Predicting Locations of Cell Towers for Network Capacity Expansion

Joy Bose; L Raghunatha Reddy; Sai Hareesh Anamandra; Sowmiyan Morri

REVIEW 4 major objections 5 minor 7 references

Predicting Locations of Cell Towers for Network Capacity Expansion

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read An AI planning loop predicts where new cell towers will best fill coverage gaps.

desk verdict A clearly written architecture sketch with zero empirical validation: everything rests on an untested coverage-prediction DNN, so the central claim is unsupported. read the letter →

arxiv 2507.19925 v1 pith:OW4C3SYD submitted 2025-07-26 cs.NI

Sowmiyan Morri , Joy Bose , L Raghunatha Reddy , Sai Hareesh Anamandra This is my paper · ORCID

classification cs.NI

keywords celltowerplacementnetworkcapacityexpansioncoveragepredictiondeepneuralspatialclusteringbudget-awareoptimizationradioplanningiterativeloop

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 proposes a machine-learning framework for deciding where to build new cell towers when a telecom network needs more coverage. The authors argue that a deep neural network trained on geographic, demographic, and infrastructural features can predict signal strength across a grid of candidate locations, and that clustering the predicted low-coverage points produces sensible candidate tower sites. A budget formula and an iterative planning loop are added so that each recommended tower is chosen against the current network state and the money available. The central claim is that this data-driven, budget-aware loop is a scalable alternative to manual drive tests and rule-based planning. The paper does not report a full end-to-end simulation; it presents the architecture and the argument for why it should work.

What carries the argument

The load-bearing object is the coverage predictor $f_\theta$, a deep neural network that maps each grid point's features—distance to nearest towers, terrain and altitude, urban/rural class, population density, frequency band, antenna type and directionality—to a predicted coverage value such as RSSI or SINR. The network is trained on labelled drive-test or crowdsourced measurements and, once trained, is assumed to generalize to new locations and hypothetical tower configurations. Around $f_\theta$, the framework builds a planning loop: extract grid points where predicted coverage $c_i$ falls below threshold $\tau$, cluster them with DBSCAN or K-means, take cluster centroids (or boundary points) as candidate sites, rank clusters by predicted coverage gain per cost, and limit the number of towers to $\lfloor B/C \rfloor$ under budget $B$ and per-tower cost $C$. The loop then adds the chosen sites to the configuration and re-runs prediction, continuing until coverage targets are met or budget is exhausted.

What would settle it

Train the proposed coverage network on drive-test or crowdsourced labels from one city, then hold out a set of locations and a set of hypothetical new tower configurations; if the model's predicted RSSI/SINR at the held-out points fails to beat a simple baseline such as distance-only path loss, or produces errors large enough to move points across the coverage threshold $\tau$, then the clusters and tower recommendations built on those predictions have no empirical foundation. A field test would be the stronger check: deploy a recommended tower and compare measured coverage after installation with the coverage the loop predicted for that configuration.

Watch

Extended reading notes

Core claim

The paper's central claim is that network capacity expansion can be planned as an automated loop: predict coverage everywhere with a trained deep network, mark points where predicted signal falls below a threshold, cluster those low-coverage points, propose cluster centroids as new tower locations, then repeat with the updated tower configuration until coverage targets are met or the budget is spent. The authors assert that this identifies low-coverage regions and recommends feasible tower placements using a deep neural network trained on geographic, demographic, and infrastructural features, and that embedding a budget constraint early in the clustering and ranking process makes the recommendations both technically and financially practical. In their telling, the framework advances radio network planning by offering a scalable, data-driven alternative to manual methods, while remaining modular and robust to missing inputs.

Load-bearing premise

The whole system relies on the assumption that a neural network trained on historical coverage measurements can accurately predict coverage at locations and tower configurations it has never seen; the paper gives no simulation or experiment that tests this generalization.

Editorial extensions

If this is right

Telecom operators could replace or reduce drive tests and manual site surveys with model-based coverage maps for candidate areas.
Budget constraints would enter planning at the clustering stage, not just at the end, so a fixed budget directly shapes how many and where towers are proposed.
Each new virtual tower changes the predicted coverage landscape, so later recommendations adapt to earlier ones instead of being computed once from a static plan.
The same framework would apply to urban densification, rural greenfield rollouts, highway corridors, and partial upgrades, since the input features are general.
Missing or low-quality data would not block the pipeline, because the preprocessing steps drop unusable samples and impute coverage labels at cell-site granularity.

Reading between the lines

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

The framework's usefulness hinges on one empirical question the paper does not answer: whether $f_\theta$ predicts coverage at new tower configurations accurately enough that the clustering step is fed real signal, not artifacts.
Because cluster centroids ignore land availability, zoning, and site acquisition costs, a practical deployment would need a feasibility filter after clustering; the paper's budget model covers only per-tower construction cost.
The same iterative loop could be extended to optimize not only locations but also antenna tilts, power levels, or frequency bands, since the predictor takes those as input features.
A cheap testable version of the idea could be built with open geospatial data and an open-source propagation simulator instead of live drive tests, which would reveal how sensitive the recommendations are to the coverage predictor's accuracy.

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

Quick take: this is a system description, not a demonstrated result. The pipeline—DNN coverage predictor, thresholding, clustering, budget loop—is clearly explained, and the authors are honest that they couldn't simulate it end to end. But the one component everything depends on, fθ, has no data, no training procedure, no accuracy numbers, not even a toy example. So the central claim that the framework identifies low-coverage regions and recommends feasible placements is unsupported.

What's genuinely there: the problem framing is sensible; drive-test and manual planning pain points are real. The iterative budget-aware loop that re-predicts coverage after each virtual tower is a reasonable way to think about sequential deployment, and it's a modest extension over prior work rather than a copy. Related work is cited and compared, including Binzer/Landstorfer, Chen et al., Qiu et al. The limitations section (VII) is candid about data heterogeneity, missing geospatial data, and computational cost. Credit where due: the authors know what they didn't do.

Soft spots, in proportion: - No experiments. This isn't a minor omission; the paper's value claim is scalability and generalizability, and there is no evidence for either. The abstract's "full-scale simulation was limited" is an understatement: there is no partial simulation, no synthetic data test, no ablation, no baseline comparison. - The load-bearing assumption is in Section IV: "Once trained, fθ can estimate coverage at any new location." Everything downstream inherits fθ's errors. The paper offers no reason to believe a DNN trained on historical drive-test data generalizes to hypothetical tower configurations, which is a different distribution. This is the kind of claim that needs held-out validation or at least a path-loss sanity check. - There is a self-referential evaluation risk: if the loop is ever tested using the same fθ that proposed the towers, improvements are predicted by the model, not measured. Section VII says components were "individually validated in prior research or internal prototypes," with no specifics. That sentence does a lot of work and cites nothing. - Manuscript hygiene: Section II contains a passage that reads like a critique of the authors' own method—saying simulations rely on an equal-user assumption and ignore water bodies—and the rest of the paper claims the opposite. Section VIII has a stray author name "Anamandra" embedded mid-sentence. These suggest a hastily assembled draft.

Bottom line: the architecture is plausible and the problem is worth solving, but as it stands this is a position paper with no empirical content. It doesn't deserve a full peer-review cycle. If you want to engage, treat it as a pointer to a research direction, not as evidence. I'd desk reject.

Referee Report

4 major / 5 minor

Summary. The paper proposes a machine-learning framework for cell-tower placement in telecom network expansion. The method combines a deep neural network (fθ) that predicts signal coverage from geographic, demographic, and infrastructural features with spatial clustering (DBSCAN/K-means) of low-coverage points, followed by budget-constrained selection of candidate tower sites in an iterative planning loop. The manuscript presents the architecture, the iterative algorithm, and qualitative arguments for its advantages, but it contains no experimental validation: no dataset, no training procedure, no simulation results, and no comparison to existing methods. Section VII explicitly acknowledges that full end-to-end simulation was not performed due to data and infrastructure limitations, and the abstract concedes that full-scale simulation was limited by data availability.

Significance. If the proposed framework were validated, it could offer a useful data-driven complement to manual drive-test-based planning, particularly because it integrates coverage prediction, spatial clustering, and budget constraints in a single iterative workflow. The modular design and explicit treatment of missing data are sensible engineering choices. However, the paper's central claim that the framework 'identifies low-coverage regions and recommends feasible tower placements' as a scalable alternative is entirely unsupported by evidence. The absence of any experimental or simulation-based evaluation, combined with the self-referential evaluation loop, means the paper currently functions as a design proposal rather than a demonstrated method. The strengths are the clarity of the proposed pipeline and the candid statement of limitations, but these do not substitute for validation.

major comments (4)

[Section VII and Abstract] The manuscript contains no experimental, simulation, or numerical validation. Section VII states that 'simulating the full system proved to be challenging' and the abstract says that 'full-scale simulation was limited by data availability.' The conclusion (Section IX) nevertheless claims that the framework 'identifies low-coverage regions and recommends feasible tower placements,' but this claim is unsupported. The paper is a system proposal; the central claim of being a scalable, data-driven alternative requires at least a synthetic or semi-synthetic evaluation to demonstrate that the pipeline works.
[Section IV, Machine Learning Module 1 and Budget Constrained Iterative Optimization] The iterative loop's step 6, 'Re-compute the predicted coverage with updated configuration,' uses the same DNN fθ that was trained on historical coverage data. The paper does not describe any retraining, fine-tuning, or physics-based correction after adding a tower. Consequently, the predicted coverage improvements are self-referential outputs of the model rather than independent measurements. Any systematic bias in fθ (for example, underestimating coverage near buildings or water bodies) propagates into every downstream recommendation and into the loop's stopping criterion, so the claimed 'adaptive' refinement is not evidence of actual coverage improvement.
[Section IV, Machine Learning Module 1] The load-bearing generalization assumption is asserted but never tested. The paper states that 'Once trained, fθ can estimate coverage at any new location given its features and current network configuration,' but it provides no dataset description, no network architecture, no hyperparameters, no train/test split, and no held-out accuracy metric. There is also no comparison with drive tests, ray tracing, or standard path-loss models. Without evidence that fθ generalizes to new configurations, the subsequent clustering and budget-based recommendations in Module 2 and Section VI are built on an unvalidated predictor.
[Section VI and Section IV, Machine Learning Module 2] The clustering and budget-constrained selection heuristics (DBSCAN/K-means, centroid versus boundary placement, pruning by 'coverage gain per cost') are described only qualitatively and are never evaluated against any baseline. The paper does not compare against random placement, existing manual plans, or a standard greedy coverage algorithm, nor does it report any sensitivity analysis for the threshold τ, the cost per tower C, or the clustering parameters. The claim that the framework is 'scalable' and 'cost-effective' requires at least a benchmark demonstration on a synthetic grid to establish that the pipeline produces sensible recommendations.

minor comments (5)

[Section II] Section II contains a duplicated paragraph: the text beginning 'From the perspective of telecom operators, installing base stations without considering inputs from the sales department...' repeats nearly verbatim the earlier paragraph that begins 'From the perspective of the Network Engineer (NE)...' and continues with the same latching, intra-circle roaming, and drive-test content.
[Section I] The final sentence of Section I says 'Section 7 includes a short discussion, and section 9 concludes the paper,' but the discussion is in Section VIII and the conclusion is in Section IX; Section VII is 'Simulation and Technical Difficulties.' The section numbering in the introduction is therefore inconsistent with the actual structure.
[Section IV, Budget Constrained Iterative Optimization] The line 'installed within a given budget, using a simple constraint:' appears twice in immediate succession, creating a duplicated fragment in the text.
[Section VIII] The sentence 'After each round of planning and simulated Anamandra deployment' appears to contain an unintended author name ('Anamandra') in the middle of the text; this is likely a copy-paste artifact and should be corrected.
[References] The reference list is inconsistent: reference [1] is formatted differently from the others, reference [2] includes an unrelated ScienceDirect link, and several references lack page numbers or DOIs. Please standardize the bibliography.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an unvalidated architecture proposal, not a derivation that reduces to its own inputs.

full rationale

The paper does not present a completed derivation or validation chain. It describes a machine-learning pipeline in which a deep neural network fθ is trained by supervised learning on labeled coverage data, and downstream steps (thresholding, clustering, budget ranking, iterative re-prediction) are explicit deterministic functions of fθ. This is a model-in-the-loop optimization scheme, not a case where a fitted parameter is renamed as a prediction or where a claimed result is equivalent to an input by construction. The abstract and Section VII explicitly state that full-scale simulation was not performed and that data availability limited validation; that is an evidentiary weakness, not circularity. No load-bearing self-citation, imported uniqueness theorem, or ansatz-smuggling citation appears in the manuscript. The fact that candidate selection and the reported coverage improvement are both based on fθ is a stated design choice rather than a hidden reduction: the paper does not claim independent ground-truth confirmation of the recommended improvements. Consequently, the appropriate circularity finding is a non-finding with score 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework rests on unverified assumptions about data availability and DNN generalization. The paper itself acknowledges the lack of simulation and data, so these assumptions are load-bearing.

free parameters (3)

neural network weights θ = unknown
Defined in Section IV as the parameters of the coverage prediction DNN, to be learned from data; no pretrained weights or architecture details are given.
coverage threshold τ = unspecified
Used in Module 2 to decide which points count as low-coverage; value is left open, so results would depend on its choice.
cost per tower C = unspecified
Used in the budget constraint Max Sites = floor(B/C); fixed cost is assumed but not quantified.

assumptions (4)

domain assumption Sufficient, clean, spatially aligned historical coverage and demographic data are available.
Assumed in Section IV Data Acquisition; the paper admits in Section VII that such data was not available, so the assumption is load-bearing.
ad hoc to paper The trained DNN can generalize coverage predictions to new tower configurations without retraining.
Invoked in Section IV Module 1: 'Once trained, fθ can estimate coverage at any new location given its features and current network configuration.' This is a strong generalization claim with no supporting evidence.
domain assumption Clustering low-coverage points and placing towers at centroids improves coverage.
Assumed in Section V clustering-based site selection; no simulation or empirical test supports this heuristic.
domain assumption Budget is a fixed per-tower cost that does not vary with location.
Assumed in Section VI; real-world tower costs vary by terrain and infrastructure.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Predicting Locations of Cell Towers for Network Capacity Expansion." pith.science (2026). https://pith.science/paper/OW4C3SYD

@misc{pith2026250719925,
  author       = {Pith},
  title        = {Pith review of: Predicting Locations of Cell Towers for Network Capacity Expansion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OW4C3SYD}},
  note         = {Machine review of arXiv:2507.19925}
}

read the original abstract

Network capacity expansion is a critical challenge for telecom operators, requiring strategic placement of new cell sites to ensure optimal coverage and performance. Traditional approaches, such as manual drive tests and static optimization, often fail to consider key real-world factors including user density, terrain features, and financial constraints. In this paper, we propose a machine learning-based framework that combines deep neural networks for signal coverage prediction with spatial clustering to recommend new tower locations in underserved areas. The system integrates geospatial, demographic, and infrastructural data, and incorporates budget-aware constraints to prioritize deployments. Operating within an iterative planning loop, the framework refines coverage estimates after each proposed installation, enabling adaptive and cost-effective expansion. While full-scale simulation was limited by data availability, the architecture is modular, robust to missing inputs, and generalizable across diverse deployment scenarios. This approach advances radio network planning by offering a scalable, data-driven alternative to manual methods.

Figures

Figures reproduced from arXiv: 2507.19925 by the authors.

**Figure 1.** Visualization of a unit area, broken into n equally spaced points on X and Y dimensions. This is represented by an n-dimensional binary vector [1 0 0 …. 1…0], with 1 standing for a cell currently existing at that point, 0 for a cell (or tower) not existing at that point. This vector we call the cell vector for that area. Similarly, there is a real valued n^2 dimensional coverage vector, which gives the coverage at e… view at source ↗

**Figure 2.** High level flowchart of the AI/ML solution to determine the optimal locations of the cells [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 4.** Illustration of the training of the machine learning model to predict locations of cell towers for network capacity expansion [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 3.** Figure 3: Architecture of an embodiment of the AI/ML solution to determine the coverage at each of the coordinates in the considered geographical area. Here the AI/ML model is represented as a deep neural network (DNN) The figure 3 presents the architecture of the deep learning …

**Figure 5.** Figure 5: Illustration of the application phase for inference of the machine learning model to predict locations of cell towers for network capacity expansion [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

7 extracted references · 7 canonical work pages

[1]

Placement optimization of aerial base stations with deep reinforcement learning

Qiu J, Lyu J, Fu L. Placement optimization of aerial base stations with deep reinforcement learning. InICC 2020 -2020 IEEE International Conference on Communications (ICC) 2020 Jun 7 (pp. 1-6). IEEE

work page 2020
[2]

Radio network planning with neural networks

Binzer T, Landstorfer FM. Radio network planning with neural networks. InVehicular Technology Conference Fall 2000. IEEE VTS Fall VTC2000. 52nd Vehicular Technology Conference (Cat. No. 00CH37152) 2000 Sep 24 (Vol. 2, pp. 811 -817). IEEE. https://www.sciencedirect.com/science/article/pii/S0308596113001900

work page 2000
[3]

Method and apparatus for Network Capacity expansion

China United Network Communications Group Co Ltd . Method and apparatus for Network Capacity expansion. US Patent US10477413B2. [Online]. Available: https://patents.google.com/patent/US10477413B2/en

work page
[4]

A novel method to determine the optimal location for a cellular tower by using LiDAR data

Bharadwaj S, Dubey R, Zafar MI, Tiwary SK, Faridi RA, Biswas S. A novel method to determine the optimal location for a cellular tower by using LiDAR data. Applied System Innovation. 2022 Feb 23;5(2):30

work page 2022
[5]

Base station planning based on region division and mean shift clustering

Chen J, Shi Y, Sun J, Li J, Xu J. Base station planning based on region division and mean shift clustering. Mathematics. 2023 Apr 21;11(8):1971

work page 2023
[6]

TelePlanNet: An AI-Driven Framework for Efficient Telecom Network Planning

Deng Z, Cai Y, Liu Q, Mu S, Lyu B, Yang Z. TelePlanNet: An AI-Driven Framework for Efficient Telecom Network Planning. arXiv preprint arXiv:2505.13831. 2025 May 20

work page Pith review arXiv 2025
[7]

Multi -Objective Deep Reinforcement Learning for 5G Base Station Placement to Support Localisation for Future Sustainable Traffic

Al-Tahmeesschi A, Talvitie J, López -Benítez M, Ahmadi H, Ruotsalainen L. Multi -Objective Deep Reinforcement Learning for 5G Base Station Placement to Support Localisation for Future Sustainable Traffic. In2024 Joint European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit) 2024 Jun 3 (pp. 493-498). IEEE

work page 2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Placement optimization of aerial base stations with deep reinforcement learning

Qiu J, Lyu J, Fu L. Placement optimization of aerial base stations with deep reinforcement learning. InICC 2020 -2020 IEEE International Conference on Communications (ICC) 2020 Jun 7 (pp. 1-6). IEEE

work page 2020

[2] [2]

Radio network planning with neural networks

Binzer T, Landstorfer FM. Radio network planning with neural networks. InVehicular Technology Conference Fall 2000. IEEE VTS Fall VTC2000. 52nd Vehicular Technology Conference (Cat. No. 00CH37152) 2000 Sep 24 (Vol. 2, pp. 811 -817). IEEE. https://www.sciencedirect.com/science/article/pii/S0308596113001900

work page 2000

[3] [3]

Method and apparatus for Network Capacity expansion

China United Network Communications Group Co Ltd . Method and apparatus for Network Capacity expansion. US Patent US10477413B2. [Online]. Available: https://patents.google.com/patent/US10477413B2/en

work page

[4] [4]

A novel method to determine the optimal location for a cellular tower by using LiDAR data

Bharadwaj S, Dubey R, Zafar MI, Tiwary SK, Faridi RA, Biswas S. A novel method to determine the optimal location for a cellular tower by using LiDAR data. Applied System Innovation. 2022 Feb 23;5(2):30

work page 2022

[5] [5]

Base station planning based on region division and mean shift clustering

Chen J, Shi Y, Sun J, Li J, Xu J. Base station planning based on region division and mean shift clustering. Mathematics. 2023 Apr 21;11(8):1971

work page 2023

[6] [6]

TelePlanNet: An AI-Driven Framework for Efficient Telecom Network Planning

Deng Z, Cai Y, Liu Q, Mu S, Lyu B, Yang Z. TelePlanNet: An AI-Driven Framework for Efficient Telecom Network Planning. arXiv preprint arXiv:2505.13831. 2025 May 20

work page Pith review arXiv 2025

[7] [7]

Multi -Objective Deep Reinforcement Learning for 5G Base Station Placement to Support Localisation for Future Sustainable Traffic

Al-Tahmeesschi A, Talvitie J, López -Benítez M, Ahmadi H, Ruotsalainen L. Multi -Objective Deep Reinforcement Learning for 5G Base Station Placement to Support Localisation for Future Sustainable Traffic. In2024 Joint European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit) 2024 Jun 3 (pp. 493-498). IEEE

work page 2024