Pith. sign in

REVIEW 4 major objections 4 minor 49 references

Achieving Hiding and Smart Anti-Jamming Communication: A Parallel DRL Approach against Moving Reactive Jammer

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper claims that splitting frequency and spreading-factor decisions into two parallel deep-reinforcement-learning networks, coupled by a shared success signal, lets a wireless link hide from and evade a moving reactive jammer, and…

desk verdict Plausible engineering result on joint frequency/spreading-factor anti-jamming, but the action-space decomposition and the MDP proof are not sound as written; the 90% gain is believable yet unverified. read the letter →

arxiv 2502.02385 v1 pith:C5PVFYZX submitted 2025-02-04 cs.IT cs.LGcs.SYeess.SYmath.IT

Yangyang Li , Yuhua Xu , Wen Li , Guoxin Li , Zhibing Feng , Songyi Liu , Jiatao Du , Xinran Li This is my paper · ORCID
classification cs.ITcs.LGcs.SYeess.SYmath.IT
keywords anti-jammingdeepreinforcementlearningmovingreactivejammerspreadspectrumfrequencyselectionactionspacedecompositionnormalizedthroughputexploration-exploitationtradeoff
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

Moving reactive jammers force a hard choice: a large spreading factor hides the signal but leaves it exposed to indiscriminate sweeping, while frequency hopping dodges sweeping but is trackable. The paper argues that a link can do both if it learns the jammer's pattern and chooses the frequency and spreading factor together. It proposes splitting the joint action space into two parallel deep reinforcement learning networks, one for frequency and one for spreading factor, coupled by rewards that both depend on communication success. In simulations against a moving UAV jammer, the scheme is reported to raise normalized throughput by nearly 90% over a standard DQN and a hybrid baseline.

What carries the argument

The parallelized network architecture: a convolutional neural network processes the spectrum waterfall, then two fully connected networks output Q-values for frequency and for spreading factor, shrinking the action space from $M\times N$ to $M+N$. The coupling is the indicator function $\delta(f,q)$: both heads' rewards use the same success signal, so each head learns about the other's choice through the joint reward. Exploration is replaced by the applying network's parameter randomness at initialization, with soft updates $\theta^- = \tau\theta + (1-\tau)\theta^-$ for stability.

What would settle it

Run a joint-action DQN with the same CNN and replay memory in the same moving-reactive-jammer scenario; if its normalized throughput approaches the reported 90% gain, the claimed advantage of parallel decomposition is not established. Also, an ablation that feeds both heads the same reward should show whether the separate reward design is the source of coordination.

Watch

Extended reading notes

Core claim

The central claim is that the joint decision of frequency and spreading factor need not be learned over all $M\times N$ combinations. A CNN reads the spectrum waterfall and two fully connected heads each output Q-values for their own dimension; the frequency head is rewarded with an indicator $\delta(f,q)$ that fires only when the joint choice achieves the demodulation threshold, while the spreading-factor head receives a rate term minus a bandwidth penalty, and both are updated with soft target parameters. The authors report that this parallelized, exploration-free scheme converges faster and reaches about 90% higher normalized throughput than DQN and DQN-control baselines, and remains effective under shadow fading, larger spreading-factor sets, and constant or random jamming.

Load-bearing premise

The paper assumes that choosing the best frequency and the best spreading factor separately, each with its own learned Q-function and reward, performs as well as choosing the best joint combination.

Editorial extensions

If this is right

  • Links can maintain communication without a model of the jammer's movement, since the agent learns online from spectrum waterfalls.
  • The decomposition lowers per-step decision cost from $M\times N$ to $M+N$ evaluations, which the paper estimates at about 3.3 billion FLOPs and roughly 0.00028 seconds per iteration.
  • Removing the $\varepsilon$-greedy mechanism removes a hyperparameter to tune, and soft updates appear sufficient to keep training stable.
  • The method remains effective with spreading factors up to 256, under shadow fading, and under constant and random jamming, suggesting the hiding-plus-evasion strategy is not tied to one propagation model.

Reading between the lines

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

  • Because separate maximization has no theoretical guarantee of reaching the joint optimum, a useful extension would be a bound relating the two independent Q-functions to the joint Q-function when the reward is not additively separable in frequency and spreading factor.
  • An ablation that gives both heads the identical scalar reward would isolate whether the performance gain comes from the network split or from the reward split.
  • A stronger comparison baseline would be a well-tuned joint-action DQN with prioritized replay or dueling architecture, since the reported 90% gain is measured against baselines that may be handicapped by the large joint action space.
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

4 major / 4 minor

Summary. The paper considers anti-jamming communication under a moving reactive jammer that launches tracking jamming upon detecting the user signal and indiscriminate sweeping jamming otherwise. The authors propose a deep reinforcement learning (DRL) scheme that decomposes the joint action space of frequency and spreading factor into two parallel Q-networks sharing a common CNN encoder, replaces epsilon-greedy exploration with the inherent randomness of the training network, and uses soft target updates. The reward functions for the two networks are coupled through the communication success indicator and throughput. Simulations compare the proposed algorithm against a DQN baseline, a DQN-based control baseline, and a version of the proposed algorithm with epsilon-greedy, reporting a roughly 90% improvement in normalized throughput. The paper also studies the effect of learning rate, larger action spaces, shadow fading, retrospective horizon, and jamming scenarios, and provides a FLOPs estimate.

Significance. The problem of jointly selecting frequency and spreading factor against a moving reactive jammer is well motivated, and the paper makes a concrete algorithmic proposal: parallel Q-networks with an interconnected reward structure. If the approach is valid, it offers a practical way to reduce the curse of dimensionality in anti-jamming DRL. The paper is commendable for describing the simulation setup in detail, reporting runtime and FLOPs, and testing several scenario variations. However, the central methodological justification is currently unsupported: the independent Q-decomposition is not proved and is in fact questionable, and the purported MDP proof contains algebraic invalidities. These issues affect the credibility of the claimed 90% throughput gain, which is the main contribution.

major comments (4)
  1. [Section III-A and III-B, Eqs. (11)-(19)] The joint action-value function for the pair (f,q) is replaced by two independent Q-functions, one for frequency and one for spreading factor, each trained with its own reward and separately maximized. The rewards in Eqs. (11) and (13) are 'interconnected' in that both depend on the other action through delta(f,q) and C_t. In the target definitions of Eqs. (16) and (18), the future term max_{f'} Q(S',f';theta-) and max_{q'} Q(S',q';theta-) ignores the other action entirely, while the one-step reward was generated using the action selected by the other network. Consequently, Q_f(S,f) is an expectation over the replay buffer's distribution of q, not the optimal joint value. Selecting f* = argmax_f Q_f and q* = argmax_q Q_q is not guaranteed to maximize the joint return; whenever the optimal f depends on q (e.g., a frequency is only safe when paired with a sufficiently large spreading factor), separate maximization can select incompatible actions. This is not a minor gap: the claimed 90% gain is attributed in the conclusion and Section IV-C to the parallel decomposition, but no coordination-free structure is demonstrated. The authors should either prove that the joint Q-function is separable in this setting or provide an empirical comparison showing that separate maximization does not degrade performance relative to joint maximization on the exact same environment.
  2. [Theorem 1, Eqs. (9)-(10)] The proof of the Markov property is not a valid derivation. In Eq. (9), step (b) equates Pr(S_{k+1} | S_k, a_k, ...) with Pr(o_k, o_{k-1}, ..., o_{k-H+1} | o_k, ..., o_0, a_k, ...) and then immediately writes Pr(o_{k+1} | o_k, ..., o_0, a_k, ...); this is not algebraically justified, and the intermediate product form is never established. In Eq. (10), the equalities again collapse the conditioning state to the observation window without defining the mapping from spectrum states to observations. Moreover, the theorem's condition 'if the set of actions A employed aligns with the set of optimal actions A*' is circular, since A* is not defined independently of the learning process. The MDP formulation is thus not proved; at best it is assumed. This weakens the theoretical basis for applying DRL, although the simulation results do not strictly depend on this proof.
  3. [Section IV-C, Fig. 6 and Eq. (24)] The headline claim of 'nearly 90% increase in normalized throughput' is not supported by statistical evidence. The figure reports a single learning curve per algorithm with no error bars, no number of random seeds, and no confidence intervals. DRL results are notoriously seed-sensitive (see [42] itself), so without multiple trials it is impossible to know whether the margin is significant or whether the baselines are merely undertuned. Additionally, Eq. (24) defines normalized throughput relative to C_Max, stated to be determined by exhaustive search, but the text does not specify whether C_Max is the maximum over the joint (f,q) space, over the frequency-only space, or over a per-time-step optimum; this must be clarified for the 90% figure to be interpretable.
  4. [Section III-B and Fig. 3] The claim that removing epsilon-greedy accelerates convergence by relying on 'the inherent randomness of the applying network' is not substantiated. Initially random network weights do provide some exploration, but as training proceeds the network output becomes deterministic, so the policy may commit to a suboptimal action before all (f,q) combinations have been adequately tried. The paper does not analyze the exploration coverage of this scheme, and the only evidence is the single learning curve in Fig. 6. A quantitative comparison with an epsilon-greedy version across multiple seeds (the paper does include such an algorithm, but no statistics) is necessary to support the exploration-free mechanism as a contribution.
minor comments (4)
  1. [Section II-A, Eq. (2)] The notation for the spectral sample s_{t-i\Delta t}^i has a redundant superscript; should be s_{t-i\Delta t}^j or similar to match the integration variable.
  2. [Section IV-C, Fig. 13 caption] The captions for Fig. 13 and Fig. 14 both say 'The performance in constant jamming'; Fig. 14 should indicate 'random jamming'.
  3. [Section II-C, Eq. (9)] The proof uses BOTH S_k and o_k without defining their relationship; the reader is left to guess how the spectrum waterfall relates to the underlying state. Please define S_t and o_t consistently.
  4. [Table II] In Table II, the 'Number of available frequencies' is set to 10 and 'Number of spreading factors' is set to 6, but the table says 'Reward Scale factor 0.2' for both eta and kappa; the text in Eq. (13) already uses kappa for the bandwidth penalty, so the table's column labeled 'Reward Scale factor' for the spreading factor network is ambiguous.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the 90% throughput gain is an empirical benchmark against an exhaustive-search upper bound, and self-citations are not load-bearing.

full rationale

The central quantitative claim (nearly 90% normalized-throughput improvement) is an empirical simulation result, not a derivation from the authors' own fitted quantities. Normalized throughput is benchmarked against C_Max, 'determined through exhaustive search or optimal parameter selection,' an external upper bound; the reward functions (11)-(13) are hand-designed objectives (success indicator and rate minus bandwidth penalty), not parameters fitted to force the reported gain. The baselines DQN and DQN-control are externally cited and independently specified. Self-citations [15], [19], [30], and [12] appear in related-work and comparison roles, but none is invoked as the sole evidence for the proposed algorithm's performance or as a uniqueness theorem. The Theorem 1 Markov proof is not a valid derivation, because it assumes the retrospective window H makes the state depend only on recent history and then concludes the Markov property; however, this is a proof-quality or correctness concern, not a reduction of the central throughput claim to its own inputs. No exhibit shows the 90% result or the action-decomposition benefit being equivalent to its inputs by construction.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central algorithm rests on several hand-set hyperparameters and unproved modeling assumptions. Most important are the decomposition of the Q-function into independent frequency and spreading factor heads, the finite-horizon Markovian approximation, and the assumption that initial network randomness suffices for exploration. No new physical entities are introduced.

free parameters (7)
  • Reward scale factor eta = 0.2
    Eq. (11) scales the frequency reward; chosen by hand in Table II, affects convergence.
  • Spreading factor bandwidth penalty kappa = 0.2
    Eq. (13) subtracts kappa/b_spread; chosen by hand in Table II, shapes spreading factor preference.
  • Soft update parameter tau = 0.2
    Eq. (20) mixes target network parameters; this is the mechanism that replaces epsilon-greedy, so its value is load-bearing.
  • Learning rate alpha = 0.001
    Table II; sensitivity is tested in Fig. 11 with values 0.002, 0.01, and 0.05.
  • Replay memory size D = 1000
    Table II; hand-set and small relative to typical DQN buffers, affecting sample diversity.
  • Update interval Nr = 16
    Table II; controls how often the applying network is refreshed.
  • Observation horizon H (waterfall length) = 200x200 input; tested at H=20-200
    Fig. 12 shows performance degrades for small H; H is a design parameter in the Markovian approximation.
assumptions (5)
  • domain assumption Finite observation horizon H makes the partially observed process Markovian.
    Theorem 1 (Section II-C) asserts this, but the proof in Eqs. (9)-(10) is invalid: it equates a transition probability to a joint observation probability, so the Markov property is not established.
  • ad hoc to paper The joint Q-function for (f,q) is separable into independent frequency and spreading factor Q-functions.
    Section III-A splits the action space and Eqs. (16)-(19) define separate targets without proof of equivalence to joint Q-learning.
  • domain assumption Jammer behavior follows a fixed detection threshold alpha_th with tracking jamming above it and indiscriminate sweeping below.
    Table I sets alpha_th=-70 dBm; the entire hiding and sweeping trade-off depends on this model.
  • ad hoc to paper Initial random network weights provide sufficient exploration in place of epsilon-greedy.
    Section III-B removes epsilon-greedy and relies on the inherent randomness of the applying network; no exploration guarantee is given.
  • domain assumption Channel propagation is adequately modeled by FSPL and FSPL plus log-normal shadowing.
    Eqs. (21)-(23); the throughput comparisons are computed under these models only.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Achieving Hiding and Smart Anti-Jamming Communication: A Parallel DRL Approach against Moving Reactive Jammer." pith.science (2026). https://pith.science/paper/C5PVFYZX

@misc{pith2026250202385,
  author       = {Pith},
  title        = {Pith review of: Achieving Hiding and Smart Anti-Jamming Communication: A Parallel DRL Approach against Moving Reactive Jammer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C5PVFYZX}},
  note         = {Machine review of arXiv:2502.02385}
}
abstract

This paper addresses the challenge of anti-jamming in moving reactive jamming scenarios. The moving reactive jammer initiates high-power tracking jamming upon detecting any transmission activity, and when unable to detect a signal, resorts to indiscriminate jamming. This presents dual imperatives: maintaining hiding to avoid the jammer's detection and simultaneously evading indiscriminate jamming. Spread spectrum techniques effectively reduce transmitting power to elude detection but fall short in countering indiscriminate jamming. Conversely, changing communication frequencies can help evade indiscriminate jamming but makes the transmission vulnerable to tracking jamming without spread spectrum techniques to remain hidden. Current methodologies struggle with the complexity of simultaneously optimizing these two requirements due to the expansive joint action spaces and the dynamics of moving reactive jammers. To address these challenges, we propose a parallelized deep reinforcement learning (DRL) strategy. The approach includes a parallelized network architecture designed to decompose the action space. A parallel exploration-exploitation selection mechanism replaces the $\varepsilon $-greedy mechanism, accelerating convergence. Simulations demonstrate a nearly 90\% increase in normalized throughput.

Figures

Figures reproduced from arXiv: 2502.02385 by the authors.

Figure 1
Figure 1. System model. mobile unmanned aerial vehicle (UAV) jammer. The moving reactive jammer unleashes high-power tracking jamming once it detects activity from the transmitter. If the jammer fails to capture the user’s signal, it proceeds to implement indiscrim￾inate jamming, such as sweeping or comb jamming. To be specific, for the transmitter, the available frequency range is defined as [fL, fU ], where fL and fU repres… view at source ↗
Figure 2
Figure 2. The illustration of the algorithm steps. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the designed mechanism and the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: The simultaneously update of our design. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The locations of transceivers and patrol route of jammer [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The performance of different algorithms. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: The Comparison of spectrum waterfalls during algorithm training and [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: The performance under larger spread factor set. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: The performance under shadow fading channel. [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: The performance comparison of different learning rates. [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: The performance in constant jamming.          " " ! [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 13
Figure 13. Figure 13: The performance in constant jamming D. Time Complexity Based on the parameters in Table II, the computational requirements of the deep networks in the simulation can be roughly estimated. The FLOPs (Floating Point Operations          #!# " [PITH_FULL…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 48 canonical work pages

  1. [30]

    Advancing secretly by an unknown path: A reinforce- ment learning-based hidden strategy for combating intelligent reactive jammer,

    W. Li et al. , “Advancing secretly by an unknown path: A reinforce- ment learning-based hidden strategy for combating intelligent reactive jammer,” IEEE Wireless Commun. Lett. , vol. 11, no. 7, pp. 1320–1324, 2022

    work page 2022

  2. [12]

    Fight against intelligent reactive jammer in mec networks: A hierarchical reinforcement learning-based hybrid hidden strategy,

    S. Liu et al., “Fight against intelligent reactive jammer in mec networks: A hierarchical reinforcement learning-based hybrid hidden strategy,” IEEE Wireless Commun. Lett. , vol. 13, no. 4, pp. 1078–1082, 2024

    work page 2024

  3. [15]

    Dynamic spectrum anti-jamming in broadband commu- nications: A hierarchical deep reinforcement learning approach,

    Y . Li et al. , “Dynamic spectrum anti-jamming in broadband commu- nications: A hierarchical deep reinforcement learning approach,” IEEE Wireless Commun. Lett., vol. 9, no. 10, pp. 1616–1619, 2020

    work page 2020

  4. [42]

    Deep reinforcement learning that matters,

    P. Henderson et al., “Deep reinforcement learning that matters,” in Proc. AAAI Conf. Artif. Intell. , vol. 32, no. 1, 2018

    work page 2018

  5. [1]

    Jamming attacks and anti-jamming strategies in wireless networks: A comprehensive survey,

    H. Pirayesh and H. Zeng, “Jamming attacks and anti-jamming strategies in wireless networks: A comprehensive survey,” IEEE Commun. Surv. Tutorials, vol. 24, no. 2, pp. 767–809, 2022

    work page 2022

  6. [2]

    Uav anti-jamming video transmissions with qoe guaran- tee: A reinforcement learning-based approach,

    L. Xiao et al., “Uav anti-jamming video transmissions with qoe guaran- tee: A reinforcement learning-based approach,” IEEE Trans. Commun. , vol. 69, no. 9, pp. 5933–5947, 2021

    work page 2021

  7. [3]

    Jammer tracking based on efficient covari- ance matrix reconstruction with iterative spatial spectrum sampling,

    S. Mohammadzadeh et al., “Jammer tracking based on efficient covari- ance matrix reconstruction with iterative spatial spectrum sampling,” IEEE Trans. Aerosp. Electron. Syst. , vol. 59, no. 6, pp. 8681–8695, 2023

    work page 2023

  8. [4]

    Event-based wireless tracking control for a wheeled mobile robot against reactive jamming attacks,

    H. Xu et al., “Event-based wireless tracking control for a wheeled mobile robot against reactive jamming attacks,”IEEE Trans. Control Netw. Syst., vol. 10, no. 4, pp. 1925–1936, 2023

    work page 1925

Show all 49 references
  1. [5]

    Anti-jamming games in multi-channel cognitive radio networks,

    Y . Wu et al. , “Anti-jamming games in multi-channel cognitive radio networks,” IEEE J. Sel. Areas Commun. , vol. 30, no. 1, pp. 4–15, 2012

    2012

  2. [6]

    Using channel hopping to increase 802.11 resilience to jamming attacks,

    V . Navda et al., “Using channel hopping to increase 802.11 resilience to jamming attacks,” in IEEE INFOCOM 2007 - 26th IEEE International Conference on Computer Communications , 2007, pp. 2526–2530

    2007

  3. [7]

    Game theory-based anti-jamming strategies for frequency hopping wireless communications,

    Y . Gao et al., “Game theory-based anti-jamming strategies for frequency hopping wireless communications,” IEEE Trans. Wireless Commun. , vol. 17, no. 8, pp. 5314–5326, 2018

    2018

  4. [8]

    Anti-jamming transmission stackelberg game with observation errors,

    L. Xiao et al. , “Anti-jamming transmission stackelberg game with observation errors,” IEEE Commun. Lett. , vol. 19, no. 6, pp. 949–952, 2015

    2015

  5. [9]

    Game-theoretic learning anti-jamming approaches in wireless networks,

    L. Jia et al. , “Game-theoretic learning anti-jamming approaches in wireless networks,” IEEE Commun. Mag. , vol. 60, no. 5, pp. 60–66, 2022

    2022

  6. [10]

    Game theory and reinforcement learning for anti-jamming de- fense in wireless communications: Current research, challenges, and solutions,

    ——, “Game theory and reinforcement learning for anti-jamming de- fense in wireless communications: Current research, challenges, and solutions,” IEEE Commun. Surv. Tutor., pp. 1–1, 2024

    2024

  7. [11]

    Spread spectrum for covert communication in ultra- violet communication system,

    Y . Yang et al. , “Spread spectrum for covert communication in ultra- violet communication system,” Opt. Express, vol. 32, no. 15, pp. 25 981– 25 994, Jul 2024

    2024

  8. [13]

    Deep reinforcement learning: A survey,

    X. Wang et al., “Deep reinforcement learning: A survey,” IEEE Trans. Neural Networks Learn. Syst. , vol. 35, no. 4, pp. 5064–5078, 2024

    2024

  9. [14]

    Anti-jamming communications using spectrum waterfall: A deep reinforcement learning approach,

    X. Liu et al., “Anti-jamming communications using spectrum waterfall: A deep reinforcement learning approach,” IEEE Commun. Lett., vol. 22, no. 5, pp. 998–1001, 2018

    2018

  10. [16]

    Deep reinforcement learning based hopping strategy for wideband anti-jamming wireless communications,

    J. Qi et al. , “Deep reinforcement learning based hopping strategy for wideband anti-jamming wireless communications,” IEEE Trans. Veh. Technol., vol. 73, no. 3, pp. 3568–3579, 2024

    2024

  11. [17]

    Uav-aided anti-jamming maritime communications: A deep reinforcement learning approach,

    K. Liu et al. , “Uav-aided anti-jamming maritime communications: A deep reinforcement learning approach,” in 2021 13th Int. Conf. Wireless Commun. Signal Process. (WCSP) , 2021, pp. 1–6

    2021

  12. [18]

    Intelligent reflecting surface assisted anti-jamming communications: A fast reinforcement learning approach,

    H. Yang et al. , “Intelligent reflecting surface assisted anti-jamming communications: A fast reinforcement learning approach,” IEEE Trans. Wireless Commun., vol. 20, no. 3, pp. 1963–1974, 2021. 13

    1963

  13. [19]

    Dynamic spectrum anti-jamming access with fast conver- gence: A labeled deep reinforcement learning approach,

    Y . Li et al., “Dynamic spectrum anti-jamming access with fast conver- gence: A labeled deep reinforcement learning approach,” IEEE Trans. Inf. Forensics Secur., vol. 18, pp. 5447–5458, 2023

    2023

  14. [20]

    Understanding deep neural function ap- proximation in reinforcement learning via \epsilon-greedy exploration,

    F. Liu, L. Viano, and V . Cevher, “Understanding deep neural function ap- proximation in reinforcement learning via \epsilon-greedy exploration,” in Adv. Neural Inf. Process. Syst., S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35. Curran Associ...

    2022

  15. [21]

    Exploration in deep rein- forcement learning: A survey,

    P. Ladosz, L. Weng, M. Kim, and H. Oh, “Exploration in deep rein- forcement learning: A survey,” Inf. Fusion, vol. 85, pp. 1–22, 2022

    2022

  16. [22]

    A survey of convolutional neural networks: Analysis, applications, and prospects,

    Z. Li et al. , “A survey of convolutional neural networks: Analysis, applications, and prospects,” IEEE Trans. Neural Networks Learn. Syst., vol. 33, no. 12, pp. 6999–7019, 2022

    2022

  17. [23]

    Intelligent spectrum anti-jamming with cognitive software-defined architecture,

    Y . Huang et al. , “Intelligent spectrum anti-jamming with cognitive software-defined architecture,” IEEE Syst. J. , vol. 17, no. 2, pp. 2686– 2697, 2023

    2023

  18. [24]

    Game theory for anti-jamming strategy in mul- tichannel slow fading iot networks,

    A. Gouissem et al. , “Game theory for anti-jamming strategy in mul- tichannel slow fading iot networks,” IEEE Internet Things J. , vol. 8, no. 23, pp. 16 880–16 893, 2021

    2021

  19. [25]

    Spatial anti-jamming scheme for internet of satellites based on the deep reinforcement learning and stackelberg game,

    C. Han et al. , “Spatial anti-jamming scheme for internet of satellites based on the deep reinforcement learning and stackelberg game,” IEEE Trans. Veh. Technol., vol. 69, no. 5, pp. 5331–5342, 2020

    2020

  20. [26]

    Power and spreading factor control in low power wide area networks,

    B. Reynders et al. , “Power and spreading factor control in low power wide area networks,” in 2017 IEEE Int. Conf. Commun. (ICC) , 2017, pp. 1–6

    2017

  21. [27]

    Two-dimensional anti-jamming communication based on deep reinforcement learning,

    G. Han et al., “Two-dimensional anti-jamming communication based on deep reinforcement learning,” in 2017 IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP) , 2017, pp. 2087–2091

    2017

  22. [28]

    Rl-based frequency hopping with block-shifted patterns: Balancing between anti-jamming performance and synchro- nization overhead,

    J. Zhang and X. Wu, “Rl-based frequency hopping with block-shifted patterns: Balancing between anti-jamming performance and synchro- nization overhead,” IEEE Trans. Veh. Technol., vol. 73, no. 1, pp. 909– 922, 2024

    2024

  23. [29]

    Anti-jamming underwater transmission with mobility and learning,

    L. Xiao et al. , “Anti-jamming underwater transmission with mobility and learning,” IEEE Commun. Lett. , vol. 22, no. 3, pp. 542–545, 2018

    2018

  24. [31]

    Reinforcement learning-based dynamic anti-jamming power control in uav networks: An effective jamming signal strength based approach,

    N. Ma et al. , “Reinforcement learning-based dynamic anti-jamming power control in uav networks: An effective jamming signal strength based approach,” IEEE Commun. Lett. , vol. 26, no. 10, pp. 2355–2359, 2022

    2022

  25. [32]

    The k-means algorithm: A comprehensive survey and performance evaluation,

    M. Ahmed, R. Seraj, and S. M. S. Islam, “The k-means algorithm: A comprehensive survey and performance evaluation,” Electronics, vol. 9, no. 8, 2020

    2020

  26. [33]

    Secure transmission scheme based on joint radar and communication in mobile vehicular networks,

    Y . Yao et al. , “Secure transmission scheme based on joint radar and communication in mobile vehicular networks,” IEEE Trans. Intell. Transp. Syst., vol. 24, no. 9, pp. 10 027–10 037, 2023

    2023

  27. [34]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,” Bell Syst. Tech. J., vol. 27, no. 3, pp. 379–423, 1948

    1948

  28. [35]

    On empirical comparisons of optimizers for deep learning,

    D. Choi et al. , “On empirical comparisons of optimizers for deep learning,” arXiv preprint arXiv:1910.05446 , 2019

    1910 arXiv

  29. [36]

    t-soft update of target network for deep reinforcement learning,

    T. Kobayashi and W. E. L. Ilboudo, “t-soft update of target network for deep reinforcement learning,” Neural Networks , vol. 136, pp. 63–71, 2021

    2021

  30. [37]

    Path loss prediction model development in a mountainous forest environment,

    B. Myagmardulam et al., “Path loss prediction model development in a mountainous forest environment,” IEEE Open J. Commun. Soc. , vol. 2, pp. 2494–2501, 2021

    2021

  31. [38]

    Modelling and simulation of rayleigh fading, path loss, and shadowing fading for wireless mobile networks,

    Z. Ren, G. Wang, Q. Chen, and H. Li, “Modelling and simulation of rayleigh fading, path loss, and shadowing fading for wireless mobile networks,” Simul. Model. Pract. Theory , vol. 19, no. 2, pp. 626–637, 2011

    2011

  32. [39]

    Shadow fading correlation in high-speed railway envi- ronments,

    R. He et al. , “Shadow fading correlation in high-speed railway envi- ronments,” IEEE Trans. Veh. Technol. , vol. 64, no. 7, pp. 2762–2772, 2015

    2015

  33. [40]

    An additive model as a physical basis for shadow fading,

    J. Salo et al., “An additive model as a physical basis for shadow fading,” IEEE Trans. Veh. Technol., vol. 56, no. 1, pp. 13–26, 2007

    2007

  34. [41]

    Asynchronous contention resolution- aided aloha in lr-fhss networks,

    J. M. de Souza Sant’Ana et al. , “Asynchronous contention resolution- aided aloha in lr-fhss networks,” IEEE Internet Things J., vol. 11, no. 9, pp. 16 684–16 692, 2024

    2024

  35. [43]

    Performance analysis for dsss with short period sequences encountering broadband interference,

    Q. Lu et al., “Performance analysis for dsss with short period sequences encountering broadband interference,” IEEE Trans. Veh. Technol., pp. 1– 13, 2024

    2024

  36. [44]

    Experimental analysis of the joint statistical properties of azimuth spread, delay spread, and shadow fading,

    A. Algans et al., “Experimental analysis of the joint statistical properties of azimuth spread, delay spread, and shadow fading,” IEEE J. Sel. Areas Commun., vol. 20, no. 3, pp. 523–531, 2002

    2002

  37. [45]

    Convolutional neural networks at constrained time cost,

    K. He and J. Sun, “Convolutional neural networks at constrained time cost,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2015

    2015

  38. [46]

    Color image to grayscale image conversion,

    C. Saravanan, “Color image to grayscale image conversion,” in 2010 Second Int. Conf. Comput. Eng. Appl. , vol. 2, 2010, pp. 196–199

    2010

  39. [47]

    Goodfellow et al., Deep Learning

    I. Goodfellow et al., Deep Learning. MIT Press, 2016

    2016

  40. [48]

    Roofline performance analysis of dnn architectures on cpu and gpu systems,

    H. C. Prashanth and M. Rao, “Roofline performance analysis of dnn architectures on cpu and gpu systems,” in 2024 25th Int. Symp. Quality Electron. Design (ISQED) , 2024, pp. 1–8

    2024

  41. [49]

    Benchmark analysis of jetson tx2, jetson nano and raspberry pi using deep-cnn,

    A. A. Süzen, B. Duman, and B. ¸ Sen, “Benchmark analysis of jetson tx2, jetson nano and raspberry pi using deep-cnn,” in 2020 Int. Congr. Human-Comput. Interact., Optim. Robot. Appl. (HORA) , 2020, pp. 1–5

    2020

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.