pith. sign in

arxiv: 2510.25105 · v2 · submitted 2025-10-29 · 💻 cs.NI

Learning-Based vs Human-Derived Congestion Control: An In-Depth Experimental Study

Mihai Mazilu , Luca Giacomoni , George Parisis This is my paper

Pith reviewed 2026-05-18 03:51 UTC · model grok-4.3

classification 💻 cs.NI
keywords congestion controlreinforcement learningTCP CubicBBRfairnessbandwidth utilizationnetwork latencysimulation study
0
0 comments X

The pith

Learning-based congestion control acquires full bandwidth with low latency but fairness fails to generalize and performance drops when bandwidth or latency changes dynamically.

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

The paper extends earlier experiments to systematically compare publicly available learning-based congestion control, including reinforcement learning variants, against established human-designed algorithms such as TCP Cubic and BBR version 3. Large-scale simulations reveal that these learning approaches can fully utilize available bandwidth while keeping latency low under stable conditions. Embedding fairness directly into training rewards works in the tested cases, yet the fairness behavior does not carry over to network conditions not encountered during training. The methods also underperform when bandwidth and end-to-end latency shift over time, although they remain robust against non-congestive packet loss. These findings matter because congestion control must adapt to evolving network technologies and traffic patterns.

Core claim

RL learning-based approaches can acquire all available bandwidth while largely maintaining low latency. Embedding fairness directly into reward functions is effective; however, the fairness properties do not generalise into unseen conditions. Existing approaches under-perform when the available bandwidth and end-to-end latency dynamically change while remaining resistant to non-congestive loss.

What carries the argument

Large-scale reproducible experimentation that directly contrasts publicly available learning-based CC implementations against TCP Cubic and BBR version 3 across controlled variations in bandwidth, latency, and loss.

If this is right

  • Fairness achieved by direct reward engineering remains effective only inside the training distribution.
  • Learning-based methods saturate bandwidth while preserving low latency in stable network settings.
  • Performance degrades when bandwidth and latency vary dynamically over time.
  • Resistance to non-congestive loss persists across the tested scenarios.
  • A reproducible evaluation methodology and public codebase enable direct comparison of future learning-based CC proposals.

Where Pith is reading between the lines

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

  • Training on a broader set of traces that include abrupt bandwidth and latency shifts could strengthen generalization.
  • The observed resistance to non-congestive loss may give learning-based methods an edge in wireless or error-prone links.
  • Hybrid designs that fall back to traditional rules during detected transitions might combine the strengths of both approaches.
  • Standardized benchmarks should explicitly include sudden parameter changes to expose generalization gaps.

Load-bearing premise

The publicly available learning-based CC implementations and the chosen simulation scenarios are representative of both state-of-the-art methods and the range of conditions encountered in real networks.

What would settle it

A new learning-based CC implementation that maintains both fairness and high performance after sudden drops in available bandwidth combined with increases in end-to-end latency would directly challenge the reported limitations.

Figures

Figures reproduced from arXiv: 2510.25105 by George Parisis, Luca Giacomoni, Mihai Mazilu.

Figure 1
Figure 1. Figure 1: Intra-RTT Fairness. Goodput ratio for two competing flows in a dumbbell topology. Bottleneck capacity is 100Mbps, both flows experience the same base RTT (shown on x-axis), buffer capacity is set to 0.2× (a), 1× (b), and 4× (c) the BDP. we focus on the effect that RTT, available bandwidth, and buffer capacity have on fairness when two flows contend for bandwidth. The first flow runs for 2000× its base RTT … view at source ↗
Figure 2
Figure 2. Figure 2: Intra-RTT Fairness. Congestion window (sending rate for Vivace and BBRv3) for two competing flows in a dumbbell topology. Bottleneck capacity is 100Mbps, base RTT is 80ms, buffer capacity is set to 0.2×, 1× and 4× the BDP. Orca Sage Astraea Vivace Cubic BBRv3 50 100 150 200 RTT (ms) 0.0 0.5 1.0 Goodput Ratio (a) Buffer Size: 0.2× BDP 50 100 150 200 RTT (ms) 0.0 0.5 1.0 Goodput Ratio (b) Buffer Size: 1× BDP… view at source ↗
Figure 3
Figure 3. Figure 3: Inter-RTT Fairness. Goodput ratio for two competing flows in a dumbbell topology. Bottleneck capacity is 100Mbps and buffer capacity is set to 0.2× (a), 1× (b), and 4× (c) the BDP of the path with the smallest RTT. Flows experience different RTTs; RTT of first flow is set to 20ms and RTT of second flow is shown on x-axis. val, while Cubic continuously performing per-ACK window adjustments [25]. When the bu… view at source ↗
Figure 4
Figure 4. Figure 4: Fairness with Bandwidth Variation. Goodput ratio for two competing flows in a dumbbell topology. Base RTT is 40ms, bottleneck bandwidth varies as shown on the x-axis, buffer capacity is set to 0.2× (a), 1× (b), and 4× (c) the BDP. schemes used to train Sage lack heuristics that enable learning fairness in inter-RTT scenarios. In the lower range of base RTT values in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Fairness in a Parking Lot Topology. Bottleneck capacity is 100Mbps, all 4 flows experience the same base RTT (shown on x-axis), buffer capacity is set to 0.2× (a), 1× (b), and 4× (c) the BDP. bottleneck flow is dominated by the other three flows. This also holds (but is less prominent) when the buffer capacity is 1× the BDP (Figure 5a). When the buffer capacity is 4× the BDP (Figure 5c), Orca comes close t… view at source ↗
Figure 6
Figure 6. Figure 6: TCP Friendliness. Goodput ratio for two competing flows (one being Cubic) in a dumbbell topology. Bottleneck capacity is 100Mbps, flows experience the same base RTT (x-axis), buffer capacity is 0.2× (a), 1× (b), and 4× (c) the BDP. Orca Sage Astraea Vivace BBRv3 0 100 0 100 0 100 0 100 0 100 0 25 50 75 100 125 150 175 200 0 100 Time (s) Goodput (Mbps) (a) TCP flow first, Buffer Size: 0.2× BDP 0 100 0 100 0… view at source ↗
Figure 7
Figure 7. Figure 7: TCP Friendliness. Goodput evolution for two competing flows (one being Cubic) in a dumbbell topology. Bottleneck capacity is 100Mbps, both flows experience the same base RTT (100ms), buffer capacity is set to 0.2×, 1×, and 4× the BDP. sizes (Figures 6a and 6b), BBRv3 dominates Cubic. When the buffer capacity is set to 4× the BDP and the base RTT values are high (Figure 6c), Cubic fills the buffer, preventi… view at source ↗
Figure 8
Figure 8. Figure 8: TCP Friendliness. Goodput ratio between the tested scheme and the average goodput achieved by the Cubic flows. Bottleneck capacity is 100Mbps, RTT is set to 30ms, buffer capacity is set to 0.2× (a), 1× (b), and 4× (c) the BDP. The number of joining Cubic flows is shown on the y-axis. set to 4× the BDP and the Orca flow starts second, it cannot capture any bandwidth at all; as explained above, Cubic has alr… view at source ↗
Figure 9
Figure 9. Figure 9: Efficiency. Aggregate network throughput and average latency for four competing flows in a dumbbell topology. Bottleneck capacity is 100Mbps, flows experience the same base RTT, buffer capacity is set to 0.2×, 1× and 4× the BDP. setups, including 2 Cubic flows, it has learnt to be more aggressive the more pressure it gets from its competitor. When the buffer capacity is 0.2× the BDP, Astraea maintains its … view at source ↗
Figure 10
Figure 10. Figure 10: Responsiveness: Cumulative Distribution of Goodput 0 100 100 120 140 160 180 200 Time (s) 0 100 100 200 Base RTT (ms) 2 4 Loss Rate (%) Sending Rate (Mbps) [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Convergence. Goodput evolution for four competing flows in a dumbbell topology. Bottleneck capacity is 100Mbps, flows experience the same base RTT of 20ms, buffer capacity is set to 0.2×, 1× and 4× the BDP. 0 100 Orca 0 100 Sage 0 100 Astraea 0 100 Vivace 0 100 Cubic 0 25 50 75 100 125 150 175 Time (s) 0 100 BBRv3 Goodput (Mbps) (a) RTT: 200ms, Buffer Size: 0.2× BDP 0 100 Orca 0 100 Sage 0 100 Astraea 0 1… view at source ↗
Figure 13
Figure 13. Figure 13: Convergence. Goodput evolution for four competing flows in a dumbbell topology. Bottleneck capacity is 100Mbps, flows experience the same base RTT of 200ms, buffer capacity is set to 0.2×, 1× and 4× the BDP. this experiment are shown in dashed lines. Orca, Sage and Astraea are resistant to random loss. This, in combination with the fact that we do not change the underlying RTT in this experiment, means th… view at source ↗
read the original abstract

Learning-based congestion control (CC), including Reinforcement-Learning, promises efficient CC in a fast-changing networking landscape, where evolving communication technologies, applications and traffic workloads pose severe challenges to human-derived, static CC algorithms. Learning-based CC is in its early days and substantial research is required to understand existing limitations, identify research challenges and, eventually, yield deployable solutions for real-world networks. In this paper, we extend our prior work and present a reproducible and systematic study of learning-based CC with the aim to highlight strengths and uncover fundamental limitations of the state-of-the-art. We directly contrast said approaches with widely deployed, human-derived CC algorithms, namely TCP Cubic and BBR (version 3). We identify challenges in evaluating learning-based CC, establish a methodology for studying said approaches and perform large-scale experimentation with learning-based CC approaches that are publicly available. We show that embedding fairness directly into reward functions is effective; however, the fairness properties do not generalise into unseen conditions. We then show that RL learning-based approaches existing approaches can acquire all available bandwidth while largely maintaining low latency. Finally, we highlight that existing the latest learning-based CC approaches under-perform when the available bandwidth and end-to-end latency dynamically change while remaining resistant to non-congestive loss. As with our initial study, our experimentation codebase and datasets are publicly available with the aim to galvanise the research community towards transparency and reproducibility, which have been recognised as crucial for researching and evaluating machine-generated policies.

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

1 major / 0 minor

Summary. This paper extends prior work to present a systematic, reproducible experimental study comparing learning-based congestion control (CC) algorithms, including reinforcement learning (RL) approaches, with human-derived algorithms such as TCP Cubic and BBR version 3. Through large-scale experimentation using publicly available implementations, the authors show that RL-based methods can acquire all available bandwidth while maintaining low latency, that embedding fairness in reward functions is effective but the fairness properties do not generalize to unseen conditions, and that these learning-based approaches under-perform when bandwidth and end-to-end latency change dynamically, although they remain resistant to non-congestive loss. The study also discusses challenges in evaluating learning-based CC and provides public code and datasets.

Significance. If the results hold, the paper makes a significant contribution by empirically demonstrating the strengths and limitations of state-of-the-art learning-based congestion control in comparison to traditional methods. The identification of generalization failures and dynamic adaptation issues is valuable for guiding future research. The public availability of the codebase and datasets is a key strength that promotes transparency and reproducibility in the field.

major comments (1)
  1. The headline finding that existing learning-based CC approaches under-perform when available bandwidth and end-to-end latency dynamically change is central to the paper's assessment of limitations. This relies on the simulation scenarios being representative of real networks. The paper should provide more details on the specific traces used for dynamic changes, including their range, variation frequency, and how they compare to real-world network conditions, to address potential concerns that the results may be testbed-specific.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive review and recommendation for major revision. We address the major comment below and will revise the manuscript to incorporate additional details on the simulation traces.

read point-by-point responses

  1. Referee: The headline finding that existing learning-based CC approaches under-perform when available bandwidth and end-to-end latency dynamically change is central to the paper's assessment of limitations. This relies on the simulation scenarios being representative of real networks. The paper should provide more details on the specific traces used for dynamic changes, including their range, variation frequency, and how they compare to real-world network conditions, to address potential concerns that the results may be testbed-specific.

    Authors: We thank the referee for highlighting this point. To strengthen the paper, we will add a dedicated subsection in the evaluation methodology describing the dynamic traces in detail. This will specify the bandwidth ranges (typically 1-100 Mbps), latency ranges (10-200 ms), variation frequencies and patterns, and direct comparisons to real-world conditions using public datasets such as those from CAIDA and M-Lab. Our publicly released codebase and datasets already contain the exact traces, enabling full inspection and reproducibility. We believe these additions will confirm the scenarios are representative rather than testbed-specific while preserving the core findings on generalization failures. revision: yes

Circularity Check

0 steps flagged

No significant circularity in this experimental study

full rationale

The paper is an empirical experimental comparison of learning-based congestion control algorithms against human-derived ones (TCP Cubic, BBR v3), with all central claims about bandwidth acquisition, latency maintenance, fairness generalization failure, and under-performance under dynamic bandwidth/latency changes resting directly on simulation measurements rather than any mathematical derivation, fitted parameters, or first-principles results. Although the abstract notes that the work extends the authors' prior research, this self-reference does not bear the load of the findings, which are supported by new large-scale experiments, publicly available implementations, and datasets. No equations, ansatzes, uniqueness theorems, or reductions to inputs by construction appear; the analysis is self-contained against external benchmarks and reproducible artifacts.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claims depend on the representativeness of the tested public implementations and the simulation scenarios used for evaluation.

axioms (2)
  • domain assumption The publicly available learning-based CC implementations accurately represent current state-of-the-art approaches.
    The study directly contrasts these implementations with Cubic and BBR.
  • domain assumption The simulated network conditions and dynamic changes capture relevant real-world behaviors.
    Large-scale experimentation relies on these scenarios to demonstrate generalization failures.

pith-pipeline@v0.9.0 · 5800 in / 1236 out tokens · 37264 ms · 2026-05-18T03:51:59.666499+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

58 extracted references · 58 canonical work pages · 1 internal anchor

  1. [1]

    Playing Atari with Deep Reinforcement Learning

    V . Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wier- stra, and M. Riedmiller, “Playing atari with deep reinforcement learn- ing,”arXiv preprint arXiv:1312.5602, 2013

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  2. [2]

    Grandmaster level in StarCraft II using multi-agent reinforcement learning,

    O. Vinyals, I. Babuschkin, W. M. Czarnecki, M. Mathieu, A. Dudzik et al., “Grandmaster level in StarCraft II using multi-agent reinforcement learning,”Nature, vol. 575, no. 7782, 2019

    work page 2019

  3. [3]

    Mastering the game of go without human knowledge,

    D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guezet al., “Mastering the game of go without human knowledge,” Nature, vol. 550, no. 7676, 2017

    work page 2017

  4. [4]

    A general reinforcement learning algorithm that masters chess, shogi, and go through self-play,

    D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez et al., “A general reinforcement learning algorithm that masters chess, shogi, and go through self-play,”Science, 2018

    work page 2018

  5. [5]

    Training language models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwrightet al., “Training language models to follow instructions with human feedback,” Advances in Neural Information Processing Systems, vol. 35, 2022

    work page 2022

  6. [6]

    Magnetic control of tokamak plasmas through deep reinforce- ment learning,

    B. D. Tracey, A. Michi, Y . Chervonyi, I. Davies, C. Paduraru, N. Lazic et al., “Magnetic control of tokamak plasmas through deep reinforce- ment learning,”Nature, vol. 602, no. 7897, 2022

    work page 2022

  7. [7]

    Reinforcement learning based routing in networks: Re- view and classification of approaches,

    Z. Mammeri, “Reinforcement learning based routing in networks: Re- view and classification of approaches,”IEEE Access, vol. 7, 2019

    work page 2019

  8. [8]

    QARC: Video quality aware rate control for real-time video streaming based on deep rein- forcement learning,

    T. Huang, R.-X. Zhang, C. Zhou, and L. Sun, “QARC: Video quality aware rate control for real-time video streaming based on deep rein- forcement learning,” inProc. of ACM MM, 2018

    work page 2018

  9. [9]

    Neural adaptive video stream- ing with pensieve,

    H. Mao, R. Netravali, and M. Alizadeh, “Neural adaptive video stream- ing with pensieve,” inProc. of ACM SIGCOMM, 2017

    work page 2017

  10. [10]

    Deep reinforcement learning for dynamic multichannel access in wireless networks,

    S. Wang, H. Liu, P. H. Gomeset al., “Deep reinforcement learning for dynamic multichannel access in wireless networks,”IEEE Transactions on Cognitive Communications and Networking, vol. 4, no. 2, 2018

    work page 2018

  11. [11]

    Deep multi-user reinforcement learning for distributed dynamic spectrum access,

    O. Naparstek and K. Cohen, “Deep multi-user reinforcement learning for distributed dynamic spectrum access,”IEEE Transactions on Wireless Communications, vol. 18, no. 1, 2018

    work page 2018

  12. [12]

    Deep reinforcement learning for cyber security,

    T. T. Nguyenet al., “Deep reinforcement learning for cyber security,” IEEE Transactions on Neural Networks and Learning Systems, 2019

    work page 2019

  13. [13]

    Reinforcement learning for IoT security: A comprehensive survey,

    A. Uprety and D. B. Rawat, “Reinforcement learning for IoT security: A comprehensive survey,”IEEE Internet of Things Journal, 2020

    work page 2020

  14. [14]

    Deep reinforcement learning for mobile edge caching: Review, new features, and open issues,

    H. Zhu, Y . Cao, W. Wang, T. Jiang, and S. Jin, “Deep reinforcement learning for mobile edge caching: Review, new features, and open issues,”IEEE Network, 2018

    work page 2018

  15. [15]

    Integrated networking, caching, and com- puting for connected vehicles: A deep reinforcement learning approach,

    Y . He, N. Zhao, and H. Yin, “Integrated networking, caching, and com- puting for connected vehicles: A deep reinforcement learning approach,” IEEE Transactions on Vehicular Technology, vol. 67, no. 1, 2017

    work page 2017

  16. [16]

    CUBIC: A New TCP-Friendly High-Speed TCP Variant,

    S. Ha, I. Rhee, and L. Xu, “CUBIC: A New TCP-Friendly High-Speed TCP Variant,”SIGOPS Oper. Syst. Rev., 2008

    work page 2008

  17. [17]

    BBR: Congestion-based congestion control: Measuring bottleneck bandwidth and round-trip propagation time,

    N. Cardwell, Y . Cheng, C. S. Gunn, S. H. Yeganeh, and V . Jacob- son, “BBR: Congestion-based congestion control: Measuring bottleneck bandwidth and round-trip propagation time,”Queue, vol. 14, no. 5, 2016

    work page 2016

  18. [18]

    Tcp ex machina: computer-generated congestion control,

    K. Winstein and H. Balakrishnan, “Tcp ex machina: computer-generated congestion control,” ser. SIGCOMM ’13. Association for Computing Machinery, 2013. [Online]. Available: https://doi.org/10.1145/2486001. 2486020

    work page doi:10.1145/2486001 2013

  19. [19]

    Pcc: re-architecting congestion control for consistent high performance,

    M. Dong, Q. Li, D. Zarchy, P. B. Godfrey, and M. Schapira, “Pcc: re-architecting congestion control for consistent high performance,” in Proceedings of the 12th USENIX Conference on Networked Systems Design and Implementation, ser. NSDI’15. USENIX Association, 2015

    work page 2015

  20. [20]

    PCC Vivace: Online-learning congestion control,

    M. Dong, T. Meng, D. Zarchy, E. Arslan, Y . Gilad, B. Godfrey, and M. Schapira, “PCC Vivace: Online-learning congestion control,” inProc. of USENIX NSDI, 2018

    work page 2018

  21. [21]

    Experimental evaluation of TCP protocols for high-speed networks,

    Y .-T. Li, D. Leith, and R. N. Shorten, “Experimental evaluation of TCP protocols for high-speed networks,”IEEE/ACM Transactions on Networking, vol. 15, no. 5, 2007

    work page 2007

  22. [22]

    TCP-Drinc: Smart congestion control based on deep reinforcement learning,

    K. Xiao, S. Mao, and J. K. Tugnait, “TCP-Drinc: Smart congestion control based on deep reinforcement learning,”IEEE Access, 2019

    work page 2019

  23. [23]

    SmartCC: A reinforcement learning approach for multipath TCP congestion control in heterogeneous networks,

    W. Li, H. Zhang, S. Gao, C. Xue, X. Wang, and S. Lu, “SmartCC: A reinforcement learning approach for multipath TCP congestion control in heterogeneous networks,”IEEE Journal on Selected Areas in Com- munications, vol. 37, no. 11, 2019

    work page 2019

  24. [24]

    A deep reinforcement learning perspective on Internet congestion control,

    N. Jay, N. Rotman, B. Godfreyet al., “A deep reinforcement learning perspective on Internet congestion control,” inProc. of ICML, 2019

    work page 2019

  25. [25]

    Classic meets modern: A pragmatic learning-based congestion control for the Internet,

    S. Abbasloo, C.-Y . Yen, and H. J. Chao, “Classic meets modern: A pragmatic learning-based congestion control for the Internet,” inProc. of ACM SIGCOMM, 2020

    work page 2020

  26. [26]

    Astraea: Towards fair and efficient learning-based congestion control,

    X. Liao, H. Tian, C. Zeng, X. Wan, and K. Chen, “Astraea: Towards fair and efficient learning-based congestion control,” inProc. of EuroSys 2024, 2024

    work page 2024

  27. [27]

    Spine: an efficient DRL-based congestion control with ultra-low overhead,

    H. Tian, X. Liao, C. Zenget al., “Spine: an efficient DRL-based congestion control with ultra-low overhead,” inProc. of ACM CoNEXT, 2022

    work page 2022

  28. [28]

    Computers can learn from the heuristic designs and master internet congestion control,

    C.-Y . Yen, S. Abbasloo, and H. J. Chao, “Computers can learn from the heuristic designs and master internet congestion control,” inProc. of ACM SIGCOMM, 2023

    work page 2023

  29. [29]

    Reproducible network experiments using container-based emulation,

    N. Handigol, B. Helleret al., “Reproducible network experiments using container-based emulation,” inProc. of CoNEXT, 2012

    work page 2012

  30. [30]

    QTCP: Adaptive congestion control with reinforcement learning,

    W. Li, F. Zhou, K. R. Chowdhury, and W. Meleis, “QTCP: Adaptive congestion control with reinforcement learning,”IEEE Transactions on Network Science and Engineering, vol. 6, no. 3, 2018

    work page 2018

  31. [31]

    Learning in situ: a randomized experiment in video streaming,

    F. Y . Yan, H. Ayers, C. Zhu, S. Fouladi, J. Hong, K. Zhang, P. Levis, and K. Winstein, “Learning in situ: a randomized experiment in video streaming,” inProc. of USENIX NSDI, 2020

    work page 2020

  32. [32]

    RayNet: A simulation platform for developing reinforcement learning-driven network protocols,

    L. Giacomoni, B. Benny, and G. Parisis, “RayNet: A simulation platform for developing reinforcement learning-driven network protocols,”CoRR, vol. abs/2302.04519, 2023

    work page arXiv 2023

  33. [33]

    ns-3 meets OpenAI Gym: The playground for machine learning in networking research,

    P. Gawłowicz and A. Zubow, “ns-3 meets OpenAI Gym: The playground for machine learning in networking research,” inACM MSWiM, 2019

    work page 2019

  34. [34]

    Reinforcement learning-based congestion control: A systematic evaluation of fairness, efficiency and responsive- ness,

    L. Giacomoni and G. Parisis, “Reinforcement learning-based congestion control: A systematic evaluation of fairness, efficiency and responsive- ness,” inProc. of IEEE INFOCOM, 2024

    work page 2024

  35. [35]

    Hybrid modeling of TCP congestion control,

    J. P. Hespanha, S. Bohacek, K. Obraczka, and J. Lee, “Hybrid modeling of TCP congestion control,” inProc. of HSCC, 2001

    work page 2001

  36. [36]

    Modelling TCP congestion control dynamics in drop-tail environments,

    R. Shorten, C. King, F. Wirth, and D. Leith, “Modelling TCP congestion control dynamics in drop-tail environments,”Automatica, 2007

    work page 2007

  37. [37]

    Towards a deeper understanding of TCP BBR congestion control,

    D. Scholz, B. Jaeger, L. Schwaighofer, D. Raumer, F. Geyer, and G. Carle, “Towards a deeper understanding of TCP BBR congestion control,” inProc. of IFIP Networking, 2018

    work page 2018

  38. [38]

    UDT: UDP-based data transfer for high- speed wide area networks,

    Y . Gu and R. L. Grossman, “UDT: UDP-based data transfer for high- speed wide area networks,”Computer Networks, vol. 51, no. 7, 2007

    work page 2007

  39. [39]

    Promises and potential of bbrv3,

    D. Zeynali, E. N. Weyulu, S. Fathalli, B. Chandrasekaran, and A. Feldmann, “Promises and potential of bbrv3,” inPassive and Active Measurement: 25th International Conference, PAM 2024, Virtual Event, March 11–13, 2024, Proceedings, Part II, 2024. [Online]. Available: https://doi.org/10.1007/978-3-031-56252-5 12

    work page doi:10.1007/978-3-031-56252-5 2024

  40. [40]

    Evaluating tcp bbrv3 performance in wired broadband networks,

    J. Gomez, E. F. Kfoury, J. Crichigno, and G. Srivastava, “Evaluating tcp bbrv3 performance in wired broadband networks,”Computer Communications, vol. 222, pp. 198–208, 2024. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0140366424001658

    work page 2024

  41. [41]

    Eagle: Refining congestion control by learning from the experts,

    S. Emara, B. Li, and Y . Chen, “Eagle: Refining congestion control by learning from the experts,” inProc of IEEE INFOCOM, 2020

    work page 2020

  42. [42]

    Pareto: Fair congestion control with online reinforcement learning,

    S. Emara, F. Wang, B. Li, and T. Zeyl, “Pareto: Fair congestion control with online reinforcement learning,”IEEE Transactions on Network Science and Engineering, vol. 9, no. 5, 2022

    work page 2022

  43. [43]

    Available: https://iperf.fr/

    “iperf3.” [Online]. Available: https://iperf.fr/

    work page

  44. [44]

    TCP testing (tcpprobe),

    S. Hemminger, “TCP testing (tcpprobe),”The Linux Foundation, http://devresources.linuxfoundation.org/shemminger/tcp, 2011

    work page 2011

  45. [45]

    SYSSTAT home page,

    S. Godard, “SYSSTAT home page,”Information and code available at http://sebastien.godard.pagesperso-orange.fr/index.html, 2015

    work page 2015

  46. [46]

    A network in a laptop: Rapid prototyping for software-defined networks,

    B. Lantz, B. Heller, and N. McKeown, “A network in a laptop: Rapid prototyping for software-defined networks,” inProc. of ACM HotNets, 2010

    work page 2010

  47. [47]

    BBR3 enabled kernel

    “BBR3 enabled kernel.” [Online]. Available: https://github.com/google/ bbr/blob/v3/net/ipv4/tcp bbr.c#L866

    work page

  48. [48]

    this dataset is used to generate the figures.” [Online]

    “The dataset contains measurements collected from many experimentals, of various cc schemes including orca, sage, astraea, pcc vivace (kernel and userspace), cubic, and bbrv3 and bbrv1 in emulated dumbbell and parking-lot topologies. this dataset is used to generate the figures.” [Online]. Available: https://figshare.com/s/97c09c17972fb0ca56b4

    work page

  49. [49]

    Code repository for the experiment and plotting scripts

    “Code repository for the experiment and plotting scripts.” [Online]. Available: https://github.com/Aruuni/mininettestbed

    work page

  50. [50]

    Bbrv3: Algorithm bug fixes and public internet deployment,

    N. C. Y . C. K. Y . D. M. S. H. Y . P. J. Y . Seung, “Bbrv3: Algorithm bug fixes and public internet deployment,” 2023. [Online]. Available: https://datatracker.ietf.org/meeting/117/materials/ slides-117-ccwg-bbrv3-algorithm-bug-fixes-and-public-internet-deployment-00

    work page 2023

  51. [51]

    Differentiated end-to-end internet services using a weighted proportional fair sharing tcp,

    J. Crowcroft and P. Oechslin, “Differentiated end-to-end internet services using a weighted proportional fair sharing tcp,” 1998. 15

    work page 1998

  52. [52]

    Mutant: learning congestion control from existing protocols via online reinforcement learning,

    L. Pappone, A. Sacco, and F. Esposito, “Mutant: learning congestion control from existing protocols via online reinforcement learning,” ser. NSDI ’25. USENIX Association, 2025

    work page 2025

  53. [53]

    Orc: Online reinforcement learning for congestion control with fast convergence,

    Y . Li, J. Huang, C. Wu, X. Zhu, and J. Wang, “Orc: Online reinforcement learning for congestion control with fast convergence,” inProceedings of the 9th Asia-Pacific Workshop on Networking, ser. APNET ’25,

    work page

  54. [54]

    Available: https://doi.org/10.1145/3735358.3735381

    [Online]. Available: https://doi.org/10.1145/3735358.3735381

    work page doi:10.1145/3735358.3735381

  55. [55]

    Achieving fairness generalizability for learning-based congestion control with jury,

    H. Tian, X. Liao, D. Sun, C. Zeng, Y . Jin, J. Zhang, X. Wan, Z. Wang, Y . Wang, and K. Chen, “Achieving fairness generalizability for learning-based congestion control with jury,” ser. EuroSys ’25. New York, NY , USA: Association for Computing Machinery, 2025. [Online]. Available: https://doi.org/10.1145/3689031.3696065

    work page doi:10.1145/3689031.3696065 2025

  56. [56]

    Experience-driven congestion control: When multi-path TCP meets deep reinforcement learning,

    Z. Xu, J. Tang, C. Yin, Y . Wang, and G. Xue, “Experience-driven congestion control: When multi-path TCP meets deep reinforcement learning,”IEEE Journal on Selected Areas in Communications, vol. 37, no. 6, 2019

    work page 2019

  57. [57]

    Multi-objective congestion control,

    Y . Ma, H. Tian, X. Liao, J. Zhang, W. Wang, K. Chen, and X. Jin, “Multi-objective congestion control,” inProc. of EuroSys, 2022

    work page 2022

  58. [58]

    Wanna make your tcp scheme great for cellular networks? let machines do it for you!

    S. Abbasloo, C.-Y . Yen, and H. J. Chao, “Wanna make your tcp scheme great for cellular networks? let machines do it for you!”IEEE Journal on Selected Areas in Communications, 2021

    work page 2021