Pith. sign in

REVIEW 4 major objections 8 minor 20 references

Graph Attention-based Decentralized Actor-Critic for Dual-Objective Control of Multi-UAV Swarms

T0 review · 4 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A decentralized actor-critic that gives coverage and battery lifetime their own critics, linked by a KL-divergence trust region, improves UAV swarm coverage and lifetime while cutting training time by half.

desk verdict GADC is a plausible MARL extension with a theory-algorithm gap around the KL trust region and experiments that need more runs and code. read the letter →

arxiv 2506.09195 v1 pith:BXJCIDIP submitted 2025-06-10 eess.SP cs.AIcs.MA

Haoran Peng , Ying-Jun Angela Zhang This is my paper
classification eess.SPcs.AIcs.MA
keywords multi-UAVswarmsgraphattentionnetworkmulti-agentreinforcementlearningactor-criticdual-objectiveoptimizationservicecoveragebatterylifetimeraytracingdigitaltwin
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

The paper proposes GADC, a decentralized multi-agent reinforcement learning method for swarms of UAVs acting as flying base stations. The method claims to improve both service coverage and network lifetime at once, without the hand-tuned weighting coefficients that conventional weighted-sum reward designs require. The authors report about 50% faster convergence than GAT-based MADDPG, and better coverage and lifetime than conventional MADDPG, in both an ideal simulator and NVIDIA Sionna's Munich ray-tracing digital twin. A theoretical bound, Lemma 1, is used to argue that the secondary lifetime objective does not degrade coverage beyond a controllable threshold.

What carries the argument

The core object is the actor-double-critic update: a primary coverage critic trained with DDPG-style temporal-difference error, followed by a secondary lifetime-aware critic whose update uses the clipped surrogate loss L_CLIP, with probability ratio F = pi_f(a|s)/pi_c(a|s) between the lifetime and coverage policies. Lemma 1 applies the Performance Difference Lemma and Pinsker's inequality to bound J_c(pi_f) - J_c(pi_c) by the expressions in Eqs. (39)-(40), assuming the KL-divergence constraint of Eq. (38) holds. This bound is what underwrites the claim that extending battery lifetime does not compromise coverage beyond a controllable threshold.

What would settle it

Run GADC's Algorithm 1 and record D_KL(pi_f || pi_c) at every secondary update; if max over states of the expected KL divergence exceeds delta at any update, the Lemma 1 bound does not apply to the trained policy, and the paper's coverage-degradation guarantee is not established. A direct test would compare measured coverage loss under the trained policy against the right-hand side of Eq. (39).

Watch

Extended reading notes

Core claim

GADC combines a two-hop graph attention network that aggregates neighboring UAV observations with an actor-double-critic architecture: a coverage-effectiveness critic updates the actor to maximize served users, then a lifetime-aware critic updates the same actor with a clipped PPO-style objective whose probability ratio relative to the coverage policy keeps the update near the coverage policy. The paper claims this structure converges in about 500 episodes versus over 700 for GAT-based MADDPG in the ideal setting, achieves full user coverage for 30 or more UAVs, and extends UAV lifetime toward the exhaustive-search optimum. The theoretical centerpiece is Lemma 1, which bounds the difference between the coverage return of the pure coverage policy and the combined dual-objective policy by a KL-divergence budget.

Load-bearing premise

The bound that lifetime updates preserve coverage assumes the trained lifetime policy actually stays within the stated KL-divergence distance delta of the coverage policy, yet the algorithm never computes or enforces that KL divergence; it only clips probability ratios with epsilon, so the guarantee may not hold for the policy actually trained.

Editorial extensions

If this is right

  • GADC removes the need to hand-tune a weighting coefficient between the two rewards; tuning epsilon instead gives a more linear and predictable coverage-lifetime trade-off.
  • The two-hop GAT aggregation lets UAVs base decisions on observed neighbor information, which the authors show is the key to outperforming conventional MADDPG in both open and urban ray-traced environments.
  • The reported 50% convergence speedup (500 vs. over 700 episodes in the ideal setting; about 400 vs. 600 in Sionna) follows from optimizing the two objectives independently rather than as a weighted sum.
  • The digital-twin experiments indicate that policies trained at small scale (5 UAVs, 40 users) can be scaled up to larger deployments (10 UAVs, 60 users) with averaged weights, suggesting the GAT architecture generalizes across swarm sizes.

Reading between the lines

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

  • Beyond the paper: the epsilon-clip mechanism is a looser proxy for the KL trust region stated in Eq. (38); directly enforcing that KL constraint with a penalty or projection would make the Lemma 1 bound applicable to the trained policy and could tighten the coverage guarantee.
  • Beyond the paper: the reported linear dependence of the trade-off on epsilon suggests epsilon could be scheduled or learned online to adapt the coverage-lifetime balance to changing user distributions, which the current fixed-epsilon experiments do not explore.
  • Beyond the paper: because digital-twin training used only 5 UAVs and testing used 10, the claimed scalability is evidence of transfer within a small regime, not proof of performance for dozens of UAVs in dense urban canyons.
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 / 8 minor

Summary. The paper studies dual-objective control of a multi-UAV swarm: primary service coverage and secondary network lifetime. The authors propose GADC, in which each UAV encodes local observations, aggregates neighbor information through a two-hop multi-head GAT plus a GRU memory unit, and trains an actor with two critics: a DDPG-style coverage critic and a PPO-style lifetime critic. Lemma 1 provides upper and lower bounds on the coverage loss when the KL divergence between the lifetime-updated policy and the coverage policy is at most delta. Experiments compare GADC with MADDPG, GAT-based MADDPG, and an approximate exhaustive search in an ideal simulator and in NVIDIA Sionna's ray-tracing environment, reporting faster convergence and improved coverage and lifetime.

Significance. If the claims held, the paper would offer a practical way to avoid weighted-sum reward tuning in multi-agent UAV control, with a concrete architectural combination (GAT observation aggregation plus dual critics) and a theoretical bound that connects the secondary lifetime objective to a controllable loss in the primary coverage objective. The proof is self-contained and follows the standard TRPO argument, and the Sionna digital-twin validation is a genuine effort toward realistic evaluation. However, the central theoretical guarantee is conditional on a KL constraint that Algorithm 1 does not enforce, the actor representation mixes incompatible DDPG and PPO assumptions, and the empirical evidence consists of single runs without error bars. These issues currently make the paper's main claims stronger than the evidence supports, though they appear addressable within the scope of a revision.

major comments (4)
  1. [§IV-D, Eqs. (38)-(40); Algorithm 1 lines 25-29] The coverage-loss bound in Lemma 1 is conditional on the KL constraint max_s E[D_KL(pi_f || pi_c)] <= delta in Eq. (38), but Algorithm 1 never computes or enforces D_KL. The secondary actor update in Eq. (16) clips the probability ratio F to [1-epsilon, 1+epsilon], which does not imply a statewise KL bound, especially after T_epi inner-loop gradient steps. The paper never specifies delta or connects it to the tuned value epsilon=0.2 in Section V. As a result, the precondition of Lemma 1 may be violated for the actually trained policy, and the claim that lifetime optimization does not compromise coverage beyond a controllable threshold is not established. Additionally, the bounds in Eqs. (39)-(40) contain the term E_{s~d_pi_c, a~pi_f}[A^c(s,a)], which is not controlled by delta, so the 'controllable threshold' wording would need further justification even under the KL assumption. This is a structural gap between the theory and the implementation, not merely a missing empirical detail.
  2. [§IV-D, Eqs. (12)-(16); §IV-A action space] The algorithm as written mixes incompatible policy representations. The action space is discrete (17 actions, Section IV-A), but the coverage update is a DDPG deterministic-policy gradient: Eq. (14) uses the chain rule through a differentiable action a=pi(s), and Eq. (13) evaluates Q_c(s_i, pi(s_i|theta)). The lifetime update in Eq. (16) instead uses a probability ratio F=pi_f(a|s)/pi_c(a|s), which requires a stochastic policy. The paper never specifies how a single actor network supplies both a differentiable action for the Q-gradient and a probability distribution for the ratio, nor how sampling from the discrete action space is handled. As written, Algorithm 1 cannot be executed as described, and the KL divergence in Lemma 1 is undefined if the actor is deterministic.
  3. [§V-A, Figs. 8-11; §V-B, Fig. 14] The empirical claims rest on single training runs with no multiple seeds, confidence intervals, or error bars. The Contributions state that GADC improves convergence speed by 50%, but no formal convergence criterion is defined, and the displayed curves do not clearly support that number: Fig. 8 shows convergence at 500 versus 'more than 700' episodes, which is not 50% faster, while Fig. 14 shows roughly 400 versus 600 episodes in a different setting. Furthermore, epsilon=0.2 is selected post hoc from the tradeoff curves in Figs. 10 and 11(b), so the claim that GADC avoids careful tradeoff tuning is overstated; epsilon remains a tuned hyperparameter, even if its effect is more linear than that of the weighting coefficient phi.
  4. [§V-A, Fig. 9 and Eq. (41)] The 'exhaustive search (ES)' baseline is implemented as 10^7 random action selections per time step, which is not an exhaustive search and has no optimality guarantee. Describing ES as approaching the optimal solution in Fig. 9 is therefore unsupported. The paper also does not report the coefficient phi used for the MADDPG baseline in Fig. 9, so it is unclear whether the weighted-sum baselines were tuned comparably to the epsilon used for GADC.
minor comments (8)
  1. [Algorithm 1 line 17; §IV-A] The rewards r^c_n(t) and r^f_n(t) are written with an agent index, but Section IV-A defines them as global quantities (sum over UAVs and minimum residual energy). Please clarify whether each agent receives the global reward or a local proxy, since this affects the decentralized claim.
  2. [Eq. (16)] The text says the actor is updated by minimizing the clipped objective, but the displayed update is theta <- theta + sigma * grad L_CLIP, which is gradient ascent. Please correct the sign or the wording.
  3. [Eq. (5)] The softmax in Eq. (5) is normalized over l in V, while the text says attention is computed over neighbors I_n. If V is the set of all UAVs, the denominator requires global information and contradicts the decentralized setting; please define the normalization set explicitly.
  4. [Eq. (19)] The statement of the Performance Difference Lemma is per-state and omits the expectation over the state distribution; as written it can be misread. Please state the lemma with the usual discounted state-distribution expectation.
  5. [Proof of Lemma 1, after Eq. (36)] The sentence 'the expected value of the TV distance under pi_c and pi_c' should read 'under pi_f and pi_c'.
  6. [§V-B] The training scale is given as five UAVs and 40 UTs near the start of the section, but later the text says '10 UAVs and 40 UTs' when comparing convergence with Fig. 8; please reconcile the numbers.
  7. [§V, Table III] Many RL hyperparameters are not reported (learning rate, batch size, replay buffer size, gamma, T_epi, epsilon schedule, network dimensions), which limits reproducibility.
  8. [Table I and Eq. (1)] The symbol D_s is listed as the maximum communication distance but is used in Eq. (1) as the distance at which the received power equals sigma_s; please define it consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity; central theory is a standard conditional bound and empirical claims are benchmarked externally.

full rationale

The paper's central theoretical result, Lemma 1 (Eqs. (20)-(21) and (39)-(40)), is a conventional performance-difference bound. It invokes Kakade's Performance Difference Lemma, applies Holder's inequality and Pinsker's inequality, and conditions the final coverage-loss bound on the explicit KL-divergence premise in Eq. (38), max E[D_KL(pi_f || pi_c)] <= delta. This is a conditional theorem, not a conclusion defined in terms of the algorithm's outputs: no parameter in the proof is fitted to the experimental results, and the bound is not equivalent to the algorithm's objective by construction. Algorithm 1's clipped-ratio PPO update (Eq. (16)) does not in fact compute or enforce Eq. (38), and delta is never set or verified, so the theoretical guarantee is not automatically transferred to the trained policy. That is a real theory-algorithm gap that weakens the support for the 'controllable threshold' claim, but it is a correctness/support issue rather than circularity: there is no equation in which the derived result reduces to its input. The empirical evaluation is self-contained against external benchmarks: GADC is compared with weighted-sum MADDPG, GAT-based MADDPG, and exhaustive search in an ideal simulator and in NVIDIA Sionna ray tracing, so the convergence and performance claims are not manufactured by construction. The only self-citation, reference [8] by the first author, appears in a background sentence on UAV connectivity and is peripheral; it does not carry GADC's architectural or performance claims. No load-bearing self-definition, fitted-input-as-prediction, or renamed-known-result step is present.

Assumptions & free parameters 1 free parameters · 7 assumptions · 0 invented entities

The central result rests on standard RL and probability inequalities, one domain assumption about the MDP, energy, and channel model, and one ad hoc assumption that the KL constraint in the theorem is satisfied by the algorithm, which is not enforced. There is one tuned hyperparameter, epsilon=0.2, selected from the same experiments used to report performance. No new physical entities, forces, or mediators are introduced.

free parameters (1)
  • epsilon (PPO clipping range) = 0.2
    Controls how far the lifetime-aware policy can deviate from the coverage policy. It is chosen post hoc after inspecting the tradeoff curves in Figs. 10 and 11, with the stated justification that 0.2 balances lifetime while preserving coverage. It is a tuned hyperparameter, not a predicted constant.
assumptions (7)
  • domain assumption The multi-UAV system is modeled as a Markov Decision Process where local observations plus graph updates form a sufficient state representation.
    Section IV.A defines the MDP and the per-UAV state observation; the sufficiency of local observations for the dual objectives is assumed without proof.
  • standard math The Performance Difference Lemma (Kakade et al.) holds for the discounted coverage objective J^c.
    Invoked in the proof before Eq. (22) and cited as reference [20].
  • standard math The total-variation distance between discounted state distributions is bounded by gamma/(1-gamma) times the expected per-state TV distance between policies (Eq. 33).
    Standard trust-region bound used to derive Eq. (34).
  • standard math Pinsker's inequality bounds the per-state TV distance by the square root of half the KL divergence (Eq. 37).
    Standard inequality; used to express the bound in terms of delta.
  • ad hoc to paper The KL constraint max E[D_KL(pi_f || pi_c)] <= delta holds during the inner-loop updates (Eq. 38).
    The algorithm uses PPO clipping, not KL projection or a KL penalty. The constraint is assumed in the theorem but never verified in Algorithm 1. This is the load-bearing gap between theory and practice.
  • domain assumption The energy model in Eq. (3) captures all relevant UAV battery consumption.
    Section III.B defines navigation, service, communication, and hovering energy terms; the lifetime reward min_n b_n(t) depends on this model.
  • domain assumption Channel thresholds sigma_o and sigma_s translate to fixed observation and service radii R_o and R_s in the ideal simulator.
    Section III.A and Eq. (1) define coverage boundaries used in the Python/Matlab simulator.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph Attention-based Decentralized Actor-Critic for Dual-Objective Control of Multi-UAV Swarms." pith.science (2026). https://pith.science/paper/BXJCIDIP

@misc{pith2026250609195,
  author       = {Pith},
  title        = {Pith review of: Graph Attention-based Decentralized Actor-Critic for Dual-Objective Control of Multi-UAV Swarms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BXJCIDIP}},
  note         = {Machine review of arXiv:2506.09195}
}
read the original abstract

This research focuses on optimizing multi-UAV systems with dual objectives: maximizing service coverage as the primary goal while extending battery lifetime as the secondary objective. We propose a Graph Attention-based Decentralized Actor-Critic (GADC) to optimize the dual objectives. The proposed approach leverages a graph attention network to process UAVs' limited local observation and reduce the dimension of the environment states. Subsequently, an actor-double-critic network is developed to manage dual policies for joint objective optimization. The proposed GADC uses a Kullback-Leibler (KL) divergence factor to balance the tradeoff between coverage performance and battery lifetime in the multi-UAV system. We assess the scalability and efficiency of GADC through comprehensive benchmarking against state-of-the-art methods, considering both theory and experimental aspects. Extensive testing in both ideal settings and NVIDIA Sionna's realistic ray tracing environment demonstrates GADC's superior performance.

Figures

Figures reproduced from arXiv: 2506.09195 by the authors.

Figure 1
Figure 1. A multi-UAV system serves ground UTs. information across the multi-UAV network, complicating the development of a collaborative navigation strategy among all UAVs. Recent studies have employed stacked graph layers to enhance UAV exploration in unobserved environments through message passing. This approach enables dynamic interaction between each UAV’s actions and their neighboring UAVs’ local observations in subsequ… view at source ↗
Figure 2
Figure 2. The information aggregation through message passing. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The illustration of observation and service coverage of the multi-UAV system. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: The observation, service, and connectivity distance. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: The architecture of the proposed GADC-based RL. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: The training processing of the proposed GADC-based decentralized actor-critic agent. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: The illustration of the message passing between UAVs. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 6
Figure 6. Figure 6: In the Actor-Critic networks, each agent’s environment [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 8
Figure 8. Figure 8: The convergence performance of different algorithms [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: The impact of the number of UAVs on different objectives. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: The impact of 𝜖 on the proposed GADC. speed and cumulative reward can be attributed to the GAT and memory units enriching the UAVs’ information [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: The trade-off performance comparison between the GAT-based MADDPG and the proposed GADC. [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: The deployment of UAVs when UTs are in an open area [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: The deployment of UAVs at the final step when in an urban scenario [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 14
Figure 14. Figure 14: The convergence performance of different algorithms [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 18 canonical work pages

  1. [1]

    Multi-UA V navigation for partially observable commu- nication coverage by graph reinforcement learning,

    Z. Ye, et al., “Multi-UA V navigation for partially observable commu- nication coverage by graph reinforcement learning,”IEEE Trans. Mob. Comput., vol. 22, no. 7, pp. 4056–4069, Jan. 2023

    work page 2023

  2. [2]

    Online trajectory and resource optimization for stochastic UA V-enabled MEC systems,

    Z. Yang, et al., “Online trajectory and resource optimization for stochastic UA V-enabled MEC systems,”IEEE Trans. Wireless Commun., vol. 21, no. 7, pp. 5629–5643, Jul. 2022

    work page 2022

  3. [3]

    Fair and energy-efficient coverage optimization for UA V placement problem in the cellular network,

    Y . Liu, et al., “Fair and energy-efficient coverage optimization for UA V placement problem in the cellular network,”IEEE Trans. Commun., vol. 70, no. 6, pp. 4222–4235, Jun. 2022

    work page 2022

  4. [4]

    Distributed algorithms for maximum lifetime routing in wireless sensor networks,

    R. Madan, et al., “Distributed algorithms for maximum lifetime routing in wireless sensor networks,”IEEE Trans. Wireless Commun., vol. 5, no. 8, pp. 2185–2193, Aug. 2006

    work page 2006

  5. [5]

    Cooperative trajectory design of multiple UA V base stations with heterogeneous graph neural networks,

    X. Zhang, et al., “Cooperative trajectory design of multiple UA V base stations with heterogeneous graph neural networks,”IEEE Trans. Wireless Commun., vol. 22, no. 3, pp. 1495–1509, Mar. 2023

    work page 2023

  6. [6]

    Trust region policy optimization,

    J. Schulman, et al., “Trust region policy optimization,” inProc. the 32nd Int. Conf. Mach. Learn. (ICML), vol. 37, Lille, France, Jul. 2015, pp. 1889–1897

    work page 2015

  7. [7]

    Proximal policy optimization algorithms,

    ——, “Proximal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    arXiv 2017

  8. [8]

    Energy harvesting reconfigurable intelligent surface for UA V based on robust deep reinforcement learning,

    H. Peng, et al., “Energy harvesting reconfigurable intelligent surface for UA V based on robust deep reinforcement learning,”IEEE Trans. Wireless Commun., vol. 22, no. 10, pp. 6826–6838, Oct. 2023

    2023

Show all 20 references
  1. [9]

    Deep reinforcement learning-based online resource management for UA V-assisted edge computing with dual connectivity,

    L. T. Hoang, et al., “Deep reinforcement learning-based online resource management for UA V-assisted edge computing with dual connectivity,” IEEE/ACM Trans. Netw., vol. 31, no. 6, pp. 2761–2776, Dec. 2023

    2023

  2. [10]

    Deep reinforcement learning-based mining task offloading scheme for intelligent connected vehicles in UA V-aided MEC,

    C. Li, et al., “Deep reinforcement learning-based mining task offloading scheme for intelligent connected vehicles in UA V-aided MEC,”ACM Trans. Des. Autom. Electron. Syst., vol. 29, no. 3, pp. 1–29, May 2024. 14

    2024

  3. [11]

    Dynamic offloading and trajectory control for UA V- enabled mobile edge computing system with energy harvesting devices,

    Z. Yang, et al., “Dynamic offloading and trajectory control for UA V- enabled mobile edge computing system with energy harvesting devices,” IEEE Trans. Wireless Commun., vol. 21, no. 12, pp. 10 515–10 528, Dec. 2022

    2022

  4. [12]

    Resource allocation for NOMA-enabled cognitive satel- lite–UA V–terrestrial networks with imperfect CSI,

    R. Liu, et al., “Resource allocation for NOMA-enabled cognitive satel- lite–UA V–terrestrial networks with imperfect CSI,”IEEE Trans. Cog. Commun. Netw., vol. 9, no. 4, pp. 963–976, Aug. 2023

    2023

  5. [13]

    Space-air-ground integrated networks: Spherical stochas- tic geometry-based uplink connectivity analysis,

    Y . Liu, et al., “Space-air-ground integrated networks: Spherical stochas- tic geometry-based uplink connectivity analysis,”IEEE J. Sel. Areas Commun., vol. 42, no. 5, pp. 1387–1402, May 2024

    2024

  6. [14]

    UA V-assisted secure uplink communications in satellite- supported IoT: Secrecy fairness approach,

    Z. Yin, et al., “UA V-assisted secure uplink communications in satellite- supported IoT: Secrecy fairness approach,”IEEE Internet Things J., vol. 11, no. 4, pp. 6904–6915, Feb. 2024

    2024

  7. [15]

    Intelligent UA V swarm cooperation for multiple targets tracking,

    L. Zhou, et al., “Intelligent UA V swarm cooperation for multiple targets tracking,”IEEE Internet Things J., vol. 9, no. 1, pp. 743–754, Jan. 2022

    2022

  8. [16]

    Efficient and secured swarm pattern multi-UA V com- munication,

    G. Raja, et al., “Efficient and secured swarm pattern multi-UA V com- munication,”IEEE Trans. Veh. Technol., vol. 70, no. 7, pp. 7050–7058, Jul. 2021

    2021

  9. [17]

    Energy-efficient UA V control for effective and fair communication coverage: A deep reinforcement learning approach,

    C. H. Liu, et al., “Energy-efficient UA V control for effective and fair communication coverage: A deep reinforcement learning approach,” IEEE J. Sel. Areas Commun., vol. 36, no. 9, pp. 2059–2070, Sep. 2018

    2018

  10. [18]

    Multi-agent reinforcement learning with policy clipping and average evaluation for UA V-assisted communication markov game,

    Z. Feng, et al., “Multi-agent reinforcement learning with policy clipping and average evaluation for UA V-assisted communication markov game,” IEEE Trans. Intell. Transp. Syst., vol. 24, no. 12, pp. 14 281–14 293, Dec. 2023

    2023

  11. [19]

    UA V swarm cooperative target search: A multi-agent reinforcement learning approach,

    Y . Hou, et al., “UA V swarm cooperative target search: A multi-agent reinforcement learning approach,”IEEE Trans. Intell. Veh., vol. 9, no. 1, pp. 568–578, Jan. 2024

    2024

  12. [20]

    Approximately optimal approximate reinforcement learning,

    S. Kakade, et al., “Approximately optimal approximate reinforcement learning,” inProc. Int. Conf. Mach. Learn., San Francisco, CA, Jul. 2002, pp. 267–274

    2002

Pith tools

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