Robust Multi-agent Communication Based on Decentralization-Oriented Adversarial Training

Boyu Wu; Fanjiang Xu; Junjie Wang; Qing Wang; Shoubin Li; Xiaofei Xie; Xuyan Ma; Yawen Wang

REVIEW 4 major objections 5 minor 35 references

Robust Multi-agent Communication Based on Decentralization-Oriented Adversarial Training

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Retraining a communication policy against an adversary that masks its critical channels makes multi-agent teams markedly more robust against attacks, and slightly more performant when no attack occurs.

desk verdict The adversary is well-specified, but the paper's core claim that DMAC retrains the communication policy has no corresponding algorithm or equation, so the reported robustness gains lack a mechanism as submitted. read the letter →

arxiv 2504.21278 v1 pith:MQHF2NR2 submitted 2025-04-30 cs.MA

Xuyan Ma , Yawen Wang , Junjie Wang , Xiaofei Xie , Boyu Wu , Shoubin Li , Fanjiang Xu , Qing Wang This is my paper · ORCID

classification cs.MA

keywords multi-agentreinforcementlearningcommunicationpolicyadversarialtrainingrobustnessdecentralizationcriticalchannelidentificationmaskingcooperativesystems

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes DMAC, a training-time method for making learned communication policies in multi-agent reinforcement learning harder to break. The central claim is that a communication policy retrained against an adversary which dynamically identifies and masks its most-used channels will spread its communication across more channels, become more decentralized, and therefore survive attacks that previously crippled it. Across two communication policies (T2MAC and I2C) and four tasks, DMAC raises post-attack win rates by 47.9% to 81.9% for T2MAC and by 54.4% to 99.0% for I2C relative to the AME and R-MACRL baselines under a learned adaptive attack, with comparable gains under a heuristic attack. Under normal conditions the retrained policies are not worse and are usually slightly better, and the standard deviation of channel communication frequencies drops, which the paper treats as the signature of decentralization. If the claim holds, DMAC is a portable training-time defense that adds no extra computation at inference time.

What carries the argument

The load-bearing object is DMAC_Adv, an adversary that assigns one masking agent to every unordered pair of target agents. At each time step it chooses $a^c_{i,j} \in \{0,1\}$, and the masked observation is $\hat{o}_{i,j} = o_{i,j} \odot a^c_{i,j}$, where a masked channel delivers null to both endpoints. The adversary's reward is $\hat{r} = 1/(\sum_{t=0}^T \gamma^t (w_1 R(s^t,a^t,s^{t+1}) + w_2 R_m(a^{ct})) + \xi)$, so maximizing it means both hurting the team's reward and masking as few channels as possible; this pressures the adversary to find the smallest set of channels whose loss is most damaging. The adversary's policy network consumes each agent's observation concatenated with a graph embedding $e_v$ built by one round of distance-weighted neighbor aggregation, $e^{k+1}_v \leftarrow \mathrm{Sum}(e^k_v, \mathrm{Mean}(\frac{1}{w(u,v)} e^k_u), u \in N(v))$. The masking policy is trained with a value-decomposed critic that enforces the Individual-Global-Max property, so the joint optimum equals per-channel optimality, and the masks it produces are used as adversarial training samples to retrain the communication policy while the target agents' decision policy stays fixed.

What would settle it

Run DMAC exactly as specified, then attack the final policy with a continuous adversarial perturbation that alters message values while leaving all channels open; if the post-attack win rate falls back to the unprotected baseline level, the claimed robustness mechanism would be disproved.

Watch

Extended reading notes

Core claim

The paper's claim is that robustness of a learned communication policy can be improved not by filtering or repairing attacked messages at inference time, but by deliberately attacking the policy during training. An adversary trained with a dual-objective reward, maximizing the drop in team reward while minimizing the number of masked channels, generates adversarial samples that close the currently critical channels. Retraining the communication policy on these samples forces it to fall back on alternative channels, shifting its usage pattern from a concentrated, hub-like structure to a decentralized one. The paper reports that this shift is what makes the policy resilient: under learned adaptive and heuristic message attacks, the win rate of policies trained with DMAC roughly doubles compared with the AME and R-MACRL defenses, while the standard deviation of per-channel communication frequency drops by about a third to a half. DMAC is framed as a portable wrapper: the target agents' decision policy stays fixed during adversarial training, and only the communication policy is retrained, so the method can be attached to any learnable communication policy without changing the deployed decision model.

Load-bearing premise

The argument rests on the assumption that robustness against binary channel masking, the only attack form used during adversarial training, transfers to the message-perturbation attacks used in evaluation; if that transfer fails, the reported gains would not be explained by the training method.

Editorial extensions

If this is right

DMAC improves post-attack win rates substantially for both tested communication policies; for example, T2MAC rises from 27.8% to 60.4% on SMAC under the learned adaptive attack, so teams hardened this way no longer collapse when one critical channel is damaged.
Because DMAC is a training wrapper over the communication policy and leaves the target agents' decision policy fixed, the defense adds no inference-time computation and can be paired with any learnable communication policy.
Under normal conditions the retrained policy remains at least as good as the original policy, with T2MAC on SMAC going from 81.2% to 83.7% win rate, so robustness is not bought by sacrificing ordinary performance.
The reduced standard deviation of per-channel communication frequency indicates the retrained policy spreads messages more evenly, so attacking the previously dominant channel no longer removes most of the team's information flow.

Reading between the lines

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

An implication the paper leaves implicit is that the same recipe could apply to communication policies with continuous-valued or directional channels, not just binary links, by training the adversary to select message dimensions to corrupt.
A testable extension is to measure decentralization directly with graph-theoretic indices such as effective resistance or spectral gap rather than only the standard deviation of channel frequencies; if decentralization is the true causal mechanism, those indices should track the robustness gain.
The paper's protocol trains the learned adaptive attack on the victim policy, but an adaptive adversary that re-optimizes after each DMAC update is not fully explored, so whether the defense survives repeated attacker re-optimization remains open.
Because the adversary is rewarded for masking as few channels as possible, DMAC implicitly assumes the target policy has a small set of critical channels to find; in regimes where all channels matter equally, the training signal may be much weaker.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper's central idea is attractive: train a masking adversary to identify critical communication channels, then adversarially retrain the communication policy to become less reliant on them, yielding a decentralized and more robust structure. That recipe is new in this combination, and the design of the adversary, with its dual reward and graph-based features, is coherent. The reported improvements over AME and R-MACRL are large and consistent across four environments, so something substantive is likely happening. But the manuscript as submitted has a load-bearing gap: the retraining of the communication policy is never actually specified. Algorithm 1, which is the only training loop given, updates only the masking policy and critic. It takes the target policy π as fixed input, executes it under masks, and returns the adversary's policy. There is no update rule for CP, no adversarial training objective, no alternation schedule, no hyperparameters for the retraining step. The sentence in Section 3.2 saying 'the communication policy is retrained in the process of adversarial training' is an assertion without any supporting pseudocode or equation, and the referenced appendix is not present. A fixed policy evaluated under masks cannot become more robust from this procedure, so Table 1's gains currently have no explained mechanism. That is the main problem, and it is not a matter of style. The secondary issues are also real but minor by comparison: single-point results with no error bars or seed counts, no statistical tests, and no random-masking or dropout control to show that the adversary's targeted masking, rather than any perturbation, drives the robustness. The evaluation attacks perturb messages while the training adversary masks channels, so the transfer needs explicit evidence. The decentralization measurements are a nice addition, but they show correlation, not the claimed causal path. Who should read this? Researchers working on robust communication in MARL would find the adversary formulation useful, and the empirical data is suggestive enough to warrant a closer look. But the paper is not ready as is. I would send it to peer review, because the idea is timely and the empirical claims are substantial, but I would make the authors specify the retraining algorithm, add seeds and confidence intervals, and include a random-masking baseline. If the code is available, a referee can verify, but the version I saw cannot be evaluated as a complete method.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes DMAC, an adversarial training method intended to improve the robustness of learned multi-agent communication policies. DMAC trains an adversary (DMAC_Adv) to identify and mask critical communication channels, then claims to retrain the target communication policy on the resulting adversarial samples so that the policy becomes less dependent on any single channel and shifts toward a decentralized communication structure. The authors evaluate DMAC on two communication policies, T2MAC and I2C, across four tasks (SMAC, Cooperative Navigation, Traffic Junction, and Predator-Prey), reporting higher win rates than AME and R-MACRL under learned adaptive and heuristic attacks, slightly improved performance in normal conditions, and reduced variance in per-channel communication frequencies.

Significance. If the mechanism were fully specified and the empirical results were reported with uncertainty quantification, DMAC would be a useful contribution: it is a training-time defense that can be combined with any learnable communication policy and adds no inference-time overhead. The adversary formulation as a MARL problem with an IGM-constrained critic is coherent, and evaluating under both learned and heuristic attacks is a sensible protocol. The authors state that source code is released, which supports reproducibility. However, the central retraining step for the communication policy is not described anywhere in the manuscript, and the empirical evidence consists of single-point win rates without seeds, error bars, or statistical tests. As a result, the strength of the claimed robustness gains cannot currently be assessed.

major comments (4)

[§3.2, Algorithm 1] The central claim that DMAC retrains the communication policy is not supported by the presented algorithm. Algorithm 1 takes the already-trained target policy π as input and updates only the masking policy πc and the critic C (line 14), while line 11 executes the fixed policy π on masked observations. There is no update rule, loss function, alternation schedule, or set of hyperparameters for retraining CP/π anywhere in Section 3.2. The sentence stating that 'the communication policy is retrained in the process of adversarial training' is an assertion without a corresponding equation or pseudocode, and the referenced appendix is not included in the manuscript. Because a fixed policy evaluated under masks cannot change its behavior, the robustness gains in Tables 1 and 2 have no mechanism in the paper as written. Please provide the complete retraining loop for CP, including the objective, the gradient update, how adversarial samples are incorporated, and how the adversary and CP interact over training iterations.
[§3.2 Eq. (1) and §4.1] The adversary's training objective uses binary channel masking, as defined in Eq. (1), while the evaluation attacks in Section 4.1 are message perturbations: the learned adaptive attack and the heuristic attack perturb or replace message content rather than masking whole channels. If masking-based adversarial training is supposed to confer robustness against perturbation attacks, the paper needs an explicit transfer argument or a control experiment, such as training with randomly masked channels or evaluating under masking attacks as well. Without such evidence, the reported robustness gains under perturbation attacks may not be attributable to the masking-based training procedure described in Section 3.2.
[§4.2, Tables 1 and 2] Tables 1 and 2 report a single win-rate value per condition with no number of seeds, no standard deviation or confidence interval, and no statistical test. The headline improvements (47.9%–81.9% for T2MAC and 54.4%–99.0% for I2C under the learned adaptive attack) are therefore not statistically established. Please report mean ± standard deviation over at least five independent runs and include pairwise significance tests or effect sizes for the comparisons against AME and R-MACRL.
[§4.2, Table 3] Table 3 reports communication-frequency statistics only for SC and PP, although the decentralization evaluation in Section 4.2 is presented as a general result for the adjusted communication policy. The claim that DMAC achieves a decentralized communication structure across all four tasks requires supporting numbers for CN and TJ as well, or an explicit statement that only two environments were analyzed in this part of the evaluation.

minor comments (5)

[General] The manuscript repeatedly refers to an appendix ('More results can be found in the appendix', including ablation studies and additional experimental details), but the arXiv submission contains no appendix. Include the appendix or remove the references to it.
[References] Several references are duplicated: Das et al. 2019 appears as [Das et al., 2019a] and [Das et al., 2019b], Sukhbaatar et al. 2016 appears as [Sukhbaatar et al., 2016a] and [Sukhbaatar et al., 2016b], and the Mitchell et al. 2020 citation is repeated in Section 2. Please consolidate these entries.
[§3.2, Eq. (3)] Equation (3) uses the reciprocal of the edge weight w(u,v) after w is defined as the distance between two agents. Please clarify whether this is inverse-distance weighting, how the normalization is performed, and why distance alone is sufficient to capture interaction strength for identifying critical communication channels; an ablation or justification would be helpful.
[§3.2, Eq. (2)] Equation (2) is ambiguous: the objective says it maximizes r̂, but the expression is written as the reciprocal of a weighted sum, with ξ preventing division by zero. Please state the exact transformation and the roles of w1, w2, and ξ, and provide their values in the experimental setup.
[Algorithm 1] The number of masking agents N = n(n−1)/2 is defined in Section 3.2 but never used explicitly in Algorithm 1. Please clarify whether each masking agent is responsible for exactly one channel and how the pairing (i,j) is assigned in the loop over training batches.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: DMAC's robustness numbers are empirical measurements, and the only self-citation is minor and non-load-bearing; a missing CP-update specification is a correctness gap, not circularity.

full rationale

The paper's claimed robustness improvements are empirical measurements (Tables 1-3) and are not derived from the equations in Section 3.2, so there is no algebraic self-definition. DMAC_Adv's objective (Eq. 2) uses the environment reward R and mask count Rm; this is an external optimization criterion, not the reported win-rate deltas. The evaluation uses separately described learned adaptive and heuristic attacks; nothing in the paper defines those attacks as DMAC_Adv itself, so the numbers are not a renamed fit. The only self-citation, [Ma et al., 2024], is used to justify graph-based feature extraction for the adversary, and it does not contain the DMAC result; no uniqueness theorem or load-bearing premise is imported from the authors' prior work, and the cited TD-loss support [Liu et al., 2023] is external. A genuine specification gap exists: Section 3.2 asserts 'the communication policy is retrained in the process of adversarial training' and Algorithm 1 inputs only the fixed target policy π and updates only πc and C (lines 14 and 16), with no CP update rule; this is a missing-support/correctness concern, not circularity. Because the one self-citation is minor and non-load-bearing, the appropriate non-circularity score is 2 rather than 0.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central method rests on a handful of standard RL machinery plus several task-specific design choices: the weighted adversary reward, distance-based graph features, fixed target policy during retraining, and the assumption that binary masking of messages is an adequate stand-in for the attack classes considered. None of these are derived from first principles; each is an engineering choice that a replicator must tune.

free parameters (3)

w1, w2 (adversary reward weights) = not reported
Equation 2 weights the target MAS reward and the number of masked channels; these weights control the critical-channel trade-off and are not stated in the main text, so they may have been tuned per task.
ξ (offset in adversary reward) = not reported
Used in Equation 2 denominator to avoid division by zero; value not given.
graph embedding iterations K and edge weight normalization = not reported
Equation 3 aggregates neighbor embeddings for K iterations with 1/w(u,v) weighting based on distance; K and distance normalization are unspecified and affect channel criticality features.

assumptions (6)

standard math IGM principle and non-negative weight constraint in Equation 5 allow Qtot to be maximized by per-agent argmax.
Invoked from Hong et al. 2022 and standard QMIX-style value decomposition; used to train DMAC_Adv in Section 3.2.
domain assumption The target agent policy π is fixed while the communication policy CP is retrained.
Algorithm 1 runs π on masked observations but never updates π; this is a design choice and a boundary condition for the method's validity.
domain assumption Masking a communication channel is equivalent to setting the message to null (Equation 1).
The entire adversarial training signal is defined by this operation; if the communication policy cannot be trained with nulled messages, the method breaks.
ad hoc to paper Distance between agents is a sufficient weight for interaction strength in the graph embedding.
Equation 3 uses 1/w(u,v) as aggregation weight where w is distance from the environment; no evidence is given that distance alone captures channel criticality.
domain assumption Decentralized networks are more robust to node/channel failure (decentralization theory).
Motivates the whole approach in Section 1; the paper treats it as background from Boss and Gralla 2023, but does not prove it for MARL communication policies.
domain assumption Win rate is a comparable task-completion metric across the four environments.
Section 4.1 states 'we uniformly use win rate' despite differing task completion definitions; comparison across environments assumes this aggregation is meaningful.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Robust Multi-agent Communication Based on Decentralization-Oriented Adversarial Training." pith.science (2026). https://pith.science/paper/MQHF2NR2

@misc{pith2026250421278,
  author       = {Pith},
  title        = {Pith review of: Robust Multi-agent Communication Based on Decentralization-Oriented Adversarial Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MQHF2NR2}},
  note         = {Machine review of arXiv:2504.21278}
}

read the original abstract

In typical multi-agent reinforcement learning (MARL) problems, communication is important for agents to share information and make the right decisions. However, due to the complexity of training multi-agent communication, existing methods often fall into the dilemma of local optimization, which leads to the concentration of communication in a limited number of channels and presents an unbalanced structure. Such unbalanced communication policy are vulnerable to abnormal conditions, where the damage of critical communication channels can trigger the crash of the entire system. Inspired by decentralization theory in sociology, we propose DMAC, which enhances the robustness of multi-agent communication policies by retraining them into decentralized patterns. Specifically, we train an adversary DMAC\_Adv which can dynamically identify and mask the critical communication channels, and then apply the adversarial samples generated by DMAC\_Adv to the adversarial learning of the communication policy to force the policy in exploring other potential communication schemes and transition to a decentralized structure. As a training method to improve robustness, DMAC can be fused with any learnable communication policy algorithm. The experimental results in two communication policies and four multi-agent tasks demonstrate that DMAC achieves higher improvement on robustness and performance of communication policy compared with two state-of-the-art and commonly-used baselines. Also, the results demonstrate that DMAC can achieve decentralized communication structure with acceptable communication cost.

Figures

Figures reproduced from arXiv: 2504.21278 by the authors.

**Figure 2.** The overview of our proposed DMAC. its action a t i based on its local observation o t i . The joint action a t = {a t 1 , ..., at n} results in the next state st+1 with the state transition probability P(st+1|st, at). Consequently, a global reward rt is obtained according to reward function R(st, at, st+1). In our problem, we focus on enhancing the robustness of communication policy CP. 3.2 DMAC We draw inspiratio… view at source ↗

**Figure 3.** The overview of DMAC Adv. (a) At each time step, DMAC Adv obtains the feature of each agent from the multi-agent environment, and outputs a decision for each communication channel whether or not to mask. (b) During the training, the masking agents’ policy network learns the masking action and individual value, and the critic network learns the total value to estimate the expected reward. The loss function is introdu… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Multiple environments used in our experiments. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: The communication frequency of the communication channel between each pair of agents in two environments. The darker the color, [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 32 canonical work pages

[1]

Deep reinforcement learning for truck-drone delivery problem

[Bi et al., 2023] Zhiliang Bi, Xiwang Guo, Jiacun Wang, Shu- jin Qin, and Guanjun Liu. Deep reinforcement learning for truck-drone delivery problem. Drones, 7:445, 07

work page 2023
[3]

Garcia Cena, Pedro F

[Cena et al., 2013] Cecilia E. Garcia Cena, Pedro F. C´ardenas, Roque Saltar ´en Pazmi ˜no, Lisandro Puglisi, and Rafael Aracil Santonja. A cooperative multi-agent robotics system: Design and modelling. Expert Syst. Appl., 40(12):4737–4748,

work page 2013
[6]

Learning individually inferred communication for multi-agent cooperation

[Ding et al., 2020] Ziluo Ding, Tiejun Huang, and Zongqing Lu. Learning individually inferred communication for multi-agent cooperation. In Advances in Neural Informa- tion Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, De- cember 6-12, 2020, virtual,

work page 2020
[11]

Learning to schedule communication in multi-agent reinforcement learning

[Kim et al., 2019] Daewoo Kim, Sangwoo Moon, David Hostallero, Wan Ju Kang, Taeyoung Lee, Kyunghwan Son, and Yung Yi. Learning to schedule communication in multi-agent reinforcement learning. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9,

work page 2019
[12]

Multi-agent game abstraction via graph attention neural network

[Liu et al., 2020] Yong Liu, Weixun Wang, Yujing Hu, Jianye Hao, Xingguo Chen, and Yang Gao. Multi-agent game abstraction via graph attention neural network. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, New York, NY, USA, February 7-12, 2020 , pages 7211–7218. AAAI Press,

work page 2020
[13]

Na 2q: Neural attention additive model for inter- pretable multi-agent q-learning

[Liu et al., 2023] Zichuan Liu, Yuanyang Zhu, and Chunlin Chen. Na 2q: Neural attention additive model for inter- pretable multi-agent q-learning. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors,International Confer- ence on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, U...

work page 2023
[14]

Multi-agent actor- critic for mixed cooperative-competitive environments

[Lowe et al., 2017] Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, Pieter Abbeel, and Igor Mordatch. Multi-agent actor- critic for mixed cooperative-competitive environments. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Sys- tems 2017, December 4-9, 2017, Long Beach, CA, USA , pages 6379–6390,

work page 2017
[16]

Enhancing multi-agent system testing with diversity- guided exploration and adaptive critical state exploitation

[Ma et al., 2024] Xuyan Ma, Yawen Wang, Junjie Wang, Xi- aofei Xie, Boyu Wu, Shoubin Li, Fanjiang Xu, and Qing Wang. Enhancing multi-agent system testing with diversity- guided exploration and adaptive critical state exploitation. In Maria Christakis and Michael Pradel, editors, Proceed- ings of the 33rd ACM SIGSOFT International Symposium on Software Tes...

work page 2024

Show all 35 references

[18]

PMAC: personalized multi-agent communication

[Meng and Tan, 2024] Xiangrui Meng and Ying Tan. PMAC: personalized multi-agent communication. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, February 20-27, 2024, Vancouver, Canada, pages 17505– 17513. AAAI Press,

2024
[19]

Gaussian process based message fil- tering for robust multi-agent cooperation in the presence of adversarial communication

[Mitchell et al., 2020] Rupert Mitchell, Jan Blumenkamp, and Amanda Prorok. Gaussian process based message fil- tering for robust multi-agent cooperation in the presence of adversarial communication. CoRR, abs/2012.00508,

2020 arXiv
[20]

Improving coordination in small-scale multi- agent deep reinforcement learning through memory-driven communication

[Pesce and Montana, 2020] Emanuele Pesce and Giovanni Montana. Improving coordination in small-scale multi- agent deep reinforcement learning through memory-driven communication. Mach. Learn. , 109(9-10):1727–1747,

2020
[21]

Deep rein- forcement learning framework for autonomous driving

[Sallab et al., 2017] Ahmad El Sallab, Mohammed Abdou, Etienne Perot, and Senthil Kumar Yogamani. Deep rein- forcement learning framework for autonomous driving. In Patrick Denny, Darnell Moore, and Buyue Zhang, editors, Autonomous Vehicles and Machines 2017, Burlingame, CA, U...

2017
[22]

The starcraft multi- agent challenge

[Samvelyan et al., 2019] Mikayel Samvelyan, Tabish Rashid, Christian Schroeder De Witt, Gregory Farquhar, Nantas Nardelli, Tim GJ Rudner, Chia-Man Hung, Philip HS Torr, Jakob Foerster, and Shimon Whiteson. The starcraft multi- agent challenge. arXiv preprint arXiv:1902.04043,

2019 arXiv
[23]

Safe, multi-agent, re- inforcement learning for autonomous driving

[Shalev-Shwartz et al., 2016] Shai Shalev-Shwartz, Shaked Shammah, and Amnon Shashua. Safe, multi-agent, re- inforcement learning for autonomous driving. CoRR, abs/1610.03295,

2016 arXiv
[25]

Learning when to communicate at scale in multiagent cooperative and competitive tasks

[Singh et al., 2019] Amanpreet Singh, Tushar Jain, and Sain- bayar Sukhbaatar. Learning when to communicate at scale in multiagent cooperative and competitive tasks. In 7th In- ternational Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 ....

2019
[26]

QTRAN: learning to factorize with transformation for cooperative multi-agent reinforcement learning

[Son et al., 2019] Kyunghwan Son, Daewoo Kim, Wan Ju Kang, David Hostallero, and Yung Yi. QTRAN: learning to factorize with transformation for cooperative multi-agent reinforcement learning. In Proceedings of the 36th Interna- tional Conference on Machine Learning, ICML 2019, ...

2019
[27]

Learning multiagent commu- nication with backpropagation

[Sukhbaatar et al., 2016a] Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. Learning multiagent commu- nication with backpropagation. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 201...

2016
[28]

Learning multiagent communica- tion with backpropagation

[Sukhbaatar et al., 2016b] Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. Learning multiagent communica- tion with backpropagation. In Advances in Neural Informa- tion Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, pages 2244–2252,

2016
[29]

Certifiably robust policy learning against adversarial multi-agent communication

[Sun et al., 2023] Yanchao Sun, Ruijie Zheng, Parisa Hassan- zadeh, Yongyuan Liang, Soheil Feizi, Sumitra Ganesh, and Furong Huang. Certifiably robust policy learning against adversarial multi-agent communication. In The Eleventh In- ternational Conference on Learning Represen...

2023
[30]

T2MAC: targeted and trusted multi-agent communication through selective engagement and evidence-driven inte- gration

[Sun et al., 2024] Chuxiong Sun, Zehua Zang, Jiabao Li, Jiangmeng Li, Xiao Xu, Rui Wang, and Changwen Zheng. T2MAC: targeted and trusted multi-agent communication through selective engagement and evidence-driven inte- gration. In Thirty-Eighth AAAI Conference on Artificial Int...

2024
[31]

Leibo, Karl Tuyls, and Thore Graepel

[Sunehag et al., ] Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vin´ıcius Flores Zam- baldi, Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z. Leibo, Karl Tuyls, and Thore Graepel. Value- decomposition networks for cooperative multi-agent learn- ...

2018
[33]

Mis- spoke or mis-lead: Achieving robustness in multi-agent communicative reinforcement learning

[Xue et al., 2022] Wanqi Xue, Wei Qiu, Bo An, Zinovi Ra- binovich, Svetlana Obraztsova, and Chai Kiat Yeo. Mis- spoke or mis-lead: Achieving robustness in multi-agent communicative reinforcement learning. In 21st Interna- tional Conference on Autonomous Agents and Multiagent S...

2022
[34]

Efficient communication in multi-agent reinforcement learning via variance based control

[Zhang et al., 2019] Sai Qian Zhang, Qi Zhang, and Jieyu Lin. Efficient communication in multi-agent reinforcement learning via variance based control. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, Neu...

2019
[35]

A survey of multi-agent deep reinforcement learning with communication

[Zhu et al., 2024] Changxi Zhu, Mehdi Dastani, and Shihan Wang. A survey of multi-agent deep reinforcement learning with communication. In Proceedings of the 23rd Interna- tional Conference on Autonomous Agents and Multiagent Systems, AAMAS 2024, Auckland, New Zealand, May 6- ...

2024
[2012]

Hausknecht and Peter Stone

[Hausknecht and Stone, 2015] Matthew J. Hausknecht and Peter Stone. Deep recurrent q-learning for partially ob- servable mdps. CoRR, abs/1507.06527,

2015 arXiv
[2013]

Tarmac: Targeted multi-agent communication

[Das et al., 2019a] Abhishek Das, Theophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Mike Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. In Agent1 Agent2 Agent3 Agent4 Agent5 Agent6 Agent7 Agent8 Agent9 Communication Frequency of each channel in...

2019
[2015]

Rethinking individual global max in cooperative multi- agent reinforcement learning

[Hong et al., 2022] Yitian Hong, Yaochu Jin, and Yang Tang. Rethinking individual global max in cooperative multi- agent reinforcement learning. In Advances in Neural Infor- mation Processing Systems 35: Annual Conference on Neu- ral Information Processing Systems 2022, NeurIP...

2022
[2016]

Learning structured communication for multi-agent reinforcement learning

[Sheng et al., 2023] Junjie Sheng, Xiangfeng Wang, Bo Jin, Wenhao Li, Jun Wang, Junchi Yan, Tsung-Hui Chang, and Hongyuan Zha. Learning structured communication for multi-agent reinforcement learning. In Proceedings of the 2023 International Conference on Autonomous Agents and...

2023
[2017]

Grey-box adver- sarial attack on communication in multi-agent reinforce- ment learning

[Ma and Li, 2023] Xiao Ma and Wu-Jun Li. Grey-box adver- sarial attack on communication in multi-agent reinforce- ment learning. In Proceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems, AAMAS 2023, London, United Kingdom, 29 May 2023 - 2...

2023
[2019]

Tarmac: Targeted multi-agent communication

[Das et al., 2019b] Abhishek Das, Th´eophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Mike Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, pages 1538–1546. PMLR,

2019
[2020]

A multi-agent reinforcement learning approach to robot soccer

[Duan et al., 2012] Yong Duan, Baoxia Cui, and Xinhe Xu. A multi-agent reinforcement learning approach to robot soccer. Artif. Intell. Rev., 38(3):193–211,

2012
[2021]

QPLEX: duplex dueling multi-agent q-learning

[Wang et al., 2021a] Jianhao Wang, Zhizhou Ren, Terry Liu, Yang Yu, and Chongjie Zhang. QPLEX: duplex dueling multi-agent q-learning. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7,

2021
[2022]

Graph convolutional reinforcement learn- ing

[Jiang et al., 2020] Jiechuan Jiang, Chen Dun, Tiejun Huang, and Zongqing Lu. Graph convolutional reinforcement learn- ing. In 8th International Conference on Learning Repre- sentations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30,

2020
[2023]

Robustness of decentralized decision-making architectures in command and control systems

[Boss and Gralla, 2023] Lewis N Boss and Erica L Gralla. Robustness of decentralized decision-making architectures in command and control systems. Systems Engineering, 26(2):149–161,

2023
[2024]

Learning agent commu- nication under limited bandwidth by message pruning

[Mao et al., 2020] Hangyu Mao, Zhengchao Zhang, Zhen Xiao, Zhibo Gong, and Yan Ni. Learning agent commu- nication under limited bandwidth by message pruning. In The Thirty-Fourth AAAI Conference on Artificial Intelli- gence, AAAI 2020, New York, NY, USA, February 7-12, 2020, p...

2020

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Deep reinforcement learning for truck-drone delivery problem

[Bi et al., 2023] Zhiliang Bi, Xiwang Guo, Jiacun Wang, Shu- jin Qin, and Guanjun Liu. Deep reinforcement learning for truck-drone delivery problem. Drones, 7:445, 07

work page 2023

[2] [3]

Garcia Cena, Pedro F

[Cena et al., 2013] Cecilia E. Garcia Cena, Pedro F. C´ardenas, Roque Saltar ´en Pazmi ˜no, Lisandro Puglisi, and Rafael Aracil Santonja. A cooperative multi-agent robotics system: Design and modelling. Expert Syst. Appl., 40(12):4737–4748,

work page 2013

[3] [6]

Learning individually inferred communication for multi-agent cooperation

[Ding et al., 2020] Ziluo Ding, Tiejun Huang, and Zongqing Lu. Learning individually inferred communication for multi-agent cooperation. In Advances in Neural Informa- tion Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, De- cember 6-12, 2020, virtual,

work page 2020

[4] [11]

Learning to schedule communication in multi-agent reinforcement learning

[Kim et al., 2019] Daewoo Kim, Sangwoo Moon, David Hostallero, Wan Ju Kang, Taeyoung Lee, Kyunghwan Son, and Yung Yi. Learning to schedule communication in multi-agent reinforcement learning. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9,

work page 2019

[5] [12]

Multi-agent game abstraction via graph attention neural network

[Liu et al., 2020] Yong Liu, Weixun Wang, Yujing Hu, Jianye Hao, Xingguo Chen, and Yang Gao. Multi-agent game abstraction via graph attention neural network. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, New York, NY, USA, February 7-12, 2020 , pages 7211–7218. AAAI Press,

work page 2020

[6] [13]

Na 2q: Neural attention additive model for inter- pretable multi-agent q-learning

[Liu et al., 2023] Zichuan Liu, Yuanyang Zhu, and Chunlin Chen. Na 2q: Neural attention additive model for inter- pretable multi-agent q-learning. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors,International Confer- ence on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, U...

work page 2023

[7] [14]

Multi-agent actor- critic for mixed cooperative-competitive environments

[Lowe et al., 2017] Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, Pieter Abbeel, and Igor Mordatch. Multi-agent actor- critic for mixed cooperative-competitive environments. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Sys- tems 2017, December 4-9, 2017, Long Beach, CA, USA , pages 6379–6390,

work page 2017

[8] [16]

Enhancing multi-agent system testing with diversity- guided exploration and adaptive critical state exploitation

[Ma et al., 2024] Xuyan Ma, Yawen Wang, Junjie Wang, Xi- aofei Xie, Boyu Wu, Shoubin Li, Fanjiang Xu, and Qing Wang. Enhancing multi-agent system testing with diversity- guided exploration and adaptive critical state exploitation. In Maria Christakis and Michael Pradel, editors, Proceed- ings of the 33rd ACM SIGSOFT International Symposium on Software Tes...

work page 2024