Predicting Drive Test Results in Mobile Networks Using Optimization Techniques

Abolfazl Diyanat; Ali Nazari; MohammadJava Taheri; MortezaAli Ahmadi

REVIEW 3 major objections 4 minor 16 references

Predicting Drive Test Results in Mobile Networks Using Optimization Techniques

T0 review · 3 major / 4 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 A constrained least-squares fit of the Friis path-loss model to nearby drive-test measurements, grouped by serving cell, predicts RSRP at untested locations well enough to reduce drive-test cost.

desk verdict The evaluation never tests the advertised use case: predictions are only at measured points, and the target is likely in the fitting set, so the central extrapolation claim is unsupported. read the letter →

arxiv 2502.09305 v1 pith:MEK5RKRH submitted 2025-02-13 cs.NI cs.AIcs.SE

MohammadJava Taheri , Abolfazl Diyanat , MortezaAli Ahmadi , Ali Nazari This is my paper · ORCID

classification cs.NIcs.AIcs.SE

keywords DriveTestMinimalMobileNetworksReceivedSignalStrengthRSRPDataPredictionNetworkOptimizationPathLoss

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

Mobile operators rely on drive tests to measure signal strength, but drive tests are costly and miss hard-to-reach areas. This paper claims that the Reference Signal Received Power (RSRP) at a location that was not drive-tested can be estimated from measurements at nearby drive-test points. The method draws a circle of radius $R$ around the target, groups the measured points inside it by their serving cell, and fits the two-parameter Friis path-loss model to each group with a constrained least-squares or maximum-likelihood optimization. It also estimates the standard deviation of shadowing noise from pairwise differences of consecutive measurements, without needing cell locations. If these claims hold, operators could fill coverage gaps in their data and reduce the number of drive-test campaigns needed.

What carries the argument

The engine of the method is a local, per-cell least-squares fit of the Friis path-loss model $P_r = P_0 - 10\beta\log_{10}(d/d_0)$ (with $d_0 = 1$) to the drive-test points inside a radius-$R$ circle around the target, subject to box constraints on $P_0$ and $\beta$ written as $Ax \le b$. This is derived as a maximum-likelihood problem under Gaussian shadowing noise and reduced to the MSE objective $\min_{P_0,\beta} \sum_i (P_i - P_0 + \beta\log_{10} d_i)^2$ when noise variances are unknown. A second piece is the shadowing-noise estimator: for consecutive measurements close enough that $d_{i+1} \approx d_i$, the pairwise power differences $P^d_i = n_{i+1} - n_i$ are Gaussian with standard deviation $\sqrt{2}\sigma$, so $\sigma$ is recovered as $\sqrt{\frac{1}{2N_k}\sum_i (P^d_i)^2}$ without knowing cell locations. The machinery converts RSRP prediction into two small estimation problems whose data needs are purely local.

What would settle it

Take a drive-test dataset, hold out an entire spatial region (for example, a street segment) with no measurements at all, fit the local Friis model only outside it, and compare predicted RSRP inside the region against a fresh measurement campaign; the central claim fails if the mean absolute error at truly unmeasured points substantially exceeds the few-dB errors reported for measured points.

Watch

Extended reading notes

Core claim

The paper's central claim is that RSRP prediction at unmeasured points reduces to a local parameter-estimation problem. Within a circle of radius $R$ around each target, measured points are grouped by serving cell, and the Friis model $P_r = P_0 - 10\beta\log_{10}(d)$ (dBm, with reference distance $d_0=1$) is fitted per cell by minimizing the squared error between measured and modeled power under box constraints on $P_0$ and $\beta$. A second contribution estimates the shadowing-noise standard deviation $\sigma$ without knowing base-station locations by taking pairwise differences of consecutive measurements along the drive-test route; because those differences are argued to be Gaussian with standard deviation $\sqrt{2}\sigma$, $\sigma$ follows from a root-mean-square formula. Using more than 600 km of drive-test data and over 20,000 points in Tehran, the paper reports average prediction errors of roughly 3.5 to 5.5 dB across parameter settings, with the smallest errors near a 50 m radius.

Load-bearing premise

The method assumes that any unmeasured location behaves exactly like the nearby measured points under one simple signal-strength-versus-distance formula, so that a fit to measured points remains valid where nothing was measured.

Editorial extensions

If this is right

Operators can estimate RSRP at locations inside a drive-tested area that were not physically measured, using only nearby measured points and known serving-cell identities.
The shadowing-noise standard deviation can be extracted from drive-test logs alone, which is useful for coverage planning and for setting confidence intervals on predictions.
Smaller selection radii, around 50 m, give the most homogeneous channel conditions and the smallest prediction error, while the minimum number of points per cell and the minimum distance from the antenna have little effect in the tested range.
Prediction error grows with shadowing noise, so the method is most accurate in low-shadowing areas and the noise estimate flags where predictions are trustworthy.
The same local-fitting idea extends in principle to other network generations and other signal-quality metrics such as RSRQ, since the underlying model is generic path loss.

Reading between the lines

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

A held-out spatial evaluation, where an entire region is omitted from fitting and predicted values are compared with a fresh measurement campaign, would directly test whether the fit extrapolates to truly unmeasured locations.
The per-route shadowing estimate could be upgraded into a per-location shadowing map by interpolating residuals between measured and fitted values, potentially sharpening predictions at unmeasured points.
Because the fit is local and inexpensive, the radius could be chosen adaptively, growing until each serving cell has enough points, trading homogeneity against sample size.

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: the paper has a real data collection effort and a clean formulation, but the evaluation does not support the abstract's claim. In eq. (1), the set Φ includes every measured point within radius R, and the target point is at distance 0, so it satisfies the condition and enters the fitting set. The paper never states that the target is withheld. Section 4.2 says the method estimates RSRP at each remaining point using surrounding points, and the error is computed against the actual RSRP at that same point. So the reported errors are plausibly in-sample residuals. Even under the charitable reading that the target is excluded, every evaluation point is a measured drive-test point; the method is never asked to predict at an unmeasured location, which is the entire motivation of the paper. The low median errors in Figures 7–9 could simply reflect interpolation in a dense point cloud, not extrapolation to unmeasured areas.

What is actually good: the data effort is substantial — over 600 km, more than 20,000 points, collected with a purpose-built Android app. The per-cell grouping followed by a bounded least-squares fit of the Friis model is a reasonable engineering baseline, and the first-difference shadowing estimator is a neat trick, even if not new. But there are no baselines: the ML methods cited in related work are never compared, so we do not know what this method buys over existing practice.

The secondary issue is real too. The proof of Theorem 2 is mathematically flawed. The discriminant in eq. (22) is −4 sin² α, so real solutions require sin α = 0, i.e., α = 0 or π, not just α = 0. More seriously, the proof sets the left-hand side to zero before solving, discarding the small but nonzero ε² term. The intuitive conclusion — consecutive distances are nearly equal when the arc is small — is approximately true under the right conditions, but the proof does not establish it. This is secondary because it affects only the shadowing estimate, not the main RSRP prediction.

Net: the central claim is untested. The paper needs an out-of-sample evaluation on truly unmeasured locations, with a held-out set of entire regions or time slices, and at least one baseline comparison. As-is, I would tell the authors to redo the evaluation before considering publication. That said, the flaw is fixable and the data is real, so a serious referee could guide the revision. I would send it to review, but with clear instructions that the circular evaluation must be addressed.

Referee Report

3 major / 4 minor

Summary. This paper proposes an optimization-based method for predicting RSRP values in 4G networks using drive test data. For a target point, the method selects all measured points within a radius R (Eq. 1), groups them by serving cell, fits a Friis path-loss model to each group by constrained least squares (Eq. 16), and uses the fitted model to estimate RSRP. A secondary contribution estimates the shadowing noise standard deviation from differences between consecutive measurements (Theorem 2, Eqs. 23-26). The method is evaluated on real drive test data from Tehran, with reported errors typically on the order of a few dB across various parameter settings.

Significance. If the claims were validated, the approach could help operators reduce drive test costs by predicting RSRP in areas that are not drive-tested. The use of a simple, interpretable model with few parameters is attractive, and the paper includes real-world measurements. However, the evaluation does not actually test the extrapolation setting advertised in the abstract, so the significance of the reported accuracy numbers is currently not established.

major comments (3)

[§4.2, Eq. (1)] The evaluation does not test the paper's central claim of predicting RSRP at locations without drive test measurements. Equation (1) defines Φ as all points within radius R that have an RSRP measurement, and a target point with a measurement satisfies the definition (distance 0 < R) unless explicitly excluded; the paper nowhere states such an exclusion. Moreover, §4.2 says the method "estimates the received power at each of the remaining points" using "surrounding points," with no hold-out procedure. The reported errors in Figures 7-9 are therefore at best interpolation residuals at measured points and, if the target is included, in-sample residuals. No experiment predicts at a genuinely unmeasured location, so the abstract's promise—estimating signal strength at locations without drive-test data—remains unsupported.
[§3.2.2–§3.2.3] The method as described cannot be applied to an unmeasured target point because it requires knowing the serving cell at the target. In §3.2.2, the points in Φ are grouped according to the serving cell at each measured point, and in §3.2.3 a separate Friis model is fitted per cell. If the target has no measurement, its serving cell is unknown, and the paper does not specify how to select among the cell-specific models (e.g., by nearest cell, strongest predicted signal, or a combination). This missing step is essential for the claimed use case, since drive-test-free locations are exactly those where the serving cell may be unknown.
[§3.2.4, Theorem 2, Eq. (22)] The proof of Theorem 2 contains a mathematical error. From Eq. (22), k = cos α ± √(cos²α − 1); the discriminant is −4 sin²α, so real solutions require sin α = 0 (α = 0 or π), not "α must be zero" as stated. At α = π the solution k = −1 is unphysical for positive distances, so the intended conclusion di+1 = di only follows for α = 0, a degenerate case. Since this theorem underlies the shadowing-noise estimator in Eqs. (23)–(26), the proof needs to be corrected or replaced with an approximate argument that justifies di+1 ≈ di when the inter-measurement displacement is small.

minor comments (4)

[§1, §3.2.3] There are typographical errors such as "Recieved" in the introduction, "isn?t" in §1.1, and unusual formatting in "Cram´ er–Rao" in §3.2.3; these should be cleaned up.
[Eqs. (3) and (6)] Equation (6) omits the factor 10 that appears in Eq. (3), so the symbol β has different meanings in the two equations; please make the notation consistent.
[Eq. (14) and §3.2.3] The constraint matrices A and b are defined twice with different orderings and signs; the final form of the constraints should be stated precisely.
[Lemmas 1] Lemma 1 is used twice with different statements (MLE optimality and Gaussianity of P^d); the second statement should be renumbered as Lemma 2.

Circularity Check

1 steps flagged · score 7.0 of 10

Evaluation protocol turns the advertised extrapolation into an in-sample fit: target points are measured points inside Φ (Eq. 1) and are not withheld, so prediction at unmeasured locations is never tested.

self definitional [Eq. (1) in §3.2.1; evaluation protocol in §4.2; Fig. 6 caption]
"Φ = {ρi|dit < R,Has RSRP measurement}, (1) ... After filtering out the points without measurements, the next step is to estimate the received power at each of the remaining points. ... the estimation error is computed as the difference between the actual RSRP values obtained from the drive test and the estimated values."

For any target ρt that has a drive-test measurement, dtt = 0 < R, so ρt satisfies Eq. (1) and belongs to Φ unless explicitly withheld; the text never states such withholding. §4.2 defines the evaluation targets as the remaining measured points and computes errors against the drive-test RSRP at those same points, while the least-squares fit (16) is trained on Φ. The reported 'prediction' is therefore at least partly an in-sample residual of a model fitted on the target itself, and even on a charitable reading where the target is excluded, no unmeasured location is ever evaluated. This contradicts the abstract's 'data from other drive test points' and §1.1's goal of estimating places where drive testing is infeasible.

full rationale

The derivation chain itself—Friis model (3), MLE/MSE fit (15)-(16), applying fitted P0 and beta to compute RSRP at a target—contains no imported uniqueness theorem and no load-bearing self-citation. Reference [2] is only the data-collection tool, not an argument. The circularity is in the evaluation: Eq. (1) defines the fitting set as all measured points within radius R of the target, and a measured target trivially satisfies the definition; §4.2 estimates power at 'each of the remaining points' and never says the target is withheld. Consequently the reported errors in Figs. 7-9 are, in the worst case, in-sample residuals, and in the best case interpolation at measured points. The paper's central use case—predicting RSRP where no measurement exists—is never exercised, so the abstract's claim is not established by the reported experiments. Separately, the proof of Theorem 2 is mathematically incorrect as written: 'For (22) to have a real solution, α must be zero' ignores the real solution at α=π; the discriminant is −4 sin² α, so real solutions occur for sin α=0. That is a correctness defect in the shadowing-noise side result, not a circularity, and does not affect the above score beyond noting it. Overall, because the central reported prediction reduces by construction to fitting/evaluating the same measured points, the circularity score is 7 rather than 0-2.

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

The central claim rests on five fitted or tuned quantities (P0, beta, R, minimum points, minimum distance) plus an estimated sigma. P0 and beta are legitimate model parameters, but the radius and thresholds are chosen empirically on the same data used for evaluation. The main axioms are standard domain assumptions: 4G RSRP, known base station locations, Gaussian shadowing, and local stationarity of the Friis model. The weakest axiom is the di+1 approximately di approximation, whose proof does not establish the claim. No invented physical entities are introduced.

free parameters (6)

P0 (reference received power per cell) = not reported for individual cells
Equation (16) fits P0 by constrained least squares for each cell group; the predicted RSRP depends directly on this fitted value.
beta (path loss exponent per cell) = not reported for individual cells
Fitted jointly with P0 in eq. (16); central to converting distance to received power.
R (selection circle radius) = swept 50-800 m, smaller radii favored
Hyperparameter controlling which points enter Phi; the paper shows it changes error and selects favorable values after observing results.
Minimum points connected to a cell = 8, 10, 12, 14
Threshold for including a cell group; varied in evaluation and affects results.
Minimum distance from serving cell antenna = 10, 15, 20, 25 m
Filter applied before fitting; treated as a model parameter in the experiments.
sigma (shadowing standard deviation) = not reported as a table
Estimated via eq. (26) and used in the error-versus-shadowing analysis; not validated against known ground truth.

assumptions (6)

domain assumption Network is 4G and only RSRP is considered.
Section 3.1 assumptions 1 and 2 scope the method to 4G RSRP; extension to other generations is asserted without proof.
domain assumption Base station locations and configurations are known.
Assumption 3 in Section 3.1; distances in eq. (4) require transmitter coordinates.
domain assumption Shadowing noise is zero-mean Gaussian, independent, and the only error source.
Eq. (6) and surrounding text; thermal noise and location errors are explicitly ignored.
domain assumption A single Friis log-distance model with fixed P0 and beta holds for all points in the selection circle.
Eq. (3) and Section 3.2.3; local stationarity is assumed but never tested at unmeasured locations.
ad hoc to paper Consecutive drive-test points are close enough that di+1 is approximately di.
Section 3.2.4 Theorem 2; the proof is flawed because the quadratic has real solutions only for alpha=0, yet the approximation is applied generally without an error bound.
domain assumption Serving cell identifiers are available for every measurement.
Section 4.1 states identifiers are obtained from available databases; grouping in eq. (2) depends on this.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Predicting Drive Test Results in Mobile Networks Using Optimization Techniques." pith.science (2026). https://pith.science/paper/MEK5RKRH

@misc{pith2026250209305,
  author       = {Pith},
  title        = {Pith review of: Predicting Drive Test Results in Mobile Networks Using Optimization Techniques},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MEK5RKRH}},
  note         = {Machine review of arXiv:2502.09305}
}

read the original abstract

Mobile network operators constantly optimize their networks to ensure superior service quality and coverage. This optimization is crucial for maintaining an optimal user experience and requires extensive data collection and analysis. One of the primary methods for gathering this data is through drive tests, where technical teams use specialized equipment to collect signal information across various regions. However, drive tests are both costly and time-consuming, and they face challenges such as traffic conditions, environmental factors, and limited access to certain areas. These constraints make it difficult to replicate drive tests under similar conditions. In this study, we propose a method that enables operators to predict received signal strength at specific locations using data from other drive test points. By reducing the need for widespread drive tests, this approach allows operators to save time and resources while still obtaining the necessary data to optimize their networks and mitigate the challenges associated with traditional drive tests.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

[1]

& Sathya, V

Sahin, O. & Sathya, V. Network quality prediction in a designated area using gps data. Journal of Network and Computer Applications 231, 104002 (2024). URL https://www.sciencedirect.com/science/ article/pii/S1084804524001796

work page 2024
[2]

Venus pioneer

Parto Ertebat Saba. Venus pioneer. https://www.pesaba.com/en/products/ detail/118/Venus-Pioneer. [Accessed 23-01-2025]

work page 2025
[3]

Skocaj, M. et al. Cellular network capac- ity and coverage enhancement with mdt data and deep reinforcement learning. Computer Communications 195, 403–415 (2022). 13 100 200 300 400 500 600 700 800 Radius (m) 8 10 12 14 16 18 20 Min Points 3.5 4.0 4.5 5.0 5.5 Average Error 3D Chart of Average Error 3.5 4.0 4.5 5.0 5.5 Average Error Fig. 9: Impact of Minim...

work page 2022
[4]

& Diamanti, R

Micheli, D. & Diamanti, R. Statistical analy- sis of interference in a real lte access network by massive collection of mdt radio mea- surement data from smartphones 1906–1916 (2019)

work page 2019
[5]

& Mota, E

Christofides, N., Campos, V., Corber´ an, A. & Mota, E. An algorithm for the Rural Post- man problem on a directed graph , 155–166 (Springer Berlin Heidelberg, 1986)

work page 1986
[6]

& Meza, O

Ar´ aoz, J., Fern´ andez, E. & Meza, O. Solving the prize-collecting rural postman problem. European Journal of Operational Research 196, 886–896 (2009)

work page 2009
[7]

M., Budiyanto, S., Silaban, F

Silalahi, L. M., Budiyanto, S., Silaban, F. A., Simanjuntak, I. U. V. & Rochendi, A. D. Improvement of quality and signal coverage lte in bali province using drive test method 376–380 (2021)

work page 2021
[8]

& Uthansakul, P

Peerajing, J. & Uthansakul, P. Multi- sector optimization of antenna tilt angle based empirical knowledge from drive test. 13th International Conference on Elec- trical Engineering/Electronics, Computer, Telecommunications and Information Tech- nology (ECTI-CON) (2016)

work page 2016

Show all 16 references

[9]

& Alienyi, D

Ojo, S., Imoize, A. & Alienyi, D. Radial basis function neural network path loss prediction model for lte networks in multitransmitter signal propagation environments. Interna- tional Journal of Communication Systems (2020)

2020
[10]

& Suzuki, H

Ostlin, E., Zepernick, H.-J. & Suzuki, H. Macrocell path-loss prediction using artifi- cial neural networks. IEEE Transactions on Vehicular Technology 59, 2735–2747 (2010)

2010
[11]

& Chris- tiansen, H

Thrane, J., Artuso, M., Zibar, D. & Chris- tiansen, H. L. Drive test minimization using deep learning with bayesian approximation 1–5 (2018)

2018
[12]

Chomsuay, D. et al. Improvement of deep learning-based reference signal received power prediction for lte communication sys- tem 1–4 (2023)

2023
[13]

Wu, S. et al. A machine learning based intel- ligent propagation model for rsrp prediction 1–5 (2022)

2022
[14]

& Maps, G

This, D. & Maps, G. Calculate dis- tance , bearing and more between. [Online], Available: Http://www.movable- type.co.uk/scripts/latlong.html 2, 1–8 (2009)

2009
[15]

& Buehrer, R

Zekavat, R. & Buehrer, R. M. Handbook of Position Location: Theory, Practice, and Advances IEEE Series on Digital & Mobile Communication (Wiley, 2019)

2019
[16]

A First Course in Probability (Pear- son Education, 2015)

Ross, S. A First Course in Probability (Pear- son Education, 2015). URL https://books. google.com/books?id=AstQCwAAQBAJ. 14 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Error 0 2 4 6 8 10Shadowing Noise Radius: 50m 5% Boundary Line (a) 0 5 10 15 20 Error 0 2 4 6 8 10 12Shadowing N...

2015

Pith tools

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

[1] [1]

& Sathya, V

Sahin, O. & Sathya, V. Network quality prediction in a designated area using gps data. Journal of Network and Computer Applications 231, 104002 (2024). URL https://www.sciencedirect.com/science/ article/pii/S1084804524001796

work page 2024

[2] [2]

Venus pioneer

Parto Ertebat Saba. Venus pioneer. https://www.pesaba.com/en/products/ detail/118/Venus-Pioneer. [Accessed 23-01-2025]

work page 2025

[3] [3]

Skocaj, M. et al. Cellular network capac- ity and coverage enhancement with mdt data and deep reinforcement learning. Computer Communications 195, 403–415 (2022). 13 100 200 300 400 500 600 700 800 Radius (m) 8 10 12 14 16 18 20 Min Points 3.5 4.0 4.5 5.0 5.5 Average Error 3D Chart of Average Error 3.5 4.0 4.5 5.0 5.5 Average Error Fig. 9: Impact of Minim...

work page 2022

[4] [4]

& Diamanti, R

Micheli, D. & Diamanti, R. Statistical analy- sis of interference in a real lte access network by massive collection of mdt radio mea- surement data from smartphones 1906–1916 (2019)

work page 2019

[5] [5]

& Mota, E

Christofides, N., Campos, V., Corber´ an, A. & Mota, E. An algorithm for the Rural Post- man problem on a directed graph , 155–166 (Springer Berlin Heidelberg, 1986)

work page 1986

[6] [6]

& Meza, O

Ar´ aoz, J., Fern´ andez, E. & Meza, O. Solving the prize-collecting rural postman problem. European Journal of Operational Research 196, 886–896 (2009)

work page 2009

[7] [7]

M., Budiyanto, S., Silaban, F

Silalahi, L. M., Budiyanto, S., Silaban, F. A., Simanjuntak, I. U. V. & Rochendi, A. D. Improvement of quality and signal coverage lte in bali province using drive test method 376–380 (2021)

work page 2021

[8] [8]

& Uthansakul, P

Peerajing, J. & Uthansakul, P. Multi- sector optimization of antenna tilt angle based empirical knowledge from drive test. 13th International Conference on Elec- trical Engineering/Electronics, Computer, Telecommunications and Information Tech- nology (ECTI-CON) (2016)

work page 2016