Pith. sign in

REVIEW 4 major objections 3 minor 34 references

Learning Bilateral Team Formation in Cooperative Multi-Agent Reinforcement Learning

T0 review · 4 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A stable matching algorithm for grouping agents generalizes better than an unstable score-based one in cooperative multi-agent learning.

desk verdict A useful empirical study that frames MARL team formation as bilateral stable matching, but the headline stability claim is confounded with the order-vs-score encoding difference and the reported win count does not match the tables. read the letter →

arxiv 2506.20039 v1 pith:E776ZHZ2 submitted 2025-06-24 cs.MA cs.AIcs.GTcs.LG

Koorosh Moslemi , Chi-Guhn Lee This is my paper · ORCID
classification cs.MAcs.AIcs.GTcs.LG
keywords multi-agentreinforcementlearningteamformationstablematchingdeferredacceptancevaluedecompositiongeneralizationdynamicagentpopulations
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 asks how cooperative agents should be grouped into teams when the population can grow beyond the training size. It treats team formation as a bilateral matching problem between two disjoint sets—leaders and followers—and learns each agent's preferences over partners from attention scores. The central result is empirical: policies trained with a stable matching algorithm (deferred acceptance using only the order of preferences) generalize better to unseen team sizes and leader/follower compositions than policies trained with an unstable score-based matching algorithm. A reader should care because team formation is usually either fixed or unilateral in multi-agent learning, and this work points to a principled, market-inspired criterion for choosing teams that affects how well learned behaviour transfers.

What carries the argument

The mechanism is the interpretation of multi-head attention scores between agents as a preference matrix, which is then consumed by one of two matching algorithms. OOM runs the many-to-one deferred acceptance algorithm on the ordinal preferences of leaders and followers, producing a stable matching; SOM instead sums the mutual raw scores and assigns each follower to the highest-scoring available leader, producing an unstable one. Around this, the framework adds a group-aware encoder-decoder to the agent utility network and a group-pooled hypernetwork, so that value decomposition is conditioned on the learned team structure. The stability property of deferred acceptance is the object that the paper credits for the generalization gap.

What would settle it

Run an unstable matching algorithm that consumes only the ordinal preference order, such as a random serial dictatorship over the same leader and follower rankings; if it matches OOM's generalization across the SMAC evaluation grid, then stability is not the operative property. Alternatively, a stable score-based variant that beats SOM would support the stability interpretation.

Watch

Extended reading notes

Core claim

The paper's central claim is that in cooperative multi-agent reinforcement learning with a dynamic population, the algorithmic property of the grouping mechanism—not just the grouping itself—shapes policy quality and, especially, generalization. Concretely, it compares 'Order Oriented Matching' (OOM), which applies deferred acceptance to the ranking of learned inter-agent preferences and yields a stable outcome, against 'Score Oriented Matching' (SOM), which greedily assigns followers to leaders by summed raw attention scores and yields no stability guarantee. Across the SMAC scenarios SZ, CSZ, and MMM, policies trained with OOM achieve higher or comparable training win rates and maintain a consistent edge over SOM and prior baselines when evaluated with larger teams (6–8 agents) than seen in training (3–5 agents). The paper interprets the stable matching's advantage as discouraging frequent partner switching, which stabilizes the learning signal.

Load-bearing premise

The paper's headline claim assumes that the generalization gap comes from matching stability, not from the fact that OOM uses only preference order while SOM uses raw scores—a confound that the paper itself flags as future work.

Editorial extensions

If this is right

  • If stability is the driver, then designing team formation around stable matching algorithms will be a useful principle for MARL tasks where team composition changes at execution time.
  • The framework requires only a maximum bound on the number of agents at training, not a fixed population, so learned policies carry over to larger leader and follower counts.
  • Attention scores, originally used for credit assignment in REFIL, can be repurposed as a learned preference signal for grouping, connecting multi-agent learning to matching-market learning.
  • The comparison shows that order information alone (OOM) can outperform raw score information (SOM), suggesting that the magnitude of attention scores may matter less than their ranking.
  • Because matching operates on the learned preference matrix independent of the learning paradigm, the approach can be applied to both value-based and policy-gradient methods.

Reading between the lines

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

  • The paper leaves the stability-versus-information confound open; a natural next experiment is an unstable ordinal matching, such as random serial dictatorship on the same preference order, to test whether the order information alone rather than stability explains the generalization gain.
  • If the stability interpretation holds, the OOM advantage should grow with task length or stochasticity, where frequent partner switching imposes a heavier learning cost; checking that trend would be a direct extension.
  • The same attention-as-preference pipeline could be tested on heterogeneous robot or traffic coordination domains, where leader–follower structure is explicit and the generalization target is genuinely unseen compositions.
  • A score-based but stable variant of matching, such as deferred acceptance with score-aware tie-breaking, would help separate the effect of stability from the effect of ordinal versus cardinal information.
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 / 3 minor

Summary. The paper introduces a framework for learning bilateral team formation in cooperative multi-agent reinforcement learning, in which agents are partitioned into leaders and followers and attention scores serve as learned preferences. Two matching algorithms are compared: Order Oriented Matching (OOM), based on deferred acceptance and thus stable, and Score Oriented Matching (SOM), a greedy score-based assignment that is unstable. The authors modify REFIL with an encoder-decoder utility network and a group-aware hypernetwork, plus auxiliary losses, to support dynamic populations. They evaluate on SMAC scenarios SZ, CSZ, and MMM, training with 3-5 agents and evaluating with 6-8 agents under varying leader counts. The central claim is that stable matching (OOM) generalizes better to unseen agent compositions than unstable matching (SOM).

Significance. If the stability-attribution claim held, it would offer a concrete design principle for team-formation mechanisms in dynamic MARL, a setting where bilateral matching has been little explored. The paper's framework itself is a useful step: it learns inter-agent preferences via attention, handles a variable number of agents, and directly compares two matching mechanisms within a controlled architecture. The authors report five-seed results on standard benchmark scenarios. However, the causal claim about stability is not cleanly isolated from the order-versus-score confound, and the headline empirical summary contains a numerical error. The contribution is nevertheless within the scope of the journal and can be made sound with focused revisions.

major comments (4)
  1. [Section 6] The claim "in 26 out of 27 evaluation compositions OOM consistently outperforms SOM" is not supported by the tables. Counting the evaluation cells in Tables 2-4 gives 21 cells (SZ: 7, CSZ: 7, MMM: 7), and in one of those cells—Table 3, CSZ, 3 leaders, 6 agents—SOM (0.486 ± 0.092) beats OOM (0.390 ± 0.020). The correct count appears to be 20 out of 21, not 26 out of 27. Please correct the count or define what is included, and report per-cell significance or effect sizes to support the word "consistently".
  2. [Sections 4.3 and 6] The OOM-versus-SOM comparison does not isolate stability as the causal mechanism. OOM applies deferred acceptance to the ordinal ranking of attention scores, while SOM uses a greedy assignment based on the sum of raw attention scores. These algorithms differ in two ways: the matching procedure (stable DA vs. greedy scoring) and the information transformation (rank order vs. cardinal values). Any observed generalization gap could be due to either factor. Since Contribution 1 claims stability is the cause, a matched condition is needed, e.g., score-ordered DA or an order-based greedy matching. The paper itself flags the order-versus-score issue only as future work (Section 6, on distracted attention); as written, the causal attribute is underdetermined.
  3. [Section 5, Tables 1-4] The "best compositions" in Table 1 appear to be selected per evaluation column from Tables 2-4 (e.g., OOM* in SZ at 8 agents is 0.364, which matches the L4 entry in Table 2), but the selection rule is never stated. If the best leader count is chosen using evaluation performance, the comparison against the baselines in Table 1—and the "6 out of 9 evaluation scenarios" claim—reflects selection on the test set. Please state the rule explicitly, or restrict the comparison to a pre-specified composition (e.g., |L|=2 used in training).
  4. [Tables 2-4] Many OOM-versus-SOM differences are within one standard deviation; for example, SZ with 3 leaders and 6 agents is 0.478 ± 0.037 vs. 0.475 ± 0.049, and CSZ with 3 leaders and 6 agents goes the other way (0.390 ± 0.020 vs. 0.486 ± 0.092). Since only mean and standard deviation over five seeds are reported, the strength of the generalization claim is hard to judge. Please add paired significance tests (e.g., Wilcoxon on the per-seed OOM-minus-SOM differences) or report confidence intervals.
minor comments (3)
  1. [Section 4.4, Eq. (4)] The definition of I(i,j) has a formatting error: "a i, aj ∈ gk" should be "a_i, a_j ∈ g_k"; similar subscript issues occur in the case expression. Please fix the notation.
  2. [Algorithm 1] The tie-breaking rule for followers with equal preferences is unspecified; stable matching outcomes can depend on tie-breaking. State how ties are broken (e.g., by follower index).
  3. [Section 4.1] The relationship between the mask tuple (M_µ, M_µ_O, M_µ_I) in MHA and the masks M_I, M_O used in Equations 2 and 3 is not explained; clarify whether the preference matrix is obtained with an all-ones mask.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the stable-matching generalization claim is an empirical comparison on held-out agent compositions, not a derivation that reduces to its inputs.

full rationale

The paper's central claim is empirical rather than derivational: OOM and SOM are fixed matching algorithms (deferred acceptance using preference order versus greedy assignment using raw mutual scores) applied to an attention-based preference matrix that is learned as part of a conventional MARL value-decomposition training loop. No equation in the paper defines OOM's evaluation outcome in terms of its inputs by construction; the generalization metric is measured on held-out agent counts (6-8 agents) after training on 3-5 agents, so the comparison is not a fitted-input-called-prediction pattern. The claimed role of stability is an interpretation of the observed difference between two concrete algorithms, not a parameter fitted to the evaluation data. The paper itself flags in Section 6 that OOM and SOM differ in whether they use relative order or raw attention scores, which is a genuine confound for the causal attribution to stability, but that is a limitation of the experimental design, not circular reasoning. There are no load-bearing self-citations: the method builds on REFIL, but that is an external baseline and architecture reference, and no uniqueness theorem or prior result by the authors is invoked to force the conclusion. The paper is self-contained against external benchmarks (SMAC, MIPI, REFIL, CollaQ, AQMIX, MAPPO), and the results are reported as empirical win rates rather than as consequences of an assumed definition. Therefore no circular step can be exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The framework rests on learned attention-based preferences (an unverified proxy for true inter-agent preference), a static stability concept applied to a dynamic setting, and an experimental identification assumption that conflates stability with score-versus-order information. These are domain assumptions, not fitted constants; the only true free parameters are small structural choices (lambda, leader count, group cap, team balancing). No new physical or mathematical entities are introduced.

free parameters (4)
  • Lambda (loss tradeoff) = 0.5
    Weight balancing the value loss LQ and the auxiliary factorized loss Laux (Equation 5). The paper states it is a default value from REFIL and not tuned, but it still shapes the learned embeddings and thus the matching preferences.
  • Number of leaders during training |L| = 2
    The leader count is set to 2 during training and varies in [2,4] at evaluation. This structural prior determines team capacities and the matching problem size; it is chosen by hand.
  • Maximum number of supported groups = 4
    A random one-hot vector of group numbers of size 4 is concatenated to hidden states (Section 5), limiting the number of teams the framework can represent.
  • Team size balancing rule = equal or nearly equal team sizes
    Section 5 states a team size balancing strategy is enforced across all matching algorithms. This is a hand-chosen constraint on the matching solution space rather than a learned quantity.
assumptions (5)
  • domain assumption Attention weights among agents are valid preference scores for matching.
    Section 4.1: 'we use attention weights returned by MHA as preferences among agents for team formation'. This equivalence is assumed, not derived.
  • domain assumption Single-timestep Gale-Shapley stability is a meaningful property for a dynamic team formation process with time-varying preferences.
    Algorithm 1 applies deferred acceptance to the current timestep's preferences; the paper gives no argument that a static stability concept transfers to sequentially changing preferences, though the conclusion depends on it.
  • domain assumption The outcome difference between OOM and SOM is attributable to stability rather than to the order-vs-score information difference.
    OOM consumes only the order of preferences, SOM consumes summed raw scores (Algorithms 1 and 2). The paper's central conclusion assigns the generalization gap to stability, while acknowledging the alternative in future work.
  • domain assumption Leader identity is fixed to the first |L| agents and is not learned.
    Section 5: 'in both training and evaluation, the first |L| agents are designated as leaders'. The result may not transfer when leadership is chosen differently.
  • standard math QMIX-style positive monotonic value decomposition is appropriate for this setting.
    Section 3.2 relies on the standard monotonic mixing function assumption from Rashid et al. 2018; this is accepted background, not a new claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Bilateral Team Formation in Cooperative Multi-Agent Reinforcement Learning." pith.science (2026). https://pith.science/paper/E776ZHZ2

@misc{pith2026250620039,
  author       = {Pith},
  title        = {Pith review of: Learning Bilateral Team Formation in Cooperative Multi-Agent Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E776ZHZ2}},
  note         = {Machine review of arXiv:2506.20039}
}
read the original abstract

Team formation and the dynamics of team-based learning have drawn significant interest in the context of Multi-Agent Reinforcement Learning (MARL). However, existing studies primarily focus on unilateral groupings, predefined teams, or fixed-population settings, leaving the effects of algorithmic bilateral grouping choices in dynamic populations underexplored. To address this gap, we introduce a framework for learning two-sided team formation in dynamic multi-agent systems. Through this study, we gain insight into what algorithmic properties in bilateral team formation influence policy performance and generalization. We validate our approach using widely adopted multi-agent scenarios, demonstrating competitive performance and improved generalization in most scenarios.

Figures

Figures reproduced from arXiv: 2506.20039 by the authors.

Figure 1
Figure 1. Modifications to agent utility network and hypernetwork. (a) An encoder-decoder structure [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 25 canonical work pages

  1. [1]

    Stable matching with uncertain linear preferences

    Haris Aziz, P \'e ter Bir \'o , Serge Gaspers, Ronald de Haan, Nicholas Mattei, and Baharak Rastegari. Stable matching with uncertain linear preferences. Algorithmica, 82: 0 1410--1433, 2020

    work page 2020

  2. [2]

    College admissions and the stability of marriage

    David Gale and Lloyd S Shapley. College admissions and the stability of marriage. The American Mathematical Monthly, 69 0 (1): 0 9--15, 1962

    work page 1962

  3. [3]

    Hypernetworks

    David Ha, Andrew M Dai, and Quoc V Le. Hypernetworks. In International Conference on Learning Representations, 2016

    work page 2016

  4. [4]

    Putting Gale & Shapley to Work: Guaranteeing Stability Through Learning

    Hadi Hosseini, Sanjukta Roy, and Duohan Zhang. Putting gale & shapley to work: Guaranteeing stability through learning. arXiv preprint arXiv:2410.04376, 2024

    work page Pith review arXiv 2024

  5. [5]

    Randomized Entity-wise Factorization for Multi-Agent Reinforcement Learning

    Shariq Iqbal, Christian A Schroeder de Witt, Bei Peng, Wendelin B \"o hmer, Shimon Whiteson, and Fei Sha. Ai-qmix: Attention and imagination for dynamic multi-agent reinforcement learning. arXiv preprint arXiv:2006.04222, 2020

    work page Pith review arXiv 2006

  6. [6]

    Randomized entity-wise factorization for multi-agent reinforcement learning

    Shariq Iqbal, Christian A Schroeder De Witt, Bei Peng, Wendelin B \"o hmer, Shimon Whiteson, and Fei Sha. Randomized entity-wise factorization for multi-agent reinforcement learning. In International Conference on Machine Learning, pp.\ 4596--4606. PMLR, 2021 a

    work page 2021

  7. [7]

    Supplementary material: Randomized entity-wise factorization for multi-agent reinforcement learning

    Shariq Iqbal, Christian A Schroeder de Witt, Bei Peng, Wendelin B \"o hmer, Shimon Whiteson, and Fei Sha. Supplementary material: Randomized entity-wise factorization for multi-agent reinforcement learning. International Conference on Machine Learning, 2021 b

    work page 2021

  8. [8]

    Improved bandits in many-to-one matching markets with incentive compatibility

    Fang Kong and Shuai Li. Improved bandits in many-to-one matching markets with incentive compatibility. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 13256--13264, 2024

    work page 2024

Show all 34 references
  1. [9]

    a fer, Andrew Wing Keung To, Kuan-Ho Lao, Murat Cubuktepe, Matthew Haley, Peter B \

    Aleksandar Krnjaic, Raul D Steleac, Jonathan D Thomas, Georgios Papoudakis, Lukas Sch \"a fer, Andrew Wing Keung To, Kuan-Ho Lao, Murat Cubuktepe, Matthew Haley, Peter B \"o rsting, et al. Scalable multi-agent reinforcement learning for warehouse logistics with robotic and hum...

    2024

  2. [10]

    Traffic control via connected and automated vehicles (cavs): An open-road field experiment with 100 cavs

    Jonathan W Lee, Han Wang, Kathy Jang, Nathan Lichtl \'e , Amaury Hayat, Matthew Bunting, Arwa Alanqary, William Barbour, Zhe Fu, Xiaoqian Gong, et al. Traffic control via connected and automated vehicles (cavs): An open-road field experiment with 100 cavs. IEEE Control Systems...

    2025

  3. [11]

    Self-improving reactive agents based on reinforcement learning, planning and teaching

    Long-Ji Lin. Self-improving reactive agents based on reinforcement learning, planning and teaching. Machine learning, 8: 0 293--321, 1992

    1992

  4. [12]

    Coach-player multi-agent reinforcement learning for dynamic team composition

    Bo Liu, Qiang Liu, Peter Stone, Animesh Garg, Yuke Zhu, and Anima Anandkumar. Coach-player multi-agent reinforcement learning for dynamic team composition. In International Conference on Machine Learning, pp.\ 6860--6870. PMLR, 2021

    2021

  5. [13]

    Competing bandits in matching markets

    Lydia T Liu, Horia Mania, and Michael Jordan. Competing bandits in matching markets. In International Conference on Artificial Intelligence and Statistics, pp.\ 1618--1628. PMLR, 2020

    2020

  6. [14]

    Human-level control through deep reinforcement learning

    Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518 0 (7540): 0 529--533, 2015

    2015

  7. [15]

    A concise introduction to decentralized POMDPs, volume 1

    Frans A Oliehoek, Christopher Amato, et al. A concise introduction to decentralized POMDPs, volume 1. Springer, 2016

    2016

  8. [16]

    Vast: Value function factorization with variable agent sub-teams

    Thomy Phan, Fabian Ritz, Lenz Belzner, Philipp Altmann, Thomas Gabor, and Claudia Linnhoff-Popien. Vast: Value function factorization with variable agent sub-teams. Advances in Neural Information Processing Systems, 34: 0 24018--24032, 2021

    2021

  9. [17]

    Qmix: Monotonic value function factorisation for deep multi-agent reinforcement learning

    T Rashid, M Samvelyan, C Schroeder de Witt, G Farquhar, J Foerster, and S Whiteson. Qmix: Monotonic value function factorisation for deep multi-agent reinforcement learning. In 35th International Conference on Machine Learning (ICML 2018). Journal of Machine Learning Research, 2018

    2018

  10. [18]

    The evolution of the labor market for medical interns and residents: a case study in game theory

    Alvin E Roth. The evolution of the labor market for medical interns and residents: a case study in game theory. Journal of political Economy, 92 0 (6): 0 991--1016, 1984

    1984

  11. [19]

    The redesign of the matching market for american physicians: Some engineering aspects of economic design

    Alvin E Roth and Elliott Peranson. The redesign of the matching market for american physicians: Some engineering aspects of economic design. American economic review, 89 0 (4): 0 748--780, 1999

    1999

  12. [20]

    Mikayel Samvelyan, Tabish Rashid, Christian Schroeder de Witt, Gregory Farquhar, Nantas Nardelli, Tim G. J. Rudner, Chia-Man Hung, Philiph H. S. Torr, Jakob Foerster, and Shimon Whiteson. The StarCraft Multi - Agent Challenge . CoRR, abs/1902.04043, 2019

    1902 arXiv

  13. [21]

    Multi-agent common knowledge reinforcement learning

    Christian Schroeder de Witt, Jakob Foerster, Gregory Farquhar, Philip Torr, Wendelin Boehmer, and Shimon Whiteson. Multi-agent common knowledge reinforcement learning. Advances in neural information processing systems, 32, 2019

    2019

  14. [22]

    Self-organized group for cooperative multi-agent reinforcement learning

    Jianzhun Shao, Zhiqiang Lou, Hongchang Zhang, Yuhang Jiang, Shuncheng He, and Xiangyang Ji. Self-organized group for cooperative multi-agent reinforcement learning. Advances in Neural Information Processing Systems, 35: 0 5711--5723, 2022

    2022

  15. [23]

    Complementary attention for multi-agent reinforcement learning

    Jianzhun Shao, Hongchang Zhang, Yun Qu, Chang Liu, Shuncheng He, Yuhang Jiang, and Xiangyang Ji. Complementary attention for multi-agent reinforcement learning. In International Conference on Machine Learning, pp.\ 30776--30793. PMLR, 2023

    2023

  16. [24]

    Value-decomposition networks for cooperative multi-agent learning

    Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinicius Zambaldi, Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z Leibo, Karl Tuyls, et al. Value-decomposition networks for cooperative multi-agent learning. arXiv preprint arXiv:1706.05296, 2017

    2017 arXiv

  17. [25]

    Rode: Learning roles to decompose multi-agent tasks

    Tonghan Wang, Tarun Gupta, Anuj Mahajan, Bei Peng, Shimon Whiteson, and Chongjie Zhang. Rode: Learning roles to decompose multi-agent tasks. arXiv preprint arXiv:2010.01523, 2020

    2010 arXiv

  18. [26]

    Bandit learning in many-to-one matching markets

    Zilong Wang, Liya Guo, Junming Yin, and Shuai Li. Bandit learning in many-to-one matching markets. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pp.\ 2088--2097, 2022

    2022

  19. [27]

    Mutual-information regularized multi-agent policy iteration

    Deheng Ye and Zongqing Lu. Mutual-information regularized multi-agent policy iteration. Advances in Neural Information Processing Systems, 36, 2024

    2024

  20. [28]

    Differential transformer

    Tianzhu Ye, Li Dong, Yuqing Xia, Yutao Sun, Yi Zhu, Gao Huang, and Furu Wei. Differential transformer. arXiv preprint arXiv:2410.05258, 2024

    2024 arXiv

  21. [29]

    The surprising effectiveness of ppo in cooperative multi-agent games

    Chao Yu, Akash Velu, Eugene Vinitsky, Jiaxuan Gao, Yu Wang, Alexandre Bayen, and Yi Wu. The surprising effectiveness of ppo in cooperative multi-agent games. Advances in Neural Information Processing Systems, 35: 0 24611--24624, 2022

    2022

  22. [30]

    Automatic grouping for efficient cooperative multi-agent reinforcement learning

    Yifan Zang, Jinmin He, Kai Li, Haobo Fu, Qiang Fu, Junliang Xing, and Jian Cheng. Automatic grouping for efficient cooperative multi-agent reinforcement learning. Advances in Neural Information Processing Systems, 36, 2024

    2024

  23. [31]

    Multi-agent collaboration via reward attribution decomposition

    Tianjun Zhang, Huazhe Xu, Xiaolong Wang, Yi Wu, Kurt Keutzer, Joseph E Gonzalez, and Yuandong Tian. Multi-agent collaboration via reward attribution decomposition. arXiv preprint arXiv:2010.08531, 2020

    2010 arXiv

  24. [32]

    Decentralized two-sided bandit learning in matching market

    YiRui Zhang and Zhixuan Fang. Decentralized two-sided bandit learning in matching market. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024

    2024

  25. [33]

    Field deployment of multi-agent reinforcement learning based variable speed limit controllers

    Yuhang Zhang, Zhiyao Zhang, Marcos Qui \ n ones-Grueiro, William Barbour, Clay Weston, Gautam Biswas, and Daniel Work. Field deployment of multi-agent reinforcement learning based variable speed limit controllers. arXiv preprint arXiv:2407.08021, 2024

    2024 arXiv

  26. [34]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.