Pith. sign in

REVIEW 4 major objections 5 minor 25 references

Multi-Agent Reinforcement Learning Scheduling to Support Low Latency in Teleoperated Driving

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Radio-resource scheduling trained by multi-agent reinforcement learning can keep teleoperated driving inside its end-to-end latency budget without compressing the sensor data.

desk verdict MAPPO+GA is the best on average latency but worst on the 95th-percentile tail in the paper's own Figure 4, so the abstract overclaims; the underlying study is solid enough to save. read the letter →

arxiv 2505.03558 v1 pith:R5HBTKCZ submitted 2025-05-06 cs.NI

Giacomo Avanzi , Marco Giordani , Michele Zorzi This is my paper · ORCID
classification cs.NI
keywords teleoperateddrivingpredictivequalityofservicemulti-agentreinforcementlearningproximalpolicyoptimizationradioresourceschedulingend-to-endlatency5GNRmmWave
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 tries to show that the radio scheduler, not data compression, is the right lever for keeping teleoperated driving within its end-to-end latency budget. It trains multiple agents with Proximal Policy Optimization to assign radio resources (OFDM symbols per 5G slot) to vehicles according to priority levels, and compares independent training (IPPO) with centralized training (MAPPO) under both proportional and greedy allocation. The central claim is that MAPPO combined with greedy allocation gives the lowest average latency and the highest probability of meeting the latency threshold, with the advantage growing as the number of vehicles and the data rate increase. A reader should care because this approach preserves the quality of the LiDAR data being transmitted while still reacting to congested or degraded network conditions.

What carries the argument

The load-bearing mechanism is the MARL scheduler operating inside the radio access network. Each vehicle is an agent whose observation bundles average SINR, uplink buffer size, required OFDM symbols, average modulation and coding index, and recent end-to-end latency; its action is a priority level k in {1,...,K}; and the reward is +1 when end-to-end latency is at or below the threshold and a penalty proportional to the violation otherwise. The scheduling strategies convert priorities into OFDM-symbol allocations: proportional allocation weights each vehicle's share by its priority and gives leftover symbols to the highest-priority vehicles, while greedy allocation gives all symbols to the highest-priority vehicle before serving the next. MAPPO is the centralized-training, decentralized-execution variant in which one actor-critic pair is trained on data from all agents, and the paper argues that this shared model coordinates agents better than IPPO, which trains each agent on its own observations.

What would settle it

Run the same MAPPO-with-greedy-allocation versus round-robin comparison across backhaul delays of 1, 5, 10, and 20 ms in the simulator, or on a hardware-in-the-loop teleoperation testbed with a measured 5G data path; if the approximately 35% latency-success advantage at five vehicles and aggressive compression disappears when the backhaul delay drops to 1 ms, the paper's claim that MARL scheduling is the decisive lever for end-to-end latency would be falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a multi-agent reinforcement-learning scheduler trained with centralized parameter sharing (MAPPO) and allocating all available OFDM symbols to the most latency-constrained vehicle (greedy allocation) outperforms both a round-robin baseline and decentralized training for teleoperated driving over millimeter-wave links. In simulations with five vehicles and the most aggressive compression setting, the probability that end-to-end latency stays under threshold is about 35% higher than with round-robin; under the least compressed setting, only MAPPO with greedy allocation keeps the median latency below the threshold, satisfying it more than 70% of the time versus about 50% for proportional allocation and 42% for round-robin. The paper also reports that this same greedy variant has the worst 95th-percentile latency-violation probability, because it starves lower-priority vehicles; the authors present the result as the best average-latency performance at the cost of fairness.

Load-bearing premise

The load-bearing premise is that the simulated environment, especially an ideal wired backhaul with a fixed 10 ms propagation delay and high-fidelity millimeter-wave channel traces from an urban mobility scenario, is representative enough that the relative ordering of schedulers carries over to real teleoperated driving; if the real backhaul delay is much smaller than 10 ms, the radio scheduler's share of the end-to-end latency budget changes and the ranking could shift.

Editorial extensions

If this is right

  • Network operators could improve teleoperated-driving latency without modifying the sensor data, since scheduling alone can lift the share of vehicles meeting a 25 ms end-to-end budget.
  • MAPPO with greedy allocation is the recommended configuration when congestion comes from many vehicles or high-rate LiDAR traffic, while under light load it offers little or no gain over round-robin.
  • The average-latency win of greedy allocation comes with a fairness cost: it produces the worst 95th-percentile latency violations, so delay-sensitive applications that care about worst cases should prefer proportional allocation or a hybrid strategy.
  • As the number of vehicles grows, all schedulers converge in average latency because the system becomes resource-limited, but the MARL schedulers still keep a higher latency-success probability than round-robin.
  • Because the action space is a small set of priority levels rather than per-symbol assignments, the approach fits existing 5G NR frame structures and remains scalable as the network grows.

Reading between the lines

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

  • The fixed 10 ms backhaul propagation delay in the simulation is a large share of the 25 ms latency budget, so I expect the relative scheduler ranking to compress if that delay is reduced; changing only the backhaul delay in the simulator would test this directly.
  • A reward function that penalizes tail latency more heavily, such as the 95th percentile rather than the average, would likely change the ranking between greedy and proportional allocation, given that the paper's own tail-violation results favor proportional allocation.
  • A natural extension is to let the number of priority levels scale with the number of vehicles or to use continuous actions, which could reduce the starvation that greedy allocation causes while keeping its average-latency benefit.
  • The RAN-level scheduler could be paired with compression as a fallback: use MARL scheduling first and activate data compression only when the predicted latency still violates the threshold, bounding the data-quality loss while keeping latency under control.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes multi-agent reinforcement learning (MARL) schedulers for teleoperated driving (TD) networks, in which each vehicle/agent asks the RAN for a priority level and the gNB allocates OFDM symbols either proportionally (PA) or greedily (GA). Two PPO-based training variants are compared: independent PPO (IPPO) and multi-agent PPO (MAPPO). The evaluation, carried out in ns-3 with mmWave traces from GEMV2 and SUMO mobility, compares MAPPO+PA, MAPPO+GA, IPPO variants, and a round-robin (RR) baseline as a function of the number of vehicles, the Draco compression configuration, and the latency threshold. The central claim, stated in the abstract and repeated in Sec. IV-B, is that MAPPO combined with GA achieves the best latency performance, especially as the number of vehicles grows, and that it maximizes the number of vehicles satisfying the latency constraint.

Significance. If understood as a claim about average/median latency and average latency-success probability, the paper makes a useful, reproducible empirical contribution: it releases source code, uses open simulators with realistic channel and mobility traces, and provides a concrete MARL formulation of RAN-level scheduling for TD. The comparison between IPPO and MAPPO under two allocation rules is also informative. However, the headline claim as written is broader than the evidence: the paper's own Fig. 4 shows that MAPPO+GA has the worst tail behavior (95th-percentile latency-violation probability) for every number of vehicles, and the conclusion explicitly concedes that many severe latency violations occur. Given that the paper motivates TD with 99%-99.999% reliability requirements, tail behavior is not a secondary concern but a decisive one. The contribution is therefore solid but needs a substantially more careful statement of what exactly is optimized and what the trade-offs are.

major comments (4)
  1. [Abstract; Sec. IV-B-b] The unqualified claim that "MAPPO, combined with GA, achieves the best results in terms of latency" is contradicted by the paper's own Fig. 4, which shows that MAPPO+GA has the highest 95th-percentile latency-violation probability for every N (3, 5, and 8). Since Sec. I motivates the work with 5GAA reliability requirements of 99% to 99.999%, the tail metric is the most safety-relevant one, not a minor caveat. Please restrict the central claim to average/median latency and to the average latency-success probability, and explicitly state in the abstract and in Sec. IV-B-b that MAPPO+GA trades fairness and worst-case violations for average performance.
  2. [Sec. IV-A-e; Figs. 3-5] All comparative conclusions are based on point estimates, but no confidence intervals, error bars, or statistical significance tests are reported despite 250 independent simulation runs per setting. For example, Sec. IV-B-b states that at N=5 MAPPO+GA improves the latency-success probability by about 35% over RR, yet no variance information is given, and Fig. 4 also reports point estimates without uncertainty. Without such information, "best" is not statistically established, especially because the reported differences in some regimes are small (e.g., the reward values in Fig. 1). Please add confidence intervals or significance tests for the key metrics, or at minimum error bars on the figures.
  3. [Sec. III-A-c, Eq. (2)] The reward function gives a constant +1 whenever latency is at or below the threshold and a linear penalty proportional to the violation size otherwise. This reward rewards reducing the average violation count but does not distinguish a moderate violation from a severe tail violation, and it provides no gradient toward avoiding extreme latency. This objective mismatch is a likely mechanism behind the Fig. 4 result, where the policy that maximizes average reward is also the one with the worst 95th-percentile violation probability. Please discuss this trade-off explicitly and consider either a tail-sensitive reward term or a per-UE penalty structure that prevents systematic starvation of lower-priority vehicles.
  4. [Sec. IV-A-a] The wired backhaul is modeled with a fixed propagation delay of 10 ms and a rate of 100 Gbps. This is a large, configuration-independent component of the 15-35 ms end-to-end latency budget, and although it may not change the relative ordering of schedulers, it can determine whether a configuration meets the threshold at all (e.g., the claim that MAPPO+GA is the only method keeping median latency below the threshold for (10,10) depends on the absolute latency budget). Please justify the 10 ms value, clarify whether it is one-way or round-trip, and ideally provide a sensitivity sweep over backhaul delay and rate to show that the qualitative conclusions are robust.
minor comments (5)
  1. [Abstract] The phrase "we prove via ns-3 simulations" is too strong for simulation-based evidence; "demonstrate" or "show" would be more appropriate.
  2. [Sec. IV-A-c] The number of priority levels K=3 is chosen empirically, but no sensitivity analysis is given. Since the greedy allocation rule starves lower-priority UEs by construction, the sensitivity of the tail results to K should at least be discussed.
  3. [Sec. IV-A-e] The end-to-end latency definition says it is measured from packet generation at the transmitter to reception, but it is not stated explicitly whether this includes the wired backhaul delay and queueing at the remote host or only the radio access segment. Please make the decomposition of the E2E latency explicit.
  4. [Fig. 3] The top subplot is described in the text as a boxplot, but the figure caption does not explain the boxplot elements (median, quartiles, whiskers). Please add a sentence to the caption.
  5. [Sec. IV-B-b] The sentence "the median latency is always lower than tau with the current settings" is stated before reporting the boxplot; please clarify that this holds for the evaluated configurations and that the mean can still exceed tau due to tail violations.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical ns-3 simulation study whose measured results do not reduce to their inputs; reward and evaluation share a latency threshold by design but are not fitted predictions, and self-citations are background only.

full rationale

This paper does not present a mathematical derivation chain whose outputs are equivalent to its inputs. The central result—that MAPPO combined with GA achieves lower average latency and higher average latency-success probability than RR—is obtained from ns-3 simulations using GEMV2 channel traces and SUMO mobility; it is measured, not fitted. The reward function in Eq. (2) uses the same latency threshold τ that later appears in the evaluation metric Pℓ≤τ and in Fig. 6, but this is standard RL objective design: the agent is trained to maximize the QoS metric being scored, and the actual latency values come from the simulator, not from the reward formula. The paper also sets K=3 based on offline simulations, but that is a hyperparameter choice, not a renamed prediction. Self-citations to prior group work ([6], [7], [8], [16]) describe the simulation scenario, the PQoS context, and the ns-3 RAN-AI framework; none of these is invoked as an external uniqueness theorem or as the sole justification for the current scheduling comparison. The abstract's unqualified claim that MAPPO+GA achieves 'the best results in terms of latency' is contradicted by Fig. 4, where MAPPO+GA has the worst 95th-percentile latency-violation probability, and the conclusion itself concedes that 'many severe latency violations are experienced.' That is an overstatement or internal-consistency issue, not circularity: the claim is broader than the evidence, but it is not derived from the evidence by construction. No load-bearing step reduces to a self-citation or to the definition of a fitted parameter.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central result is an empirical simulation comparison. There are no new particles, forces, or formal objects. The quantitative behavior depends on two hand-chosen parameters (K and the reward penalty scale) plus the standard ns-3 model stack; these are the main ad hoc inputs.

free parameters (2)
  • Number of priority levels (K) = 3
    Chosen empirically via offline simulations (Sec. IV-A-c); determines the granularity of the action space and affects how much differentiation between UEs is possible.
  • Reward penalty scaling coefficient = 1/100 (in Eq. 2)
    The penalty term -(ell - tau)/100 is hand-selected; it controls the magnitude of the penalty relative to the +1 success reward and shapes the learned policy.
assumptions (4)
  • domain assumption The 5G NR mmWave channel and mobility models implemented in ns-3, SUMO, and GEMV2 faithfully represent a teleoperated driving RAN.
    All results are generated from these simulations (Sec. II-A); if the channel or mobility models are unrepresentative, the latency numbers and scheduler rankings may not transfer to real networks.
  • standard math The MDP/Dec-POMDP formulation and the PPO clipped surrogate learning rule are appropriate for this RAN scheduling problem.
    Sec. III relies on standard RL assumptions of Markov transitions and on the PPO objective in Eq. (3); these are accepted background results.
  • domain assumption The LiDAR traffic model with Draco compression configurations (q,c) is representative of teleoperated driving data.
    Sec. II-A-c fixes frame rate and compression; the conclusions are tied to these data sizes and rates.
  • domain assumption The reward function in Eq. (2) correctly encodes the QoS objective.
    The agent optimizes this reward; if the reward is misspecified, the learned scheduling policy would not reflect the true latency objective.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Agent Reinforcement Learning Scheduling to Support Low Latency in Teleoperated Driving." pith.science (2026). https://pith.science/paper/R5HBTKCZ

@misc{pith2026250503558,
  author       = {Pith},
  title        = {Pith review of: Multi-Agent Reinforcement Learning Scheduling to Support Low Latency in Teleoperated Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R5HBTKCZ}},
  note         = {Machine review of arXiv:2505.03558}
}
read the original abstract

The teleoperated driving (TD) scenario comes with stringent Quality of Service (QoS) communication constraints, especially in terms of end-to-end (E2E) latency and reliability. In this context, Predictive Quality of Service (PQoS), possibly combined with Reinforcement Learning (RL) techniques, is a powerful tool to estimate QoS degradation and react accordingly. For example, an intelligent agent can be trained to select the optimal compression configuration for automotive data, and reduce the file size whenever QoS conditions deteriorate. However, compression may inevitably compromise data quality, with negative implications for the TD application. An alternative strategy involves operating at the Radio Access Network (RAN) level to optimize radio parameters based on current network conditions, while preserving data quality. In this paper, we propose Multi-Agent Reinforcement Learning (MARL) scheduling algorithms, based on Proximal Policy Optimization (PPO), to dynamically and intelligently allocate radio resources to minimize E2E latency in a TD scenario. We evaluate two training paradigms, i.e., decentralized learning with local observations (IPPO) vs. centralized aggregation (MAPPO), in conjunction with two resource allocation strategies, i.e., proportional allocation (PA) and greedy allocation (GA). We prove via ns-3 simulations that MAPPO, combined with GA, achieves the best results in terms of latency, especially as the number of vehicles increases.

Figures

Figures reproduced from arXiv: 2505.03558 by the authors.

Figure 1
Figure 1. Average reward at the end of the training for IPPO and [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Average reward at the beginning (top) and at end of the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. The 95-percentile of the latency-violation probability vs. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: Average latency, reward and latency-success probability vs. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png]
Figure 5
Figure 5. Figure 5: Average latency, reward and latency-success probability vs. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Average reward at the end of the training vs. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages

  1. [1]

    Toward 6G Networks: Use Cases and Technologies,

    M. Giordani, M. Polese, M. Mezzavilla, S. Rangan, and M. Zorzi, “Toward 6G Networks: Use Cases and Technologies,” IEEE Commu- nications Magazine, vol. 58, no. 3, pp. 55–61, Mar. 2020

    work page 2020

  2. [2]

    Toward automated vehicle teleoperation: Vision, opportuni- ties, and challenges,

    T. Zhang, “Toward automated vehicle teleoperation: Vision, opportuni- ties, and challenges,” IEEE Internet of Things Journal , vol. 7, no. 12, pp. 11 347–11 354, Dec. 2020

    work page 2020

  3. [3]

    C-V2X Use Cases V olume II: Examples and Service Level Requirements,

    5GAA, “C-V2X Use Cases V olume II: Examples and Service Level Requirements,” White Paper, 2020

    work page 2020

  4. [4]

    Millimeter-wave vehicular communication to support massive automotive sensing,

    J. Choi, V . Va, N. Gonzalez-Prelcic, R. Daniels, C. R. Bhat, and R. W. Heath, “Millimeter-wave vehicular communication to support massive automotive sensing,” IEEE Communications Magazine , vol. 54, no. 12, pp. 160–167, Dec. 2016

    work page 2016

  5. [5]

    Predictive Quality of Service (PQoS): The Next Frontier for Fully Autonomous Systems,

    M. Boban, M. Giordani, and M. Zorzi, “Predictive Quality of Service (PQoS): The Next Frontier for Fully Autonomous Systems,” IEEE Network, vol. 35, no. 6, pp. 104–110, Nov/Dec 2021

    work page 2021

  6. [6]

    A Reinforcement Learning Framework for PQoS in a Teleoperated Driving Scenario,

    F. Mason, M. Drago, T. Zugno, M. Giordani, M. Boban, and M. Zorzi, “A Reinforcement Learning Framework for PQoS in a Teleoperated Driving Scenario,” in IEEE Wireless Communications and Networking Conference (WCNC), 2022

    work page 2022

  7. [7]

    Towards Decentralized Predictive Quality of Service in Next-Generation Vehicular Networks,

    F. Bragato, T. Lotta, G. Ventura, M. Drago, F. Mason, M. Giordani, and M. Zorzi, “Towards Decentralized Predictive Quality of Service in Next-Generation Vehicular Networks,” in IEEE Information Theory and Applications Workshop (ITA), 2023

    work page 2023

  8. [8]

    Federated Reinforcement Learning to Optimize Teleoperated Driving Networks

    F. Bragato, M. Giordani, and M. Zorzi, “Federated Reinforcement Learning to Optimize Teleoperated Driving Networks,” in IEEE Global Communications Conference , 2024. [Online]. Available: https: //arxiv.org/abs/2410.02312

    work page Pith review arXiv 2024

Show all 25 references
  1. [9]

    Deep-Learning-Based Wireless Resource Allocation With Application to Vehicular Networks,

    L. Liang, H. Ye, G. Yu, and G. Y . Li, “Deep-Learning-Based Wireless Resource Allocation With Application to Vehicular Networks,”Proceed- ings of the IEEE , vol. 108, no. 2, pp. 341–356, Feb. 2020

    2020

  2. [10]

    Knowledge-Assisted Deep Reinforcement Learning in 5G Scheduler Design: From Theoretical Framework to Implementation,

    Z. Gu, C. She, W. Hardjawana, S. Lumb, D. McKechnie, T. Essery, and B. Vucetic, “Knowledge-Assisted Deep Reinforcement Learning in 5G Scheduler Design: From Theoretical Framework to Implementation,” IEEE Journal on Selected Areas in Communications , vol. 39, no. 7, pp. 2014–20...

    2014

  3. [11]

    5G Resource Scheduling for Low-latency Communication: A Reinforcement Learning Approach,

    Q. Huang and M. Kadoch, “5G Resource Scheduling for Low-latency Communication: A Reinforcement Learning Approach,” in IEEE 92nd Vehicular Technology Conference (VTC2020-Fall), 2020

    2020

  4. [12]

    End-to-End Simulation of 5G mmWave Networks,

    M. Mezzavilla, M. Zhang, M. Polese, R. Ford, S. Dutta, S. Rangan, and M. Zorzi, “End-to-End Simulation of 5G mmWave Networks,” IEEE Communications Surveys & Tutorials , vol. 20, no. 3, pp. 2237–2263, Thirdquarter 2018

    2018

  5. [13]

    Recent devel- opment and applications of SUMO - Simulation of Urban MObility,

    D. Krajzewicz, J. Erdmann, M. Behrisch, and L. Bieker, “Recent devel- opment and applications of SUMO - Simulation of Urban MObility,” International Journal On Advances in Systems and Measurements , vol. 5, no. 3&4, pp. 128–138, Dec. 2012

    2012

  6. [14]

    Geometry-based vehicle-to- vehicle channel modeling for large-scale simulation,

    M. Boban, J. Barros, and O. K. Tonguz, “Geometry-based vehicle-to- vehicle channel modeling for large-scale simulation,” IEEE Transactions on Vehicular Technology, vol. 63, no. 9, pp. 4146–4164, Nov. 2014

    2014

  7. [15]

    (2017) Draco 3D Data Compression

    Google. (2017) Draco 3D Data Compression. [Online]. Available: https://google.github.io/draco/

    2017

  8. [16]

    Artificial Intelligence in Vehicular Wireless Networks: A Case Study Using ns-3,

    M. Drago, T. Zugno, F. Mason, M. Giordani, M. Boban, and M. Zorzi, “Artificial Intelligence in Vehicular Wireless Networks: A Case Study Using ns-3,” in Proceedings of the 2022 ACM Workshop on Ns-3 , 2022

    2022

  9. [17]

    NR and NG-RAN Overall Description (Release 15),

    3GPP, “NR and NG-RAN Overall Description (Release 15),” TS 38.300, 2018

    2018

  10. [18]

    S. V . Albrecht, F. Christianos, and L. Sch¨afer, Multi-Agent Reinforcement Learning: Foundations and Modern Approaches . MIT Press, 2024

    2024

  11. [19]

    The complexity of decentralized control of Markov decision processes,

    D. S. Bernstein, S. Zilberstein, and N. Immerman, “The complexity of decentralized control of Markov decision processes,” in Proceedings of the Sixteenth Conference on Uncertainty in Artificial Intelligence , 2000

    2000

  12. [20]

    Proximal policy optimization algorithms,

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

    2017 arXiv

  13. [21]

    Trust region policy optimization,

    J. Schulman, S. Levine, P. Abbeel, M. Jordan, and P. Moritz, “Trust region policy optimization,” in Proceedings of the 32nd International Conference on Machine Learning , 2015

    2015

  14. [22]

    High- dimensional continuous control using generalized advantage estimation,

    J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel, “High- dimensional continuous control using generalized advantage estimation,” in Proceedings of the International Conference on Learning Represen- tations (ICLR), 2016

    2016

  15. [23]

    Is independent learning all you need in the starcraft multi-agent challenge?

    C. S. de Witt, T. Gupta, D. Makoviichuk, V . Makoviychuk, P. H. S. Torr, M. Sun, and S. Whiteson, “Is independent learning all you need in the starcraft multi-agent challenge?” 2020. [Online]. Available: https://arxiv.org/abs/2011.09533

    2020 arXiv

  16. [24]

    The surprising effectiveness of PPO in cooperative multi-agent games,

    C. Yu, A. Velu, E. Vinitsky, J. Gao, Y . Wang, A. Bayen, and Y . Wu, “The surprising effectiveness of PPO in cooperative multi-agent games,” in Proceedings of the 36th International Conference on Neural Information Processing Systems, 2022

    2022

  17. [25]

    An introduction to centralized training for decentralized execution in cooperative multi-agent reinforcement learning,

    C. Amato, “An introduction to centralized training for decentralized execution in cooperative multi-agent reinforcement learning,” 2024. [Online]. Available: https://arxiv.org/abs/2409.03052

    2024 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.