Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

PAGNet: Pluggable Adaptive Generative Networks for Information Completion in Multi-Agent Communication

T0 review · 4 major / 6 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 PAGNet claims that learning information-level communication weights and generating global-state representations from weighted local observations improves cooperative multi-agent reinforcement learning while decoupling communication…

desk verdict New architecture for MARL communication with credible experiments, but the headline 'decoupled training' claim is contradicted by its own update rule. read the letter →

arxiv 2502.03845 v1 pith:5FOA4SN5 submitted 2025-02-06 cs.MA

Zhuohui Zhang , Bin Cheng , Zhipeng Wang , Yanmin Zhou , Gang Li , Ping Lu , Bin He , Jie Chen This is my paper · ORCID
classification cs.MA
keywords multi-agentreinforcementlearningcommunicationgenerativeadversarialnetworkpluggablearchitectureinformation-levelweightsglobalstategenerationvaluedecompositionpartialobservability
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 argues that cooperative multi-agent reinforcement learning (MARL) is held back by communication models that specify who talks to whom but not which pieces of information matter, and by the need to train the communication module through the same sparse task rewards as the policy. PAGNet instead learns an information-level weight for each agent's observation, uses those weights to filter what each agent receives, and feeds the filtered observations through a generative network that reconstructs a global-state representation. That generated state replaces the true global state in the value-mixing network, while the communication and generation losses are computed separately from the reinforcement-learning reward. The authors report faster convergence and higher final returns than communication baselines on level-based foraging, the sparse Hallway coordination task, and eight StarCraft micro-management maps, with the gap largest on the hard maps.

What carries the argument

The load-bearing object is the information-level weight $W_t$, a per-agent, per-observation-dimension mask in $[0,1]$ computed by scaled dot-product attention over local observations with positional encodings. It is applied through the noise-mixing equation $x = (1-W)\cdot M + W\cdot\varepsilon$, which simultaneously implements timing, target, and content selection. The second piece is the adaptive generative network: a U-Net information-completion generator paired with a 1D-convolutional GAN discriminator, trained by a weighted MSE term that keeps generated states consistent with local observations plus a GAN term that makes them look real to the discriminator. These two modules together output the generated global state $\hat{s}_t$ that enters the mixing network, while a Transformer-based decoder turns the weighted information into individual Q-values. This machinery carries the argument because the three losses, communication MSE, GAN, and TD, touch disjoint parameter groups, which is what makes the communication module pluggable.

What would settle it

Replace the Gaussian noise $\varepsilon$ in Eq. (3) with zeros or with a learned token while keeping everything else fixed: if performance does not drop, the noise-mixing content model is not what drives PAGNet's gains. Also run the full architecture with the true global state $s_t$ substituted for the generated $\hat{s}_t$ in the mixing network: if the generated-state version does not approach the true-state version's win rate, the information-completion claim is not carrying the result.

Watch

Extended reading notes

Core claim

On its own terms, PAGNet's central discovery is that information-level communication can be reduced to a learned weight vector $W_t \in [0,1]^{n\times l}$ over the received observations, computed by scaled dot-product attention, and that this weight can do double duty: it selects what each agent should attend to, and it degrades irrelevant messages by mixing them with Gaussian noise. The resulting weighted information $x_i^t = (1 - W_t^{i,-i}) M_t^{i,-i} + W_t^{i,-i} \varepsilon_t$ is passed through a U-Net generator trained adversarially with a global discriminator, producing $\hat{s}_t = G(o_t, W_t)$. The paper's claim is that $\hat{s}_t$ is a faithful surrogate for the true global state in the TD target, so centralized training no longer needs access to $s_t$, and the same $W_t$ can be reused at execution time. The decoupling claim is that the generator's MSE-plus-GAN loss updates the communication module without waiting for sparse rewards, which is why the pretrained variant converges faster.

Load-bearing premise

Everything rests on the belief that mixing received observations with Gaussian noise according to $x = (1-W)M + W\varepsilon$ is a faithful model of what an agent should receive; if that noise-mixing equation is the wrong model, the generated global states, the mixing network inputs, and the policies built on them are all corrupted.

Editorial extensions

If this is right

  • MARL training can proceed without a hand-defined global state, since the trained generator supplies $\hat{s}_t$ from local observations alone.
  • Communication learning no longer waits for sparse task rewards; with an offline dataset, the weight network and generator can be pretrained and then attached to a new policy.
  • The learned weights give a per-timestep, per-information audit of what each agent actually uses from others, which the paper shows rises when enemies are first spotted and when allies fall.
  • The same architecture extends to reduced observation ranges: the paper shows that shrinking agents' sight range from 9 to 1 slows convergence but leaves the converged win rate largely intact.

Reading between the lines

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

  • If the Eq. (3) noise term is doing real work, an ablation replacing $\varepsilon$ with a zero vector or a learned constant should hurt; the paper does not report such an ablation, so this is the first test I would run.
  • Because the generator is trained to fill in out-of-view information, the same pluggable module could be applied to sensor-dropout or missing-observation settings, where the 'information completion' label is literal rather than metaphorical.
  • The per-dimension nature of $W$ suggests a direct path to bandwidth-constrained communication: threshold $W$ at execution time and transmit only the entries above the threshold, turning continuous weights into a sparse protocol.
  • The claim that generated states outperform raw messages in the mixing network implies a comparison against feeding true global states into the same architecture; the gap between generated and true states would quantify how much information completion still loses.
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 / 6 minor

Summary. The paper proposes PAGNet, a pluggable generative communication module for cooperative multi-agent reinforcement learning. PAGNet learns information-level communication weights (Eqs. 3–5), generates global-state representations from weighted local observations using a U-Net generator and a GAN discriminator (Eqs. 6–8), and integrates these representations and weights into a QMIX-style value decomposition through Eq. (9). The architecture is evaluated on Level-Based Foraging, Hallway, and eight SMAC maps, with comparisons to QMIX, DGN, G2ANet, and MASIA, plus ablations PAGNet FC and PAGNet PT. The paper claims superior performance, faster convergence, improved scalability, and that the communication module can be trained without sparse reinforcement-learning rewards.

Significance. If the empirical claims hold, PAGNet would be a useful contribution to MARL with communication: the information-level weighting and generative completion of global states are plausible mechanisms for improving coordination under partial observability, and the interpretability analysis of learned weights and generated states is a valuable qualitative addition. The paper provides clear architecture tables, a detailed hyperparameter description, and ablations for the weight network and pretraining. However, the significance is currently limited by three factors: the central 'decoupled training' claim is internally inconsistent with the stated update rule; the noise-mixing communication model in Eq. (3) is unvalidated; and no code, data, or numeric results are provided, so the reported advantages cannot be verified or reproduced from the manuscript alone.

major comments (4)
  1. [§IV-C, Algorithm 1] The headline claim that the communication module is decoupled from sparse RL rewards (Sec. I, contribution 3, and Sec. IV-B) is not supported by the training procedure as written. Equation (9) places W_t and G(o_t, W_t) inside Qtot, and Algorithm 1 line 12 updates θ, φ, and ψ after computing the RL loss. Without an explicit stop-gradient, detached copy, or separate optimizer groups, the information-level weight network and the adaptive generative network receive gradients from the TD error. Please specify the exact gradient flow (e.g., stop-gradient on W_t and on the generator output in Eq. (9), or a parameter-group separation in the optimizer), or soften the decoupling claim. This is load-bearing because contribution 3 is used to justify the pluggable and efficiency advantages of PAGNet.
  2. [§IV-A, Eq. (3)] The noise-mixing communication model x_t^i = (1 - W_t^{i,-i}) · M_t^{i,-i} + W_t^{i,-i} · ε_t, with ε_t ~ N(0,I), is introduced without theoretical justification and is not isolated by any ablation. Since this weighted information feeds both the adaptive generative network and the Transformer-based decoder (and hence Qtot), the validity of the whole architecture depends on this model being correct or at least harmless. Please add an ablation comparing Eq. (3) with alternatives such as raw weighted summation and a learned gate without Gaussian corruption, and provide a rationale (information-theoretic or empirical) for the noise term.
  3. [§V-F-b, §V-B] The experimental evidence is not reproducible from the manuscript. No code, data, or numeric result tables are provided; all results are shown as learning curves with 95% confidence intervals. In addition, hyperparameters are selected by maximizing the average evaluation over three seeds (Sec. V-F-b), which risks overfitting to those seeds, and the manuscript does not report per-seed variance or evaluation on a held-out seed. Please release code and checkpoints, include numeric tables of final performance with standard errors, and clarify how the hyperparameter selection affects the reported curves.
  4. [§IV-B, Eq. (6)] The MSE loss in Eq. (6) uses the observation function O in the expression O(G(o_t, W_t)), but O was defined in Sec. III-A as a mapping from global states and agent indices to observations. As written, it is ambiguous which agent's observations are being reconstructed and how the generated global state is mapped back to local observations. This matters because Eq. (6) is one of only two losses that train the weight and generator networks. Please define the reconstruction operator precisely.
minor comments (6)
  1. [§V-A-c] "NQD" should be "NDQ" to match the method name introduced in [16].
  2. [Fig. 8 caption] The caption contains the typo "sigh-trange"; it should read "sight range".
  3. [§V-H] The sentence "we conducte additional experiments" contains a typo; it should read "we conducted additional experiments".
  4. [§V-D] The figure caption "Ally lossed" should read "Ally lost".
  5. [§V-B] The phrase "the the SMAC" contains a duplicated article; it should read "the SMAC".
  6. [§V-A-c] SMAC scenario names are inconsistent: the text uses "1o2r vs 4r" and "1o10b vs 1r", whereas the standard naming convention (e.g., in the SMAC benchmark) uses underscores, such as "1o_2r_vs_4r" and "1o_10b_vs_1r"; please standardize the naming for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PAGNet's equations define losses and standard TD updates; benchmark gains are empirical, not fitted into the claimed conclusions, and no self-citation is load-bearing.

full rationale

PAGNet is an empirical architecture paper rather than a derivation paper, so there is no equation chain whose conclusion is assumed in its premises. Equation (3) defines the weighted communication representation, Eq. (6) trains the generator to reconstruct local observations, Eqs. (7)-(8) combine MSE and GAN losses, and Eq. (9) is the standard QMIX TD loss augmented with W_t and G(o_t,W_t). None of these equations defines the learned weights as the minimizer of Eq. (9), and the reported performance gains come from benchmark comparisons rather than from quantities fitted to those benchmarks. The text states that the information-level weight network is trained with the adaptive generative network via Eq. (8), and while Algorithm 1 line 12 says 'Update θ, φ and ψ' without explicitly separating optimizers, this is an implementation and gradient-flow ambiguity rather than a circular derivation. The interpretability analysis is post hoc visualization and is not a falsifiable prediction, but that is selection behavior, not circular reasoning. No self-citation is load-bearing: the only possibly overlapping reference is related-work context, and no uniqueness theorem or prior result is invoked to force the architecture. Therefore the correct circularity finding is no significant circularity.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central method rests on tuned hyperparameters (alpha and architecture sizes), standard MARL/RL assumptions, and two ad hoc design choices: raw local observations as messages and the noise-mixing rule in Eq. (3). The GAN component also assumes adversarial training works in practice. No new physical entity is introduced.

free parameters (7)
  • alpha (GAN loss weight) = 0.0004 (Table III; search range 0.0003-0.0005 in Table IV)
    Balances the GAN loss against the MSE observation loss in Eq. (8); chosen by hyperparameter search.
  • weight hidden dim = 64 (Table III; range 32/64/128)
    Capacity of the information-level weight network; tuned per benchmark.
  • weight dropout = 0.1 (Table III; range 0/0.1/0.2)
    Regularization for the weight network; tuned per benchmark.
  • decoder attention heads = 4 (Table III; range 2/4/6)
    Number of heads in the transformer decoder; tuned per benchmark.
  • attention decoder dim = 128 (Table III; range 32/64/128)
    Transformer decoder embedding dimension; tuned per benchmark.
  • number of transformer layers = 2 (Table III; range 1/2/4)
    Depth of the transformer decoder; tuned per benchmark.
  • discriminator embedding dim = 128 (Table III; range 32/64/128)
    Capacity of the GAN discriminator; tuned per benchmark.
assumptions (6)
  • domain assumption Cooperative MARL tasks are modeled as Dec-POMDPs with shared reward and communication messages M.
    Used throughout Section III-A; standard for this paper's problem class but not proven for each benchmark.
  • standard math The value-based CTDE framework with a monotonic mixing network (QMIX-style) can represent the optimal joint Q-function.
    The RL loss in Eq. (9) assumes IGM/value-decomposition conditions; this is standard in the cited QMIX/QPLEX literature.
  • domain assumption GAN minimax training (Eqs. 7-8) converges to a useful generator rather than mode collapse.
    No convergence guarantee is provided; the method relies on adversarial training stability in practice.
  • ad hoc to paper Local observations are sufficient communication content; sending raw observations as messages is lossless enough.
    Section IV-A redefines communication information as local observations; this is a design choice, not derived.
  • ad hoc to paper The noise-injection communication model in Eq. (3) is beneficial or at least harmless.
    No ablation or theoretical justification is offered for mixing messages with Gaussian noise.
  • domain assumption True global states are available during centralized training to train the discriminator.
    Required by Eq. (7) and Figure 3; this limits the claimed reduction in dependence on global states.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PAGNet: Pluggable Adaptive Generative Networks for Information Completion in Multi-Agent Communication." pith.science (2026). https://pith.science/paper/5FOA4SN5

@misc{pith2026250203845,
  author       = {Pith},
  title        = {Pith review of: PAGNet: Pluggable Adaptive Generative Networks for Information Completion in Multi-Agent Communication},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5FOA4SN5}},
  note         = {Machine review of arXiv:2502.03845}
}
read the original abstract

For partially observable cooperative tasks, multi-agent systems must develop effective communication and understand the interplay among agents in order to achieve cooperative goals. However, existing multi-agent reinforcement learning (MARL) with communication methods lack evaluation metrics for information weights and information-level communication modeling. This causes agents to neglect the aggregation of multiple messages, thereby significantly reducing policy learning efficiency. In this paper, we propose pluggable adaptive generative networks (PAGNet), a novel framework that integrates generative models into MARL to enhance communication and decision-making. PAGNet enables agents to synthesize global states representations from weighted local observations and use these representations alongside learned communication weights for coordinated decision-making. This pluggable approach reduces the computational demands typically associated with the joint training of communication and policy networks. Extensive experimental evaluations across diverse benchmarks and communication scenarios demonstrate the significant performance improvements achieved by PAGNet. Furthermore, we analyze the emergent communication patterns and the quality of generated global states, providing insights into operational mechanisms.

Figures

Figures reproduced from arXiv: 2502.03845 by the authors.

Figure 1
Figure 1. The structure of PAGNet is differentiated by various colors representing different networks: green for the information-level weight network, purple [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Structure of information-level weight network. To provide a clearer demonstration, we set the number of agents to five and use the weight [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Structure of adaptive generative network. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: (a) Illustrations of LBF tasks. (b) Average returns on LBF. (c) Illustrations of Hallway tasks. (d) Average battle win rate on Hallway. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Performance comparisons with baselines on SMAC. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 7
Figure 7. Figure 7: Overall, the generated global states exhibit temporal [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 6
Figure 6. Figure 6: Information-level weight visualization and analysis along the MARL trajectory. The line graph’s y-axis (Average Communication Weight (%)) is on [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Information completion visualization and analysis along the MARL trajectory. In the generated state and true state, the numbers above the agents [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Performance comparisons with varying sights (SR means the sigh [PITH_FULL_IMAGE:figures/full_fig_p013_8.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. Value-Aware Prediction for Robust Multi-Agent Coordination Under Communication Loss

    cs.MA 2026-07 conditional novelty 4.0 of 10

    Advantage-weighted NLL for observation imputation in MARL prevents performance collapse under high communication loss in 3 of 5 MPE tasks.

Reference graph

Works this paper leans on

51 extracted references · 42 canonical work pages · cited by 1 Pith paper

  1. [1]

    Smarts: An open-source scalable multi-agent RL training school for autonomous driving,

    M. Zhou, J. Luo, J. Villella, Y . Yang, D. Rusu, J. Miao et al. , “Smarts: An open-source scalable multi-agent RL training school for autonomous driving,” in Proc. Conf. on Robot Learn. , London, UK, Nov. 2021, pp. 323–352

    work page 2021

  2. [2]

    Deep reinforcement learning for swarm systems,

    M. H ¨uttenrauch, A. ˇSoˇsi´c, and G. Neumann, “Deep reinforcement learning for swarm systems,” J. Mach. Learn. Res. , vol. 20, no. 54, pp. 1–31, Feb. 2019

    work page 2019

  3. [3]

    Smart grid for industry using multi-agent reinforcement learning,

    M. Roesch, C. Linder, R. Zimmermann, A. Rudolf, A. Hohmann, and G. Reinhart, “Smart grid for industry using multi-agent reinforcement learning,” Appl. Sci. , vol. 10, no. 19, pp. 6900–6920, Oct. 2020

    work page 2020

  4. [4]

    Mo-mix: Multi-objective multi- agent cooperative decision-making with deep reinforcement learning,

    T. Hu, B. Luo, C. Yang, and T. Huang, “Mo-mix: Multi-objective multi- agent cooperative decision-making with deep reinforcement learning,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 45, no. 10, pp. 098–112, Oct. 2023

    work page 2023

  5. [5]

    Biases for emergent communication in multi-agent reinforcement learning,

    T. Eccles, Y . Bachrach, G. Lever, A. Lazaridou, and T. Graepel, “Biases for emergent communication in multi-agent reinforcement learning,” in Proc. Adv. Neural Inf. Process. Syst. , Vancouver, Canada, Dec. 2019, pp. 2635–2645

    work page 2019

  6. [6]

    Learning to ground multi-agent communication with autoencoders,

    T. Lin, J. Huh, C. Stauffer, S. N. Lim, and P. Isola, “Learning to ground multi-agent communication with autoencoders,” in Proc. Adv. Neural Inf. Process. Syst. , Vancouver, Canada, Dec. 2021, pp. 1445–1455. JOURNAL OF LATEX CLASS FILES 14

    work page 2021

  7. [7]

    The surprising effectiveness of ppo in cooperative multi-agent games,

    C. Yu, A. Velu, E. Vinitsky, J. Gao, Y . Wang, A. Bayen, and Y . Wu, “The surprising effectiveness of ppo in cooperative multi-agent games,” in Proc. Adv. Neural Inf. Process. Syst. , New Orleans, USA, Nov. 2022, pp. 611–624

    work page 2022

  8. [8]

    Interaction pattern disentangling for multi-agent reinforcement learning,

    S. Liu, J. Song, Y . Zhou, N. Yu, K. Chen, Z. Feng et al. , “Interaction pattern disentangling for multi-agent reinforcement learning,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 46, no. 1, pp. 104–119, Jan. 2024

    work page 2024

Show all 51 references
  1. [9]

    Counterfactual multi-agent policy gradients,

    J. Foerster, G. Farquhar, T. Afouras, N. Nardelli, and S. Whiteson, “Counterfactual multi-agent policy gradients,” in Proc. AAAI Conf. Artif. Intell., New Orleans, USA, Feb. 2018, pp. 2974–2982

    2018

  2. [10]

    Multi-agent actor-critic for mixed cooperative-competitive environ- ments,

    R. Lowe, Y . I. Wu, A. Tamar, J. Harb, O. Pieter Abbeel, and I. Mordatch, “Multi-agent actor-critic for mixed cooperative-competitive environ- ments,” in Proc. Adv. Neural Inf. Process. Syst. , Long Beach, USA, Dec. 2017, pp. 6379–6390

    2017

  3. [11]

    Towards understanding cooperative multi-agent q-learning with value factorization,

    J. Wang, Z. Ren, B. Han, J. Ye, and C. Zhang, “Towards understanding cooperative multi-agent q-learning with value factorization,” in Proc. Adv. Neural Inf. Process. Syst., Vancouver, Canada, Dec. 2021, pp. 1783– 1793

    2021

  4. [12]

    Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning,

    Q. Li, Z. Peng, L. Feng, Q. Zhang, Z. Xue, and B. Zhou, “Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 45, no. 3, pp. 461–475, Jul. 2022

    2022

  5. [13]

    Monotonic value function factorisation for deep multi- agent reinforcement learning,

    T. Rashid, M. Samvelyan, C. S. De Witt, G. Farquhar, J. Foerster, and S. Whiteson, “Monotonic value function factorisation for deep multi- agent reinforcement learning,” Journal of Machine Learning Research , vol. 21, no. 178, pp. 1–51, Aug. 2020

    2020

  6. [14]

    Value-decomposition networks for cooperative multi-agent learning,

    P. Sunehag, G. Lever, A. Gruslys, W. M. Czarnecki, V . Zambaldi, M. Jaderberg et al. , “Value-decomposition networks for cooperative multi-agent learning,” arXiv preprint arXiv:1706.05296 , 2017

    2017 arXiv

  7. [15]

    Contrasting central- ized and decentralized critics in multi-agent reinforcement learning,

    X. Lyu, Y . Xiao, B. Daley, and C. Amato, “Contrasting central- ized and decentralized critics in multi-agent reinforcement learning,” arXiv:2102.04402, 2021

    2021 arXiv

  8. [16]

    Learning nearly decom- posable value functions via communication minimization,

    T. Wang, J. Wang, C. Zheng, and C. Zhang, “Learning nearly decom- posable value functions via communication minimization,” in Proc. Int. Conf. Learn. Represent. , Addis Ababa, Ethiopia, Apr. 2020, pp. 1–15

    2020

  9. [17]

    Qplex: Duplex dueling multi-agent q-learning,

    J. Wang, Z. Ren, T. Liu, Y . Yu, and C. Zhang, “Qplex: Duplex dueling multi-agent q-learning,” in Proc. Int. Conf. Learn. Represent. , Addis Ababa, Ethiopia, Apr. 2020, pp. 21–32

    2020

  10. [18]

    Learning individually inferred commu- nication for multi-agent cooperation,

    Z. Ding, T. Huang, and Z. Lu, “Learning individually inferred commu- nication for multi-agent cooperation,” in Proc. Adv. Neural Inf. Process. Syst., Vancouver, Canada, Dec. 2020, pp. 652–663

    2020

  11. [19]

    Multi- agent concentrative coordination with decentralized task representation,

    L. Yuan, C. Wang, J. Wang, F. Zhang, F. Chen, C. Guan et al. , “Multi- agent concentrative coordination with decentralized task representation,” in Proc. Int. Joint Conf. Artif. Intell. , Vienna, Austria, Jul. 2022, pp. 2852–2858

    2022

  12. [20]

    Learning to communicate with deep multi-agent reinforcement learning,

    J. Foerster, I. A. Assael, N. De Freitas, and S. Whiteson, “Learning to communicate with deep multi-agent reinforcement learning,” in Proc. Adv. Neural Inf. Process. Syst. , Barcelona, Spain, Dec. 2016, pp. 2137– 2145

    2016

  13. [21]

    Learning multiagent commu- nication with backpropagation,

    S. Sukhbaatar, A. Szlam, and R. Fergus, “Learning multiagent commu- nication with backpropagation,” in Proc. Adv. Neural Inf. Process. Syst. , Barcelona, Spain, Dec. 2016, pp. 2244–2252

    2016

  14. [22]

    Learning attentional communication for multi-agent cooperation,

    J. Jiang and Z. Lu, “Learning attentional communication for multi-agent cooperation,” in Proc. Adv. Neural Inf. Process. Syst. , Montr´eal, Canada, Dec. 2018, pp. 7254–7264

    2018

  15. [23]

    Learning when to com- municate at scale in multiagent cooperative and competitive tasks,

    A. Singh, T. Jain, and S. Sukhbaatar, “Learning when to com- municate at scale in multiagent cooperative and competitive tasks,” arXiv:1812.09755, 2018

    2018 arXiv

  16. [24]

    Graph convolutional reinforce- ment learning,

    J. Jiang, C. Dun, T. Huang, and Z. Lu, “Graph convolutional reinforce- ment learning,” arXiv:1810.09202, 2018

    2018 arXiv

  17. [25]

    Multi-agent graph-attention communication and teaming,

    Y . Niu, R. R. Paleja, and M. C. Gombolay, “Multi-agent graph-attention communication and teaming,” in Proc. Int. Conf. Auton. Agents Multia- gent Syst. , London, UK, May 2021, pp. 764–772

    2021

  18. [26]

    Multi-agent game abstraction via graph attention neural network,

    Y . Liu, W. Wang, Y . Hu, J. Hao, X. Chen, and Y . Gao, “Multi-agent game abstraction via graph attention neural network,” in Proc. AAAI Conf. Artif. Intell. , New York, USA, Feb. 2020, pp. 7211–7218

    2020

  19. [27]

    Efficient multi-agent communication via self-supervised information aggregation,

    C. Guan, F. Chen, L. Yuan, C. Wang, H. Yin, Z. Zhang, and Y . Yu, “Efficient multi-agent communication via self-supervised information aggregation,” in Proc. Adv. Neural Inf. Process. Syst. , New Orleans, USA, Dec. 2022, pp. 1020–1033

    2022

  20. [28]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Proc. Adv. Neural Inf. Process. Syst. , Vancouver, Canada, Dec. 2020, pp. 6840–6851

    2020

  21. [29]

    Mildly conservative q-learning for offline reinforcement learning,

    J. Lyu, X. Ma, X. Li, and Z. Lu, “Mildly conservative q-learning for offline reinforcement learning,” in Proc. Adv. Neural Inf. Process. Syst. , New Orleans, USA, Dec. 2022, pp. 1255–1266

    2022

  22. [30]

    Learning multi-agent communication through structured attentive reasoning,

    M. Rangwala and R. Williams, “Learning multi-agent communication through structured attentive reasoning,” in Proc. Adv. Neural Inf. Pro- cess. Syst. , Dec. 2020, pp. 325–336

    2020

  23. [31]

    The starcraft multi-agent challenge,

    M. Samvelyan, T. Rashid, C. S. De Witt, G. Farquhar, N. Nardelli, T. G. Rudner et al. , “The starcraft multi-agent challenge,” arXiv:1902.04043, 2019

    1902 arXiv

  24. [32]

    Social networks and cooperation in hunter-gatherers,

    C. L. Apicella, F. W. Marlowe, J. H. Fowler, and N. A. Christakis, “Social networks and cooperation in hunter-gatherers,” Nature, vol. 481, no. 7382, pp. 497–501, Jan. 2012

    2012

  25. [33]

    Structured cooperative reinforcement learning with time-varying composite action space,

    W. Li, X. Wang, B. Jin, D. Luo, and H. Zha, “Structured cooperative reinforcement learning with time-varying composite action space,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 44, no. 11, pp. 8618–8634, Nov. 2022

    2022

  26. [34]

    Succinct and robust multi-agent communication with temporal message control,

    S. Q. Zhang, Q. Zhang, and J. Lin, “Succinct and robust multi-agent communication with temporal message control,” in Proc. Adv. Neural Inf. Process. Syst. , Vancouver, Canada, Dec. 2020, pp. 271–282

    2020

  27. [35]

    Deterministic policy gradient algorithms,

    D. Silver, G. Lever, N. Heess, T. Degris, D. Wierstra, and M. Riedmiller, “Deterministic policy gradient algorithms,” in Proc. Int. Conf. Mach. Learn., Beijing, China, Jun. 2017, pp. 387–395

    2017

  28. [36]

    Qtran: Learning to factorize with transformation for cooperative multi-agent reinforcement learning,

    K. Son, D. Kim, W. J. Kang, D. E. Hostallero, and Y . Yi, “Qtran: Learning to factorize with transformation for cooperative multi-agent reinforcement learning,” in Proc. Int. Conf. Mach. Learn. , Long Beach, USA, Jul. 2019, pp. 2026–2035

    2019

  29. [37]

    Tarmac: Targeted multi-agent communication,

    A. Das, T. Gervet, J. Romoff, D. Batra, D. Parikh, M. Rabbat et al. , “Tarmac: Targeted multi-agent communication,” in Proc. Int. Conf. Mach. Learn. , Long Beach, USA, Jul. 2019, pp. 1538–1547

    2019

  30. [38]

    Robust multi- agent communication with graph information bottleneck optimization,

    S. Ding, W. Du, L. Ding, J. Zhang, L. Guo, and B. An, “Robust multi- agent communication with graph information bottleneck optimization,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 46, no. 5, pp. 3096–3107, Nov. 2023

    2023

  31. [39]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair et al. , “Generative adversarial nets,” in Proc. Adv. Neural Inf. Process. Syst., Montr ´eal, Canada, Dec. 2014, pp. 2672–2680

    2014

  32. [40]

    Unsupervised representation learning with deep convolutional generative adversarial networks,

    A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with deep convolutional generative adversarial networks,” arXiv:1511.06434, 2015

    2015 arXiv

  33. [41]

    Photo-realistic single image super-resolution using a generative adversarial network,

    C. Ledig, L. Theis, F. Husz ´ar, J. Caballero, A. Cunningham, A. Acosta et al. , “Photo-realistic single image super-resolution using a generative adversarial network,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog- nit., Honolulu, USA, Jul. 2017, pp. 4681–4690

    2017

  34. [42]

    Progressive growing of gans for improved quality, stability, and variation,

    T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive growing of gans for improved quality, stability, and variation,” arXiv:1710.10196, 2017

    2017 arXiv

  35. [43]

    Banach wasserstein gan,

    J. Adler and S. Lunz, “Banach wasserstein gan,” in Proc. Adv. Neural Inf. Process. Syst. , Montreal, Canada, Dec. 2018, pp. 6743–6752

    2018

  36. [44]

    High- resolution image inpainting using multi-scale neural patch synthesis,

    C. Yang, X. Lu, Z. Lin, E. Shechtman, O. Wang, and H. Li, “High- resolution image inpainting using multi-scale neural patch synthesis,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , Honolulu, USA, Jul. 2017, pp. 6721–6729

    2017

  37. [45]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , Honolulu, USA, Jul. 2017, pp. 2242–2251

    2017

  38. [46]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare et al. , “Human-level control through deep reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533, Feb. 2015

    2015

  39. [47]

    An information theory perspective on variance-invariance- covariance regularization,

    R. Shwartz-Ziv, R. Balestriero, K. Kawaguchi, T. G. Rudner, and Y . LeCun, “An information theory perspective on variance-invariance- covariance regularization,” in Proc. Adv. Neural Inf. Process. Syst. , New Orleans, USA, Dec. 2024, pp. 1–34

    2024

  40. [48]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez et al., “Attention is all you need,” inProc. Adv. Neural Inf. Process. Syst., Long Beach, USA, Dec. 2017, pp. 5998–6008

    2017

  41. [49]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Proc. Med. Image Comput. Comput. Assisted Intervention , Munich, Germany, Oct. 2015, pp. 234– 241

    2015

  42. [50]

    Mish: A self regularized non-monotonic activation function,

    D. Misra, “Mish: A self regularized non-monotonic activation function,” arXiv:1908.08681, 2019

    1908 arXiv

  43. [51]

    Bench- marking multi-agent deep reinforcement learning algorithms in cooper- ative tasks,

    G. Papoudakis, F. Christianos, L. Sch ¨afer, and S. V . Albrecht, “Bench- marking multi-agent deep reinforcement learning algorithms in cooper- ative tasks,” arXiv:2006.07869, 2020

    2006 arXiv

Pith tools

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