pith. sign in

arxiv: 2604.16356 · v1 · submitted 2026-03-18 · 💻 cs.NI

ML and Smartphones Assisted Real-Time Uplink Performance Prediction in 5G Cellular System

Md Mahfuzur Rahman , Jareen Shuva , Nishith Tripathi , Lingjia Liu , Jeffrey Reed This is my paper

Pith reviewed 2026-05-15 09:15 UTC · model grok-4.3

classification 💻 cs.NI
keywords 5G uplinkperformance predictionmachine learningsmartphone measurementsthroughput forecastingBLER predictionCOTS devicesreal-time estimation
0
0 comments X

The pith

Machine learning models trained on commercial smartphone measurements can forecast 5G uplink throughput and block error rates in real time.

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

The paper demonstrates that data collected from ordinary Google Pixel 7a phones in a controlled 5G testbed can train standard regression algorithms to predict uplink throughput and BLER across indoor, outdoor, stationary, and mobile conditions. A sympathetic reader would care because this approach relies only on widely available consumer devices and off-the-shelf ML techniques rather than specialized test equipment, opening a route to practical performance monitoring in live networks. The work covers multiple traffic patterns including video streaming and bulk transfers under both line-of-sight and obstructed propagation to build the training sets. Five algorithms are compared on the same datasets to identify which deliver usable accuracy for these two target metrics.

Core claim

By instrumenting two Pixel 7a devices to log CQI, MCS, TTI, throughput, and BLER while running srsRAN at 3.4 GHz, the authors generate labeled datasets under varied mobility, interference, and traffic conditions; supervised models trained on these measurements achieve reliable prediction of uplink throughput and BLER, establishing that COTS smartphones plus common ML methods suffice for real-time 5G uplink performance estimation.

What carries the argument

Supervised regression models (linear regression, decision tree, random forest, XGBoost, LightGBM) trained directly on smartphone-reported radio metrics such as CQI and MCS to output predicted throughput and BLER.

If this is right

  • Uplink performance forecasting becomes possible with only consumer handsets instead of dedicated measurement hardware.
  • Network operators could use aggregated phone reports to estimate cell-level throughput and reliability without additional probes.
  • Real-time prediction enables applications such as dynamic scheduling adjustments based on forecasted BLER before errors occur.
  • The same data-collection pipeline works for both video and bulk-transfer traffic in LOS and nLOS environments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Smartphone-based prediction could be embedded in future UE firmware to let devices self-optimize transmit power or handover decisions using local forecasts.
  • Extending the approach to downlink metrics or multi-cell interference scenarios would require only additional label collection rather than new hardware.
  • If models remain accurate across software-defined and commercial stacks, crowdsourced phone measurements could supplement drive-test data for coverage mapping.

Load-bearing premise

The srsRAN testbed with its chosen traffic patterns and mobility scenarios produces radio conditions that match those of commercial 5G networks and that the learned models will generalize to unseen deployments.

What would settle it

Deploy the same trained models on live commercial 5G base stations and measure whether the predicted throughput and BLER deviate by more than a few percent from simultaneous ground-truth measurements taken under comparable user density and mobility.

Figures

Figures reproduced from arXiv: 2604.16356 by Jareen Shuva, Jeffrey Reed, Lingjia Liu, Md Mahfuzur Rahman, Nishith Tripathi.

Figure 1
Figure 1. Figure 1: Overview of the standalone 5G experimental testbed [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ookla uplink testing across local and long-distance [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: YouTube-based uplink workload generation. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: iPerf uplink throughput measurements. 5.1.3 YouTube Testing. To examine performance under real-world application-driven workloads, we generated traffic using YouTube in both indoor and outdoor configurations. Our experiments fo￾cused on uplink-intensive use cases, including 4K and 8K streaming uploads. Unlike the constant-rate behavior of iPerf, YouTube pro￾duces highly bursty traffic governed by adaptive … view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of regression models across MSE, [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Scatter plots comparing LightGBM predictions with [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Histogram of prediction errors for LightGBM mod [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
read the original abstract

We propose a machine learning (ML) and smartphone-assisted framework for uplink performance prediction in a private, realistic 5G cellular system using real-time measurements in both indoor and outdoor settings. This work presents a comprehensive data-driven evaluation of 5G performance prediction using a controllable software-defined radio test environment. The experimental platform is built on srsRAN 5G NR stack running on a Dell workstation configured as a gNB and 5G core operating at 3.4 GHz. Two commercial Google Pixel 7a devices are instrumented to capture uplink metrics, including channel quality indicator (CQI), modulation and coding scheme (MCS), throughput, transmission time interval (TTI), and block error rate (BLER). Different types of traffic are generated using industry-standard tools such as Ookla and iperf, spanning stationary, pedestrian, and mobility cases under both line-of-sight (LOS) and non-line-of-sight (nLOS) propagation environments. Additional datasets include YouTube video sessions and global server endpoints to introduce variability in path characteristics. The resulting measurements, including multi-UE interference conditions, serve as training data for several supervised regression models. Five learning algorithms-linear regression, decision tree, random forest, XGBoost, and LightGBM-are benchmarked for prediction accuracy. The study shows that reliable forecasting of throughput and BLER is feasible using only COTS smartphones and widely available ML methods, offering a practical pathway for real-world 5G network performance estimation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a machine learning framework that uses real-time uplink measurements (CQI, MCS, TTI, throughput, BLER) collected from two commercial Pixel 7a smartphones in a private srsRAN 5G NR testbed to train and benchmark five supervised regressors (LR, DT, RF, XGBoost, LightGBM). Experiments span indoor/outdoor, LOS/nLOS, stationary/pedestrian/mobility scenarios with varied traffic (Ookla, iperf, YouTube) and claim that reliable forecasting of throughput and BLER is feasible with only COTS devices, offering a practical pathway for real-world 5G performance estimation.

Significance. If the empirical results hold under broader conditions, the work demonstrates a low-cost, smartphone-only method for uplink prediction that could be deployed by operators or researchers without specialized test equipment. The use of real measurements from commercial UEs and systematic benchmarking of multiple models across mobility and traffic patterns provides concrete evidence of feasibility within the controlled setting.

major comments (2)
  1. [Abstract and §5] Abstract and §5 (Results): The central claim that the approach offers a 'practical pathway for real-world 5G network performance estimation' rests on generalization from a single srsRAN private testbed (Dell workstation at 3.4 GHz, two Pixel 7a UEs, controlled traffic). No cross-deployment validation on commercial networks is reported, so the learned mappings may capture srsRAN-specific artifacts (MCS-to-BLER curves, TTI timing, scheduling) rather than universal 5G behavior.
  2. [§4] §4 (Experimental Setup): The evaluation uses a single test environment with fixed multi-UE interference and specific propagation conditions; absence of any hold-out across different gNB implementations, channel models, or operator deployments makes it impossible to quantify how prediction error would increase outside the collected distribution.
minor comments (2)
  1. [§4.3] §4.3: Specify the exact train/test split ratios, whether k-fold cross-validation was performed, and whether error bars or standard deviations accompany the reported accuracy metrics for each model.
  2. [Figures 4-6] Figures 4-6: Add confidence intervals or per-run variability to the throughput and BLER prediction plots so readers can assess consistency across the mobility and traffic cases.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the scope limitations of our evaluation. We have revised the manuscript to qualify our claims accordingly while preserving the core contribution of demonstrating ML-based uplink prediction using only COTS smartphones in a realistic controlled 5G testbed.

read point-by-point responses

  1. Referee: [Abstract and §5] Abstract and §5 (Results): The central claim that the approach offers a 'practical pathway for real-world 5G network performance estimation' rests on generalization from a single srsRAN private testbed (Dell workstation at 3.4 GHz, two Pixel 7a UEs, controlled traffic). No cross-deployment validation on commercial networks is reported, so the learned mappings may capture srsRAN-specific artifacts (MCS-to-BLER curves, TTI timing, scheduling) rather than universal 5G behavior.

    Authors: We agree that the results are obtained from a single private srsRAN testbed and that no cross-deployment validation on commercial networks is included. Although the testbed employs standard 5G NR protocols and commercial Pixel 7a UEs, srsRAN-specific scheduling and MCS-to-BLER mappings may be present. In the revised manuscript we have removed the phrasing 'practical pathway for real-world 5G network performance estimation' from the abstract and Section 5, replacing it with a more precise statement that the work shows reliable prediction is feasible within a controlled, standards-compliant environment using only COTS devices. We now explicitly note cross-network validation as future work. revision: yes

  2. Referee: [§4] §4 (Experimental Setup): The evaluation uses a single test environment with fixed multi-UE interference and specific propagation conditions; absence of any hold-out across different gNB implementations, channel models, or operator deployments makes it impossible to quantify how prediction error would increase outside the collected distribution.

    Authors: We acknowledge that the evaluation is performed in one test environment. In the revised manuscript we have added a new limitations paragraph in Section 4 that explicitly states the single-environment constraint and notes that prediction error may increase under different gNB implementations or channel models. Within the studied domain we retain the systematic coverage of indoor/outdoor, LOS/nLOS, stationary/pedestrian/mobility, and varied traffic patterns to demonstrate robustness under those conditions. revision: yes

Circularity Check

0 steps flagged

No circularity: standard supervised ML on independent test data

full rationale

The paper collects real-time uplink measurements (CQI, MCS, throughput, TTI, BLER) from Pixel 7a devices on an srsRAN testbed under controlled traffic and mobility, then trains off-the-shelf regressors (LR, DT, RF, XGBoost, LightGBM) to forecast throughput and BLER. No equations, self-definitions, or derivations are shown that would make the reported predictions equivalent to the inputs by construction. Training and evaluation use standard train/test splits on the collected dataset; the models are not fitted to a subset and then renamed as predictions of the same quantity. No self-citations are invoked to justify uniqueness or to smuggle ansatzes. The feasibility claim rests on empirical accuracy numbers obtained from held-out measurements, which is self-contained and externally falsifiable.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The work relies on standard supervised learning assumptions and the representativeness of the testbed data for real-world generalization.

pith-pipeline@v0.9.0 · 5587 in / 905 out tokens · 48739 ms · 2026-05-15T09:15:29.852167+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

  1. [1]

    E.A. Abdiel. 2022. Forecasting LTE Network Throughput for Optimizing Opera- tional and Business Aspects. https://medium.com/@earlyanabdiel/forecasting- lte-network-throughput-for-optimizing-operational-and-business-aspect- a9599a565d6a

    work page 2022

  2. [2]

    Abbas Al-Thaedan, Zaenab Shakir, Ahmed Yaseen Mjhool, Ruaa Alsabah, Ali Al-Sabbagh, Monera Salah, and Josko Zec. 2023. Downlink throughput prediction using machine learning models on 4G-LTE networks.International Journal of Information Technology15, 6 (2023), 2987–2993

    work page 2023

  3. [3]

    Alves, Joao Guilherme A

    Ruan P. Alves, Joao Guilherme A. da Silva Alves, Mikael R. Camelo, Wilker O. de Feitosa, Victor F. Monteiro, and Fco. Rodrigo P. Cavalcanti. 2024. Experimental Comparison of 5G SDR Platforms: srsRAN x OpenAirInterface.arXiv preprint arXiv:2406.01485(2024). https://arxiv.org/abs/2406.01485

    work page arXiv 2024

  4. [4]

    Jonny Kong, Y

    Omar Basit, Phuc Dinh, Imran Khan, Z. Jonny Kong, Y. Charlie Hu, Dimitrios Koutsonikolas, Myungjin Lee, and Chaoyue Liu. 2024. On the Predictability of Fine-Grained Cellular Network Throughput Using Machine Learning Models. In2024 IEEE 21st International Conference on Mobile Ad-Hoc and Smart Systems (MASS). 47–56. https://doi.org/10.1109/MASS62177.2024.00018

    work page doi:10.1109/mass62177.2024.00018 2024

  5. [5]

    Hao-Hsuan Chang, Nima Mohammadi, Ramin Safavinejad, Yang Yi, and Lingjia Liu. 2024. Dyna-ESN: Efficient Deep Reinforcement Learning for Partially Observ- able Dynamic Spectrum Access.IEEE Transactions on Wireless Communications (2024)

    work page 2024

  6. [6]

    S Chang and A Baliga. 2021. Development of machine learning-based radio propagation models and benchmarking for mobile networks.J. Stud. Res10 (2021), 1–12

    work page 2021

  7. [7]

    Clemson University. 2022. Machine Learning and Deep Learning for Through- put Prediction. https://people.computing.clemson.edu/~jmarty/projects/ lowLatencyNetworking/papers/AI-ML/MLAppliedToNetworks/Machine_ Learning_and_Deep_Learning_for_Throughput_Prediction.pdf

    work page 2022

  8. [8]

    T.M. Dias, A. Oliveira, L.G. Gonçalves, and J.F. Martins-Filho. 2022. RSRP Pre- diction on LTE Network Testbed Using a Software Defined Radio Platform. In XXXIV Simpósio Brasileiro de Telecomunicações e Processamento de Sinais (SBrT). https://biblioteca.sbrt.org.br/articlefile/3582.pdf

    work page 2022

  9. [9]

    Habiba Elsherbiny and M. Abbas. 2020. 4G LTE Network Throughput Modelling and Prediction.Queen’s Telecommunications Research Lab(2020)

    work page 2020

  10. [10]

    Habiba Elsherbiny, Ahmad M Nagib, Hatem Abou-zeid, Hazem M Abbas, Hos- sam S Hassanein, Aboelmagd Noureldin, Akram Bin Sediq, and Gary Boudreau

    work page

  11. [11]

    InGLOBECOM 2020-2020 IEEE Global Communications Conference

    4G LTE network data collection and analysis along public transportation routes. InGLOBECOM 2020-2020 IEEE Global Communications Conference. IEEE, 1–6

    work page 2020

  12. [12]

    Engin Eyceyurt and J. Egi. 2022. Machine-Learning-Based Uplink Throughput Prediction from Physical Layer Measurements.Electronics11, 8 (2022), 1227. https://doi.org/10.3390/electronics11081227

    work page doi:10.3390/electronics11081227 2022

  13. [13]

    Eric Forbes. 2023. A Closer Look at LTE Throughput Testing with srsRAN. https: //ewf-engineering.com/a-closer-look-at-lte-throughput-testing-with-srsran/

    work page 2023

  14. [14]

    Koenig et al

    N. Koenig et al. 2021. Throughput Prediction Using Machine Learning in LTE and 5G Networks. https://people.computing.clemson.edu/~jmarty/projects/ lowLatencyNetworking/papers/AI-ML/MLAppliedToNetworks/Throughput_ Prediction_using_Machine_Learning_in_LTE_and_5G_Networks.pdf

    work page 2021

  15. [15]

    Dimitar Minovski, Niclas Ögren, Karan Mitra, and Christer Åhlund. 2023. Throughput Prediction Using Machine Learning in LTE and 5G Networks.IEEE Transactions on Mobile Computing22, 3 (2023), 1825–1840. https://doi.org/10. 1109/TMC.2021.3099397

    work page arXiv 2023

  16. [16]

    Iyengar, Akshay C

    Neha Mishra, Yamini V. Iyengar, Akshay C. Raikar, Nikitha Thomas, Sabarish K. Moorthy, Jiangqi Hu, Zhiyuan Zhao, Nicholas Mastronarde, Elizabeth S. Bentley, Michael J. Medley, and Zhangyu Guan. 2023. Demo Abstract: Scaling Out srsRAN Through Interfacing Wirelessly srsENB with srsEPC. InIEEE INFOCOM 2023 - IEEE Conference on Computer Communications Worksho...

    work page doi:10.1109/infocomwkshps58353.2023.10154848 2023

  17. [17]

    OpenAirInterface Software Alliance. 2024. OpenAirInterface 5G Wireless Imple- mentation. https://www.openairinterface.org/. Accessed: 2025-05-07

    work page 2024

  18. [18]

    Perveen, M.H

    H. Perveen, M.H. Zafar, S.G. Abbas, S. Rehman, I. Ahmad, and M.M. Rathore. 2022. Dynamic Traffic Forecasting and Fuzzy-Based Optimized Admission Control in Federated 5G-Open RAN Networks.Neural Computing and Applications34 (2022), 2925–2940. https://doi.org/10.1007/s00521-021-06206-0

    work page doi:10.1007/s00521-021-06206-0 2022

  19. [19]

    Michele Polese, Leonardo Bonati, Salvatore D’Oro, Pedram Johari, Davide Villa, Sakthivel Velumani, Rajeev Gangula, Maria Tsampazi, Clifton Paul Robinson, Gabriele Gemmi, Andrea Lacava, Stefano Maxenti, Hai Cheng, and Tommaso Melodia. 2024. Colosseum: The Open RAN Digital Twin.IEEE Open Journal of the Communications Society5 (2024), 5452–5466. https://doi....

    work page doi:10.1109/ojcoms 2024

  20. [20]

    Zahran, Cormac J

    Darijo Raca, Ahmed H. Zahran, Cormac J. Sreenan, Rakesh K. Sinha, Emir Hale- povic, Rittwik Jana, and Vijay Gopalakrishnan. 2020. On Leveraging Machine ML and Smartphones Assisted Real-Time Uplink Performance Prediction in 5G Cellular System and Deep Learning for Throughput Prediction in Cellular Networks: Design, Performance, and Challenges.IEEE Communic...

    work page doi:10.1109/mcom.001.1900394 2020

  21. [21]

    Rehmani et al

    A. Rehmani et al. 2023. Machine Learning for Wireless Network Throughput Prediction.ScholarWorks @ UTRGV(2023)

    work page 2023

  22. [22]

    Shanshan Shen and Hai Li. 2023. A Platform Based on srsRAN for Security Research in LTE Network. In2023 4th Information Communication Technologies Conference (ICTC). 40–43

    work page 2023

  23. [23]

    Upadhyaya and Aly S

    S.D. Sinha et al. 2022. Prototyping Next Generation O-RAN Research Testbeds with SDRs.arXiv preprint arXiv:2205.13178(2022). https://arxiv.org/pdf/2205. 13178

    work page arXiv 2022

  24. [24]

    2024.srsRAN 4G Documentation

    Software Radio Systems (SRS). 2024.srsRAN 4G Documentation. https://docs. srsran.com/projects/4g/en/latest [Online; accessed May 6, 2025]

    work page 2024

  25. [25]

    Adrian Sonnert. 2018. Predicting inter-frequency measurements in an LTE network using supervised machine learning: a comparative study of learning algorithms and data processing techniques

    work page 2018

  26. [26]

    Teixeira and J

    A. Teixeira and J. Timóteo. 2021. A Predictive Resource Allocation for Wireless Communications Systems.SN Computer Science2, 2 (2021), 1–14. https://doi. org/10.1007/s42979-021-00854-8

    work page doi:10.1007/s42979-021-00854-8 2021

  27. [27]

    2025.Fundamentals of O-RAN

    Nishith D Tripathi and Vijay K Shah. 2025.Fundamentals of O-RAN. John Wiley & Sons

    work page 2025

  28. [28]

    Shuochao Yao, Shaohan Hu, Yiran Zhao, Aston Zhang, and Tarek Abdelzaher

    work page

  29. [29]

    InProceedings of the 26th international conference on world wide web

    Deepsense: A unified deep learning framework for time-series mobile sensing data processing. InProceedings of the 26th international conference on world wide web. 351–360

    work page

  30. [30]

    Xiubin Ye et al. 2022. Design and Performance Study of 4G Communication System Based on srsRAN. InProc. SPIE 12474, Sixth International Conference on Photonics and Optical Engineering. 124742B. https://doi.org/10.1117/12.2653437

    work page doi:10.1117/12.2653437 2022