Pith. sign in

REVIEW 2 major objections 5 minor 29 references

MUTE unlearns low-return messages so multi-agent teams keep almost full performance at a fraction of the bandwidth.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 02:14 UTC pith:F5BR6HA3

load-bearing objection Solid systems paper: post-hoc value-guided unlearning from a dense expert beats joint MI/gating training for sparse MARL communication, with real multi-benchmark evidence and a conditional bound. the 2 major comments →

arxiv 2607.03473 v1 pith:F5BR6HA3 submitted 2026-07-03 cs.MA

MUTE: Return-Preserving Communication Unlearning for Efficient Multi-Agent Coordination

Rui Zuo , Qinwei Huang , Mingyang Li , Zhenhang Zhang , Simon Khan , Qinru Qiu This is my paper
classification cs.MA
keywords multi-agent reinforcement learningcommunication efficiencymachine unlearningcounterfactual message valuevalue-guided sparsityDec-POMDPbehavioral anchoring
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Cooperative multi-agent reinforcement learning often needs agents to talk under partial observability, but real bandwidth limits make dense messaging impractical. Most prior methods compress or gate messages by information-theoretic proxies such as mutual information; those proxies can keep statistically salient messages that do not actually raise team return. MUTE instead first trains an unrestricted expert policy, then treats communication reduction as machine unlearning: it estimates each message’s Counterfactual Message Value (the drop in joint action-value if that message is masked) and systematically unlearns only the low-value channels. A dual loss enforces sparsity while anchoring agents to the original expert’s local value functions, and a theorem bounds how much return can fall. Across Hallway, Traffic Junction, SMAC and SMACv2 maps the method typically cuts communication 80–90 percent (sometimes near 99 percent) while matching or nearly matching full-communication baselines.

Core claim

Communication reduction in cooperative MARL can be cast as return-preserving unlearning: after training an unrestricted joint policy, an attention-based Message Value Estimator scores each message by its Counterfactual Message Value, low-value messages are driven toward zero by an L1 sparsity loss, and a behavioral-anchoring loss keeps individual Q-functions close to the expert, yielding a controllable performance gap and large bandwidth savings.

What carries the argument

Counterfactual Message Value (CMV) — the difference in joint action-value with versus without a given message — estimated in one attention pass by the Message Value Estimator and used to select the redundant set that is unlearned under the dual sparsity-plus-anchoring objective of Theorem 4.2.

Load-bearing premise

The joint value function is assumed smooth enough in the continuous message embeddings, and single-message counterfactual scores plus anchoring are assumed to catch the messages that matter even when several messages interact.

What would settle it

On a map where several messages are jointly necessary but each alone looks low-value, check whether MUTE’s estimated Counterfactual Message Values still keep at least one of them and whether the observed return gap stays inside the bound of Theorem 4.2; a large, unexplained drop would falsify the claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper introduces MUTE, a three-stage framework that first trains an unrestricted full-communication MARL expert (value-based CTDE backbone), then fits an attention-based Message Value Estimator to approximate Counterfactual Message Value (CMV = difference in Q*_tot with vs. without a given message), and finally unlearns low-CMV messages via the dual objective L_sparse (L1 on messages below an adaptive threshold) + β L_anchor (behavioral anchoring of individual Q_i to the frozen expert). A performance-difference bound (Theorem 4.2) is derived under L-smoothness of Q*_tot and bounded estimation errors. Empirically, MUTE is evaluated on Hallway, Traffic Junction, SMAC, SMAC-Communication and SMACv2, reporting 80–99 % communication reduction while matching or nearly matching full-communication baselines, with dropout-AUC robustness, component ablations, hyper-parameter sweeps and empirical checks that estimation errors remain small.

Significance. If the empirical results hold under broader scrutiny, MUTE offers a practically useful alternative to information-theoretic or gating-based communication sparsification: by decoupling expert coordination learning from subsequent value-guided pruning, it avoids the unstable multi-objective trade-offs that plague joint training. The CMV estimator, dual unlearning loss, and explicit return-gap bound constitute clear technical contributions relative to SMS (Shapley), NDQ/IMAC (mutual information) and IC3Net/TarMAC. Strengths that raise the paper above typical MARL communication work include the multi-environment evaluation with 95 % CIs over five seeds, the dropout-AUC table, the budget-allocation and L1-vs-L2 ablations, the empirical verification of the bound assumptions (Appendix I), and the release of anonymous code. These elements make the central claim falsifiable and reproducible.

major comments (2)
  1. [Theorem 4.2, Lemma A.2, §6] Theorem 4.2 / Lemma A.2 and Assumption A.1: the return-gap certificate rests on L-smoothness of the joint action-value with respect to the continuous concatenated message vector together with a first-order Taylor remainder that bounds multi-message interaction residuals. The Limitations section already acknowledges that single-message CMV does not enumerate subsets. Because the bound is presented as a guarantee of controlled degradation, an additional empirical diagnostic (e.g., measured return gap under simultaneous masking of the entire M_red set versus the sum of individual CMVs on held-out trajectories) would make the theoretical claim more robust and would directly address the higher-order interaction concern.
  2. [Algorithm 1, §4.3, Appendix I] Algorithm 1 and §4.3: the Message Value Estimator is trained once on the unrestricted expert and then frozen; during the unlearning phase only θ and ψ are updated. Although Appendix I reports that “MVE loss” stays small, it is not stated whether this loss is recomputed against fresh counterfactual targets under the evolving (sparsified) message distribution or merely the original Phase-1 targets. Clarifying this point, and if necessary re-estimating CMV periodically or with a small online update, is load-bearing for the claim that the messages being driven to zero remain the truly low-value ones.
minor comments (5)
  1. [Abstract, §5.2] Abstract and §5.2 claim “80 % to 90 %” reduction while the Hallway results and text repeatedly cite up to 99 %. Align the abstract wording with the strongest reported figures.
  2. [Title page] Author email for Qinru Qiu is listed as qiqiu@syr.com (missing .edu). Correct for reproducibility.
  3. [Appendix K] Appendix K title contains the typo “Compuatational”. Fix throughout.
  4. [Figure 2, §5.1] Figure 2 caption and main text should more explicitly remind the reader that the triangles mark the 2 M-step expert performance that MUTE subsequently sparsifies; this avoids any residual confusion about training-budget fairness.
  5. [Definition 4.1] In Definition 4.1 the notation Q*_tot(τ, m) ≔ Q*_tot(τ, a) is convenient but slightly overloads the second argument; a short clarifying sentence would help readers.

Circularity Check

0 steps flagged

No significant circularity: MUTE trains an unrestricted expert first, estimates CMV from frozen counterfactual Q* queries, then unlearns under an independent anchoring regularizer; the performance bound is conditional on stated assumptions rather than tautological.

full rationale

The derivation chain is sequential and non-circular. An unrestricted joint policy and critic Q*_tot are first obtained by ordinary TD training (Eq. 1). Counterfactual Message Value targets are then computed by explicit masking-and-resampling queries against that frozen critic (Definition 4.1 / Eq. 3), and an attention MVE is fitted to those targets (Eq. 5). Unlearning subsequently minimizes an L1 sparsity penalty on the low-CMV set together with a separate behavioral-anchoring term that keeps individual Q_i close to the frozen expert (Eqs. 6–8). The performance gap bound (Theorem 4.2) follows from the Performance Difference Lemma plus an L-smoothness assumption and bounded estimation errors; it is an upper bound under those hypotheses, not an identity that forces the reported win rates. Hyper-parameters λ and β are selected by ablation (Appendix H), not fitted so as to reproduce the final metrics. Self-citations are to standard external MARL baselines (QMIX, MAPPO, MASIA backbone, COMA-style counterfactuals) and do not load-bear the central claim. The Limitations section already notes that single-message CMV does not enumerate higher-order subsets; that is a genuine modeling gap, not circularity. Empirical checks (Appendix I, Figure 4) further show that the bounded-error and anchoring conditions hold in practice. Consequently the paper’s strongest claim—value-guided unlearning yields large bandwidth reduction with near-baseline return—is supported by independent experimental evidence rather than by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 3 invented entities

The central claim rests on standard Dec-POMDP/CTDE machinery plus three paper-specific modeling choices: L-smoothness of Q in message space, bounded MVE and critic errors, and the adequacy of single-message CMV plus anchoring for multi-message redundancy. Free knobs λ, β, and the 2M/0.5M/1.5M schedule materially affect reported sparsity–return tradeoffs. Invented constructs are CMV, MVE, and the MUTE unlearning objective—tools rather than new physical entities, with only internal empirical handles.

free parameters (4)
  • sparsity threshold multiplier λ (default 1.0 · μ_MVE)
    Defines M_red and thus which messages receive L1 pressure; Appendix H shows win rate and comm rate swing sharply for λ ∈ {0.1,0.5,2.0}.
  • anchoring weight β (default 10)
    Balances L_sparse vs L_anchor; β=1 collapses performance, β=50 resists unlearning (Tables 7–8).
  • three-stage step budget (2M pretrain / 0.5M MVE / 1.5M unlearn)
    Appendix M shows early unlearning (0.5M/1M pretrain) destroys win rate under fixed 4M total budget; schedule is load-bearing for the reported numbers.
  • EMA momentum τ_ema = 0.01 for μ_MVE
    Adaptive threshold baseline for redundancy; listed as a training hyperparameter without external calibration.
axioms (4)
  • ad hoc to paper Joint action-value Q*_tot is continuously differentiable and L-smooth in the concatenated continuous message vector (Assumption A.1).
    Used to bound the multi-message interaction remainder in Lemma A.2 and close Theorem 4.2; not a standard Dec-POMDP axiom.
  • domain assumption MVE error |V_ϕ(m)−ΔQ(m)| ≤ δ and critic gap ||Q*_tot−Q_tot|| ≤ ϵ remain bounded during unlearning.
    Stated as standard MARL practice; partially checked empirically in Appendix I but not proved.
  • domain assumption Cooperative task is a Dec-POMDP under CTDE with broadcast messages and value factorization (QMIX-style Q_tot).
    Problem formulation §3; standard in the literature the paper builds on.
  • standard math Performance Difference Lemma applies to the unlearned vs expert joint policies.
    Invoked at the start of the Theorem 4.2 proof (Kakade & Langford).
invented entities (3)
  • Counterfactual Message Value (CMV) ΔQ_i no independent evidence
    purpose: Scalar score of a broadcast message’s marginal contribution to joint return via masked counterfactual actions.
    Defined in Def. 4.1 / Eq. 3; internal to the method, validated only by downstream win-rate and dropout AUC.
  • Message Value Estimator (attention MVE V_ϕ) no independent evidence
    purpose: Single-forward-pass approximation of all agents’ CMVs to avoid O(N(N−1)) critic queries.
    Trained with MSE to CMV targets (Eq. 5); no external benchmark of CMV accuracy beyond task metrics.
  • MUTE dual-objective unlearning loss L_sparse + β L_anchor no independent evidence
    purpose: Progressively suppress low-CMV messages while anchoring individual Q_i to the frozen expert.
    Eqs. 6–8; success measured only inside the paper’s training loop.

pith-pipeline@v1.1.0-grok45 · 29742 in / 3674 out tokens · 32062 ms · 2026-07-12T02:14:28.294781+00:00 · methodology

0 comments
read the original abstract

Inter-agent communication is critical for coordinating Multi-Agent Reinforcement Learning (MARL) agents under partial observability to perform effectively in cooperative games; however, real-world bandwidth constraints demand sparse interactions. Prior approaches primarily address this trade-off by optimizing information-theoretic surrogates. We argue that these statistical proxies are fundamentally misaligned with the true objective: a message can be highly informative yet irrelevant to the joint return of the task. In this work, we propose Message Unlearning for Targeted Efficiency (MUTE), a framework that views communication reduction as a value-guided machine unlearning problem. MUTE rigorously quantifies the Counterfactual Message Value using an attention-based estimator, and systematically unlearns the transmission of low-value messages from a policy trained without any communication constraints. This is achieved through a dual-objective mechanism that enforces communication sparsity while preserving the return of the original joint policy. We derive a theoretical upper bound on the performance gap induced by this sparsification, guaranteeing controlled return degradation. We also empirically evaluate MUTE on various complex multi-agent environments, achieving 80% to 90% bandwidth reduction while maintaining performance comparable to state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2607.03473 by Mingyang Li, Qinru Qiu, Qinwei Huang, Rui Zuo, Simon Khan, Zhenhang Zhang.

Figure 1
Figure 1. Figure 1: Impact of communication reduc￾tion on performance during MARL training. The performance degradation highlights the compromise in coordination when communi￾cation is restricted indiscriminately. Moreover, as discussed in Section 1, by optimizing information-theoretic objectives without directly tar￾geting the joint return, these methods [13, 5, 14] risk prioritizing messages that are merely informative over… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of performance and communication efficiency against baselines across 10 [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of test win rate under varying message dropout rates against baselines across [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study of w/o Anchoring on SMAC MMM2 (left) and w/o MVE on 1o_10b_vs_1r (right). The w/o MVE variant was pre-trained for 2M steps and 0.5M steps for unlearning; while MUTE trained for extra 0.5M steps for MVE, that causes the curve to shift by 0.5M steps. w/o MVE fails to maintain perfor￾mance as sparsity increases. While MUTE maintains a win rate of ≈80% even after the communication rate drops sig… view at source ↗
Figure 5
Figure 5. Figure 5: Experimental environments used in this work. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Training curves of Win Rate and Communication Rate between the original message-only [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Training curves of Win Rate and Communication Rate between the MUTE and [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of performance and communication [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Training curves of Mean Test Win Rate and Communication Rate for different budget [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 2 linked inside Pith

  1. [1]

    Multi-agent reinforcement learning for autonomous vehicles: A survey.Autonomous Intelligent Systems, 2 (1):27, 2022

    Joris Dinneweth, Abderrahmane Boubezoul, René Mandiau, and Stéphane Espié. Multi-agent reinforcement learning for autonomous vehicles: A survey.Autonomous Intelligent Systems, 2 (1):27, 2022

    2022

  2. [2]

    Mikayel Samvelyan, Tabish Rashid, Christian Schroeder de Witt, Gregory Farquhar, Nantas Nardelli, Tim G. J. Rudner, Chia-Man Hung, Philip H. S. Torr, Jakob Foerster, and Shimon Whiteson. The starcraft multi-agent challenge. InProceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems, AAMAS ’19, page 2186–2188, Richland, ...

    2019

  3. [3]

    Multi-agent reinforcement learning as a rehearsal for decentralized planning.Neurocomputing, 190:82–94, 2016

    Landon Kraemer and Bikramjit Banerjee. Multi-agent reinforcement learning as a rehearsal for decentralized planning.Neurocomputing, 190:82–94, 2016

    2016

  4. [4]

    Monotonic value function factorisation for deep multi-agent reinforcement learning.Journal of Machine Learning Research, 21(178):1–51, 2020

    Tabish Rashid, Mikayel Samvelyan, Christian Schroeder De Witt, Gregory Farquhar, Jakob Foerster, and Shimon Whiteson. Monotonic value function factorisation for deep multi-agent reinforcement learning.Journal of Machine Learning Research, 21(178):1–51, 2020

    2020

  5. [5]

    Multi-agent incentive communication via decentralized teammate modeling

    Lei Yuan, Jianhao Wang, Fuxiang Zhang, Chenghe Wang, Zongzhang Zhang, Yang Yu, and Chongjie Zhang. Multi-agent incentive communication via decentralized teammate modeling. InProceedings of the AAAI conference on artificial intelligence, volume 36, pages 9466–9474, 2022

    2022

  6. [6]

    The surprising effectiveness of ppo in cooperative multi-agent games.Advances in neural information processing systems, 35:24611–24624, 2022

    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, 35:24611–24624, 2022

    2022

  7. [7]

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

    Ryan Lowe, YI WU, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, ed- itors,Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. UR...

    2017

  8. [8]

    Learning to communicate with deep multi-agent reinforcement learning.Advances in neural information processing systems, 29, 2016

    Jakob Foerster, Ioannis Alexandros Assael, Nando De Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning.Advances in neural information processing systems, 29, 2016

    2016

  9. [9]

    Learning multiagent communication with backpropa- gation.Advances in neural information processing systems, 29, 2016

    Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropa- gation.Advances in neural information processing systems, 29, 2016

    2016

  10. [10]

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

    Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to communicate at scale in multiagent cooperative and competitive tasks. InInternational Conference on Learning Representations

  11. [11]

    Efficient multi-agent communication via self-supervised information aggregation.Advances in Neural Information Processing Systems, 35:1020–1033, 2022

    Cong Guan, Feng Chen, Lei Yuan, Chenghe Wang, Hao Yin, Zongzhang Zhang, and Yang Yu. Efficient multi-agent communication via self-supervised information aggregation.Advances in Neural Information Processing Systems, 35:1020–1033, 2022

    2022

  12. [12]

    T2mac: Targeted and trusted multi-agent communication through selective engagement and evidence-driven integration

    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 integration. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 15154–15163, 2024. 10

    2024

  13. [13]

    Learning nearly decom- posable value functions via communication minimization.arXiv preprint arXiv:1910.05366, 2019

    Tonghan Wang, Jianhao Wang, Chongyi Zheng, and Chongjie Zhang. Learning nearly decom- posable value functions via communication minimization.arXiv preprint arXiv:1910.05366, 2019

    Pith/arXiv arXiv 1910

  14. [14]

    Learning efficient multi-agent communication: An information bottleneck approach

    Rundong Wang, Xu He, Runsheng Yu, Wei Qiu, Bo An, and Zinovi Rabinovich. Learning efficient multi-agent communication: An information bottleneck approach. InInternational conference on machine learning, pages 9908–9918. PMLR, 2020

    2020

  15. [15]

    Multi-agent reinforcement learning for autonomous driving: A survey.arXiv preprint arXiv:2408.09675, 2024

    Ruiqi Zhang, Jing Hou, Florian Walter, Shangding Gu, Jiayi Guan, Florian Röhrbein, Yali Du, Panpan Cai, Guang Chen, and Alois Knoll. Multi-agent reinforcement learning for autonomous driving: A survey.arXiv preprint arXiv:2408.09675, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Efficient multi-agent communication via shapley message value

    Di Xue, Lei Yuan, Zongzhang Zhang, and Yang Yu. Efficient multi-agent communication via shapley message value. InIJCAI, pages 578–584, 2022

    2022

  17. [17]

    A value for n-person games

    Lloyd S Shapley et al. A value for n-person games. 1953

    1953

  18. [18]

    Dop: Off- policy multi-agent decomposed policy gradients

    Yihan Wang, Beining Han, Tonghan Wang, Heng Dong, and Chongjie Zhang. Dop: Off- policy multi-agent decomposed policy gradients. InInternational conference on learning representations, 2020

    2020

  19. [19]

    Towards making systems forget with machine unlearning

    Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE symposium on security and privacy, pages 463–480. IEEE, 2015

    2015

  20. [20]

    Eternal sunshine of the spotless net: Selective forgetting in deep networks

    Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9304–9312, 2020

    2020

  21. [21]

    A survey of machine unlearning.ACM Transactions on Intelligent Systems and Technology, 16(5):1–46, 2025

    Thanh Tam Nguyen, Thanh Trung Huynh, Zhao Ren, Phi Le Nguyen, Alan Wee-Chung Liew, Hongzhi Yin, and Quoc Viet Hung Nguyen. A survey of machine unlearning.ACM Transactions on Intelligent Systems and Technology, 16(5):1–46, 2025

    2025

  22. [22]

    Smacv2: An improved benchmark for cooperative multi-agent reinforcement learning.Advances in Neural Information Processing Systems, 36:37567–37593, 2023

    Benjamin Ellis, Jonathan Cook, Skander Moalla, Mikayel Samvelyan, Mingfei Sun, Anuj Mahajan, Jakob Foerster, and Shimon Whiteson. Smacv2: An improved benchmark for cooperative multi-agent reinforcement learning.Advances in Neural Information Processing Systems, 36:37567–37593, 2023

    2023

  23. [23]

    Tarmac: Targeted multi-agent communication

    Abhishek Das, Théophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Mike Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. InInternational Conference on machine learning, pages 1538–1546. PMLR, 2019

    2019

  24. [24]

    Towards true lossless sparse communication in multi-agent systems

    Seth Karten, Mycal Tucker, Siva Kailas, and Katia Sycara. Towards true lossless sparse communication in multi-agent systems. In2023 ieee international conference on robotics and automation (icra), pages 7191–7197. IEEE, 2023

    2023

  25. [25]

    Rescom: Reward- shaped curriculum for efficient multi-agent communication learning

    Xinghai Wei, Tingting Yuan, Jie Yuan, Dongxiao Liu, and Xiaoming Fu. Rescom: Reward- shaped curriculum for efficient multi-agent communication learning. InProceedings of the 24th International Conference on Autonomous Agents and Multiagent Systems, pages 2153–2161, 2025

    2025

  26. [26]

    Counterfactual multi-agent policy gradients

    Jakob Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counterfactual multi-agent policy gradients. InProceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  27. [27]

    Trust region policy optimisation in multi-agent reinforcement learning

    Jakub Grudzien Kuba, Ruiqing Chen, Muning Wen, Ying Wen, Fanglei Sun, Jun Wang, and Yaodong Yang. Trust region policy optimisation in multi-agent reinforcement learning. In International Conference on Learning Representations, 2022

    2022

  28. [28]

    Rgmcomm: Return gap minimization via discrete communications in multi-agent reinforcement learning

    Jingdi Chen, Tian Lan, and Carlee Joe-Wong. Rgmcomm: Return gap minimization via discrete communications in multi-agent reinforcement learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17327–17336, 2024

    2024

  29. [29]

    surrounded and reflect

    Sham Kakade and John Langford. Approximately optimal approximate reinforcement learning. InProceedings of the nineteenth international conference on machine learning, pages 267–274, 2002. 11 Appendix Table of Contents Contents A Proof of Theorem 4.2 13 B Complete Algorithm for MUTE 15 C Environments Details 15 C.1 StarCraft Multi-Agent Challenge (SMAC) . ...

    2002