Pith. sign in

REVIEW 3 major objections 5 minor 33 references

Decentralized Multi-agent Reinforcement Learning for Resilient Critical Infrastructures

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Decentralized MARL is structurally aligned with resilient critical infrastructures, but only if credit assignment and communication are solved.

desk verdict A clear, honest position paper arguing decentralized MARL is structurally aligned with infrastructure resilience, but the robustness claim in §III.B is asserted without evidence and tensions with the communication bottleneck the paper itself identifies in §IV.B. read the letter →

arxiv 2607.18359 v1 pith:DCYMT7SW submitted 2026-07-20 cs.MA cs.LG

Minghui Ding , Evangelos Pournaras This is my paper · ORCID
classification cs.MAcs.LG
keywords decentralizedmulti-agentreinforcementlearningcriticalinfrastructuresresiliencecreditassignmentcommunicationcentralizedtrainingwithexecutioninteraction-drivenadaptationadaptiveprotocols
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that decentralized multi-agent reinforcement learning is not just a distributed fallback to centralized training, but a paradigm whose properties—scalability, privacy, local autonomy, robustness to partial failures, and interaction-driven adaptation—match the operational needs of critical infrastructures like traffic networks and smart grids. Yet the paper insists that this structural alignment is conditional: without solving credit assignment (how agents attribute system-level outcomes to their own actions) and communication (how agents coordinate under realistic bandwidth, latency, and reliability limits), decentralized MARL will not deliver resilience in practice. It then proposes a research agenda centered on structure-aware, causality-aware, and resilience-aware credit assignment, on communication serving both coordination and attribution, and on safe, timely, and recoverable learning under deployment constraints. A sympathetic reader would take away that the value of decentralized MARL for infrastructure resilience depends on whether these two bottlenecks are overcome, not on decentralization alone.

What carries the argument

The central object is the structural correspondence argument between decentralized MARL's properties and critical infrastructure requirements, developed by decomposing each agent's learning cycle into three stages: local interaction and learning, information exchange, and aggregation and adaptation. This decomposition shows how peer-to-peer parameter exchange and weighted averaging let agents coordinate without centralized data aggregation. The two load-bearing mechanisms are credit assignment—attributing system-level rewards to individual actions despite strong coupling between agents—and adaptive communication—selective, timely exchanges that support both coordination during execution and

What would settle it

A concrete falsifying observation would be a realistic infrastructure testbed, such as a traffic network with a mid-corridor incident, where a decentralized MARL controller using standard parameter-sharing and local rewards fails to recover system-wide throughput because agents cannot attribute congestion changes to their own actions, or because adaptive communication fails under bandwidth limits—showing that structural alignment does not translate to resilience without additional mechanisms.

Watch

Extended reading notes

Core claim

The paper's central claim is that decentralized MARL should be understood as a paradigm structurally aligned with the requirements of resilient critical infrastructures. It establishes this by analyzing how decentralized MARL's properties—scalability to many agents, selective information exchange that supports privacy and local autonomy, robustness to partial failures without a central point of failure, and learning driven by mutual interaction through a shared environment—correspond to the operational demands of critical infrastructures. The paper further distinguishes decentralized MARL from decentralized federated learning, arguing that only interaction-driven learning can capture how dis

Load-bearing premise

The paper assumes that the properties it attributes to decentralized MARL—scalability, privacy-preserving selective exchange, robustness under partial failures, and interaction-driven adaptation—are intrinsic to the paradigm itself, rather than contingent on specific algorithms, environments, or deployment conditions; this alignment is argued qualitatively without empirical evidence.

Editorial extensions

If this is right

  • If the alignment argument holds, critical infrastructure operators should treat decentralized MARL as a candidate architecture for distributed control of large-scale, privacy-sensitive systems, rather than defaulting to centralized training.
  • The distinction from decentralized federated learning implies that infrastructure resilience research must prioritize learning procedures that let agents adapt to each other's changing behavior, not just share data or model updates.
  • Credit assignment mechanisms should exploit infrastructure topology—limiting attribution to physically or functionally connected agents—and should explicitly reward resilience-relevant outcomes such as preserving essential functions or supporting recovery.
  • Communication must be designed as a constrained resource whose content, timing, recipients, and usage adapt as agents' policies evolve, serving both coordination and credit assignment.
  • Evaluation of decentralized MARL for infrastructure must include disruption scenarios such as communication failures, partial controller failures, abnormal operating conditions, and malicious attacks, measuring safety, timeliness, and recoverability rather than only ideal-condition performance.
  • Simple safety and fallback mechanisms—conservative actions, smaller policy updates, and reliance on trusted local information—will be needed before decentralized MARL can be deployed in real infrastructures.

Reading between the lines

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

  • Editorial extension: If credit assignment and communication are truly the binding constraints, then progress in decentralized MARL for infrastructure will be driven more by advances in representation learning for causal inference and by adaptive communication protocols than by scaling up existing CTDE algorithms.
  • Editorial extension: The paper's distinction between decentralized MARL and decentralized federated learning suggests a testable hypothesis: in infrastructure domains where disturbances do not propagate through component interactions, decentralized federated learning may match decentralized MARL, whereas in tightly coupled domains like traffic or power networks, interaction-driven learning should
  • Editorial extension: One could derive a concrete benchmark—a traffic corridor with an induced incident—where a decentralized MARL agent that uses topology-aware credit assignment should recover faster and avoid congestion spread compared to one using only local rewards; such an experiment would directly probe the paper's central conditions.
  • Editorial extension: The emphasis on resilience-aware credit assignment implies that reward design itself is insufficient; a system that correctly assigns credit for local efficiency may still fail to learn recovery behaviors unless the reward signal explicitly accounts for recovery and essential-function preservation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This position paper argues that decentralized multi-agent reinforcement learning (MARL) should be viewed as structurally aligned with the requirements of resilient critical infrastructures, beyond being a distributed variant of centralized training with decentralized execution (CTDE). It identifies four aligned properties—scalability, privacy/local autonomy, robustness, and interaction-driven adaptation—distinguishes decentralized MARL from decentralized federated learning, and then argues that credit assignment and communication are the two central conditions for practical feasibility. It concludes with a three-part research agenda: structure-/causality-/resilience-aware credit assignment, communication for both coordination and credit assignment, and safe/timely/recoverable learning. The paper is conceptual and does not present experiments or quantitative comparisons.

Significance. The paper fills a gap in the MARL-for-infrastructure literature by offering a principled framing rather than an algorithm. Its distinction between decentralized MARL and federated learning is insightful, and the proposed research agenda is concrete. If the structural-alignment thesis holds, it could influence how the community evaluates decentralized MARL. However, the core claims are stated without empirical support, and one internal tension undercuts the robustness argument. The paper's value is therefore as a research manifesto that needs validation.

major comments (3)
  1. [III.B vs. IV.B] Section III.B asserts that because decentralized MARL policies rely only on local observations and limited neighbor communication, the system 'can degrade more gracefully under local disturbances, disrupted communication, and partial failures.' This conflates distributed execution with robustness to communication loss. The methods cited in Section II.B ([15]–[18]) rely on ongoing exchange of policy/value parameters during training; if those links fail, the learning updates themselves are compromised. Section IV.B then states that communication is 'a fundamental bottleneck' and that impaired communication causes agents to act on stale/partial knowledge that 'can slow adaptation, allow local disruptions to propagate, and hinder recovery.' The two sections are in tension. Since robustness is one of the four claimed structural alignments, the paper should either soften the III.B claim to a c
  2. [III.A] The abstract and Section III.A state that decentralized MARL 'scales naturally to large numbers of agents.' This is asserted without supporting evidence or analysis. Decentralized learning can also suffer from increasing communication overhead as the number of neighbors grows, and from non-stationarity that increases with the number of interacting agents. Whether the approach scales depends on the specific algorithm, network topology, and communication budget; the cited references [19], [20] do not directly demonstrate this claim for resilient critical infrastructures. The paper should qualify the claim or cite comparative scalability studies.
  3. [IV] Section IV introduces 'two challenges are especially central'—credit assignment and communication—but no explicit criterion is given for why these two outrank e.g. safety, sample efficiency, or non-stationarity. The paper's own Section V adds a third direction on safety/timeliness/recoverability, which suggests the framing as exactly two central conditions is not essential. The authors should explain their selection criterion, or explicitly present the third as a boundary condition rather than an additional challenge.
minor comments (5)
  1. [II.B / figures] The figure captions contain formatting issues: 'agenti∈ {1, . . . , N}selectsan actionai' and 'agentiinteracts' should be 'agent i', 'action a_i', etc. The LaTeX spacing has been lost in the PDF.
  2. [References] Reference [15]: 'T. Bas ¸ar' should be 'T. Başar' or 'T. Basar' with no space between the stem and the diacritic.
  3. [III.B] The phrase 'can degrade more gracefully' is a quantitative claim without a metric or threshold. Suggest replacing with 'may degrade gracefully' or define a measure (e.g., control performance under link failure).
  4. [IV.A] The traffic-corridor example is illustrative but does not provide quantitative evidence that incorrect credit assignment prevents recovery. A short example from the literature or a reference to an empirical study would strengthen the motivation.
  5. [V.A] The proposal to 'construct localized counterfactual estimates' is plausible but leaves open how this is achieved in a fully decentralized setting where agents lack access to the joint action. A brief discussion of existing counterfactual methods (e.g., COMA) and their adaptation would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is a qualitative alignment thesis with no derivation chain that reduces to its own inputs.

full rationale

The paper argues that decentralized MARL is structurally aligned with resilient critical infrastructures by identifying shared properties (scalability, privacy, robustness, interaction-driven adaptation). This is a substantive conceptual argument, not a derivation in which an output is equivalent to an input by construction. No fitted parameters are presented as predictions, no uniqueness theorem is imported to force a choice, and no ansatz is smuggled in through self-citation. The self-citations by the authors (e.g., [5], [6], [21], [23], [25], [27], [33]) are contextual references to prior work on cascading failures, decentralized learning, and smart grids; they do not bear the logical weight of the central claim. The potential tension between the Section III.B claim that decentralized MARL 'can degrade more gracefully' under communication disruption and the Section IV.B warning that communication reliability 'cannot be assumed' is an empirical and consistency concern about the strength of the robustness assumption, not a circularity: the claims are distinct and testable rather than definitionally equivalent. The paper is self-aware that structural alignment alone is insufficient, explicitly conditioning its thesis on solving credit assignment and communication. Thus no circular step can be quoted and exhibited, and the appropriate finding is no significant circularity.

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

The paper has no free parameters and invents no new entities. Its argument rests on several domain assumptions about the nature of critical infrastructure resilience and the inherent properties of decentralized MARL. These assumptions are plausible but unverified.

assumptions (4)
  • domain assumption Critical infrastructures require resilience, defined as the ability to sustain or recover essential functions under disruptions.
    Used in the introduction to motivate the entire argument. It is a reasonable but unstated value judgment about what resilience means.
  • domain assumption Decentralized MARL inherently possesses properties of scalability, privacy preservation, robustness to failures, and interaction-driven adaptation.
    Section III asserts these properties as inherent to the paradigm, but they are not proven or empirically demonstrated. The argument depends on this being true.
  • domain assumption Resilience in interdependent infrastructures depends on interaction-driven adaptation among components, not just isolated local optimization.
    Section III.C introduces this to distinguish MARL from federated learning. It is a plausible but not established premise.
  • domain assumption Credit assignment and communication are the two central bottlenecks for practical decentralized MARL.
    Section IV argues this qualitatively, but the centrality of these two challenges is asserted rather than derived from a formal analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decentralized Multi-agent Reinforcement Learning for Resilient Critical Infrastructures." pith.science (2026). https://pith.science/paper/DCYMT7SW

@misc{pith2026260718359,
  author       = {Pith},
  title        = {Pith review of: Decentralized Multi-agent Reinforcement Learning for Resilient Critical Infrastructures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DCYMT7SW}},
  note         = {Machine review of arXiv:2607.18359}
}
read the original abstract

Critical infrastructures are increasingly distributed, interdependent, and exposed to evolving disruptions, making resilience a central requirement for their operation and control. This paper argues that decentralized multi-agent reinforcement learning (MARL) should be understood not merely as a distributed alternative to centralized training with decentralized execution but as a paradigm structurally aligned with the requirements of resilient critical infrastructures. This perspective is grounded in an analysis of the properties of decentralized MARL and the requirements of critical infrastructures, including scalability to large numbers of agents, support for privacy and local autonomy, robustness to failures, and interaction-driven adaptation among interdependent components. However, structural alignment alone is insufficient for practical deployment. This paper identifies credit assignment and communication as two central conditions for its practical feasibility. Credit assignment determines whether local learning remains aligned with system-level objectives, while communication determines whether coordination can be learned and maintained under realistic operational constraints. Building on these challenges, this paper proposes a research agenda focused on structure-aware, causality-aware, and resilience-aware credit assignment; communication for both coordination and credit assignment; and safe, timely, and recoverable decentralized learning under deployment constraints. Overall, this paper reframes decentralized MARL as a promising but conditional foundation for resilient critical infrastructures.

Figures

Figures reproduced from arXiv: 2607.18359 by the authors.

Figure 2
Figure 2. Local Interaction and Learning: During training, each agent learns a local decision-making rule that maps its own observations to actions, using feedback from the environment. Through repeated interactions, the agent collects sequences of local observations, selected actions, and received rewards, and uses this experience to improve its decision rule over time. In practice, this decision rule is typically represente… view at source ↗
Figure 1
Figure 1. Decentralized MARL with peer-to-peer communication. During [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Iterative learning cycle of an individual agent in decentralized MARL. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figures from the paper (1 more)
Figure 3
Figure 3. Figure 3: Decentralized federated learning over distributed local datasets. Each [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 1 canonical work pages

  1. [15]

    Fully decentralized multi-agent reinforcement learning with networked agents,

    K. Zhang, Z. Yang, H. Liu, T. Zhang, and T. Bas ¸ar, “Fully decentralized multi-agent reinforcement learning with networked agents,” inInterna- tional Conference on Machine Learning, 2018, pp. 5872–5881

    2018

  2. [18]

    Neighbor-based decentralized training strategies for multi-agent reinforcement learning,

    N. Malucelli, D. Domini, G. Aguzzi, and M. Viroli, “Neighbor-based decentralized training strategies for multi-agent reinforcement learning,” inProceedings of the 40th ACM/SIGAPP Symposium on Applied Com- puting, 2025, pp. 1250–1257

    2025

  3. [19]

    Multi-agent reinforcement learning: A selective overview of theories and algorithms,

    K. Zhang, Z. Yang, and T. Bas ¸ar, “Multi-agent reinforcement learning: A selective overview of theories and algorithms,” inHandbook of Reinforcement Learning and Control. Springer, 2021, pp. 321–384

    2021

  4. [20]

    Efficient and scalable reinforcement learning for large-scale network control,

    C. Ma, A. Li, Y . Du, H. Dong, and Y . Yang, “Efficient and scalable reinforcement learning for large-scale network control,”Nature Machine Intelligence, vol. 6, no. 9, pp. 1006–1020, 2024

    2024

  5. [1]

    Multi-agent deep reinforcement learning for large-scale traffic signal control,

    T. Chu, J. Wang, L. Codec `a, and Z. Li, “Multi-agent deep reinforcement learning for large-scale traffic signal control,”IEEE Transactions on Intelligent Transportation Systems, vol. 21, no. 3, pp. 1086–1095, Mar. 2019

    2019

  6. [2]

    Multi-agent reinforcement learning for active voltage control on power distribution networks,

    J. Wang, W. Xu, Y . Gu, W. Song, and T. C. Green, “Multi-agent reinforcement learning for active voltage control on power distribution networks,” inAdvances in Neural Information Processing Systems, vol. 34, 2021, pp. 3271–3284

    2021

  7. [3]

    Deep reinforcement learning challenges and opportunities for urban water systems,

    A. Negm, X. Ma, and G. Aggidis, “Deep reinforcement learning challenges and opportunities for urban water systems,”Water Research, vol. 253, p. 121145, 2024

    2024

  8. [4]

    Guide to industrial control systems (ICS) security,

    K. Stouffer, J. Falco, and K. Scarfone, “Guide to industrial control systems (ICS) security,” National Institute of Standards and Technology, NIST Special Publication 800-82, 2011

    2011

Show all 33 references
  1. [5]

    Cascading failures in interconnected power-to-water networks,

    E. Pournaras, R. Taormina, M. Thapa, S. Galelli, V . Palleti, and R. Kooij, “Cascading failures in interconnected power-to-water networks,”ACM SIGMETRICS Performance Evaluation Review, vol. 47, no. 4, pp. 16– 20, 2020

    2020

  2. [6]

    Measuring network relia- bility and repairability against cascading failures,

    M. Thapa, J. Espejo-Uribe, and E. Pournaras, “Measuring network relia- bility and repairability against cascading failures,”Journal of Intelligent Information Systems, vol. 52, no. 3, pp. 573–594, 2019

    2019

  3. [7]

    A game-theoretical approach to cyber-security of critical infrastructures based on multi-agent reinforcement learning,

    M. Panfili, A. Giuseppi, A. Fiaschetti, H. B. Al-Jibreen, A. Pietrabissa, and F. D. Priscoli, “A game-theoretical approach to cyber-security of critical infrastructures based on multi-agent reinforcement learning,” in2018 26th Mediterranean Conference on Control and Automation...

    2018

  4. [8]

    Prioritizing postdisaster recovery of transportation infrastructure systems using multiagent reinforcement learning,

    P. Ghannad, Y .-C. Lee, and J. O. Choi, “Prioritizing postdisaster recovery of transportation infrastructure systems using multiagent reinforcement learning,”Journal of Management in Engineering, vol. 37, no. 1, p. 04020100, 2021

    2021

  5. [9]

    Review on modeling and simulation of interdependent critical infrastructure systems,

    M. Ouyang, “Review on modeling and simulation of interdependent critical infrastructure systems,”Reliability Engineering & System Safety, vol. 121, pp. 43–60, 2014

    2014

  6. [10]

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

    R. Lowe, Y . I. Wu, A. Tamar, J. Harb, P. Abbeel, and I. Mordatch, “Multi-agent actor-critic for mixed cooperative-competitive environ- ments,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    2017

  7. [11]

    Cross-border information sharing for critical infrastructure resilience: Requirements and platform architecture,

    B. Petrenj, M. Piraina, G. Feletti, P. Trucco, V . Urbano, and S. Gelmi, “Cross-border information sharing for critical infrastructure resilience: Requirements and platform architecture,” inInternational Conference on Information Systems for Crisis Response and Management, 202...

    2021

  8. [12]

    Asynchronous methods for deep rein- forcement learning,

    V . Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu, “Asynchronous methods for deep rein- forcement learning,” inInternational Conference on Machine Learning, 2016, pp. 1928–1937

    2016

  9. [13]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” 2017, arXiv:1707.06347. [Online]. Available: https://doi.org/10.48550/arXiv.1707.06347

    2017 doi arXiv

  10. [14]

    Human-level control through deep reinforcement learning,

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

    2015

  11. [16]

    A communication-efficient decentralized actor-critic algorithm,

    X. Ren, N. Bastianello, T. Parisini, and A. A. Malikopou- los, “A communication-efficient decentralized actor-critic algorithm,” 2025, arXiv:2510.19199. [Online]. Available: https://doi.org/10.48550/arXiv.2510.19199

    2025 doi

  12. [17]

    Communication-efficient actor-critic methods for homogeneous markov games,

    D. Chen, Y . Li, and Q. Zhang, “Communication-efficient actor-critic methods for homogeneous markov games,” inInternational Conference on Learning Representations, 2022

    2022

  13. [21]

    Holarchic struc- tures for decentralized deep learning: A performance analysis,

    E. Pournaras, S. Yadhunathan, and A. Diaconescu, “Holarchic struc- tures for decentralized deep learning: A performance analysis,”Cluster Computing, vol. 23, no. 1, pp. 219–240, 2020

    2020

  14. [22]

    Security and privacy challenges in the smart grid,

    P. McDaniel and S. McLaughlin, “Security and privacy challenges in the smart grid,”IEEE Security & Privacy, vol. 7, no. 3, pp. 75–77, 2009

    2009

  15. [23]

    Decentralized collective learning for self-managed sharing economies,

    E. Pournaras, P. Pilgerstorfer, and T. Asikis, “Decentralized collective learning for self-managed sharing economies,”ACM Transactions on Autonomous and Adaptive Systems (TAAS), vol. 13, no. 2, pp. 1–33, 2018

    2018

  16. [24]

    Resilience of critical infras- tructure elements and its main factors,

    D. Rehak, P. Senovsky, and S. Slivkova, “Resilience of critical infras- tructure elements and its main factors,”Systems, vol. 6, no. 2, p. 21, 2018

    2018

  17. [25]

    Optimization under attack: Resilience, vulnerability, and the path to collapse,

    A. Aldawsari and E. Pournaras, “Optimization under attack: Resilience, vulnerability, and the path to collapse,”Future Generation Computer Systems, p. 108017, 2025

    2025

  18. [26]

    Learning decentralized traffic signal controllers with multi- agent graph reinforcement learning,

    Y . Zhang, Z. Yu, J. Zhang, L. Wang, T. H. Luan, B. Guo, and C. Yuen, “Learning decentralized traffic signal controllers with multi- agent graph reinforcement learning,”IEEE Transactions on Mobile Computing, vol. 23, no. 6, pp. 7180–7195, Jun. 2023

    2023

  19. [27]

    Self-repairable smart grids via online coordination of smart transformers,

    E. Pournaras and J. Espejo-Uribe, “Self-repairable smart grids via online coordination of smart transformers,”IEEE Transactions on Industrial Informatics, vol. 13, no. 4, pp. 1783–1793, 2016

    2016

  20. [28]

    A survey on federated learning,

    C. Zhang, Y . Xie, H. Bai, B. Yu, W. Li, and Y . Gao, “A survey on federated learning,”Knowledge-Based Systems, vol. 221, p. 106775, 2021

    2021

  21. [29]

    Gossip learning as a decen- tralized alternative to federated learning,

    I. Heged ´us, G. Danner, and M. Jelasity, “Gossip learning as a decen- tralized alternative to federated learning,” inDistributed Applications and Interoperable Systems: 19th IFIP WG 6.1 International Conference, DAIS 2019, Proceedings, ser. Lecture Notes in Computer Science. ...

    2019

  22. [30]

    Counterfactual multi-agent policy gradients,

    J. N. Foerster, G. Farquhar, T. Afouras, N. Nardelli, and S. Whiteson, “Counterfactual multi-agent policy gradients,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 32, no. 1, 2018

    2018

  23. [31]

    Assign- ing credit with partial reward decoupling in multi-agent proximal policy optimization,

    A. Kapoor, B. Freed, H. Choset, and J. Schneider, “Assign- ing credit with partial reward decoupling in multi-agent proximal policy optimization,” 2024, arXiv:2408.04295. [Online]. Available: https://arxiv.org/abs/2408.04295

    2024 arXiv

  24. [32]

    A survey of multi-agent deep reinforcement learning with communication,

    C. Zhu, M. Dastani, and S. Wang, “A survey of multi-agent deep reinforcement learning with communication,”Autonomous Agents and Multi-Agent Systems, vol. 38, no. 1, p. 4, 2024

    2024

  25. [33]

    Discrete-choice multi-agent optimization: Decentralized hard constraint satisfaction for smart cities,

    S. Majumdar, C. Qin, and E. Pournaras, “Discrete-choice multi-agent optimization: Decentralized hard constraint satisfaction for smart cities,” inInternational Conference on Autonomous Agents and Multiagent Systems. Springer, 2023, pp. 60–76

    2023

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.