arxiv: 2604.13038 · v1 · submitted 2026-02-20 · 📡 eess.SP

Recognition: no theorem link

Uncertainty-Weighted Experience Replay for Continual MIMO Channel Prediction

Muhammad Jazib Qamar , Muhammad Hamza Nawaz , Messaoud Ahmed Ouameur , Ayesha Mohsin , Miloud Bagaa

Authors on Pith no claims yet

Pith reviewed 2026-05-15 20:24 UTC · model grok-4.3

classification 📡 eess.SP

keywords continual learningexperience replayMIMO channel predictionuncertainty estimationMonte-Carlo dropoutLSTMCSI prediction6G

0 comments

The pith

Uncertainty-weighted experience replay stabilizes generalization in continual MIMO channel prediction.

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

The paper introduces Uncertainty-Weighted Experience Replay to handle non-stationary MIMO channels by feeding model uncertainty back into which past samples are replayed and how heavily they are weighted during training. A lightweight LSTM with Monte-Carlo dropout produces both the channel prediction and an estimate of its own variance; that variance then scales the reconstruction loss for each replayed example. A sympathetic reader cares because wireless environments change continuously with mobility, so models must keep adapting without forgetting earlier conditions or requiring ever-larger memory. The reported results show validation NMSE staying near 0 dB and a 0.93 correlation between predicted uncertainty and actual error, suggesting the uncertainty signal is reliable enough to guide learning.

Core claim

The Uncertainty-Weighted Experience Replay (UW-ER) framework employs a lightweight LSTM with Monte-Carlo dropout to estimate predictive variance for each sample, which is then used to adaptively weight the reconstruction loss during replay-based training. On a UMi-Dense MIMO dataset generated from a 3GPP-consistent stochastic model, this yields stable generalization with validation NMSE centered near 0 dB and a correlation of r = 0.93 between predicted uncertainty and reconstruction error. The LARS-based replay policy further enables competitive performance at smaller memory budgets compared to standard reservoir replay.

What carries the argument

Uncertainty-Weighted Experience Replay (UW-ER) that uses predictive variance from Monte-Carlo dropout in an LSTM to scale the reconstruction loss on replayed samples.

If this is right

Validation NMSE remains centered near 0 dB across continual updates on changing channels.
Predicted uncertainty correlates at r = 0.93 with actual reconstruction error.
LARS replay matches reservoir performance while using less memory.
Stability improves without raising computational cost per update.
The approach scales to adaptive 6G systems that must track CSI in real time.

Where Pith is reading between the lines

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

The same uncertainty-weighted replay loop could be tested on other sequential wireless tasks such as beam prediction or interference forecasting.
Smaller memory footprints make the method attractive for edge devices that cannot store large replay buffers.
Well-calibrated uncertainty opens the door to risk-aware resource allocation that trusts predictions only when variance is low.
Direct comparison against other continual-learning regularizers on the same 3GPP dataset would isolate how much gain comes from the uncertainty weighting itself.

Load-bearing premise

Monte-Carlo dropout supplies a reliable estimate of predictive uncertainty whose magnitude actually tracks reconstruction error on non-stationary MIMO channels.

What would settle it

A new set of non-stationary channel traces where the correlation between MC-dropout variance and true prediction error falls well below 0.7, or where validation NMSE drifts far from 0 dB under continued updates, would show that the weighting does not deliver the claimed robustness.

Figures

Figures reproduced from arXiv: 2604.13038 by Ayesha Mohsin, Messaoud Ahmed Ouameur, Miloud Bagaa, Muhammad Hamza Nawaz, Muhammad Jazib Qamar.

**Figure 1.** Figure 1: CDF of validation NMSE for UW-ER. V. RESULTS This section evaluates the proposed Uncertainty–Weighted Experience Replay (UW-ER) framework on the 3GPP UMiDense continual-learning CSI stream. We demonstrate that UW-ER provides (i) higher accuracy, (ii) significantly improved calibration, (iii) stronger robustness across frequency, and (iv) superior stability under non-stationarity when compared with state… view at source ↗

**Figure 5.** Figure 5: and 6 show predicted channel magnitude maps. UWER preserves the smooth frequency evolution and TX/RX structure observed in the true CSI. This distinguishes it from transformer-based predictors [8], which often oversmooth or [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

**Figure 6.** Figure 6: Channel magnitude(case B) [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗

**Figure 3.** Figure 3: Overall NMSE histogram [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 7.** Figure 7: Per-RB NMSE (dB). distort fine-grained patterns when updated continually with limited memory. D. Frequency-Wise Robustness The per-RB NMSE curve in [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗

**Figure 4.** Figure 4: Uncertainty calibration C. Preservation of Spatial–Frequency Structure [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

read the original abstract

In dynamic wireless environments, accurate channel state information (CSI) prediction remains challenging due to non-stationary fading, mobility. This paper proposes an Uncertainty-Weighted Experience Replay (UW-ER) framework that integrates model uncertainty into the replay sampling process to improve robustness in online CSI prediction. A lightweight LSTM architecture with Monte-Carlo dropout is employed to estimate predictive variance, which is then used to adaptively weight the reconstruction loss for each training sample. The proposed method is evaluated on a UMi-Dense MIMO channel dataset generated using a stochastic fading model consistent with 3GPP standards. Results show that UW-ER achieves stable generalization, with validation NMSE centered near 0 dB and a strong correlation (r = 0.93) between predicted uncertainty and reconstruction error, indicating well-calibrated confidence estimates. Ablation studies demonstrate that the LARS-based replay policy achieves competitive performance with smaller memory budgets compared to conventional reservoir replay. Overall, the UW-ER approach improves continual channel learning stability without increasing computational complexity, offering a scalable solution for future 6G adaptive communication systems.

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.

Desk Editor's Note private letter to a colleague

The paper combines MC dropout uncertainty with experience replay in a lightweight LSTM for continual MIMO CSI prediction and gets a solid r=0.93 correlation on simulated 3GPP channels, but the evaluation leaves open whether that calibration holds on true future shifts.

read the letter

The core contribution is a practical tweak: use Monte-Carlo dropout to estimate predictive variance in an LSTM, then weight the replay sampling or loss by that uncertainty so the model focuses on harder, more uncertain channel samples during continual updates. On the UMi-Dense dataset generated from the stochastic 3GPP model, they report validation NMSE near 0 dB and that strong correlation between predicted uncertainty and reconstruction error. The LARS-based replay also beats plain reservoir sampling at smaller memory sizes, which matters for edge deployment. That part is cleanly executed and directly relevant to online channel tracking in 6G settings where mobility keeps shifting the distribution. The ablation on memory budgets is a useful detail that practitioners can actually use. The work stays within standard LSTM and continual-learning tools rather than inventing new architectures, so the novelty is in the domain-specific weighting rather than a new framework. The main soft spot is the evaluation. Everything is on synthetic fading traces; there are no real over-the-air measurements or hardware traces to check whether the uncertainty estimates survive RF impairments, antenna calibration drift, or more abrupt mobility changes. The r=0.93 figure is reported without clear confirmation that it was measured on a strictly held-out future window after the continual updates, which raises the exact concern in the stress-test note about possible self-reinforcement from the replay policy itself. MC dropout in recurrent models also tends to underestimate epistemic uncertainty under distribution shift, so the calibration claim would need tighter checks such as proper scoring rules or comparison against ensemble or Bayesian baselines. The paper does not appear to have circular reasoning or invented entities, and the math is standard. This is the sort of targeted incremental paper that wireless-ML groups working on CSI prediction would want to see in review so the methods and data splits can be examined in detail. I would bring it to a reading group to discuss the uncertainty calibration metrics, but I would not cite it in my own work without seeing stronger held-out results. It is solid enough to deserve peer review rather than a desk reject.

Referee Report

3 major / 1 minor

Summary. The paper proposes an Uncertainty-Weighted Experience Replay (UW-ER) framework for continual MIMO channel prediction that integrates Monte-Carlo dropout uncertainty estimates from a lightweight LSTM into adaptive replay sampling and loss weighting. Evaluated on a 3GPP UMi-Dense stochastic fading dataset, it claims stable generalization (validation NMSE near 0 dB), strong calibration (r=0.93 correlation between predicted uncertainty and reconstruction error), and competitive performance with LARS-based replay under smaller memory budgets compared to reservoir sampling.

Significance. If the central empirical claims hold under rigorous verification, the work would be moderately significant for 6G adaptive systems by demonstrating a practical way to stabilize online CSI prediction in non-stationary environments without added complexity. The explicit use of predictive variance for replay weighting and the reported uncertainty-error calibration are strengths that could inform continual learning in wireless applications; however, the absence of detailed baselines, splits, and statistical reporting limits immediate impact.

major comments (3)

[Abstract] Abstract: the reported r=0.93 correlation between predicted uncertainty and reconstruction error is not stated to have been computed on a strictly held-out future temporal window after all continual updates; without this, the statistic may be inflated by the uncertainty-weighted sampling itself preferentially retaining high-uncertainty samples.
[Abstract] Abstract: no experimental setup details (data generation parameters, train/validation/test splits, number of continual tasks, baselines such as standard reservoir replay or EWC, or statistical significance tests) are provided, preventing verification of the NMSE-near-0 dB and stable-generalization claims.
[Abstract] Abstract: the assumption that Monte-Carlo dropout in an LSTM yields predictive variance that reliably tracks true reconstruction error under mobility-induced distribution shifts is load-bearing for the calibration claim, yet no justification or ablation against alternative uncertainty estimators (e.g., deep ensembles) is indicated.

minor comments (1)

[Abstract] Abstract: the term 'LARS-based replay policy' is introduced without definition or citation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and commit to revisions that improve clarity and rigor without altering the core contributions.

read point-by-point responses

Referee: [Abstract] Abstract: the reported r=0.93 correlation between predicted uncertainty and reconstruction error is not stated to have been computed on a strictly held-out future temporal window after all continual updates; without this, the statistic may be inflated by the uncertainty-weighted sampling itself preferentially retaining high-uncertainty samples.

Authors: We thank the referee for this important clarification. The reported correlation was computed on the validation set encountered during the continual updates. To eliminate any potential bias from the replay mechanism, we will recompute the correlation on a strictly held-out test set drawn from a future temporal window after all updates have completed. The revised statistic and associated methodology will appear in the updated manuscript. revision: yes
Referee: [Abstract] Abstract: no experimental setup details (data generation parameters, train/validation/test splits, number of continual tasks, baselines such as standard reservoir replay or EWC, or statistical significance tests) are provided, preventing verification of the NMSE-near-0 dB and stable-generalization claims.

Authors: We agree that the abstract is too terse. In the revision we will expand it to include the 3GPP UMi-Dense stochastic fading parameters, the temporal train/validation/test splits across the sequence of mobility scenarios, the number of continual tasks, explicit comparison to reservoir sampling and EWC, and reporting of mean NMSE with standard deviation over repeated runs. revision: yes
Referee: [Abstract] Abstract: the assumption that Monte-Carlo dropout in an LSTM yields predictive variance that reliably tracks true reconstruction error under mobility-induced distribution shifts is load-bearing for the calibration claim, yet no justification or ablation against alternative uncertainty estimators (e.g., deep ensembles) is indicated.

Authors: Monte-Carlo dropout was selected for its negligible memory and compute overhead in an online setting. We will insert a concise justification in the methods section, supported by references to its established use for recurrent models under distribution shift. We will also add a direct comparison against deep ensembles (with associated complexity trade-offs) to the experimental results. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical results on generated data are independent measurements

full rationale

The paper presents an empirical UW-ER method using LSTM with MC dropout, evaluated on a synthetically generated 3GPP UMi MIMO dataset. Reported statistics (validation NMSE near 0 dB, r=0.93 correlation between uncertainty and error) are measured outcomes on held-out validation data rather than quantities derived by construction from fitted parameters or self-citations. No equations, self-definitional steps, or load-bearing self-citations appear in the abstract or description that would reduce the central claims to inputs. The correlation is presented as an observed calibration result, not a renamed fit.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on abstract; no explicit free parameters, axioms, or invented entities are detailed beyond standard neural network assumptions.

axioms (1)

domain assumption Monte-Carlo dropout approximates Bayesian predictive uncertainty
Invoked to estimate predictive variance for weighting replay samples.

pith-pipeline@v0.9.0 · 5505 in / 1147 out tokens · 24990 ms · 2026-05-15T20:24:18.382815+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

[1]

Impact of channel aging on massive mimo vehicular networks in non-isotropic scattering scenar- ios,

H. Li, L. Ding, Y . Wang, P. Wu, and Z. Wang, “Impact of channel aging on massive mimo vehicular networks in non-isotropic scattering scenar- ios,” in2021 IEEE Global Communications Conference (GLOBECOM). IEEE, 2021, pp. 1–6

work page 2021
[2]

Neural network-based fading channel pre- diction: A comprehensive overview,

W. Jiang and H. D. Schotten, “Neural network-based fading channel pre- diction: A comprehensive overview,”IEEE Access, vol. 7, pp. 118 112– 118 124, 2019

work page 2019
[3]

Deep learning-based channel prediction in realistic vehicular communications,

J. Joo, M. C. Park, D. S. Han, and V . Pejovic, “Deep learning-based channel prediction in realistic vehicular communications,”IEEE Access, vol. 7, pp. 27 846–27 858, 2019

work page 2019
[4]

Recurrent neural networks with long short-term memory for fading channel prediction,

W. Jiang and H. D. Schotten, “Recurrent neural networks with long short-term memory for fading channel prediction,” in2020 IEEE 91st vehicular technology conference (VTC2020-Spring). IEEE, 2020, pp. 1–5

work page 2020
[5]

Transformer-based channel prediction for rate-splitting multiple access- enabled vehicle-to-everything communication,

S. Zhang, S. Zhang, Y . Mao, L. K. Yeung, B. Clerckx, and T. Q. Quek, “Transformer-based channel prediction for rate-splitting multiple access- enabled vehicle-to-everything communication,”IEEE Transactions on Wireless Communications, 2024

work page 2024
[6]

Lstm: A search space odyssey,

K. Greff, R. K. Srivastava, J. Koutn ´ık, B. R. Steunebrink, and J. Schmid- huber, “Lstm: A search space odyssey,”IEEE transactions on neural networks and learning systems, vol. 28, no. 10, pp. 2222–2232, 2016

work page 2016
[7]

Gate-variants of gated recurrent unit (gru) neural networks,

R. Dey and F. M. Salem, “Gate-variants of gated recurrent unit (gru) neural networks,” in2017 IEEE 60th international midwest symposium on circuits and systems (MWSCAS). IEEE, 2017, pp. 1597–1600

work page 2017
[8]

Accurate channel prediction based on transformer: Making mobility negligible,

H. Jiang, M. Cui, D. W. K. Ng, and L. Dai, “Accurate channel prediction based on transformer: Making mobility negligible,”IEEE Journal on Selected Areas in Communications, vol. 40, no. 9, pp. 2717–2732, 2022. [Online]. Available: https://oa.ee.tsinghua.edu.cn/ dailinglong/publications/paper/Accurate Channel Prediction Based on Transformer Making Mobility ...

work page 2022
[9]

Generating high dimensional user-specific wireless channels using diffusion models,

T. Lee, J. Park, H. Kim, and J. G. Andrews, “Generating high dimensional user-specific wireless channels using diffusion models,”

work page
[10]

Available: https://arxiv.org/abs/2409.03924

[Online]. Available: https://arxiv.org/abs/2409.03924

work page arXiv
[11]

Llm4cp: Adapting large language models for channel prediction,

B. Liu, X. Liu, S. Gao, X. Cheng, and L. Yang, “Llm4cp: Adapting large language models for channel prediction,”Journal of Communications and Information Networks, vol. 9, no. 2, pp. 113–125, 2024

work page 2024
[12]

Overcoming catastrophic forgetting in neural networks,

J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska et al., “Overcoming catastrophic forgetting in neural networks,”Pro- ceedings of the national academy of sciences, vol. 114, no. 13, pp. 3521–3526, 2017

work page 2017
[13]

Continual learning through synaptic intelligence,

F. Zenke, B. Poole, and S. Ganguli, “Continual learning through synaptic intelligence,” inInternational conference on machine learning. PMLR, 2017, pp. 3987–3995

work page 2017
[14]

Learning without forgetting,

Z. Li and D. Hoiem, “Learning without forgetting,”IEEE transactions on pattern analysis and machine intelligence, vol. 40, no. 12, pp. 2935– 2947, 2017

work page 2017
[15]

Expe- rience replay for continual learning,

D. Rolnick, A. Ahuja, J. Schwarz, T. Lillicrap, and G. Wayne, “Expe- rience replay for continual learning,”Advances in neural information processing systems, vol. 32, 2019

work page 2019
[16]

Revisiting fundamentals of experience replay,

W. Fedus, P. Ramachandran, R. Agarwal, Y . Bengio, H. Larochelle, M. Rowland, and W. Dabney, “Revisiting fundamentals of experience replay,” inInternational Conference on Machine Learning (ICML), ser. PMLR, vol. 119, 2020, pp. 3061–3071. [Online]. Available: https://proceedings.mlr.press/v119/fedus20a/fedus20a.pdf

work page 2020
[17]

Uncertainty prioritized experience replay,

R. Carrasco-Davis, S. Lee, C. Clopath, and W. Dabney, “Uncertainty prioritized experience replay,”arXiv preprint arXiv:2506.09270, 2025

work page arXiv 2025
[18]

Uncertainty-based dynamic weighted experience replay for human-in-the-loop deep reinforcement learning,

Y . Ye, H. Zhouet al., “Uncertainty-based dynamic weighted experience replay for human-in-the-loop deep reinforcement learning,” inArtificial Intelligence and Human-Computer Interaction, 2025, p. –

work page 2025
[19]

Spatial wireless channel prediction under location uncertainty,

L. S. Muppirisettyet al., “Spatial wireless channel prediction under location uncertainty,”Wireless Personal Communications, vol. –, no. –, pp. –, 2015

work page 2015
[20]

Quadriga: A 3-d multi-cell channel model with time evolution for enabling virtual field trials,

S. Jaeckel, L. Raschkowski, K. B ¨orner, and L. Thiele, “Quadriga: A 3-d multi-cell channel model with time evolution for enabling virtual field trials,”IEEE transactions on antennas and propagation, vol. 62, no. 6, pp. 3242–3256, 2014

work page 2014
[21]

Study on channel model for frequencies from 0.5 to 100 GHz (rel-16),

3GPP, “Study on channel model for frequencies from 0.5 to 100 GHz (rel-16),” ETSI, Tech. Rep. TR 38.901 v16.1.0, 2020, accessed Nov. 2025. [Online]. Available: https://www.etsi.org/deliver/etsi tr/ 138900 138999/138901/16.01.00 60/tr 138901v160100p.pdf

work page 2020
[22]

Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,

Y . Gal and Z. Ghahramani, “Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,” inProceedings of the 33rd International Conference on Machine Learning (ICML), ser. PMLR, vol. 48, 2016, pp. 1050–1059. [Online]. Available: https://proceedings.mlr.press/v48/gal16.pdf

work page 2016
[23]

A statistical theory of regularization-based continual learning,

X. Zhao, H. Wang, W. Huang, and W. Lin, “A statistical theory of regularization-based continual learning,”arXiv preprint arXiv:2406.06213, 2024

work page arXiv 2024