Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Age and Power Minimization via Meta-Deep Reinforcement Learning in UAV Networks

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Meta-initialized DQN adapts UAV to new AoI-power goals in ~100 episodes

desk verdict A plausible MAML-DQN extension for UAV AoI/power trade-offs whose central 'fast adaptation' claim is undercut by an unreported train/test λ split. read the letter →

arxiv 2501.14603 v1 pith:6G67EMGJ submitted 2025-01-24 cs.LG cs.AI

Sankani Sarathchandra , Eslam Eldeeb , Mohammad Shehab , Hirley Alves , Konstantin Mikhaylov , Mohamed-Slim Alouini This is my paper · ORCID
classification cs.LGcs.AI
keywords AgeofinformationMeta-learningDeepreinforcementlearningUAVnetworksMAMLIoTTrajectoryoptimizationTransmissionpower
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 argues that a meta-learned initialization for a deep Q-network can make a UAV data-collector quickly adapt when the operator changes how it weighs information freshness against device battery consumption, and that this adaptation costs far fewer training episodes than learning each weighting from scratch. The target is a power-limited IoT network served by a flying UAV that decides both its trajectory and which device transmits each slot. Task variation comes from a scalar lambda in the reward that trades age-of-information against transmission power. The paper combines DQN with model-agnostic meta-learning (MAML), claiming faster convergence, better adaptation to new objectives, and lower combined AoI and power than random initialization. A sympathetic reader would take this as evidence that meta-learning can reuse experience across wireless multi-objective control tasks rather than retraining per objective.

What carries the argument

The mechanism is MAML's two-loop optimization applied to the DQN's Q-network weights. The inner loop performs task-specific gradient descent on the Bellman-residual loss for each sampled lambda task, producing task-adapted weights; the outer loop aggregates those per-task losses into a meta-loss and updates the shared initialization. The DQN supplies off-policy experience replay and a target network to stabilize the Q-learning updates, while the reward in equation (6) encodes the AoI-power trade-off via lambda. This combined structure is what carries the argument: the learned initialization is supposed to sit in a parameter region from which any new lambda objective can be reached in a few gradient steps and a small number of episodes.

What would settle it

Run the same MAML-vs-random comparison on lambda values held out from the training set and outside the training range, then check whether the MAML advantage survives; if the gap narrows or reverses outside the training range, the generalization claim is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that a MAML-trained Q-network initialization, compared to a randomly initialized DQN, reaches a better or equal AoI-plus-power trade-off in far fewer episodes when the UAV is asked to serve a new weighting of the two objectives. In the reported simulations, the MAML-based initialization converges after roughly 100 episodes while random initialization needs the whole 500-episode budget to approach it, and across 11 test environments the MAML agent achieves lower combined AoI and transmission power for both 5- and 10-device grids. The authors frame this as scalability across varying objective functions: the same initial weights, after a few gradient updates, adapt to a new lambda where a from-scratch agent must explore for a long time.

Load-bearing premise

The paper's fast-adaptation claim rests on testing environments whose lambda values were not seen during meta-training, yet the paper never states that its 11 test values are disjoint from its 10 training values; if they overlap, the advantage could come from remembering training rewards rather than generalizing to new objectives.

Editorial extensions

If this is right

  • Because the MAML initialization converges in roughly 100 episodes instead of 500, an operator could retrain a UAV quickly whenever the priority between data freshness and battery consumption changes.
  • The same Q-network architecture worked for both 5- and 10-device environments, suggesting the learned initialization is not tied to one particular device count.
  • With at least six sampled training tasks, the 10-device scenario converges well using only 30 episodes, so task diversity matters more than episode count for fast adaptation.
  • Meta-testing rewards improve as the number of training tasks increases, indicating that broader task coverage strengthens the resulting initialization's ability to generalize to new objectives.
  • Random initialization eventually reaches similar AoI-power regions when given enough episodes, so the practical benefit of the meta-learned initialization is reduced training time and computational cost rather than a better asymptotic solution.

Reading between the lines

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

  • The paper leaves implicit a strong generalization claim: if the 11 test lambda values are truly disjoint from the 10 training values, the results suggest the initialization learns an objective prior that transfers across reward scalings; if they overlap, the speed-up could be reward memorization rather than true generalization.
  • A testable extension is to apply the same MAML-over-DQN pattern to continuous-action deep RL such as DDPG or PPO, which would let the UAV choose arbitrary headings and speeds instead of five discrete actions.
  • One could also probe what the meta-learned initialization encodes by inspecting whether the Q-network's early layers learn a generic AoI state representation while later layers encode the lambda weighting; that would tell whether the initialization is genuinely task-agnostic.
  • Since random initialization catches up with enough episodes, the strongest practical implication may be computational savings, and a direct benchmark of wall-clock time and energy per converged policy would make that benefit concrete.
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.

Referee Report

3 major / 6 minor

Summary. The paper studies a UAV-assisted IoT data collection problem in which a UAV moves over a grid and schedules one IoT device per time slot to minimize a weighted combination of age-of-information (AoI) and transmission power. The authors formulate the problem as P1 in Eq. (4a), cast it as an MDP with the reward function in Eq. (6), and propose combining DQN with MAML so that the Q-network initial weights are meta-trained across tasks generated by varying the trade-off parameter λ. The experimental section compares the proposed meta-DQN against a randomly initialized DQN in terms of meta-training and meta-testing rewards, convergence speed, and the achieved AoI/power trade-off. The central claim is that the MAML-initialized DQN converges faster, adapts to new objectives with few episodes, and attains lower combined AoI and power than traditional DQN.

Significance. If the central claim is borne out, the paper makes a modest but useful contribution: it shows that a standard meta-learning wrapper (MAML) can initialize a DQN for a UAV trajectory and scheduling problem so that re-training on a new AoI/power trade-off is faster than training from scratch. The paper is clearly written in terms of the system model, the MDP components, and the algorithms, and it includes a parameter table that makes the simulations largely reproducible. The main strength is that the proposed method is compared against a conventional random-initialization DQN baseline, which is the natural traditional deep RL reference for this setting. However, the empirical evidence currently has three load-bearing gaps: the reward function does not match the stated optimization objective, the meta-testing tasks are not shown to be unseen, and the simulation results are reported without any statistical characterization. These gaps need to be addressed before the fast-adaptation claim can be accepted.

major comments (3)
  1. [§II-A, Eq. (4a) and §III-A3, Eq. (6)] The reward function in Eq. (6) does not correspond to the objective P1 in Eq. (4a). P1 minimizes (1/T) Σ_t [Σ_d ω A_d(t) + (λ/D) Σ_d P_d(t)], so on a per-step basis the AoI term has coefficient 1 and the power term has coefficient λ/D. The reward in Eq. (6) is r(t) = -(1/D) Σ_d ω A_d(t) - λ Σ_d P_d(t), which has AoI coefficient 1/D and power coefficient λ. Thus the DQN is trained to maximize a different scalarization of the two objectives, and the trade-off controlled by λ in the reward is not the same trade-off as in P1. The authors should either change the reward to match P1 (up to an affine scaling and the discount factor) or justify why the mismatch is inconsequential for the experimental comparisons. As written, Fig. 6 reports the trade-off for a different objective than the one the paper claims to solve.
  2. [§V, meta-training and meta-testing protocol] The meta-testing protocol does not establish that the test tasks are unseen. The text states that 10 meta-training objective functions with varying λ values are used and that 11 different test environments with different λ values are used, but it never lists the λ values nor states that the 11 test λ values are disjoint from the 10 training λ values. Fig. 4 reports results for λ=300; if λ=300 is one of the training values, then the fast convergence in Fig. 4 and the per-task results in Fig. 5 measure memorization of a familiar reward rather than generalization to a new objective. The authors must report the exact λ sets and confirm that the test values are held out. Without this, the central claim of fast adaptation to new objectives is not empirically demonstrated.
  3. [§V, all simulation figures] All reported curves and trade-off points appear to come from single runs, with no error bars, confidence intervals, or multiple seeds. Reinforcement learning results, especially with random initialization, can vary substantially across runs, and the claimed advantages in convergence speed and final reward need to be quantified over several independent runs. The authors should provide the number of seeds and report mean/median curves with variance bands or, at minimum, error bars on the key points in Figs. 3–6.
minor comments (6)
  1. [§III-B, Eq. (8)] The symbol α is used both for the Q-learning update step in Eq. (8) and for the adaptation learning rate in Eq. (9) and Table II. These are different quantities in the proposed meta-RL algorithm; please use distinct notation, e.g., η for the Q-learning step.
  2. [Algorithm 3, lines 14–18] The nested loop 'for each episode do' inside the task loop has no episode count or termination condition, and it is unclear whether the evaluation loss uses a frozen θ'_i or continues updating. Please specify the number of evaluation episodes and clarify the role of this loop.
  3. [Fig. 6] The x-axis is labeled 'Age-of-information' without units; since AoI is a time quantity, please state whether the plotted value is the average AoI per device and per episode, and similarly clarify whether the power is averaged over time slots and devices.
  4. [§V, paragraph on Fig. 4] The text says 'setting λ = 300 in an environment with 5 devices and 10 devices, respectively,' but Fig. 4 appears to contain four curves; please clarify which curve corresponds to which configuration and whether both configurations use λ=300.
  5. [§III-A3, text after Eq. (6)] The sentence describing the Bellman term states 'γ Q(s(t+1), a(t+1))' while Eq. (8) uses max_a Q(s(t+1), a); the text should match the equation.
  6. [§V, last paragraph] The statement that random initialization 'can reach the same regions with training more episodes' is not supported by any displayed experiment; either provide evidence for this claim or soften it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: the central claim is an empirical comparison of MAML-initialized vs. randomly initialized DQNs, and the cited self-references are not load-bearing.

full rationale

The paper's central claim is empirical: a MAML-initialized DQN converges faster and achieves better reward than a randomly initialized DQN across λ-weighted AoI/power tasks. This is evaluated in Section V by comparing learning curves and per-task rewards. No equation is fitted to produce these results; the reward in (6) and problem P1 in (4) are aligned by design, but the algorithm's performance is not derived from the reward by construction. The few self-citations ([6], [9], [28], [29], [31]) supply standard system-model equations and are not load-bearing: the channel gain in (1) and reward in (6) are standard and independently stated, and MAML is also cited to Finn et al. [15]. The reviewer concern that the 11 meta-testing λ values are not shown to be disjoint from the 10 meta-training values is an experimental-validity gap, not circularity: even if the test tasks overlap, that affects the interpretation of 'unseen' generalization, but the comparison itself does not reduce by definition to its inputs. Under the hard rule requiring an exhibited reduction, no step qualifies.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the standard LoS channel and AoI models, on the chosen hyperparameters, and on an unspecified task distribution over lambda. The reward mismatch between Eq. (4a) and Eq. (6) means the algorithm minimizes a slightly different objective than the one stated.

free parameters (5)
  • Task scaling factor lambda = varied; e.g., 300 in Fig. 4
    Defines the trade-off between AoI and transmission power in the reward. The paper varies lambda over 10 training tasks and 11 test tasks, but the specific values are not listed, so the reader cannot reconstruct the task distribution.
  • Adaptation learning rate alpha = 0.0001 (Table II)
    Controls the inner-loop MAML gradient step size; central to the fast-adaptation claim.
  • Meta-learning rate beta = 0.0001 (Table II)
    Controls the outer-loop meta-update step size.
  • Discount factor gamma = 0.99 (Table II)
    Determines how much the agent weights future rewards in the DQN return.
  • Few-shot budget Emax = 50 or 100 episodes (Fig. 3)
    The constraint (4b) is the entire basis for claiming few-shot adaptation; the results are shown for these two budget values.
assumptions (4)
  • domain assumption Line-of-sight channel model with channel gain g_u,d(t) = g0 / (hu^2 + r_u,d(t)^2) holds for all devices and times.
    Used in Eq. (1) to compute transmit power in Eq. (2); assumes no blockage or fading.
  • domain assumption AoI resets to 1 on successful service and increments by 1 otherwise, capped at Amax.
    Eq. (3) defines the state evolution; this is the standard AoI model but assumes instantaneous transmission with no errors.
  • ad hoc to paper The reward function in Eq. (6) is a faithful optimization signal for the P1 objective in Eq. (4a).
    The paper introduces this reward to train the RL agent, but its scaling does not match P1 exactly, so the agent is not optimizing the stated objective.
  • domain assumption The MDP transition probabilities are determined by the grid movement rules and uniform random device placements; the agent has full observability of the AoI vector and UAV position.
    The state space in Section III-A assumes full observability of all device ages and the UAV position.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Age and Power Minimization via Meta-Deep Reinforcement Learning in UAV Networks." pith.science (2026). https://pith.science/paper/6G67EMGJ

@misc{pith2026250114603,
  author       = {Pith},
  title        = {Pith review of: Age and Power Minimization via Meta-Deep Reinforcement Learning in UAV Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6G67EMGJ}},
  note         = {Machine review of arXiv:2501.14603}
}
read the original abstract

Age-of-information (AoI) and transmission power are crucial performance metrics in low energy wireless networks, where information freshness is of paramount importance. This study examines a power-limited internet of things (IoT) network supported by a flying unmanned aerial vehicle(UAV) that collects data. Our aim is to optimize the UAV flight trajectory and scheduling policy to minimize a varying AoI and transmission power combination. To tackle this variation, this paper proposes a meta-deep reinforcement learning (RL) approach that integrates deep Q-networks (DQNs) with model-agnostic meta-learning (MAML). DQNs determine optimal UAV decisions, while MAML enables scalability across varying objective functions. Numerical results indicate that the proposed algorithm converges faster and adapts to new objectives more effectively than traditional deep RL methods, achieving minimal AoI and transmission power overall.

Figures

Figures reproduced from arXiv: 2501.14603 by the authors.

Figure 1
Figure 1. The considered system model operating within a grid environment, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. A visualization of the proposed meta-RL algorithm using MAML and DQNs. The left block represents the interaction between the DQN agent and the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Meta-training performance: (a) the convergence of the meta-losses, (b) the convergence of the rewards as a function of meta-training epochs tested over [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Achieved rewards after 30 episodes tested over 11 different test environments. alize [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Testing the proposed meta-RL algorithm over [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Resilient UAV Trajectory Planning via Few-Shot Meta-Offline Reinforcement Learning

    cs.RO 2025-02 conditional novelty 4.0 of 10

    A hybrid meta-offline reinforcement learning algorithm trains a UAV to minimize data age and transmission power from static datasets and adapts to new tasks in under 40 epochs.

Reference graph

Works this paper leans on

32 extracted references · 26 canonical work pages · cited by 1 Pith paper

  1. [1]

    Multi-agent deep reinforcement learning to manage connected autonomous vehicles at tomorrow’s inter- sections,

    G.-P. Antonio and C. Maria-Dolores, “Multi-agent deep reinforcement learning to manage connected autonomous vehicles at tomorrow’s inter- sections,” IEEE Transactions on V ehicular Technology, vol. 71, no. 7, pp. 7033–7043, 2022

    work page 2022

  2. [2]

    Distributed learning methodologies for massive machine type commu- nication,

    M. V . Da Silva, E. Eldeeb, M. Shehab, H. Alves, and R. D. Souza, “Distributed learning methodologies for massive machine type commu- nication,” Authorea Preprints, 2024

    work page 2024

  3. [3]

    A perspective on time toward wireless 6G,

    P. Popovski, F. Chiariotti, K. Huang, A. E. Kalør, M. Kountouris, N. Pappas, and B. Soret, “A perspective on time toward wireless 6G,” Proceedings of the IEEE , vol. 110, no. 8, pp. 1116–1146, 2022

    work page 2022

  4. [4]

    White Paper on Critical and Massive Machine Type Communication Towards 6G

    N. H. Mahmood, S. B ¨ocker, A. Munari, F. Clazzer, I. Moerman, K. Mikhaylov, O. Lopez, O.-S. Park, E. Mercier, H. Bartz et al. , “White paper on critical and massive machine type communication towards 6G,” arXiv preprint arXiv:2004.14146 , 2020

    work page Pith review arXiv 2004

  5. [5]

    Age of information: A new concept, metric, and tool,

    A. Kosta, N. Pappas, and V . Angelakis, “Age of information: A new concept, metric, and tool,” F oundations and Trends in Networking, Now Publishers, Inc. , 2017

    2017

  6. [6]

    Traffic learning and proactive UA V trajectory planning for data uplink in markovian IoT models,

    E. Eldeeb, M. Shehab, and H. Alves, “Traffic learning and proactive UA V trajectory planning for data uplink in markovian IoT models,”IEEE Internet of Things Journal , vol. 11, no. 8, pp. 13 496–13 508, 2024

    2024

  7. [7]

    Traffic prediction and fast uplink for hidden markov IoT models,

    E. Eldeeb, M. Shehab, A. E. Kalø r, P. Popovski, and H. Alves, “Traffic prediction and fast uplink for hidden markov IoT models,” IEEE Internet of Things Journal , vol. 9, no. 18, pp. 17 172–17 184, 2022

    work page 2022

  8. [8]

    A Tutorial on UA Vs for Wireless Networks: Applications, Challenges, and Open Problems,

    M. Mozaffari, W. Saad, M. Bennis, Y .-H. Nam, and M. Debbah, “A Tutorial on UA Vs for Wireless Networks: Applications, Challenges, and Open Problems,” IEEE Communications Surveys & Tutorials , vol. 21, no. 3, pp. 2334–2360, 2019

    2019

Show all 32 references
  1. [9]

    Multi-UA V path learning for age and power optimization in IoT with UA V battery recharge,

    E. Eldeeb, J. M. de Souza Sant’Ana, D. E. P ´erez, M. Shehab, N. H. Mahmood, and H. Alves, “Multi-UA V path learning for age and power optimization in IoT with UA V battery recharge,” IEEE Transactions on V ehicular Technology, vol. 72, no. 4, pp. 5356–5360, 2022

    2022

  2. [10]

    Machine learning for large-scale optimization in 6G wireless networks,

    Y . Shi, L. Lian, Y . Shi, Z. Wang, Y . Zhou, L. Fu, L. Bai, J. Zhang, and W. Zhang, “Machine learning for large-scale optimization in 6G wireless networks,” IEEE Communications Surveys & Tutorials , vol. 25, no. 4, pp. 2088–2132, 2023

    2023

  3. [11]

    Deep reinforcement learning for internet of things: A comprehensive survey,

    W. Chen, X. Qiu, T. Cai, H.-N. Dai, Z. Zheng, and Y . Zhang, “Deep reinforcement learning for internet of things: A comprehensive survey,” IEEE Communications Surveys & Tutorials , vol. 23, no. 3, pp. 1659– 1692, 2021

    2021

  4. [12]

    Applications of deep reinforcement learning in commu- nications and networking: A survey,

    N. C. Luong, D. T. Hoang, S. Gong, D. Niyato, P. Wang, Y .-C. Liang, and D. I. Kim, “Applications of deep reinforcement learning in commu- nications and networking: A survey,” IEEE Communications Surveys & Tutorials, vol. 21, no. 4, pp. 3133–3174, 2019

    2019

  5. [13]

    Human-level control through deep reinforce- ment learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. Rusu, J. Veness, M. Bellemare, A. Graves, M. Riedmiller, A. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis, “Human-level control through deep reinforc...

    2015

  6. [14]

    Energy optimization and age of information enhancement in multi-uav networks using deep reinforcement learning,

    J. Kim, S. Park, and H. Park, “Energy optimization and age of information enhancement in multi-uav networks using deep reinforcement learning,” Electronics Letters, vol. 60, no. 20, p. e70063, 2024

    2024

  7. [15]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    S. L. Chelsea Finn, Pieter Abbeel, “Model-agnostic meta-learning for fast adaptation of deep networks,” 34th International Conference on Machine Learning, vol. 70, pp. 1126–1135, 2017. 10

    2017

  8. [16]

    Wireless communications with unmanned aerial vehicles: opportunities and challenges,

    Y . Zeng, R. Zhang, and T. J. Lim, “Wireless communications with unmanned aerial vehicles: opportunities and challenges,” IEEE Commu- nications Magazine , vol. 54, no. 5, pp. 36–42, 2016

    2016

  9. [17]

    Aoi-minimal trajectory planning and data collection in uav-assisted wireless powered iot networks,

    H. Hu, K. Xiong, G. Qu, Q. Ni, P. Fan, and K. B. Letaief, “Aoi-minimal trajectory planning and data collection in uav-assisted wireless powered iot networks,” IEEE Internet of Things Journal , vol. 8, no. 2, pp. 1211– 1223, 2021

    2021

  10. [18]

    Multi- objective optimization for uav-assisted wireless powered iot networks based on extended ddpg algorithm,

    Y . Yu, J. Tang, J. Huang, X. Zhang, D. K. C. So, and K.-K. Wong, “Multi- objective optimization for uav-assisted wireless powered iot networks based on extended ddpg algorithm,” IEEE Transactions on Communi- cations, vol. 69, no. 9, pp. 6361–6374, 2021

    2021

  11. [19]

    Toward autonomous multi-uav wireless network: A survey of reinforcement learning-based approaches,

    Y . Bai, H. Zhao, X. Zhang, Z. Chang, R. J ˜A¤ntti, and K. Yang, “Toward autonomous multi-uav wireless network: A survey of reinforcement learning-based approaches,” IEEE Communications Surveys & Tutorials , vol. 25, no. 4, pp. 3038–3067, 2023

    2023

  12. [20]

    Uav path planning for wireless data harvesting: A deep reinforcement learning approach,

    H. Bayerlein, M. Theile, M. Caccamo, and D. Gesbert, “Uav path planning for wireless data harvesting: A deep reinforcement learning approach,” in GLOBECOM 2020 - 2020 IEEE Global Communications Conference, 2020, pp. 1–6

    2020

  13. [21]

    Deep rl-based trajectory planning for aoi minimization in uav-assisted iot,

    C. Zhou, H. He, P. Yang, F. Lyu, W. Wu, N. Cheng, and X. Shen, “Deep rl-based trajectory planning for aoi minimization in uav-assisted iot,” in 2019 11th International Conference on Wireless Communications and Signal Processing (WCSP) , 2019, pp. 1–6

    2019

  14. [22]

    Uav trajectory planning in wireless sensor networks for energy consumption minimization by deep reinforcement learning,

    B. Zhu, E. Bedeer, H. H. Nguyen, R. Barton, and J. Henry, “Uav trajectory planning in wireless sensor networks for energy consumption minimization by deep reinforcement learning,” IEEE Transactions on V ehicular Technology, vol. 70, no. 9, pp. 9540–9554, 2021

    2021

  15. [23]

    Deep reinforcement learning-based resource allocation in cooperative uav-assisted wireless networks,

    P. Luong, F. Gagnon, L.-N. Tran, and F. Labeau, “Deep reinforcement learning-based resource allocation in cooperative uav-assisted wireless networks,” IEEE Transactions on Wireless Communications , vol. 20, no. 11, pp. 7610–7625, 2021

    2021

  16. [24]

    Meta-learning for wireless communications: A survey and a comparison to gnns,

    B. Zhao, J. Wu, Y . Ma, and C. Yang, “Meta-learning for wireless communications: A survey and a comparison to gnns,” IEEE Open Journal of the Communications Society , vol. 5, pp. 1987–2015, 2024

    1987

  17. [25]

    Meta-learning based optimization for large scale wireless systems,

    R. C. Loli and B. Clerckx, “Meta-learning based optimization for large scale wireless systems,” arXiv preprint arXiv:2407.01823 , 2024

    2024 arXiv

  18. [26]

    Meta-reinforcement learning for trajectory design in wireless uav networks,

    Y . Hu, M. Chen, W. Saad, H. V . Poor, and S. Cui, “Meta-reinforcement learning for trajectory design in wireless uav networks,” in GLOBECOM 2020 - 2020 IEEE Global Communications Conference , 2020, pp. 1–6

    2020

  19. [27]

    Continual meta- reinforcement learning for uav-aided vehicular wireless networks,

    R. Marini, S. Park, O. Simeone, and C. Buratti, “Continual meta- reinforcement learning for uav-aided vehicular wireless networks,” in ICC 2023 - IEEE International Conference on Communications , 2023, pp. 5664–5669

    2023

  20. [28]

    Age minimization in massive IoT via UA V swarm: A multi-agent reinforcement learning approach,

    E. Eldeeb, M. Shehab, and H. Alves, “Age minimization in massive IoT via UA V swarm: A multi-agent reinforcement learning approach,” in2023 IEEE 34th Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC) . IEEE, 2023, pp. 1–6

    2023

  21. [29]

    Traffic learning and proactive UA V trajectory planning for data uplink in markovian IoT models,

    ——, “Traffic learning and proactive UA V trajectory planning for data uplink in markovian IoT models,” IEEE Internet of Things Journal , vol. 11, no. 8, pp. 13 496–13 508, 2024

    2024

  22. [30]

    Deep reinforcement learning: A brief survey,

    K. Arulkumaran, M. P. Deisenroth, M. Brundage, and A. A. Bharath, “Deep reinforcement learning: A brief survey,” IEEE Signal Processing Magazine, vol. 34, no. 6, pp. 26–38, Nov. 2017

    2017

  23. [31]

    Semantic meta-split learning: A TinyML scheme for few-shot wireless image classification,

    E. Eldeeb, M. Shehab, H. Alves, and M.-S. Alouini, “Semantic meta-split learning: A TinyML scheme for few-shot wireless image classification,”

  24. [2024]

    Available: https://arxiv.org/abs/2409.12978

    [Online]. Available: https://arxiv.org/abs/2409.12978

    arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.