Pith. sign in

REVIEW 4 major objections 5 minor 20 references

Multi-Agent Transformer for Queue-Level XR Traffic Scheduling in TSN Networks

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

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

Pith's one-line read A multi-agent transformer that gives every TSN queue its own RL agent can schedule co-located AR and semantic-XR traffic, reducing queueing latency by up to 71.42% and failure rates by up to 83.2% in simulation.

desk verdict Applies MAT to queue-level TSN scheduling for mixed XR traffic; the idea is sensible, but the paper overclaims and the inference-latency numbers undercut the practical story. read the letter →

arxiv 2608.05340 v1 pith:T57UPAQJ submitted 2026-08-05 cs.NI cs.AI

Marcos Carvalho , Fatih Temiz , Shavbo Salehi , Melike Erol-Kantarci , Daniel F. Macedo This is my paper · ORCID
classification cs.NIcs.AI
keywords Time-SensitiveNetworkingMulti-AgentReinforcementLearningTransformerExtendedRealityTrafficSchedulingMobileEdgeComputingQueue-LevelControlLatencyReduction
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

This paper argues that in a time-sensitive network carrying mixed XR traffic, scheduling decisions are best made at the queue level, with each queue acting as an independent agent. The authors' contribution is to show that a multi-agent transformer, which uses attention over other queues' observations and previously chosen actions, can coordinate these queue agents more effectively than centralized reinforcement learning or heuristic rules. In their simulated MEC-TSN environment with eight queues serving AR and semantic-XR applications under hard deadlines, the MAT scheduler reduces 95th-percentile queueing latency by up to 71.42% and deadline-failure rates by up to 83.2% relative to baselines. If this holds beyond simulation, it gives XR network operators a concrete way to protect millisecond-level semantic traffic without starving ordinary AR flows.

What carries the argument

The load-bearing mechanism is the multi-agent transformer (MAT) architecture, originally a sequence-modeling treatment of MARL. Its observation encoder uses multi-head self-attention over all queue observations, producing latent representations that mix each queue's state with the state of the others; its auto-regressive action decoder then generates each queue's desired time-slot fraction conditioned on actions already generated for other queues. The generated fractions are normalized to the 1 ms cycle and sorted in descending order to form the schedule, with ties broken by queue index. Training uses an actor–critic objective with a team reward that penalizes deadline violations quadratically once the ratio of oldest-packet age to deadline exceeds one, plus a bandwidth-utilization penalty.

What would settle it

Measure the trained MAT policy's end-to-end performance on a real TSN-capable NIC or switch with the 1 ms cycle and the same XR workload; if the inference plus scheduling overhead cannot fit in the cycle on deployment hardware, the claimed latency and reliability gains cannot be delivered.

Watch

Extended reading notes

Core claim

The central claim is that inter-queue dependencies are the key to good XR scheduling, and the multi-agent transformer is an effective way to learn them. Each TSN queue is an agent that outputs a continuous fraction of the fixed 1 ms cycle; schedules are formed by ordering queues by their assigned fractions, so the policy implicitly learns a dynamic priority order. The transformer encoder processes all queue observations at once, and the auto-regressive decoder generates actions one queue at a time, conditioning each choice on those already made. The authors report that this setup consistently beats PPO, A2C, HAPPO, and backlog/deadline heuristics on reliability, and that it moves SeAR queues from roughly two-nines to three-nines reliability.

Load-bearing premise

The learned scheduler must be able to compute a new schedule inside the fixed 1 ms cycle on the hardware that really runs the TSN network; the paper's own measurements show one tested GPU taking 1.26 ms, longer than the cycle.

Editorial extensions

If this is right

  • Queue-level MAT scheduling protects SeAR traffic's 1 ms deadline without degrading AR queues, something the backlog and deadline heuristics fail to do.
  • The learned policy generalizes across link speeds from 10 Gbps to 100 Gbps, keeping reliability high where the deadline heuristic needs extra capacity to catch up.
  • Moving from centralized single-agent RL to queue-level agent decomposition can reduce worst-case queueing latency by up to 71.42% in high-contention XR scenarios.
  • The scheduler's reliance on high-end GPUs (0.98–1.26 ms per inference on tested hardware) means its real-time use in a 1 ms cycle is marginal, a limitation the paper explicitly acknowledges.

Reading between the lines

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

  • Because the paper only tests synthetic Johnson's SU traffic, a natural extension is to run the trained policy on trace-driven or live XR traffic; if it generalizes, queue-level attention scheduling becomes a candidate for production TSN gate control.
  • The 1 ms inference bottleneck invites a distillation step the paper leaves implicit: a distilled or pruned policy could run on NIC hardware, turning the simulation result into a deployable real-time scheduler.
  • The attention patterns over queue observations could be analyzed to extract human-readable priority rules; if those rules approximate the learned policy, they might be compiled into a deterministic scheduler with no GPU requirement.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes a queue-level multi-agent reinforcement learning scheduler for XR traffic in MEC-TSN networks, where each TSN queue is an agent and the policy is realized with the Multi-Agent Transformer (MAT). Each agent outputs a fraction of the 1 ms TSN cycle, the durations are normalized to fill the cycle, and the service order is the descending order of allocated slot durations with index-based tie-breaking. The authors formulate the scheduling problem as a constrained optimization, cast it as a multi-agent POMDP with a team reward that penalizes the maximum AOP-to-deadline ratio and bandwidth waste, and evaluate the approach against PPO, A2C, HAPPO, a backlog heuristic, and a deadline heuristic. Simulated results report lower 95th-percentile queueing latency and higher per-queue service reliability, with up to 71.42% latency reduction and up to 83.2% failure-rate reduction, together with GPU inference-time measurements.

Significance. The paper addresses a timely and practically relevant problem: scheduling heterogeneous XR flows with different deadline criticality inside a TSN switch. Modeling each queue as an agent and using attention to capture inter-queue dependencies is a reasonable design direction, and the comparison against multiple single-agent, MARL, and heuristic baselines is appropriate for a first evaluation. The inclusion of inference-time measurements on GPUs is valuable because it directly probes the system's ability to respect the 1 ms TSN cycle. However, the significance is conditional: the central claim of consistent outperformance is contradicted by at least one row of Table II, and the reported inference latency leaves essentially no margin for the intended live scheduling loop. If these issues are resolved and the results survive a more rigorous statistical treatment, the paper would make a useful contribution to RL-based TSN scheduling.

major comments (4)
  1. [Section IV.B, Table II] The statement that MAT 'consistently outperforms all baselines' is contradicted by the q1 row: MAT achieves 99.6697% reliability while the deadline-aware heuristic achieves 99.6723%, so MAT has a higher failure rate than DAH. The arrow notation for q1 is also unexplained, and the claimed 'up to 83.2% reduction in failure rate' does not match the best-baseline comparison in Table II for q2 (DAH, 99.7719%) or for q5 (HAPPO, 99.8838%) when recomputed from the failure rates. The authors should correct the table, clarify the baseline used for each arrow, and soften or qualify the 'consistently outperforms' claim in the abstract and Section IV.B.
  2. [Section II.D and Section II.E, Eq. (6a) vs. Eq. (7)] The formal objective Eq. (6a) minimizes a sum over cycles, time slots, and queues of lambda1 * (AOP/deadline) plus lambda2 * bandwidth waste, while the implemented team reward Eq. (7) is -max_i Phi_i(t) - beta * average bandwidth waste, with Phi_i(t) being a piecewise linear/quadratic function of rho_i(t). These are structurally different objectives: the max operator replaces the sum, and the quadratic penalty for rho_i(t)>1 has no counterpart in Eq. (6a). Since the reported performance is measured primarily through AOP and reliability, the relationship between the reward actually optimized and the formal problem objective must be justified. At minimum, the authors should reconcile Eq. (6a) with Eq. (7) or explain why the max-of-piecewise-quadratic reward is the intended surrogate.
  3. [Section II.C and Section IV.B, inference-latency measurements] The paper assumes a fixed scheduling cycle C_d = 1 ms and claims practical significance for TSN scheduling. Yet the authors' own measurements show MAT inference latency of 1.26 ± 0.037 ms on an RTX 5090, which already exceeds the cycle, and 0.9777 ± 0.0039 ms on an RTX PRO 6000 S, leaving less than 3% of the cycle for observation collection, preprocessing, action dissemination, and gate updates. No tail-latency statistics, no end-to-end scheduling-loop budget, and no accounting for these overheads are provided. Section V concedes that 'future work may investigate advanced knowledge and policy distillation techniques to further reduce inference latency.' This means the current solution cannot reliably compute a schedule within the intended TSN cycle, so the simulated latency and reliability gains are not yet deliverable in a live deployment. The authors should either conduct a complete latency-budget analysis, modify the hardware/algorithm so that inference reliably fits in the cycle, or clearly reframe the contribution as an offline scheduling proof-of-concept rather than a real-time scheduler.
  4. [Section IV.A and Section IV.B, statistical evidence] All results are reported as means over three seeds, and the reliability differences between MAT and the best baseline are often very small (for example, q0: 99.7284 vs 99.7021, and q4: 99.6699 vs 99.6551). The text claims MAT 'significantly outperforms' the baselines, but no confidence intervals, paired comparisons, or hypothesis tests are provided. Given the small effect sizes and the q1 inconsistency, the 'significantly outperforms' claim is not statistically supported. The authors should add per-seed results, confidence intervals, or a statistical test, and adjust the language of significance accordingly.
minor comments (5)
  1. [Section II.C, Eq. (4)] The notation is confusing: T is used both as the number of time slots and as a set of slot-to-queue assignments, and the sum runs over t_j but the indexing in Eq. (6a) introduces t_k without a clear mapping to the slot durations.
  2. [Algorithm 1, lines 12-14 and 23-26] The action decoding loop is described as 'Input a_0, ..., a_m and infer a_{m+1}' for m=0,1,...,Q, but if m starts at 0 then a_0 has not yet been generated; the loop bound and the initial condition should be clarified. In addition, L_Encoder(phi) is computed twice (lines 23 and 26), which is likely a typo.
  3. [Table II and Section IV.B] The sentence 'the difference between MAT and the baselines can reach a shift from two-nines to three-nines reliability' is imprecise because several baselines already operate in the three-nines range; consider describing the exact failure-rate reductions instead.
  4. [Figure 4] Figure 4 reports reliability at different network speeds but does not show error bars or indicate the number of seeds, unlike Figure 2 and Table II; this makes it hard to assess the stability of the generalization claim.
  5. [Section II.B, Eq. (1)] The notation t(f)start and the sum index in the IPI formula are not typeset cleanly, and the definition of t_s(p_j^{(i)}) in Eq. (2) would benefit from an explicit statement that it is the service start time observed in the simulator.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central result is an empirical MARL benchmark; reward alignment is standard objective design and self-citations are not load-bearing.

full rationale

The paper makes no first-principles derivation whose conclusion is equivalent to its assumptions. Its claim that MAT improves latency and reliability is an empirical comparison of trained policies against PPO, A2C, HAPPO, and two heuristics in a simulator. The reward function (Eq. 7) directly includes the AOP-deadline ratio and bandwidth waste, and the evaluation metrics are p95 AOP and deadline reliability; this alignment is normal RL objective design, not a fitted-parameter prediction or a self-definitional 'prediction' of the reward itself. Baselines are trained or evaluated under the same setting, so the comparison is not forced. The paper's self-citations ([1]-[3]) support motivational framing and the assumption of a TSN-capable NIC; neither is the source of the claimed result. The MAT architecture and HAPPO/PPO/A2C baselines are cited to external work [13], [19], [20], and no uniqueness theorem or ansatz is imported from the authors' prior work. The measured inference latency on RTX 5090 exceeding the 1 ms cycle is a practical deployment limitation and a correctness-risk concern, but it is not circular reasoning. Accordingly, no circular step meets the evidence threshold.

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

The paper's central empirical result rests mainly on modeling choices (queue-per-application isolation, Johnson's SU traffic, hard deadlines, the sort-by-duration service order) and on selected reward weights. No new physical entities are introduced. The experiment is a simulation with no released artifacts, so most assumptions cannot be checked independently.

free parameters (4)
  • reward penalty weight beta = 0.2
    Chosen in Section IV.A. Balances the deadline-violation penalty against the bandwidth-utilization penalty; no sensitivity analysis is reported, and the reported latency and failure metrics are directly optimized by this reward.
  • deadline-violation scaling lambda = 4
    Section IV.A. Sets the quadratic penalty once AOP exceeds the deadline; changing it changes how conservative the learned policy is, and no sensitivity study is shown.
  • AR deadline fraction = 0.5 x IFI, giving 5.55 ms and 6.94 ms
    Section IV.A assumes half of the end-to-end latency budget is consumed before the TSN scheduling domain. This arbitrary split changes which packets count as failures.
  • SeAR deadline = 1 ms
    Section IV.A imposes a single-cycle hard deadline for SeAR queues; the 1 ms value coincides with the cycle duration, making SeAR scheduling maximally constrained by the design.
assumptions (5)
  • domain assumption Each TSN queue serves exactly one XR application with strict traffic isolation on a fixed NIC queue set.
    Section II.A defines Q and says 'each queue serves a specific XR application to ensure strict isolation.' If queues are shared, the per-queue agent decomposition and its independence assumptions fail.
  • domain assumption IFI and IPI for XR traffic follow Johnson's SU distributions.
    Section II.B invokes [15] for Johnson's SU but does not give the fitted parameters, so the simulated traffic mix is not fully specified.
  • ad hoc to paper Service order within a cycle is the descending order of allocated time-slot durations, with ties resolved by higher queue index.
    Section II.C defines this ordering as part of the schedule model. It is a design choice, not a stated IEEE 802.1Qbv requirement, and the whole schedule representation depends on it.
  • domain assumption All packets of an XR frame must be transmitted before the next frame arrives, and SeAR deadlines are stricter than AR deadlines.
    Section II.D adopts the hard-deadline model from [17] and fixes SeAR at 1 ms. The evaluation metrics are computed against these deadlines.
  • domain assumption The omitted MAT loss functions, value function, and GAE definitions are exactly those of the original MAT paper [13].
    Algorithm 1 and Section III.B defer L_Encoder, L_Decoder, value function, and joint advantage to [13]; the training procedure, and therefore the results, depend on those definitions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Agent Transformer for Queue-Level XR Traffic Scheduling in TSN Networks." pith.science (2026). https://pith.science/paper/T57UPAQJ

@misc{pith2026260805340,
  author       = {Pith},
  title        = {Pith review of: Multi-Agent Transformer for Queue-Level XR Traffic Scheduling in TSN Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T57UPAQJ}},
  note         = {Machine review of arXiv:2608.05340}
}
read the original abstract

Time-Sensitive Networking (TSN) and Mobile Edge Computing (MEC) hold strong potential for enabling ultra-reliable low-latency communication for time-sensitive applications, such as eXtended Reality (XR). However, the widespread adoption of XR introduces significant challenges due to co-located services in MEC environments, leading to contention for shared network resources. Moreover, XR traffic types have distinct characteristics and criticality in terms of timing requirements, further increasing the complexity and dynamics of such environments. Although reinforcement learning has shown promise for TSN scheduling optimization in dynamic network scenarios, existing approaches rely on centralized or high-level multi-agent designs and are typically tailored to periodic and predictable industrial traffic, limiting their applicability to XR workloads. As a result, these approaches suffer from (i) limited ability to capture inter-queue dependencies due to coarse-grained control, and (ii) poor adaptability to highly dynamic and heterogeneous XR traffic. To address these gaps, we propose a multi-agent reinforcement learning approach for queue-level XR traffic scheduling. We adopt the multi-agent transformer (MAT) to model inter-queue dependencies via attention over agents' observations and actions, enabling implicit coordination across heterogeneous co-located XR applications. Our simulation results show that the proposed method outperforms baselines, achieving up to 71.42% latency reduction and up to 83.2% reduction in failure rate, while consistently achieving high reliability across all queues.

Figures

Figures reproduced from arXiv: 2608.05340 by the authors.

Figure 1
Figure 1. Overall System Model s ∗ = arg min s∈S X C c=1 "X T k=0 X Q i=0 λ1 gi,tk,c Dqi + X tk∈c λ2 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. 95th percentile of queueing latency To further assess performance, we evaluate overall service reliability, defined as the percentage of packets meeting their deadlines and expressed in “nines,” as reported in Table II. Table II: Service reliability (%) across the queues. Arrows indicate the percentage reduction in failure rate compared to the best baseline method. MAT HAPPO PPO A2C BAH DAH q0 99.7284(0.01) ↑8.8% 99… view at source ↗
Figure 2
Figure 2. Average Cumulative Reward In contrast, HAPPO exhibits higher variance due to its ran￾domly ordered sequential updates, which hinder dependency capture. A2C achieves a significantly higher final reward, which can be explained by the fact that A2C avoids the conser￾vative updates induced by PPO’s clipped objective. However, A2C is limited by its centralized decision-making, which does not explicitly model the interact… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Service reliability (%) under different network speed [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 12 canonical work pages

  1. [15]

    An extended reality offloading ip traffic dataset and models,

    D. G. Mor ´ın, D. Medda, A. Iossifides, P. Chatzimisios, A. G. Armada, A. Villegas, and P. Per ´ez, “An extended reality offloading ip traffic dataset and models,”IEEE Transactions on Mobile Computing, vol. 23, no. 6, pp. 6820–6834, 2023

    2023

  2. [1]

    Edge learning via federated split decision transformers for metaverse resource allocation,

    F. Temiz, S. Salehi, and M. Erol-Kantarci, “Edge learning via federated split decision transformers for metaverse resource allocation,” inIEEE International Conference on Communications (ICC). IEEE, 2026

    work page 2026

  3. [2]

    Self-play ensemble q-learning enabled resource allocation for network slicing,

    S. Salehi, P. E. Iturria-Rivera, M. Elsayed, M. Bavand, R. Gaigalas, Y . Ozcan, and M. Erol-Kantarci, “Self-play ensemble q-learning enabled resource allocation for network slicing,” inGLOBECOM 2024-2024 IEEE Global Communications Conference. IEEE, 2024, pp. 4528– 4533

    work page 2024

  4. [3]

    Performance analysis of the integra- tion of dynamic cloud computing environments and tsn networks,

    M. Carvalho and D. F. Macedo, “Performance analysis of the integra- tion of dynamic cloud computing environments and tsn networks,” in 2025 IEEE 11th International Conference on Network Softwarization (NetSoft). IEEE, 2025, pp. 406–414

    2025

  5. [4]

    IEEE Standard for local and metropolitan area networks—bridges and bridged networks—amendment 25: Enhancements for scheduled traffic,IEEE Standard 802.1qbv-2015, 2016, pp. 1–57,

    “IEEE Standard for local and metropolitan area networks—bridges and bridged networks—amendment 25: Enhancements for scheduled traffic,IEEE Standard 802.1qbv-2015, 2016, pp. 1–57,” 2016

    2015

  6. [5]

    Enhancing mobile immersive streaming experience via deadline-aware scheduling and learning-enhanced congestion control,

    L. Luo, Y . Zhou, J. Shen, H. Luo, W. Wu, and H. Yu, “Enhancing mobile immersive streaming experience via deadline-aware scheduling and learning-enhanced congestion control,”IEEE Internet of Things Journal, 2025

    work page 2025

  7. [6]

    A survey of schedul- ing algorithms for the time-aware shaper in time-sensitive networking (tsn),

    T. St ¨uber, L. Osswald, S. Lindner, and M. Menth, “A survey of schedul- ing algorithms for the time-aware shaper in time-sensitive networking (tsn),”Ieee Access, vol. 11, pp. 61 192–61 233, 2023

    2023

  8. [7]

    Configuring the ieee 802.1 q time-aware shaper with deep reinforcement learning,

    A. Roberty, Q. Besnard, S. B. H. Said, F. Ridouard, H. Bauer, and A. Geniet, “Configuring the ieee 802.1 q time-aware shaper with deep reinforcement learning,” inNOMS 2024-2024 IEEE Network Operations and Management Symposium. IEEE, 2024, pp. 1–7

    2024

Show all 20 references
  1. [8]

    Mitigation of scheduling violations in time- sensitive networking using deep deterministic policy gradient,

    B. Zhou and L. Cheng, “Mitigation of scheduling violations in time- sensitive networking using deep deterministic policy gradient,” inPro- ceedings of the 4th FlexNets Workshop on Flexible Networks Artificial Intelligence Supported Network Flexibility and Agility, 2021, pp. 32–37

    2021

  2. [9]

    Cooperative resource allocation and traffic scheduling for iiot controllers in edge clouds: A hierarchical reinforcement learning approach,

    A. A. Saruhan, M. Abuibaid, A. H. Ghorab, M. St-Hilaire, and I. Moldov´an, “Cooperative resource allocation and traffic scheduling for iiot controllers in edge clouds: A hierarchical reinforcement learning approach,” in2025 International Conference on Smart Applications, Commu...

    2025

  3. [10]

    Towards distributed flow scheduling in ieee 802.1 qbv time-sensitive networks,

    M. Guo, S. He, C. Gu, X. Guo, J. Chen, T. Gao, and T. Wang, “Towards distributed flow scheduling in ieee 802.1 qbv time-sensitive networks,” ACM Transactions on Sensor Networks, vol. 20, no. 5, pp. 1–30, 2024

    2024

  4. [11]

    Multi-agent reinforcement learning-based routing and scheduling models in time-sensitive networking for internet of vehicles communi- cations between transportation field cabinets,

    S. Garcia-Cant ´on, C. Ruiz de Mendoza, C. Cervell ´o-Pastor, and S. Sal- lent, “Multi-agent reinforcement learning-based routing and scheduling models in time-sensitive networking for internet of vehicles communi- cations between transportation field cabinets,”Applied Science...

    2025

  5. [12]

    Sharp: A study on safe heterogeneous agent reinforcement learning paradigm for 5g-tsn traffic scheduling,

    S. Tan, C. Dong, and W. Wen, “Sharp: A study on safe heterogeneous agent reinforcement learning paradigm for 5g-tsn traffic scheduling,” in 2025 3rd International Conference On Mobile Internet, Cloud Comput- ing and Information Security (MICCIS). IEEE, 2025, pp. 1–9

    2025

  6. [13]

    Multi-agent reinforcement learning is a sequence modeling problem,

    M. Wen, J. Kuba, R. Lin, W. Zhang, Y . Wen, J. Wang, and Y . Yang, “Multi-agent reinforcement learning is a sequence modeling problem,” Advances in Neural Information Processing Systems, vol. 35, pp. 16 509–16 521, 2022

    2022

  7. [14]

    Semantic communi- cations in networked systems: A data significance perspective,

    E. Uysal, O. Kaya, A. Ephremides, J. Gross, M. Codreanu, P. Popovski, M. Assaad, G. Liva, A. Munari, B. Soretet al., “Semantic communi- cations in networked systems: A data significance perspective,”IEEE Network, vol. 36, no. 4, pp. 233–240, 2022

    2022

  8. [16]

    Methodology and infrastructure for tsn-based reproducible network experiments,

    M. Bosk, F. Rezabek, K. Holzinger, A. G. Marino, A. A. Kane, F. Fons, J. Ott, and G. Carle, “Methodology and infrastructure for tsn-based reproducible network experiments,”IEEE Access, vol. 10, pp. 109 203– 109 239, 2022

    2022

  9. [17]

    Pdu-set scheduling algorithm for xr traffic in multi-service 5g-advanced networks,

    P. Paymard, S. Paris, A. Amiri, T. E. Kolding, F. S. Moya, and K. I. Pedersen, “Pdu-set scheduling algorithm for xr traffic in multi-service 5g-advanced networks,” inICC 2024-IEEE International Conference on Communications. IEEE, 2024, pp. 758–763

    2024

  10. [18]

    Multi-agent transformer approach for collaborative task offloading and resource optimization in noma-based vehicular edge computing,

    Z. Zhou, Z. Zhang, J. J. Zeng, and H.-C. Chao, “Multi-agent transformer approach for collaborative task offloading and resource optimization in noma-based vehicular edge computing,”Ad Hoc Networks, p. 104222, 2026

    2026

  11. [19]

    Asynchronous methods for deep rein- forcement learning,

    V . Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu, “Asynchronous methods for deep rein- forcement learning,” inInternational conference on machine learning. PmLR, 2016, pp. 1928–1937

    2016

  12. [20]

    Trust region policy optimisation in multi-agent reinforcement learning,

    J. G. Kuba, R. Chen, M. Wen, Y . Wen, F. Sun, J. Wang, and Y . Yang, “Trust region policy optimisation in multi-agent reinforcement learning,” arXiv preprint arXiv:2109.11251, 2021

    2021 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.