Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

CH-MARL: Constrained Hierarchical Multiagent Reinforcement Learning for Sustainable Maritime Logistics

T0 review · 5 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A real-time constraint layer that raises a penalty whenever cumulative emissions exceed a global cap, paired with fairness-adjusted rewards, is claimed to let hierarchical vessel and port agents cut emissions and fuel use without…

desk verdict Nice assembly of known ideas, but the experiments never test the claimed architecture or baselines, so the headline results are unsupported. read the letter →

arxiv 2502.02060 v1 pith:3VSTYHAG submitted 2025-02-04 cs.AI cs.MA

Saad Alqithami This is my paper · ORCID
classification cs.AIcs.MA
keywords multi-agentreinforcementlearningconstrainedhierarchicalmaritimelogisticsemissioncapsfairnessinmultiagentsystemsprimal-dualoptimizationgreenhousegasreduction
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 aims to establish that a constrained hierarchical multiagent reinforcement learning (MARL) framework can simultaneously reduce fuel use and greenhouse gas emissions in maritime logistics, keep total emissions under a hard cap, and spread costs fairly across vessels. The mechanism is a real-time constraint layer built on the Lagrangian $L(\pi, \lambda) = E[R] - \lambda(E[C] - \kappa)$: whenever cumulative emissions exceed the cap, the dual variable $\lambda$ grows and every agent's reward is penalized, pushing the fleet back into compliance. Fairness is handled by a reward term that penalizes large disparities in per-agent fuel cost and delay. Validation is reported in a digital-twin simulation of eight ports and five vessels, where the configuration with cap, fairness, and storms reaches mean emissions of 4.07 versus 4.73 for the unconstrained baseline, at the cost of a lower (more negative) reward. The authors claim this makes the framework scalable and transferable to other constrained, multi-objective industrial domains.

What carries the argument

The load-bearing object is the Lagrangian $L(\pi, \lambda) = E[R] - \lambda(E[C] - \kappa)$, cut into a primal update for the policy and a dual update for the constraint price $\lambda$. Each time cumulative emissions exceed the cap $C_{\max}$, $\lambda$ is increased and each agent receives a penalty $-\lambda\cdot e_i$ proportional to its own emissions, so a global constraint is enforced through local reward shaping under partial observability. On top of this sits a nested constrained Markov decision process hierarchy: high-level agents choose routes, emission budgets, and schedules at a coarse timescale, and low-level agents refine these into speed, berth, and cargo-handling actions. A fairness term $\Delta_{\mathrm{fair}}$ built from a Gini-style disparity measure adjusts individual rewards, with the claimed guarantee tied to $\beta$-scaling of a deviation penalty. These pieces together carry the paper's compliance and equity claims.

What would settle it

Run the reported digital-twin training under the full configuration for many seeds while logging cumulative emissions at every step; the framework predicts cumulative emissions never exceed the stated cap ($C_{\max} = 800$) by more than a small tolerance. A single episode in which the overshoot exceeds that tolerance, or a head-to-head run where a flat constrained MARL agent matches CH-MARL's emissions and throughput, would undercut the central claim.

Watch

Extended reading notes

Core claim

On its own terms, the paper's claim is that primal-dual penalty updates combined with a two-level hierarchy and fairness-aware rewards give a policy that satisfies a global emission constraint and an equity criterion. The theoretical statements are Propositions 4.1 through 4.4: gradient-based primal-dual updates converge to a policy with $E[C] \leq \kappa$, the dynamic enforcement layer keeps violations within a tolerance $\epsilon$, the nested high-level and low-level policies converge to a locally optimal joint solution, and a max-min fairness objective with an adjusted reward $R'_i = R_i - \beta\Delta_i$ keeps the fairness metric above a chosen threshold $\delta$. Empirically, the reported runs show that enabling the emission cap pulls cumulative emissions down, from 4.73 in the base run to 4.07 in the full configuration, while the full configuration also absorbs storms and partial observability; the cost appears as a more negative reward rather than a loss of operational feasibility.

Load-bearing premise

Everything the paper proves about staying under the emission cap depends on the assumption that the expected emission cost is a smooth, convex function of the policy, and none of the experiments verify that for neural-network policies or real nonlinear emission dynamics.

Editorial extensions

If this is right

  • If correct, a single online penalty layer can keep a fleet's cumulative emissions under a regulatory cap without requiring any agent to know the global state, making IMO-style caps implementable in decentralized operations.
  • Fairness-shaped rewards would prevent optimization from consistently routing costs onto smaller vessels, because large per-agent disparities are directly penalized in the learning signal.
  • The hierarchical decomposition splits strategic decisions such as routes and emission budgets from operational control such as speed and berthing, which would let the framework scale to larger port networks than flat MARL.
  • The same constrained MARL recipe transfers to other settings with a shared numeric budget, such as smart grids, urban traffic, and supply chains, where aggregate limits rather than per-agent rules define compliance.
  • The reported configuration with cap, fairness, and storms achieves the lowest emissions while trading some immediate reward, indicating that added constraints can buy environmental and equity gains at a measurable operational cost.

Reading between the lines

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

  • A direct head-to-head between CH-MARL and a flat single-layer constrained MARL agent with identical rewards would isolate whether the claimed gains come from the hierarchy or from the constraint and fairness shaping alone; the paper does not run that comparison.
  • Because the constraint layer only requires a scalar cumulative emission total and a per-agent share, the same dual-penalty gate could be lifted into any shared-budget domain, such as water quotas or road-congestion pricing, where a global cap is the compliance object.
  • The fairness guarantee as stated depends on knowing the optimal reward $R_{\mathrm{optimal}}$; in fully decentralized deployments that value would have to be estimated online, and the paper does not specify how the threshold $\delta$ would be set in practice.
  • The reported emissions reduction is modest in absolute magnitude and the reward penalty from constraints is noticeable, so a real deployment would need to translate the trade-off into fuel costs and IMO compliance deadlines to judge its practical value.
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

5 major / 5 minor

Summary. The paper proposes CH-MARL, a hierarchical multi-agent reinforcement learning framework that integrates a strategic high-level layer, an operational low-level layer, a primal-dual constraint-enforcement mechanism for global emission caps, and fairness-aware reward shaping. The authors formalize a maritime logistics problem as a constrained Markov decision process, present several convergence propositions, and describe a simulated digital twin environment with 8 ports and 5 vessels. Experiments are reported for four configurations (Runs A–D) that vary the presence of emission caps, fairness penalties, partial observability, and storms. The central claim is that CH-MARL outperforms baseline methods by reducing emissions and fuel consumption without compromising operational throughput, while maintaining fairness.

Significance. If the claims were validated, combining hierarchical MARL with dynamic constraint enforcement and fairness shaping for a real-world logistics domain would be a worthwhile contribution. The manuscript gives a clear problem formulation and outlines a plausible testbed, and the conceptual architecture is easy to follow. However, the empirical section does not test the proposed multi-agent hierarchy, does not evaluate any of the three baselines defined in Section 5.3, omits the promised throughput and fairness KPIs, and contains a likely data error in Table 1. The theoretical propositions depend on unverified convexity and smoothness assumptions. As presented, the central comparative and operational claims are not supported, so the significance of the contribution cannot be assessed.

major comments (5)
  1. [5.3, 6, Table 1] The three baselines defined in Section 5.3 (decentralized MARL without caps/fairness, a super-agent, and a hierarchical method without caps/fairness) are never evaluated. Section 6 reports only four self-configurations of the proposed feature set (Runs A–D), and Table 1 lists only reward and emissions for these runs. Consequently, the abstract's claim that 'CH-MARL outperforms baseline methods' is not tested. Moreover, the KPIs 'operational throughput' (Section 5.2d) and fairness indices (Section 5.2c) are never reported, so the claims about maintaining throughput and ensuring fairness are unsupported.
  2. [Section 6, first paragraph] The experiments train 'a single proximal policy optimization agent' rather than the hierarchical multi-agent architecture described in Section 4.2 and Algorithm 1. The experimental setup includes no high-level strategic agents, no low-level operational agents, and no multi-agent coordination. Even the ablation results (Runs A–D) therefore do not evaluate CH-MARL as proposed; they evaluate a single-agent variant with different reward terms. The authors need to either run the full multi-agent algorithm or explicitly state that the experiments are on a simplified single-agent proxy and justify why that is sufficient for the claims.
  3. [Table 1] In Table 1, the emission values for Run A and Run B are exactly the negatives of the reward values (-4.730401 and -4.733810). This strongly suggests that the emissions column is a duplicate of the reward magnitude rather than an independent measurement, which would invalidate the reported emission reductions. The authors should correct this and provide independent emissions data, ideally with standard deviations across seeds.
  4. [Section 4.1, Proposition 4.1] Proposition 4.1 assumes L(π, λ) is continuously differentiable and E[C] is convex with respect to π to guarantee convergence to a constraint-satisfying policy. These assumptions are not verified for neural-network policies and nonlinear emission dynamics. The proof is a sketch that invokes a saddle-point argument, and the conditions under which the gradients exist and the saddle point is attained are not established. The same problem affects Propositions 4.2–4.4, which assert bounded violations and fairness guarantees without stating the required smoothness constants, learning-rate bounds, or penalty scales. The theoretical support for the claimed constraint compliance and fairness is therefore not provided.
  5. [Sections 4.2.2, 4.2.3, and 6] The reported empirical outcomes are largely by construction: Runs B and D incorporate a penalty -λ(E[C]-κ) whenever cumulative emissions exceed Cmax (Section 4.2.2), so reduced emissions relative to Run A are a direct consequence of the objective function. Similarly, the fairness term in Section 4.2.3 penalizes disparities by design. Without comparisons to baselines that optimize the same rewards with alternative mechanisms, these results do not demonstrate that CH-MARL's specific integration of hierarchical decision-making, constraint enforcement, and fairness shaping is effective.
minor comments (5)
  1. [Section 1.4] The paper organization paragraph misnumbers the sections: it says Section 6 outlines the experimental setup and that results are also presented in Section 6, but the setup is in Section 5 and the results are in Section 6.
  2. [Appendix A.2] The policy update formula is incomplete: 'π ← π + ηπ∇πL(π,' is truncated.
  3. [General formatting] The running header 'Chmarl —DRAFT: U NDER -REVIEW' and the title formatting (e.g., 'C ONSTRAINED') are inconsistent with a polished journal submission.
  4. [Sections 3.3 and Algorithm 1] Section 3.3 defines the reward as r_i^t = R_cost + R_emission + R_fair, but Algorithm 1 introduces r_constraint and δ_fair without a clear mapping to these terms; the notation should be unified.
  5. [Section 6] The text states that 'fairness can induce lower but more balanced rewards' but no fairness metric is quantified anywhere; adding a column or figure with the Gini coefficient or similar measure would support this statement.

Circularity Check

2 steps flagged · score 6.0 of 10

Emission and fairness 'results' are the reward/constraint objective by construction: Table 1's emissions column equals reward magnitude, and Proposition 4.4's fairness guarantee restates the reward shaping; the promised baseline comparison is absent.

  1. fitted input called prediction [Section 4.2.2 and Section 6, Table 1]
    "All agents experience a penalty proportional to λ and their incremental emissions, driving them toward compliance. ... Table 1 lists each run’s mean results at the last training iteration. Run A (Base) -4.730401 4.730401; Run B (Cap) -4.733810 4.733810."

    The emission-reduction finding is the optimization objective itself. The reward in Algorithm 1 is modified with rconstraint,i_t = -λ * e(i)_t whenever cumulative emissions exceed the cap, so a policy maximizing this adjusted reward is, by construction, reducing penalized emissions. The Table 1 'Emissions' column for Runs A and B is exactly the magnitude of the 'Reward' column, confirming that 'emissions' is not an independently measured KPI but the same number as the reward. Therefore the reported 'emission caps pull emissions downward' is a restatement of the reward design, not an independent prediction.

  2. self definitional [Proposition 4.4 and Appendix B]
    "Proposition 4.4 (Fairness Metric Guarantees). If the max-min fairness criterion F = mini Ri Roptimal is integrated into the reward structure, then the system can guaranteeF ≥ δ for a chosen threshold δ, provided the fairness penalty is scaled appropriately."

    The 'guarantee' is definitional: the fairness-adjusted reward R'_i = R_i - βΔ_i is defined with Δ_i penalizing deviations from Roptimal, and the proof in Appendix B simply states that this penalty reduces the gap and concludes F ≥ δ. No independent convergence or learning argument is supplied; the fairness outcome is assumed because the fairness term was put into the reward. The paper's later claim that 'fairness metrics indicate minimal disparities' is therefore an input of the reward design, not a derived result.

full rationale

The paper's theoretical skeleton (Lagrangian primal-dual updates, hierarchical RL) is not itself circular and relies on external, standard references rather than a self-citation chain. However, the central empirical claims of emission reduction and fairness are built into the objective. Section 4.2.2 adds a penalty -λ(E[C]-κ) to the reward, and Section 4.2.3 adds a fairness term; Section 6 then reports that capped runs have lower emissions and fairness-adjusted runs have different rewards. The exact equality between Reward and Emissions for Runs A and B in Table 1 makes the reduction explicit: the 'Emissions' KPI is the negative of the reward, so reporting lower emissions is reporting better optimization of the reward, not an external outcome. Proposition 4.4 similarly guarantees fairness only because the fairness metric is written into the reward, with no non-definitional proof. These are genuine circular reductions of the paper's headline environmental and equity outcomes. Separately, the claimed comparison against baselines is unsupported rather than circular: Section 5.3 defines decentralized, super-agent, and hierarchical baselines, but Section 6 reports only feature ablations (Runs A-D) and never reports the baselines' reward, emissions, throughput, fairness, or constraint violations. The paper also trains 'a single proximal policy optimization agent', which does not instantiate the hierarchical multiagent architecture of Section 4.2. These are validity concerns and are noted for context, but they do not by themselves change the circularity score. Overall, the central 'predictions' reduce by construction to the reward and constraint terms, so a score of 6 is appropriate.

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

The central claims rest on hand-set simulation parameters (emission cap 800, fairness weight 0.1, storm probability 0.2, observation mask 0.5) and on unverified convexity/smoothness assumptions for the convergence proofs. No new physical entity is introduced. The empirical evaluation is a self-comparison of four configurations, so the framework is not externally benchmarked.

free parameters (5)
  • Emission cap Cmax = 800 (simulation units)
    Hand-set for Runs B and D; the constraint-enforcement layer and all compliance claims are defined relative to this value.
  • Fairness penalty weight beta/gamma_i = 0.1 (default)
    Controls the trade-off between efficiency and equity; the paper states it can be tuned, and no sensitivity analysis is reported.
  • PPO learning rates and discount factor = alpha_actor=5e-4, alpha_critic=1e-3, gamma=0.99
    Standard hyperparameters chosen by hand; results may depend on them and no sensitivity analysis is given.
  • Storm probability = 0.2 per time step
    Arbitrary weather disruption probability in Runs C and D; directly affects reward and emissions outcomes.
  • Partial observability mask fraction = 0.5
    Half of state variables randomly masked at each step; central to the partial-observability claims.
assumptions (4)
  • ad hoc to paper E[C] is convex with respect to the policy and L is continuously differentiable
    Invoked in Proposition 4.1 to prove convergence to a constraint-satisfying policy, but not verified for neural network policies or nonlinear emissions; the proof of Proposition 4.2 also depends on this.
  • standard math Primal-dual gradient updates converge to a saddle point
    Appendix A uses the classical Lagrangian saddle-point argument, but the standard conditions are not met in this nonconvex deep RL setting; the paper treats convergence as given.
  • domain assumption Discrete time steps with finite weather scenarios and fixed mechanical failure probabilities
    Section 3.6 explicitly adopts these simplifications; all experiments are only meaningful inside this stylized model.
  • domain assumption Agents are cooperative or semi-cooperative
    Section 3.6 excludes competition; real shipping alliances have competing incentives, which could break the fairness and compliance results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CH-MARL: Constrained Hierarchical Multiagent Reinforcement Learning for Sustainable Maritime Logistics." pith.science (2026). https://pith.science/paper/3VSTYHAG

@misc{pith2026250202060,
  author       = {Pith},
  title        = {Pith review of: CH-MARL: Constrained Hierarchical Multiagent Reinforcement Learning for Sustainable Maritime Logistics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3VSTYHAG}},
  note         = {Machine review of arXiv:2502.02060}
}
read the original abstract

Addressing global challenges such as greenhouse gas emissions and resource inequity demands advanced AI-driven coordination among autonomous agents. We propose CH-MARL (Constrained Hierarchical Multiagent Reinforcement Learning), a novel framework that integrates hierarchical decision-making with dynamic constraint enforcement and fairness-aware reward shaping. CH-MARL employs a real-time constraint-enforcement layer to ensure adherence to global emission caps, while incorporating fairness metrics that promote equitable resource distribution among agents. Experiments conducted in a simulated maritime logistics environment demonstrate considerable reductions in emissions, along with improvements in fairness and operational efficiency. Beyond this domain-specific success, CH-MARL provides a scalable, generalizable solution to multi-agent coordination challenges in constrained, dynamic settings, thus advancing the state of the art in reinforcement learning.

Figures

Figures reproduced from arXiv: 2502.02060 by the authors.

Figure 1
Figure 1. Conceptual overview of CH-MARL, highlighting the division between high-level (strategic) and low-level [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Reward curves (mean ± std over three seeds) for each run. Emission caps and fairness both tend to decrease the max reward–additional constraints on vessel behaviors [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Emissions curves (mean ± std) across training. Runs with active caps (Run B, Run D) converge to lower fuel usage, while storms and fairness can introduce variability. These outcomes underscore the multi-objective nature of CH-MARL: emission caps enforce sustainability, fairness can induce lower but more balanced rewards, and storms introduce realistic unpredictability. Despite these constraints, all runs consistentl… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Multi-Year Maintenance Planning for Large-Scale Infrastructure Systems: A Novel Network Deep Q-Learning Approach

    math.OC 2025-07 conditional novelty 5.0 of 10

    A deep Q-learning framework with cost-normalized rewards and an annual budget-allocation linear program outperforms LP and GA baselines on a 68,800-segment pavement network.

Reference graph

Works this paper leans on

23 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [1]

    Third imo ghg study 2014

    Tristan Smith, Jukka-Pekka Jalkanen, Ben Anderson, James J Corbett, Jasper Faber, Satoshi Hanayama, Eoin O’Keeffe, Sarah Parker, Lars Johansson, Laura Aldous, et al. Third imo ghg study 2014. Technical report, International Maritime Organization (IMO), London, UK, 2014

    work page 2014

  2. [2]

    Reducing greenhouse gas emissions from ships

    International Maritime Organization (IMO). Reducing greenhouse gas emissions from ships. Technical report, IMO, London, UK, 2020

    work page 2020

  3. [3]

    Initial imo strategy on reduction of ghg emissions from ships, 2018

    International Maritime Organization (IMO). Initial imo strategy on reduction of ghg emissions from ships, 2018. Resolution MEPC.304(72)

    work page 2018

  4. [4]

    A survey on multi-agent reinforcement learning: Foundations, applications, and learning environments

    Pablo Hernandez-Leal, Bilal Kartal, and Matthew E Taylor. A survey on multi-agent reinforcement learning: Foundations, applications, and learning environments. arXiv preprint arXiv:1908.03963, 2019

    arXiv 1908

  5. [5]

    Learning to communicate with deep multi-agent reinforcement learning

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

    work page 2016

  6. [6]

    Cooperative multi-agent learning: The state of the art

    Liviu Panait and Sean Luke. Cooperative multi-agent learning: The state of the art. Autonomous agents and multi-agent systems, 11 0 (3): 0 387--434, 2005

    work page 2005

  7. [7]

    Learning with opponent-learning awareness

    Jakob Foerster, Richard Chen, Maruan Al-Shedivat, Shimon Whiteson, Pieter Abbeel, and Igor Mordatch. Learning with opponent-learning awareness. In Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems, pages 122--130, 2018

    work page 2018

  8. [8]

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

    Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. Advances in neural information processing systems, 30: 0 6379--6390, 2017

    work page 2017

Show all 23 references
  1. [9]

    Mastering chess and shogi by self-play with a general reinforcement learning algorithm

    David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering chess and shogi by self-play with a general reinforcement learning algorithm. arXiv preprint arXiv:1712.01815, 2017

    2017 arXiv

  2. [10]

    Optimal and approximate q-value functions for decentralized pomdps

    Frans A Oliehoek, Matthijs TJ Spaan, and Nikos Vlassis. Optimal and approximate q-value functions for decentralized pomdps. In Proceedings of the 2008 International Conference on Autonomous Agents and Multiagent Systems, volume 2, pages 1315--1322, 2008

    2008

  3. [11]

    Cooperative multi-agent control using deep reinforcement learning

    Jayesh K Gupta, Maxim Egorov, and Mykel Kochenderfer. Cooperative multi-agent control using deep reinforcement learning. In International Conference on Autonomous Agents and Multiagent Systems, pages 66--83, 2017

    2017

  4. [12]

    Deep recurrent q-learning for partially observable mdps

    Matthew Hausknecht and Peter Stone. Deep recurrent q-learning for partially observable mdps. arXiv preprint arXiv:1507.06527, 2015

    2015 arXiv

  5. [13]

    Learning multiagent communication with backpropagation

    Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. Learning multiagent communication with backpropagation. In Advances in neural information processing systems, volume 29, 2016

    2016

  6. [14]

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

    Kaiqing Zhang, Zhuoran Yang, and Tamer Ba s ar. Multi-agent reinforcement learning: A selective overview of theories and algorithms. Handbook of reinforcement learning and control, pages 321--384, 2021

    2021

  7. [15]

    Autonomous agents modelling other agents: A comprehensive survey and open problems

    Stefano V Albrecht and Peter Stone. Autonomous agents modelling other agents: A comprehensive survey and open problems. Artificial intelligence, 258: 0 66--95, 2018

    2018

  8. [16]

    Multi-agent reinforcement learning for markov routing games: A new modeling paradigm for dynamic traffic assignment

    Zhenyu Shou, Xu Chen, Yongjie Fu, and Xuan Di. Multi-agent reinforcement learning for markov routing games: A new modeling paradigm for dynamic traffic assignment. Transportation Research Part C: Emerging Technologies, 137: 0 103560, 2022

    2022

  9. [17]

    Constrained markov decision processes

    Eitan Altman. Constrained markov decision processes. Stochastic modeling and applied probability, 7, 1999

    1999

  10. [18]

    Constrained policy optimization

    Joshua Achiam, David Held, Aviv Tamar, and Pieter Abbeel. Constrained policy optimization. In International conference on machine learning, pages 22--31. PMLR, 2017

    2017

  11. [19]

    The bargaining problem

    John F Nash. The bargaining problem. Econometrica: Journal of the Econometric Society, pages 155--162, 1950

    1950

  12. [20]

    Approximately fair allocations of indivisible goods

    Richard J Lipton, Evangelos Markakis, Elchanan Mossel, and Amin Saberi. Approximately fair allocations of indivisible goods. In Proceedings of the 5th ACM conference on Electronic commerce, pages 125--131, 2004

    2004

  13. [21]

    Cooperation and fairness in multi-agent reinforcement learning

    Jasmine Jerry Aloor, Siddharth Nagar Nayak, Sydney Dolan, and Hamsa Balakrishnan. Cooperation and fairness in multi-agent reinforcement learning. Journal on Autonomous Transportation Systems, 2 0 (2): 0 1--25, 2024

    2024

  14. [22]

    A Theory of Justice

    John Rawls. A Theory of Justice. Harvard University Press, 1971

    1971

  15. [23]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.