pith. sign in

arxiv: 2606.18308 · v1 · pith:VQYAC7O4new · submitted 2026-06-16 · 💻 cs.LG · cs.AI

TRIDENT: Breaking the Hybrid-Safety-Physics Coupling for Provably Safe Multi-Agent Reinforcement Learning

Zijie Meng , Ziwei Li , Yufei Liu , Zhiyu Li , Jiyuan Liu , Wenhua Nie , Bingcai Wei , Miao Zhang This is my paper

Pith reviewed 2026-06-27 01:51 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords multi-agent reinforcement learningsafe reinforcement learninghybrid action spacesconstrained Nash equilibriumphysics-informed learningLyapunov methodstrust region methods
0
0 comments X

The pith

TRIDENT co-designs three modules to cancel the bias cycle between hybrid actions, safety constraints, and physics dynamics in multi-agent RL.

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

The paper shows that hybrid discrete-continuous actions, hard training-time safety constraints, and physics-governed dynamics create a directed cycle of biases that defeats standard modular MARL combinations, formalized as a three-way coupling lemma. TRIDENT counters this with a Richardson-Romberg gradient correction, Lyapunov-constrained sequential trust-region updates, and a physics-informed residual critic. The authors prove O~(1/sqrt(K)) convergence to a constrained Nash equilibrium and an O(sqrt(K)) cumulative-violation bound. Empirical tests on multi-UAV computing, autonomous intersections, and hybrid SMAC show large reductions in training violations alongside reward gains.

Core claim

The three features form a directed cycle of biases that any naive composition of off-the-shelf modules cannot escape; TRIDENT's three co-designed components cancel each leak, delivering the stated convergence rate to constrained Nash equilibrium and violation bound.

What carries the argument

The three-way coupling lemma that formalizes the bias cycle, together with the three co-designed components that cancel the leaks: Richardson-Romberg gradient correction, Lyapunov-constrained sequential trust-region update, and physics-informed residual critic.

If this is right

  • Training-time safety violations drop substantially compared with MADDPG and MACPO while reward improves over unconstrained baselines.
  • The convergence and violation bounds apply to constrained Nash equilibria in hybrid-action cyber-physical systems.
  • The framework covers multi-UAV mobile-edge computing, autonomous intersection management, and hybrid SMAC environments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the coupling lemma holds beyond the tested domains, similar co-design may be required for other hybrid safety settings in reinforcement learning.
  • Removing any one of the three TRIDENT components should reintroduce measurable bias in at least one of the three features.
  • The approach could be tested on single-agent problems to check whether the cycle requires multiple agents.

Load-bearing premise

The three features form a directed cycle of biases that defeats any naive composition of off-the-shelf modules.

What would settle it

A demonstration that some modular combination of existing methods already achieves comparable safety and performance without the co-design would falsify the necessity of the TRIDENT components.

Figures

Figures reproduced from arXiv: 2606.18308 by Bingcai Wei, Jiyuan Liu, Miao Zhang, Wenhua Nie, Yufei Liu, Zhiyu Li, Zijie Meng, Ziwei Li.

Figure 1
Figure 1. Figure 1: Three-way coupling. Red wavy arrows: the bias-leakage cycle of any naive composition; green dashed arrows: the three co-designed mechanisms in TRIDENT that cancel each leak (Lemma 1). Lemma 1 (Bias Propagation in Naive Composi￾tion). For a baseline using a Gumbel-Softmax esti￾mator with bias βGS, a Lagrangian or trust-region safety step of magnitude ηs, and a critic with MSE ϵQ, the per-iteration constrain… view at source ↗
Figure 2
Figure 2. Figure 2: System architecture of TRIDENT. The framework resolves the three-way coupling of hybrid actions, physics, and safety via three co-designed modules. Solid arrows denote forward passes; blue dashed arrows denote gradient flows. (A) SHA (Structured Hybrid Actor): Uses a bilevel conditional policy and Straight-Through Gradient Correction (STGC) across two temperatures (τ, τ0) to reduce discrete gradient bias t… view at source ↗
Figure 3
Figure 3. Figure 3: Single-UAV obstacle avoidance, sequential snapshots. Pink: executed continuous trajectory; green brackets: discrete waypoint cells chosen by the hybrid policy; red polygons: no-fly zones. TRIDENT routes between hazards without any forbidden region, providing a visual instance of the safety bound in Theorem 3. is exactly the F3→F1 leakage of Lemma 1 (Ng et al., 1999; Cao et al., 2024); decomposing the criti… view at source ↗
Figure 4
Figure 4. Figure 4: then exposes the two multi-agent regimes our framework supports: heterogeneous cruise as￾signment (left), where each UAV jointly selects a discrete role and a continuous trajectory—the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
read the original abstract

Safe coordination in networked cyber-physical systems forces learning algorithms to simultaneously handle hybrid discrete-continuous actions, hard training-time safety constraints, and physics-governed dynamics. We show that these three features form a directed cycle of biases that defeats any naive composition of off-the-shelf modules, and formalize this as a three-way coupling lemma. We then introduce TRIDENT, the first MARL framework whose three components are co-designed to cancel each leak: a Richardson-Romberg gradient correction reducing Gumbel-Softmax bias from O(tau) to O(tau^2), a Lyapunov-constrained sequential trust-region update enforcing per-iterate feasibility, and a physics-informed residual critic that decomposes value rather than reward. We prove an O~(1/sqrt(K)) convergence rate to a constrained Nash equilibrium and an O(sqrt(K)) cumulative-violation bound. On multi-UAV mobile-edge computing, autonomous intersection management, and a hybrid SMAC variant, TRIDENT cuts training-time violations by 95.5% over MADDPG and 76.3% over MACPO, while improving reward by 13.5% over the strongest unconstrained baseline.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper claims that hybrid discrete-continuous actions, hard training-time safety constraints, and physics-governed dynamics form a directed cycle of biases formalized as a three-way coupling lemma that defeats naive MARL module compositions. It introduces TRIDENT, whose three co-designed components (Richardson-Romberg gradient correction reducing Gumbel-Softmax bias, Lyapunov-constrained sequential trust-region update, and physics-informed residual critic) break this cycle. The work proves an O~(1/sqrt(K)) convergence rate to a constrained Nash equilibrium and an O(sqrt(K)) cumulative-violation bound, and reports empirical results on multi-UAV mobile-edge computing, autonomous intersection management, and hybrid SMAC showing 95.5% and 76.3% reductions in training-time violations versus MADDPG and MACPO plus 13.5% reward improvement over the strongest unconstrained baseline.

Significance. If the coupling lemma is valid and the convergence/violation proofs hold under the stated assumptions, the result would be significant for safe MARL in cyber-physical systems: it supplies the first co-designed framework that explicitly cancels the three-way interaction, together with non-asymptotic rates and large empirical violation reductions on relevant tasks. The machine-checked or fully expanded proofs (if supplied) and the parameter-free character of the rates would further strengthen the contribution.

major comments (3)
  1. [Abstract / §3] Abstract and presumed §3 (three-way coupling lemma): the lemma is load-bearing for the necessity of TRIDENT's co-design, yet the manuscript supplies neither its formal statement, the directed-cycle construction, nor its proof; without these it is impossible to verify whether standard modular combinations already avoid the cycle or whether the lemma reduces to the same safety constraints it is meant to enforce.
  2. [Proofs (not shown in supplied text)] Proof section (convergence and violation bounds): the O~(1/sqrt(K)) rate to constrained Nash and O(sqrt(K)) cumulative-violation bound are asserted without derivation steps, explicit assumption list, or equation references; this prevents checking whether the rates are circular (i.e., reduce by construction to quantities defined by the fitted parameters or the same safety constraints) as flagged in the stress-test.
  3. [Experiments] Empirical evaluation: the 95.5% / 76.3% violation reductions and 13.5% reward gain are reported without raw data, per-seed statistics, or ablation isolating each TRIDENT component; this leaves open whether the gains are attributable to breaking the coupling or to other implementation details.
minor comments (2)
  1. [Abstract] The O~ notation in the convergence claim should be expanded to show the precise logarithmic factors and the dependence on the number of agents.
  2. [Related work] A comparison table placing TRIDENT against prior safe MARL methods (MADDPG, MACPO, etc.) with respect to the three features would clarify the novelty claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

Thank you for the constructive and detailed review. We address each major comment point-by-point below. We will revise the manuscript to improve clarity, expand derivations, and add requested empirical details where the current presentation is insufficient.

read point-by-point responses

  1. Referee: [Abstract / §3] Abstract and presumed §3 (three-way coupling lemma): the lemma is load-bearing for the necessity of TRIDENT's co-design, yet the manuscript supplies neither its formal statement, the directed-cycle construction, nor its proof; without these it is impossible to verify whether standard modular combinations already avoid the cycle or whether the lemma reduces to the same safety constraints it is meant to enforce.

    Authors: The three-way coupling lemma is stated and proved in Section 3. It formalizes a directed cycle in which hybrid-action bias (from Gumbel-Softmax) induces safety-constraint violations that corrupt the physics residual, which in turn feeds back into the value estimator and re-amplifies the original bias; the cycle cannot be broken by independent module composition. We will move the full formal statement, cycle diagram, and complete proof into the main text (or a dedicated subsection) in the revision so that readers can directly inspect the construction and verify it does not collapse to the safety constraints alone. revision: yes

  2. Referee: [Proofs (not shown in supplied text)] Proof section (convergence and violation bounds): the O~(1/sqrt(K)) rate to constrained Nash and O(sqrt(K)) cumulative-violation bound are asserted without derivation steps, explicit assumption list, or equation references; this prevents checking whether the rates are circular (i.e., reduce by construction to quantities defined by the fitted parameters or the same safety constraints) as flagged in the stress-test.

    Authors: The full proofs, assumption list (Section 4), and non-circularity arguments appear in the appendix. The O(1/sqrt(K)) rate follows from the bias reduction to O(tau^2) under Richardson-Romberg, the per-iterate feasibility of the Lyapunov trust-region, and the residual decomposition that isolates physics error; the violation bound is obtained by telescoping the feasibility constraint across iterates. We will insert the key derivation steps and cross-references into the main proof sketch in the revision to make verification immediate without requiring the appendix. revision: yes

  3. Referee: [Experiments] Empirical evaluation: the 95.5% / 76.3% violation reductions and 13.5% reward gain are reported without raw data, per-seed statistics, or ablation isolating each TRIDENT component; this leaves open whether the gains are attributable to breaking the coupling or to other implementation details.

    Authors: We will add the complete per-seed raw data, standard-error statistics, and component-wise ablation tables (gradient correction alone, trust-region alone, residual critic alone, and all pairs) to the supplementary material. These ablations isolate the contribution of each co-designed element and show that only the joint configuration achieves the reported violation reductions and reward improvement, confirming the gains arise from breaking the coupling rather than ancillary implementation choices. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper introduces its own three-way coupling lemma as a formalization of feature interactions and derives convergence rates O~(1/sqrt(K)) and violation bounds O(sqrt(K)) under stated assumptions. No quoted step reduces a claimed prediction or result to a fitted parameter, self-citation chain, or definitional tautology by construction. The lemma and proofs are presented as independent contributions rather than imported or renamed from prior self-work. Empirical gains are reported separately from the theoretical claims. This is the normal case of a paper whose central derivation does not collapse to its inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the Richardson-Romberg correction, Lyapunov constraint, and residual critic are named but their status as standard background versus paper-specific inventions cannot be determined.

pith-pipeline@v0.9.1-grok · 5761 in / 1485 out tokens · 48272 ms · 2026-06-27T01:51:11.393431+00:00 · methodology

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

84 extracted references · 6 linked inside Pith

  1. [1]

    Proceedings of the 34th International Conference on Machine Learning (ICML) , pages=

    Constrained Policy Optimization , author=. Proceedings of the 34th International Conference on Machine Learning (ICML) , pages=. 2017 , organization=

    2017

  2. [3]

    Al-Hourani, Akram and Kandeepan, Sithamparanathan and Lardner, Simon , journal=. Optimal. 2014 , publisher=

    2014

  3. [4]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Safe Reinforcement Learning via Shielding , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  4. [5]

    SIAM Journal on Mathematics of Data Science , volume=

    On the Effectiveness of Richardson Extrapolation in Data Science , author=. SIAM Journal on Mathematics of Data Science , volume=. 2021 , publisher=

    2021

  5. [7]

    Annual Review of Control, Robotics, and Autonomous Systems , volume=

    Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning , author=. Annual Review of Control, Robotics, and Autonomous Systems , volume=. 2022 , publisher=

    2022

  6. [8]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    A Lyapunov-Based Approach to Safe Reinforcement Learning , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  7. [9]

    Proceedings of the International Conference on Autonomous Agents and Multiagent Systems (AAMAS) , year=

    Safe Multi-Agent Reinforcement Learning via Shielding , author=. Proceedings of the International Conference on Autonomous Agents and Multiagent Systems (AAMAS) , year=

  8. [10]

    arXiv preprint arXiv:1903.01344 , year=

    Hybrid actor-critic reinforcement learning in parameterized action space , author=. arXiv preprint arXiv:1903.01344 , year=

    Pith/arXiv arXiv 1903

  9. [11]

    arXiv preprint arXiv:1903.04959 , year=

    Deep multi-agent reinforcement learning with discrete-continuous hybrid action spaces , author=. arXiv preprint arXiv:1903.04959 , year=

    Pith/arXiv arXiv 1903

  10. [12]

    Journal of Machine Learning Research , volume=

    A Comprehensive Survey on Safe Reinforcement Learning , author=. Journal of Machine Learning Research , volume=

  11. [13]

    Artificial Intelligence Review , volume=

    Multi-Agent Deep Reinforcement Learning: A Survey , author=. Artificial Intelligence Review , volume=. 2022 , publisher=

    2022

  12. [17]

    International Conference on Learning Representations (ICLR) , year=

    Categorical Reparameterization with Gumbel-Softmax , author=. International Conference on Learning Representations (ICLR) , year=

  13. [18]

    2019 International Conference on Robotics and Automation (ICRA) , pages=

    Residual Reinforcement Learning for Robot Control , author=. 2019 International Conference on Robotics and Automation (ICRA) , pages=. 2019 , organization=

    2019

  14. [19]

    Nature Reviews Physics , volume=

    Physics-informed machine learning , author=. Nature Reviews Physics , volume=. 2021 , publisher=

    2021

  15. [20]

    International Conference on Learning Representations (ICLR) , year=

    Trust Region Policy Optimisation in Multi-Agent Reinforcement Learning , author=. International Conference on Learning Representations (ICLR) , year=

  16. [21]

    International Conference on Learning Representations (ICLR) , year=

    Global Convergence of Multi-Agent Policy Gradient in Markov Potential Games , author=. International Conference on Learning Representations (ICLR) , year=

  17. [22]

    International Conference on Machine Learning (ICML) , year=

    Constrained Variational Policy Optimization for Safe Reinforcement Learning , author=. International Conference on Machine Learning (ICML) , year=

  18. [23]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  19. [24]

    International Conference on Learning Representations (ICLR) , year=

    The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables , author=. International Conference on Learning Representations (ICLR) , year=

  20. [25]

    Proceedings of the Sixteenth International Conference on Machine Learning (ICML) , pages=

    Policy Invariance Under Reward Transformations: Theory and Application to Reward Shaping , author=. Proceedings of the Sixteenth International Conference on Machine Learning (ICML) , pages=

  21. [26]

    International Conference on Learning Representations (ICLR) , year=

    Rao-Blackwellizing the Straight-Through Gumbel-Softmax Gradient Estimator , author=. International Conference on Learning Representations (ICLR) , year=

  22. [27]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Peng, Bei and Rashid, Tabish and de Witt, Christian Schroeder and Kamienny, Pierre-Alexandre and Torr, Philip HS and B. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  23. [28]

    Learning for Dynamics and Control Conference (L4DC) , pages=

    Physics-Informed Model-Based Reinforcement Learning , author=. Learning for Dynamics and Control Conference (L4DC) , pages=. 2023 , organization=

    2023

  24. [30]

    Philosophical Transactions of the Royal Society A , volume=

    The Approximate Arithmetical Solution by Finite Differences of Physical Problems Involving Differential Equations, with an Application to the Stresses in a Masonry Dam , author=. Philosophical Transactions of the Royal Society A , volume=. 1911 , publisher=

    1911

  25. [31]

    Samvelyan, Mikayel and Rashid, Tabish and de Witt, Christian Schroeder and Farquhar, Gregory and Nardelli, Nantas and Rudner, Tim GJ and Hung, Chia-Man and Torr, Philip HS and Foerster, Jakob and Whiteson, Shimon , journal=. The

  26. [32]

    Cold Analysis of

    Shekhovtsov, Alexander , booktitle=. Cold Analysis of. 2023 , organization=

    2023

  27. [34]

    Responsive Safety in Reinforcement Learning by

    Stooke, Adam and Achiam, Joshua and Abbeel, Pieter , booktitle=. Responsive Safety in Reinforcement Learning by. 2020 , organization=

    2020

  28. [35]

    Study on Channel Model for Frequencies from 0.5 to 100

  29. [36]

    Multi-Agent Deep Reinforcement Learning-Based Trajectory Planning for Multi-

    Wang, Liang and others , journal=. Multi-Agent Deep Reinforcement Learning-Based Trajectory Planning for Multi-. 2021 , publisher=

    2021

  30. [37]

    IEEE Access , volume=

    Hybrid policy learning for multi-agent pathfinding , author=. IEEE Access , volume=. 2021 , publisher=

    2021

  31. [38]

    The Surprising Effectiveness of

    Yu, Chao and Velu, Akash and Vinitsky, Eugene and Gao, Jiajun and Wang, Yu and Bayen, Alexandre and Wu, Yi , booktitle=. The Surprising Effectiveness of

  32. [39]

    Foundations and Trends in Machine Learning , volume=

    Decentralized Multi-Agent Reinforcement Learning , author=. Foundations and Trends in Machine Learning , volume=. 2021 , publisher=

    2021

  33. [40]

    2021 , organization=

    Zhou, Ming and Luo, Jun and Villella, Julian and Yang, Yaodong and Rusu, David and Miao, Jiayu and Zhang, Weinan and Alban, Montgomery and Fadakar, Iman and Chen, Zheng and others , booktitle=. 2021 , organization=

    2021

  34. [41]

    2023 , publisher=

    Zhou, Yong and others , journal=. 2023 , publisher=

    2023

  35. [42]

    International Conference on Learning Representations , volume=

    Physics-regulated deep reinforcement learning: Invariant embeddings , author=. International Conference on Learning Representations , volume=

  36. [43]

    ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Make a Game: A Novel Paradigm for Interactive Game Rendering , author=. ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2026 , organization=

    2026

  37. [44]

    2026 , eprint=

    Decoupling Semantics from Distortions: Multi-Scale Two-Stream Vision-Language Alignment for AI-Generated Image Quality Assessment , author=. 2026 , eprint=

    2026

  38. [45]

    Science China Information Sciences , volume=

    Orpaint: a zero-shot inpainting model for oracle bone inscription rubbings with visual mamba block , author=. Science China Information Sciences , volume=. 2025 , publisher=

    2025

  39. [46]

    arXiv preprint arXiv:2606.13432 , year=

    OmniDirector: General Multi-Shot Camera Cloning without Cross-Paired Data , author=. arXiv preprint arXiv:2606.13432 , year=

    Pith/arXiv arXiv

  40. [47]

    arXiv preprint arXiv:2606.11670 , year=

    ARGUS: Stacked Multi-View Identity Mosaic Injection for Subject-Preserving Video Generation , author=. arXiv preprint arXiv:2606.11670 , year=

    Pith/arXiv arXiv

  41. [48]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    SynPo: Boosting Training-Free Few-Shot Medical Segmentation via High-Quality Negative Prompts , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2025 , organization=

    2025

  42. [49]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    Robust Single Image Sand Removal by Leveraging Uncertainty-aware SAM Priors and Prompt Learning with Refined Perceptual Loss , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  43. [50]

    Achiam, D

    J. Achiam, D. Held, A. Tamar, and P. Abbeel. Constrained policy optimization. In ICML, 2017

    2017

  44. [51]

    Ackermann, V

    J. Ackermann, V. Gabler, T. Osa, and M. Sugiyama. Reducing overestimation bias in multi-agent domains using double centralized critics. arXiv:1910.01465, 2019

    arXiv 1910

  45. [52]

    Al-Hourani, S

    A. Al-Hourani, S. Kandeepan, and S. Lardner. Optimal LAP altitude for maximum coverage. IEEE WCL, 3(6), 2014

    2014

  46. [53]

    Alshiekh et al

    M. Alshiekh et al. Safe reinforcement learning via shielding. In AAAI, 2018

    2018

  47. [54]

    F. Bach. On the effectiveness of Richardson extrapolation in data science. SIAM J.\ Math.\ Data Sci., 3(4), 2021

    2021

  48. [55]

    Banerjee, K

    C. Banerjee, K. Nguyen, C. Fookes, and G. Karniadakis. A survey on physics-informed reinforcement learning. arXiv:2309.01909, 2023

    arXiv 2023

  49. [56]

    Brunke et al

    L. Brunke et al. Safe learning in robotics. Annu.\ Rev.\ Contr.\ Robot.\ Auton.\ Syst., 5, 2022

    2022

  50. [57]

    Cao et al

    H. Cao et al. Physics-regulated deep reinforcement learning: Invariant embeddings. In ICLR, 2024

    2024

  51. [58]

    Chow et al

    Y. Chow et al. A Lyapunov-based approach to safe RL. In NeurIPS, 2018

    2018

  52. [59]

    Elsayed-Aly et al

    I. Elsayed-Aly et al. Safe multi-agent RL via shielding. In AAMAS, 2021

    2021

  53. [60]

    Fan et al

    Z. Fan et al. Hybrid actor-critic reinforcement learning in parameterized action space. In IJCAI, 2019

    2019

  54. [61]

    Fu et al

    H. Fu et al. Deep multi-agent RL with discrete-continuous hybrid action spaces. In IJCAI, 2019

    2019

  55. [62]

    Garc \' a and F

    J. Garc \' a and F. Fern \'a ndez. A comprehensive survey on safe RL. JMLR, 16, 2015

    2015

  56. [63]

    Gronauer and K

    S. Gronauer and K. Diepold. Multi-agent deep RL: A survey. Artif.\ Intell.\ Rev., 55, 2022

    2022

  57. [64]

    Gu et al

    S. Gu et al. Multi-agent constrained policy optimisation. arXiv:2110.02793, 2021

    arXiv 2021

  58. [65]

    Gu et al

    S. Gu et al. Safe multi-agent RL with convergence to generalized Nash equilibrium. arXiv:2411.15036, 2024

    arXiv 2024

  59. [66]

    Huh and I

    S. Huh and I. Yang. Safe RL for probabilistic reachability: A Lyapunov approach. arXiv:2002.10126, 2020

    arXiv 2002

  60. [67]

    E. Jang, S. Gu, and B. Poole. Categorical reparameterization with Gumbel-softmax. In ICLR, 2017

    2017

  61. [68]

    Johannink et al

    T. Johannink et al. Residual reinforcement learning for robot control. In ICRA, 2019

    2019

  62. [69]

    Karniadakis et al

    G. Karniadakis et al. Physics-informed machine learning. Nat.\ Rev.\ Phys., 3, 2021

    2021

  63. [70]

    Kuba et al

    J. Kuba et al. Trust region policy optimisation in multi-agent reinforcement learning. In ICLR, 2022

    2022

  64. [71]

    Leonardos et al

    S. Leonardos et al. Global convergence of multi-agent policy gradient in Markov potential games. In ICLR, 2022

    2022

  65. [72]

    Liu et al

    Y. Liu et al. Constrained variational policy optimization for safe RL. In ICML, 2022

    2022

  66. [73]

    Lowe et al

    R. Lowe et al. Multi-agent actor-critic for mixed cooperative-competitive environments. In NeurIPS, 2017

    2017

  67. [74]

    Maddison, A

    C. Maddison, A. Mnih, and Y. Teh. The concrete distribution. In ICLR, 2017

    2017

  68. [75]

    A. Ng, D. Harada, and S. Russell. Policy invariance under reward transformations. In ICML, 1999

    1999

  69. [76]

    Paulus, C

    M. Paulus, C. Maddison, and A. Krause. Rao-Blackwellizing the straight-through Gumbel-softmax gradient estimator. In ICLR, 2021

    2021

  70. [77]

    Peng et al

    B. Peng et al. FACMAC: Factored multi-agent centralised policy gradients. In NeurIPS, 2021

    2021

  71. [78]

    Ramesh and B

    R. Ramesh and B. Ravindran. Physics-informed model-based RL. In L4DC, 2023

    2023

  72. [79]

    A. Ray, J. Achiam, and D. Amodei. Benchmarking safe exploration in deep RL. arXiv:1910.01708, 2019

    Pith/arXiv arXiv 1910

  73. [80]

    Richardson

    L. Richardson. The approximate arithmetical solution by finite differences. Phil.\ Trans.\ Roy.\ Soc.\ A, 210, 1911

    1911

  74. [81]

    Samvelyan et al

    M. Samvelyan et al. The StarCraft multi-agent challenge. arXiv:1902.04043, 2019

    arXiv 1902

  75. [82]

    Shekhovtsov

    A. Shekhovtsov. Cold analysis of Rao-Blackwellized straight-through Gumbel-softmax. In ICML, 2023

    2023

  76. [83]

    Silver et al

    T. Silver et al. Residual policy learning. arXiv:1812.06298, 2018

    Pith/arXiv arXiv 2018

  77. [84]

    Stooke, J

    A. Stooke, J. Achiam, and P. Abbeel. Responsive safety in reinforcement learning by PID Lagrangian methods. In ICML, 2020

    2020

  78. [85]

    Study on channel model for frequencies from 0.5 to 100 GHz

    3GPP . Study on channel model for frequencies from 0.5 to 100 GHz. 3GPP TR 38.901, 2020

    2020

  79. [86]

    Wang et al

    L. Wang et al. Multi-agent deep RL-based trajectory planning for multi-UAV assisted MEC. IEEE TCCN, 7(1), 2021

    2021

  80. [87]

    Wang et al

    J. Wang et al. Hybrid policy optimization for multi-agent reinforcement learning. arXiv:2206.10485, 2022

    arXiv 2022

Showing first 80 references.