under performance constraints via carrier shutdown wi th Bayesian learning,

L. Maggi, C. Mihailescu, Q. Cao, A. Tetich, S. Khan, S. Aa ltonen, R. Koblitz, M. Holma, S. Macchi, M. E. Ruggieri et al. , “Energy sav- ings under performance constraints via carrier shutdown wi th Bayesian learning,” in Proc. EuCNC, 2023, pp. 1–6

2023
[28]

Joint uplink- downlink capacity and coverage optimization via site-spec iﬁc learning of antenna settings,

E. Tekgul, T. Novlan, S. Akoum, and J. G. Andrews, “Joint uplink- downlink capacity and coverage optimization via site-spec iﬁc learning of antenna settings,” IEEE Trans. Wireless Commun. , 2024

2024
[29]

Towards mobil ity management with multi-objective Bayesian optimization,

E. de Carvalho, A. V . Rial, and G. Geraci, “Towards mobil ity management with multi-objective Bayesian optimization,” in Proc. IEEE WCNC , 2023, pp. 1–6

2023
[30]

A tutorial on Bayesian optimization,

P . I. Frazier, “A tutorial on Bayesian optimization,” arXiv:1807.02811, 2018

2018 arXiv
[31]

Study on artiﬁcial intelligence (AI)/ machine learning (ML) for NR air interface (Release 18),

3GPP TR 38.843, “Study on artiﬁcial intelligence (AI)/ machine learning (ML) for NR air interface (Release 18),” Jun. 2023

2023
[32]

Sionna RT: Differentiable ray tra cing for radio propagation modeling,

J. Hoydis, F. A. Aoudia, S. Cammerer, M. Nimier-David, N . Binder, G. Marcus, and A. Keller, “Sionna RT: Differentiable ray tra cing for radio propagation modeling,” arXiv:2303.11103, 2023

2023 arXiv
[33]

Evolved Universal Terr estrial Radio Ac- cess (EUTRA); Further advancements for E-UTRA physical lay er aspects,

3GPP Technical Report 36.814, “Evolved Universal Terr estrial Radio Ac- cess (EUTRA); Further advancements for E-UTRA physical lay er aspects,” Mar. 2017

2017
[34]

Mobility enhancements in heterogeneo us networks,

3GPP TR 36.839, “Mobility enhancements in heterogeneo us networks,” Aug. 2012

2012
[35]

Mobility manage ment challenges in 3GPP heterogeneous networks,

D. Lopez-Perez, I. Guvenc, and X. Chu, “Mobility manage ment challenges in 3GPP heterogeneous networks,” IEEE Communications Magazine , vol. 50, no. 12, pp. 70–78, 2012

2012
[36]

Radio resource control; protocol spec iﬁcation,

3GPP TS 36.331, “Radio resource control; protocol spec iﬁcation,” Dec. 2011

2011
[37]

“evolved universal terrestrial radio access (e-utra) and evolved universal terrestrial radio access network (e-utr an),

3GPP TS 36.300, ““evolved universal terrestrial radio access (e-utra) and evolved universal terrestrial radio access network (e-utr an),” Oct. 2011

2011
[38]

Scalable global optimization via local Bayesian optimiza tion,

D. Eriksson, M. Pearce, J. Gardner, R. D. Turner, and M. P oloczek, “Scalable global optimization via local Bayesian optimiza tion,” NeurIPS, vol. 32, 2019

2019
[39]

High-dimensional Bayes ian optimization with sparse axis-aligned subspaces,

D. Eriksson and M. Jankowiak, “High-dimensional Bayes ian optimization with sparse axis-aligned subspaces,” in Uncertainty in Artiﬁcial Intelli- gence. PMLR, 2021, pp. 493–503

2021
[40]

Computationally efﬁcient hi gh-dimensional Bayesian optimization via variable selection,

Y . Shen and C. Kingsford, “Computationally efﬁcient hi gh-dimensional Bayesian optimization via variable selection,” arXiv:2109.09264, 2021

2021 arXiv
[41]

Spatially consistent air-to-ground channel modeling via generative neural networks,

A. Giuliani, R. Nikbakht, G. Geraci, S. Kang, A. Lozano, and S. Rangan, “Spatially consistent air-to-ground channel modeling via generative neural networks,” IEEE Commun. Lett. , vol. 13, no. 4, pp. 1158–1162, 2024

2024
[42]

Understanding UA V cellular communications: From existing networks to massive MIMO,

G. Geraci, A. Garcia-Rodriguez, L. Galati-Giordano, D . López-Pérez, and E. Björnson, “Understanding UA V cellular communications: From existing networks to massive MIMO,” IEEE Access , 2018

2018
[43]

Y . Zeng, I. Guvenc, R. Zhang, G. Geraci, and D. W. Matolak , UA V Communications for 5G and Beyond . John Wiley & Sons, 2020

2020
[44]

Integrating terrestrial and non-terrestrial networks: 3D opportuniti es and challenges,

G. Geraci, D. López-Pérez, M. Benzaghta, and S. Chatzin otas, “Integrating terrestrial and non-terrestrial networks: 3D opportuniti es and challenges,” IEEE Commun. Mag. , 2023

2023
[45]

Opt imizing cellular networks for UA V corridors via quantization theory,

S. Karimi-Bidhendi, G. Geraci, and H. Jafarkhani, “Opt imizing cellular networks for UA V corridors via quantization theory,” IEEE Trans. Wireless Commun., 2024

2024
[46]

Mat hematical cell deployment optimization for capacity and coverage of groun d and UA V users,

S. Karimi-Bidhendi, G. Geraci, and H. Jafarkhani, “Mat hematical cell deployment optimization for capacity and coverage of groun d and UA V users,” arXiv:2502.00928, 2025

2025 arXiv
[47]

Cellular net- work design for UA V corridors via data-driven high-dimensi onal Bayesian optimization,

M. Benzaghta, G. Geraci, D. López-Pérez, and A. V alcarc e, “Cellular net- work design for UA V corridors via data-driven high-dimensi onal Bayesian optimization,” arXiv:2504.05176, 2025

2025 arXiv
[48]

An overview of the 3GPP study on artiﬁcial intel ligence for 5G new radio,

X. Lin, “An overview of the 3GPP study on artiﬁcial intel ligence for 5G new radio,” arXiv:2308.05315, 2023

2023 arXiv
[49]

R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction, 2nd edition . MIT press Cambridge, 2018

2018
[50]

Human-level control through deep reinforcement learning ,

V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. V eness , M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ost rovski et al., “Human-level control through deep reinforcement learning ,” nature, vol. 518, no. 7540, pp. 529–533, 2015

2015
[51]

Proximal policy optimization algorithms,

J. Schulman, F. Wolski, P . Dhariwal, A. Radford, and O. K limov, “Proximal policy optimization algorithms,” arXiv preprint arXiv:1707.06347 , 2017

2017 arXiv
[52]

Massive MIMO for aerial highways: Enhancing cell selectio n via SSB beams optimization,

M. Bernabè, D. López-Pèrez, N. Piovesan, G. Geraci, and D. Gesbert, “Massive MIMO for aerial highways: Enhancing cell selectio n via SSB beams optimization,” IEEE Open Journal of the Communications Society , vol. 5, pp. 3975–3996, 2024

2024
[53]

UA V communications in integrated terrestrial and non-terrest rial networks,

M. Benzaghta, G. Geraci, R. Nikbakht, and D. López-Pére z, “UA V communications in integrated terrestrial and non-terrest rial networks,” in Proc. IEEE Globecom , 2022, pp. 3706–3711

2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Capacity an d power con- sumption of multi-layer 6G networks using the upper mid-ban d,

D. Lopez-Perez, N. Piovesan, and G. Geraci, “Capacity an d power con- sumption of multi-layer 6G networks using the upper mid-ban d,” in Proc. IEEE ICC , 2025

work page 2025

[2] [2]

The future of wireless broadband in the peak smartphone era : 6G, Wi-Fi 7, and Wi-Fi 8,

E. Oughton, G. Geraci, M. Polese, M. Ghosh, W. Webb, and D. Bubley, “The future of wireless broadband in the peak smartphone era : 6G, Wi-Fi 7, and Wi-Fi 8,” IEEE Wireless Commun. Mag. , 2025

work page 2025

[3] [3]

Incorporating mobil ity prediction in handover procedure for frequent-handover mitigation in small-cell networks,

S. M. Shahid, J.-H. Na, and S. Kwon, “Incorporating mobil ity prediction in handover procedure for frequent-handover mitigation in small-cell networks,” IEEE Trans. Network Sci. and Eng. , 2024

work page 2024

[4] [4]

User mobility evaluation for 5G small cell networks based on individual mobility model,

X. Ge, J. Y e, Y . Y ang, and Q. Li, “User mobility evaluation for 5G small cell networks based on individual mobility model,” IEEE J. on Sel. Areas Commun., 2016

work page 2016

[5] [5]

Geometry-based analysis of op timal han- dover parameters for self-organizing networks,

M.-T. Nguyen and S. Kwon, “Geometry-based analysis of op timal han- dover parameters for self-organizing networks,” IEEE Trans. Wireless Commun., 2020

work page 2020

[6] [6]

A comprehensive survey on mobility management in 5G heteroge neous networks: Architectures, challenges and solutions,

E. Gures, I. Shayea, A. Alhammadi, M. Ergen, and H. Mohama d, “A comprehensive survey on mobility management in 5G heteroge neous networks: Architectures, challenges and solutions,” IEEE Access , 2020

work page 2020

[7] [7]

What wil l the future of UA V cellular communications be? A ﬂight from 5G to 6 G,

G. Geraci, A. Garcia-Rodriguez, M. M. Azari, A. Lozano, M . Mezzavilla, S. Chatzinotas, Y . Chen, S. Rangan, and M. Di Renzo, “What wil l the future of UA V cellular communications be? A ﬂight from 5G to 6 G,” IEEE Commun. Surveys Tuts. , vol. 24, no. 3, pp. 1304–1335, 2022

work page 2022

[8] [8]

Designing cellular networks for UA V corridors via Bayesian optimizat ion,

M. Benzaghta, G. Geraci, D. López-Pérez, and A. V alcarce , “Designing cellular networks for UA V corridors via Bayesian optimizat ion,” in Proc. IEEE Globecom , 2023, pp. 4552–4557

work page 2023