Pith. sign in

REVIEW 5 major objections 6 minor 36 references

SABR: A Stable Adaptive Bitrate Framework Using Behavior Cloning Pretraining and Reinforcement Learning Fine-Tuning

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read SABR proposes a two-stage behavior-cloning plus reinforcement-learning training scheme that achieves the best average QoE rank on wide-distribution and out-of-distribution network traces.

desk verdict Clean, useful ABR paper with a real but modest contribution; the best-average-rank claim needs variance reporting before it fully lands. read the letter →

arxiv 2509.10486 v1 pith:UWFKTFYM submitted 2025-08-30 cs.NI cs.AIcs.LGcs.MM

Pengcheng Luo , Yunyang Zhao , Bowen Zhang , Genke Yang , Boon-Hee Soong , Chau Yuen This is my paper · ORCID
classification cs.NIcs.AIcs.LGcs.MM
keywords adaptivebitrateABRbehaviorcloningdirectpreferenceoptimizationPPOreinforcementlearningvideostreamingout-of-distributiongeneralization
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 addresses a known failure of learning-based adaptive bitrate (ABR) controllers: policies trained on one network trace distribution degrade when the distribution widens, and they do not transfer to unseen conditions. The authors propose SABR, a two-stage framework that first pretrains the bitrate policy by behavior cloning from an MPC-style beam-search expert using a step-wise Direct Preference Optimization (DPO) loss, then fine-tunes it with Proximal Policy Optimization (PPO) in a simulator. To test generalization, they assemble two benchmarks, ABRBench-3G and ABRBench-4G+, whose training sets deliberately span a wide range of public traces and whose test sets hold out whole trace families as out-of-distribution. Across these benchmarks, SABR achieves the lowest average rank compared with Pensieve, Comyco, NetLLM, RobustMPC, and other baselines, claiming both stronger generalization and more stable wide-distribution training. If the claim holds, the pretrain-then-finetune recipe that succeeded in language models also carries over to low-level network control.

What carries the argument

The step-wise DPO loss (Eq. 2) is the key mechanism for the pretraining stage: it takes state-action pairs, treats the beam-search expert action as the preferred action and a random alternative as the dispreferred one, and directly maximizes the log-likelihood margin between them, biased by a reference model. This converts imitation learning into a preference-style objective that gives a stable base policy. The fine-tuning stage uses PPO with GAE advantages, clipping, and entropy regularization (with entropy weight set to 0), initialized from that base policy. The benchmarks are the second key object: trace sets are partitioned so that entire families are held out as OOD test sets, preservin

What would settle it

Re-run the four learning-based methods with, say, 20 seeds per condition, record per-seed average rank, and compute a confidence interval on the rank differences. If SABR's rank advantage overlaps zero at the 95% level on either benchmark or on the OOD sets, the central claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the two-stage pretraining/fine-tuning paradigm from LLM alignment can be transplanted to ABR control. The BC stage uses a step-wise DPO loss to match a beam-search expert's actions at each state, giving the policy a good initialization without needing a reward model; the PPO stage then explores beyond the expert to improve adaptation. The authors construct ABRBench-3G and ABRBench-4G+ from public traces (FCC, Lumos, Puffer, Oboe, HSR, Ghent, Lab, Solis) with disjoint OOD splits, and report that SABR obtains the best average rank on both benchmarks and on the held-out OOD sets, beating Pensieve, Comyco, and NetLLM. They interpret this as evidence that the pre

Load-bearing premise

The claimed advantage rests on comparing average QoE over ten training runs without reporting variance or confidence intervals; if the spread across runs is large relative to the small rank differences in Tables IV-VI, the best-average-rank conclusion may not be statistically meaningful.

Editorial extensions

If this is right

  • ABR policies can be trained on a deliberately broad and mixed corpus of network traces without collapsing, so service providers could pretrain once on accumulated logs and then fine-tune per deployment.
  • The OOD split protocol gives the community a concrete way to measure generalization: report per-trace-set QoE and average rank, not just pooled averages masked by high-bandwidth sets.
  • If SABR's ranking advantage reproduces, the LLM-style two-stage recipe becomes a viable starting point for other networking control tasks with discrete actions.
  • The step-wise DPO variant provides an alternative to plain supervised behavior cloning for discrete-action control, one that needs neither a reward model nor full-trajectory rollouts.

Reading between the lines

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

  • The paper leaves variance unreported; a natural extension is to report confidence intervals across the ten training runs so readers can judge whether the average-rank lead is real.
  • One could test the same pretrain-then-finetune scheme on congestion control or scheduling tasks, where expert trajectories are also available via model-predictive control.
  • Because the OOD sets include high-speed rail (HSR), Ghent, and Lab traces, the benchmarks could be extended with even more distinct regimes, such as satellite or millimeter-wave links, to stress the generalization claim harder.
  • The step-wise DPO for BC could be compared head-to-head with plain supervised behavior cloning to isolate what the preference margin adds over simple imitation.
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 / 6 minor

Summary. The paper proposes SABR, a two-stage ABR training framework: a DPO-based behavior-cloning pretraining stage that learns from expert/random preference pairs, followed by PPO fine-tuning. It also introduces two benchmarks, ABRBench-3G and ABRBench-4G+, built from public network traces, with separate test and OOD sets. The evaluation compares SABR against Pensieve, Comyco, NetLLM, and several heuristics, reporting QoE averaged over ten training runs and an average-rank metric (Eq. 7). SABR attains the best average rank on both benchmarks (1.8 on ABRBench-3G, 1.7 on ABRBench-4G+) and on the OOD sets (2.0), which the authors interpret as evidence of better generalization and more stable learning.

Significance. If the empirical claim is statistically sound, the paper makes a useful contribution: a pretrain-then-finetune recipe that exploits wide-distribution traces, plus two publicly released benchmarks that separate in-distribution and OOD evaluation. The release of code and datasets is a concrete strength. However, the central claim rests on small average-rank differences over few trace sets, with no variance reporting or significance testing, and no ablation isolating the contribution of the pretraining stage. The significance of the result is therefore conditional on the statistical robustness of the reported comparisons.

major comments (5)
  1. [§VI-A, Tables IV-VI] The core claim—best average rank and better generalization—is not supported without measures of variability. The paper reports only QoE averaged over ten training runs, with no standard deviation, standard error, or confidence intervals. Several reported margins are very small: e.g., in Table V on Lumos 5G, SABR (1832.14) is below Comyco (1835.42); in Table IV on FCC-16, SABR (36.68) is only 0.12 above RobustMPC (36.56). These differences are likely within run-to-run variance of RL-based ABR training. Since average rank (Eq. 7) is computed over only 5, 3, and 3 trace sets, a single rank flip can change the conclusion. The authors should report per-seed QoE distributions, paired tests across seeds, or per-seed rank distributions, and state whether the reported averages are medians with error bars. Without this, the 'stability' and 'generalization' claims are not statistically grounded.
  2. [§VI-A, Tables IV-VI] The comparison with learning-based baselines is under-specified. Section VI-A states that Pensieve, Comyco, and NetLLM were each trained ten times, but it does not describe the training configuration, hyperparameters, or whether the authors used the official implementations, modified versions, or retuned hyperparameters. For a fair comparison, the paper must state the exact codebases, the network architecture (e.g., whether the 48-dimensional flattened state was used for all methods), training budgets, and any modifications. Without this, the reported rank ordering could reflect baseline under-tuning rather than a genuine advantage of SABR.
  3. [§III, Tables IV-VI] The paper's central conceptual claim is that the two-stage recipe (BC pretraining + RL fine-tuning) improves stability and generalization, but no ablation is provided. There is no comparison against a PPO-only baseline trained from random initialization on the same wide-distribution traces, nor a DPO-pretraining-only baseline without PPO fine-tuning. Such ablations are necessary to attribute the observed performance to the pretraining stage. Without them, the results only show that this particular SABR configuration works, not that the proposed framework's two stages are load-bearing.
  4. [§IV, Table I] The benchmark construction has an internal inconsistency that should be fixed. The text says 'in FCC-18, 75% of traces are allocated to the training set, while the remaining 30% are used for testing'—the percentages should sum to 100 (presumably 75%/25%). In addition, the row 'Training Same with test 1828' in Table I is unclear: it seems to indicate the training set is composed of the same trace sets as the test sets, but the split ratio and the relationship between training and test traces are not precisely defined. The authors should provide an exact description of how traces are split and confirm that no OOD trace appears in training.
  5. [§III-A, Eq. (2)] The DPO-step loss uses randomly sampled actions as the 'loser' action al, but DPO's theoretical grounding assumes preference pairs with meaningfully different quality. Random actions may provide a degenerate or noisy preference signal, especially because π_ref is the initial model and random actions can have arbitrarily low probability. The paper does not report any diagnostic of the pretraining stage (e.g., expert-action likelihood, imitation accuracy, or return under the base model before PPO). The authors should either justify the random-negative choice with an analysis or provide evidence that the pretraining stage indeed produces a good base policy.
minor comments (6)
  1. [§IV] Typo/inconsistency: '75% ... remaining 30%' should read 75%/25% or the intended split should be restated.
  2. [Table III] The entropy coefficient c2 is set to 0.0, but Eq. (5) includes an entropy regularization term. Please clarify whether entropy regularization is effectively disabled and why.
  3. [Throughout] The term 'behavior cloning' is used for DPO-based pretraining, but DPO is not conventional supervised behavior cloning. Please clarify the relationship and cite prior uses of DPO in sequential decision-making.
  4. [Algorithm 1] The procedure says 'Randomly select an alternative action al_t ≠ aw_t' but does not specify the distribution (e.g., uniform over the other five bitrates). Please specify for reproducibility.
  5. [Eq. (6)] The placement of the summation index in the smoothness term is typographically awkward; please format the term clearly.
  6. [Table I/II] The row label 'Training Same with test' is confusing; consider renaming to 'Training (same sources as test)' and providing a count of unique training traces.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the derivation is self-contained aside from one minor non-load-bearing self-citation.

full rationale

Reviewing the derivation chain, SABR's two-stage training is defined from standard RL objectives (DPO in Eq. 2, PPO in Eqs. 3-5) and evaluated with the standard QoE metric (Eq. 6). The training/test/OOD split is explicit (Section IV), and the reported average ranks (Eq. 7) are computed from measured QoE on held-out and OOD trace sets. No parameter is fitted to the evaluation metric, and no benchmark trace is used to define the algorithm's objective. The expert data for BC pretraining comes from Comyco's beam search and DAGGER-style collection, but that is an input to training, not the target being predicted; SABR is then compared against Comyco on the same held-out/OOD trace sets. There is no equation that reduces to itself by construction, and the 'best average rank' claim is an empirical comparison, not a definitional consequence. The only self-citation is reference [17] (authors' prior PPO scheduling work), used to support the generic statement that PPO is stable in discrete tasks; it is not load-bearing and no uniqueness claim or ansatz is imported from it. The seed-variance concern in the reader's take is a legitimate statistical-correctness issue, but it is not a circularity issue. Verdict: no significant circularity.

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

The central result depends on the simulator's fidelity, the QoE metric choice, the usefulness of random-action negatives in the DPO step, and the breadth of the training traces. These are standard domain assumptions or ad hoc choices that are not independently verified. No new physical entities are introduced.

free parameters (5)
  • DPO scaling beta = 0.1
    Chosen by hand; balances pretraining update strength (Table III).
  • PPO critic loss weight c1 = 0.5
    Chosen by hand; controls value loss in the PPO objective (Eq. 5).
  • PPO entropy weight c2 = 0.0
    Set to zero; no exploration bonus in fine-tuning (Table III).
  • QoE rebuffering penalty mu = 4.3 (3G), 40 (4G+)
    Taken from prior ABR work, not fitted to the new benchmarks; affects the reward all methods optimize (Section VI-A).
  • QoE smoothness penalty delta = 1
    Fixed as in Pensieve and Comyco (Section VI-A).
assumptions (4)
  • domain assumption The ABR simulator accurately represents real-world video streaming dynamics
    All training and evaluation happen in a trace-driven simulator (Section VI-A); if the simulator is not faithful, the rankings may not transfer to deployment.
  • domain assumption The QoE metric (Eq. 6) with q(R)=R is the correct objective for streaming quality
    Linear quality mapping is used by prior work and not justified for the new benchmarks (Section VI-A).
  • ad hoc to paper Random actions provide a useful negative signal for DPO step-wise learning
    Eq. 2 compares expert action against a randomly sampled alternative; the quality of this contrast depends on the action space and may not carry useful preference information (Section III-A).
  • domain assumption The 'Same with test' training sets are broad enough to represent wide-distribution conditions
    The claim of wide-distribution training relies on the trace counts and ranges in Tables I-II, but no distributional analysis is provided (Section IV).

how reviews work

0 comments
Cite this review

Pith. "Pith review of SABR: A Stable Adaptive Bitrate Framework Using Behavior Cloning Pretraining and Reinforcement Learning Fine-Tuning." pith.science (2026). https://pith.science/paper/UWFKTFYM

@misc{pith2026250910486,
  author       = {Pith},
  title        = {Pith review of: SABR: A Stable Adaptive Bitrate Framework Using Behavior Cloning Pretraining and Reinforcement Learning Fine-Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UWFKTFYM}},
  note         = {Machine review of arXiv:2509.10486}
}
read the original abstract

With the advent of 5G, the internet has entered a new video-centric era. From short-video platforms like TikTok to long-video platforms like Bilibili, online video services are reshaping user consumption habits. Adaptive Bitrate (ABR) control is widely recognized as a critical factor influencing Quality of Experience (QoE). Recent learning-based ABR methods have attracted increasing attention. However, most of them rely on limited network trace sets during training and overlook the wide-distribution characteristics of real-world network conditions, resulting in poor generalization in out-of-distribution (OOD) scenarios. To address this limitation, we propose SABR, a training framework that combines behavior cloning (BC) pretraining with reinforcement learning (RL) fine-tuning. We also introduce benchmarks, ABRBench-3G and ABRBench-4G+, which provide wide-coverage training traces and dedicated OOD test sets for assessing robustness to unseen network conditions. Experimental results demonstrate that SABR achieves the best average rank compared with Pensieve, Comyco, and NetLLM across the proposed benchmarks. These results indicate that SABR enables more stable learning across wide distributions and improves generalization to unseen network conditions.

Figures

Figures reproduced from arXiv: 2509.10486 by the authors.

Figure 1
Figure 1. An overview of ABR. As the user base continues to expand, video streaming service providers accumulate massive volumes of network data on a daily basis. This wealth of data presents unprecedented opportunities for analyzing user behavior and optimizing streaming strategies, while also providing a solid foundation for applying artificial intelligence (AI) techniques to ABR re￾search. AI approaches such as deep learni… view at source ↗
Figure 2
Figure 2. Proposed SABR framework: BC pretraining + RL fine-tuning. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 26 canonical work pages

  1. [1]

    Improving language understanding by generative pre-training,

    A. Radford, K. Narasimhan, T. Salimans, I. Sutskever et al., “Improving language understanding by generative pre-training,” 2018

    2018

  2. [2]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Pro- ceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , 2019, pp. 4171–4186

    2019

  3. [3]

    Training language models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray et al. , “Training language models to follow instructions with human feedback,” Advances in neural information processing systems , vol. 35, pp. 27 730–27 744, 2022

    2022

  4. [4]

    Direct preference optimization: Your language model is secretly a reward model,

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn, “Direct preference optimization: Your language model is secretly a reward model,” Advances in neural information processing systems, vol. 36, pp. 53 728–53 741, 2023

    2023

  5. [5]

    Prox- imal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,” arXiv preprint arXiv:1707.06347, 2017

    arXiv 2017

  6. [6]

    Neural adaptive video stream- ing with pensieve,

    H. Mao, R. Netravali, and M. Alizadeh, “Neural adaptive video stream- ing with pensieve,” in Proceedings of the conference of the ACM special interest group on data communication , 2017, pp. 197–210

    work page 2017

  7. [7]

    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,” in International conference on machine learning . PmLR, 2016, pp. 1928–1937

    work page 2016

  8. [8]

    Comyco: Quality-aware adaptive video streaming via imitation learning,

    T. Huang, C. Zhou, R.-X. Zhang, C. Wu, X. Yao, and L. Sun, “Comyco: Quality-aware adaptive video streaming via imitation learning,” in Proceedings of the 27th ACM international conference on multimedia , 2019, pp. 429–437

    work page 2019

Show all 36 references
  1. [9]

    Optimizing adaptive video streaming with human feedback,

    T. Huang, R.-X. Zhang, C. Wu, and L. Sun, “Optimizing adaptive video streaming with human feedback,” in Proceedings of the 31st ACM International Conference on Multimedia , 2023, pp. 1707–1718

    2023

  2. [10]

    Genet: Automatic curriculum generation for learning adaptation in networking,

    Z. Xia, Y . Zhou, F. Y . Yan, and J. Jiang, “Genet: Automatic curriculum generation for learning adaptation in networking,” in Proceedings of the ACM SIGCOMM 2022 Conference , 2022, pp. 397–413

    2022

  3. [11]

    Curriculum learning,

    Y . Bengio, J. Louradour, R. Collobert, and J. Weston, “Curriculum learning,” in Proceedings of the 26th annual international conference on machine learning , 2009, pp. 41–48

    2009

  4. [12]

    Netllm: Adapting large language models for networking,

    D. Wu, X. Wang, Y . Qiao, Z. Wang, J. Jiang, S. Cui, and F. Wang, “Netllm: Adapting large language models for networking,” in Proceed- ings of the ACM SIGCOMM 2024 Conference , 2024, pp. 661–678

    2024

  5. [13]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen et al., “Lora: Low-rank adaptation of large language models.” ICLR, vol. 1, no. 2, p. 3, 2022

    2022

  6. [14]

    A reduction of imitation learning and structured prediction to no-regret online learning,

    S. Ross, G. Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in Proceedings of the fourteenth international conference on artificial intelligence and statistics. JMLR Workshop and Conference Proceedings, 2011, ...

    2011

  7. [15]

    Comyco for linear-based QoE (comyco-lin),

    T. Huang, “Comyco for linear-based QoE (comyco-lin),” https://github. com/godka/comyco-lin, 2025

    2025

  8. [16]

    Mujoco: A physics engine for model- based control,

    E. Todorov, T. Erez, and Y . Tassa, “Mujoco: A physics engine for model- based control,” in 2012 IEEE/RSJ international conference on intelligent robots and systems . IEEE, 2012, pp. 5026–5033

    2012

  9. [17]

    Multi-resource constrained dynamic workshop scheduling based on proximal policy optimisation,

    P. C. Luo, H. Q. Xiong, B. W. Zhang, J. Y . Peng, and Z. F. Xiong, “Multi-resource constrained dynamic workshop scheduling based on proximal policy optimisation,” International journal of production re- search, vol. 60, no. 19, pp. 5937–5955, 2022

    2022

  10. [18]

    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,” arXiv preprint arXiv:1506.02438 , 2015

    2015 arXiv

  11. [19]

    A variegated look at 5g in the wild: performance, power, and qoe implications,

    A. Narayanan, X. Zhang, R. Zhu, A. Hassan, S. Jin, X. Zhu, X. Zhang, D. Rybkin, Z. Yang, Z. M. Mao et al., “A variegated look at 5g in the wild: performance, power, and qoe implications,” in Proceedings of the 2021 ACM SIGCOMM 2021 Conference , 2021, pp. 610–625

    2021

  12. [20]

    Lumos5g: Mapping and predicting commercial mmwave 5g throughput,

    A. Narayanan, E. Ramadan, R. Mehta, X. Hu, Q. Liu, R. A. Fezeu, U. K. Dayalan, S. Verma, P. Ji, T. Li et al. , “Lumos5g: Mapping and predicting commercial mmwave 5g throughput,” in Proceedings of the ACM internet measurement conference , 2020, pp. 176–193

    2020

  13. [21]

    Commute path bandwidth traces from 3g networks: Analysis and applications,

    H. Riiser, P. Vigmostad, C. Griwodz, and P. Halvorsen, “Commute path bandwidth traces from 3g networks: Analysis and applications,” in Proceedings of the 4th ACM Multimedia Systems Conference , 2013, pp. 114–118

    2013

  14. [22]

    Raw data - measuring broadband america 2016,

    Federal Communications Commission, “Raw data - measuring broadband america 2016,” https://www.fcc. gov/reports-research/reports/measuring-broadband-america/ raw-data-measuring-broadband-america-2016, 2016

    2016

  15. [23]

    Raw data releases - measuring broadband america 2018,

    ——, “Raw data releases - measuring broadband america 2018,” https: //www.fcc.gov/oet/mba/raw-data-releases, 2018

    2018

  16. [24]

    Pitree: Practical implementation of abr algorithms using decision trees,

    Z. Meng, J. Chen, Y . Guo, C. Sun, H. Hu, and M. Xu, “Pitree: Practical implementation of abr algorithms using decision trees,” in Proceedings of the 27th ACM International Conference on Multimedia , 2019, pp. 2431–2439

    2019

  17. [25]

    Oboe: Auto-tuning video abr algo- rithms to network conditions,

    Z. Akhtar, Y . S. Nam, R. Govindan, S. Rao, J. Chen, E. Katz-Bassett, B. Ribeiro, J. Zhan, and H. Zhang, “Oboe: Auto-tuning video abr algo- rithms to network conditions,” in Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication , 2018, pp. 44–58

    2018

  18. [26]

    Improving gen- eralization for neural adaptive video streaming via meta reinforcement learning,

    N. Kan, Y . Jiang, C. Li, W. Dai, J. Zou, and H. Xiong, “Improving gen- eralization for neural adaptive video streaming via meta reinforcement learning,” in Proceedings of the 30th ACM international conference on multimedia, 2022, pp. 3006–3016

    2022

  19. [27]

    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,” in 17th USENIX Symposium on Networked Systems Design and Implementation (NSDI 20) , 2020, pp. 495–511

    2020

  20. [28]

    Lumos: Towards better video streaming qoe through accurate throughput prediction,

    G. Lv, Q. Wu, W. Wang, Z. Li, and G. Xie, “Lumos: Towards better video streaming qoe through accurate throughput prediction,” in IEEE INFOCOM 2022-IEEE Conference on Computer Communications. IEEE, 2022, pp. 650–659

    2022

  21. [29]

    dash.js: Mpeg-dash reference client,

    DASH Industry Forum, “dash.js: Mpeg-dash reference client,” https:// github.com/Dash-Industry-Forum/dash.js, 2012

    2012

  22. [30]

    Big buck bunny,

    Blender Foundation, “Big buck bunny,” https://peach.blender.org/, 2008

    2008

  23. [31]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al. , “Pytorch: An imperative style, high-performance deep learning library,” Advances in neural information processing systems , vol. 32, 2019

    2019

  24. [32]

    Stable-baselines3: Reliable reinforcement learning implementa- tions,

    A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dor- mann, “Stable-baselines3: Reliable reinforcement learning implementa- tions,” Journal of machine learning research , vol. 22, no. 268, pp. 1–8, 2021

    2021

  25. [33]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

    2014 arXiv

  26. [34]

    A control-theoretic approach for dynamic adaptive video streaming over http,

    X. Yin, A. Jindal, V . Sekar, and B. Sinopoli, “A control-theoretic approach for dynamic adaptive video streaming over http,” in Proceed- ings of the 2015 ACM conference on special interest group on data communication, 2015, pp. 325–338

    2015

  27. [35]

    Bola: Near-optimal bitrate adaptation for online videos,

    K. Spiteri, R. Urgaonkar, and R. K. Sitaraman, “Bola: Near-optimal bitrate adaptation for online videos,” IEEE/ACM transactions on net- working, vol. 28, no. 4, pp. 1698–1711, 2020

    2020

  28. [36]

    Quetra: A queuing theory approach to dash rate adaptation,

    P. K. Yadav, A. Shafiei, and W. T. Ooi, “Quetra: A queuing theory approach to dash rate adaptation,” in Proceedings of the 25th ACM international conference on Multimedia , 2017, pp. 1130–1138

    2017

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.