Fair-Aurora: Comparing Fairness Strategies for Reinforcement Learning-Based Congestion Control in Multi-Flow Environments Thomas Mbrice, Yuyu Liu
arXiv:2605.19909v1 [cs.NI] 19 May 2026
CSE 534 Final Report Abstract
rate by interacting with a simulated network environment, directly optimizing for throughput while penalizing latency and loss. Aurora [1], proposed by Jay et al. at ICML 2019, demonstrated that a PPO-trained controller can track time-varying link capacity with 93% utilization, nearly 30 percentage points above CUBIC on the same link. TCP BBR (or just BBR) [5] occupies a middle ground: a model-based heuristic that estimates bottleneck bandwidth and propagation RTT to set rates near the bandwidth-delay product, achieving higher utilization than CUBIC without the training complexity of RL.
Reinforcement learning (RL) has emerged as a promising paradigm for Internet congestion control, achieving higher link utilization than classical heuristics. However, RL-based controllers trained in single-flow environments are not guaranteed to share bandwidth equitably when de- ,ployed in multi-flow networks. This paper investigates the fairness properties of Aurora [1], a state-of-theart deep RL congestion controller, and evaluates three post-hoc fairness strategies that preserve Aurora’s RL architecture: reward shaping (Strategy A), observation augmentation (Strategy B), and loss-sensitivity tuning (Strategy C). Using a custom shared-bottleneck simulator and Jain’s fairness index as the primary metric, we find that modest reward shaping achieves the best fairness while preserving aggregate throughput. All strategies maintain the total bandwidth budget with fairness being achieved through redistribution, not reduction. Beyond the 2-flow homo- geneous setting, an extended evaluation across mixed Aurora-CUBIC competition and dynamic flow entry/exit scenarios shows that Strategy C’s losssensitivity emerges as the most TCP-friendly mechanism, while Strategy B is the most stable through dynamic flow-set changes.
1
May 2026
1.1
The Fairness Problem in RL Congestion Control Despite its single-flow performance advantages, Aurora and similar RL controllers introduce a fundamental fairness concern when multiple flows share a bottleneck. Classical TCP protocols were designed with fairness as an explicit goal: AIMD dynamics provably converge to equal rates among competing flows over time [3]. RL controllers optimized for single-flow environments carry no such guarantee. The mechanism for RL-induced unfairness is welldocumented. When a flow trained in isolation encounters shared-queue congestion signals, its response depends entirely on what loss and latency patterns appeared during training. If the training distribution does not include the signature of multi-flow competition - gradually rising latency followed by correlated loss across competing flows - the controller interprets shared-queue signals through a single-flow lens. One flow may hold its rate while another retreats, leading to persistent bandwidth asymmetry (unfairness). Liao et al. [2] systematically characterized this problem in their Astraea system (EuroSys 2024). Under a staggered-start scenario, where one Aurora flow establishes a high sending rate before a second flow joins. In this case, Jain’s fairness index drops below 0.5, indicating that the incumbent flow holds more than twice the bandwidth of the newcomer.
Introduction
The Internet’s performance depends critically on congestion control (CC): the set of algorithms that regulate how quickly flows inject data into the network. For years, transport protocols such as TCP have governed this regulation using hand-crafted heuristics, additive increase, multiplicative decrease (AIMD), and cubic window growth (in the case of TCP CUBIC) that are provably stable and fair under idealized conditions [9]. However, these classical schemes leave substantial link capacity unused. On time-varying wireless links and high-bandwidthdelay-product paths, CUBIC can achieve link utilization as low as 60–70%, leaving significant performance on the table. Deep reinforcement learning (RL) offers an alternative: train a neural network to control sending 1
This is a stronger unfairness signal than most classical CC protocols produce; even aggressive TCP variants typically converge to equal shares within tens of RTTs. Yen et al. [4] further showed that RL-based CC can exhibit pronounced RTT unfairness on WiFi links, where flows with shorter round-trip times accumulate more reward signal per unit time and crowd out higher-latency flows. This RTT-unfairness effect is orthogonal to the rate-competition effect studied here, but both stem from the same root cause: an objective function that rewards throughput without accounting for competing flows.
Table 1: Aurora RL Formulation Parameters Component Value Observation Action Reward Policy Algorithm Episode Train links
1.2 Fairness as a Systems Property Jain’s fairness index [6] provides a principled scalar measure for evaluating bandwidth allocation across N flows: 2 PN i h i=1 xi 1 (1) J = PN 2 ∈ N , 1 N · i=1 xi
tuning, while preserving its aggregate throughput advantage over classical protocols? Concretely, we ask: Does a self-restraint reward penalty (Strategy A) cause the ego flow to yield bandwidth to competitors, and does J improve monotonically with penalty strength? Does richer observational context (Strategy B) produce cooperative behavior without an explicit penalty? Does increasing loss sensitivity (Strategy C) produce AIMD-like backoff that incidentally improves fairness? Is there a regime of over-penalization where stronger fairness incentives produce worse outcomes? And do fairness improvements come at the cost of aggregate throughput?
where xi is the throughput of flow i. A value of J=1 indicates perfectly equal shares; J=1/N indicates complete monopolization by one flow. The index measures relative balance, independent of the absolute scale of throughputs. A fairness-aware CC protocol should target high J without sacrificing aggregate throughput. This Pareto property, that fair redistribution without shrinkage is the standard by which our three strategies are evaluated.
3
1.3 Contributions This paper makes the following contributions: (1) We reproduce Aurora’s single-flow performance on a time-varying link, confirming 93.8% utilization against TCP CUBIC’s 65.5%.
Methodology
3.1 Simulator We conduct all experiments using a simulator implemented in pure Python: a simplified reproduction of the code accompanying the Aurora paper [1] with an added multi-flow shared-bottleneck environment. This is not a traditional complex network simulator which models the protocol stack and use C++. A “packet” in our environment is a (send time, flow id) pair; the bottleneck is one FIFO queue with deterministic service time 1/B; control decisions are taken once per monitor interval (MI).
(2) We build a shared-bottleneck multi-flow simulator reproducing the early-bird competitive dynamics described by Astraea. (3) We implement and evaluate three fairness strategies that augment Aurora without replacing its RL architecture: reward shaping, state augmentation, and loss-coefficient tuning.
3.2 RL Formulation We replicate the Aurora formulation from Jay et al. [1] exactly; Table 1 summarizes key parameters. All models are trained to 1.6M environment steps; base Aurora trains in approximately 20 minutes at ∼1300 fps on a single machine.
(4) We characterize the over-penalization failure mode, a counter-intuitive regime where increasing a fairness penalty worsens fairness, and explain its mechanism via the early-bird stagger interaction. 2
Last k=10 MI stats: (lat grad, lat ratio, send ratio) → 30 floats a ∈ [−1, 1]; xt = xt−1 (1+αa) if a ≥ 0, else xt−1 /(1−αa), α=0.025 10·tput − 1000·lat − 2000·loss MLP 30→32→16→1, tanh PPO, γ=0.99, lr 3×10−4 , 4 envs 400 monitor intervals BW 100–500 pps, lat 50–500 ms, queue 2–2981 pkts, loss 0–5%
Research Question
3.3 Phase 1: Single-Flow Baselines We compare two protocols on a time-varying link to reproduce Figure 3 of Jay et al. [1]. Table 2 reports summary statistics over the 25-second trace,
Can Aurora’s fairness be improved through targeted training modifications: reward shaping, state augmentation, or loss-sensitivity 2
alongside the values reported in the original Aurora paper for comparison. Our reproduction matches the paper to within ∼10% on every metric, confirming that the training pipeline produces a single-flow Aurora controller of comparable quality to the published baseline.
3.4.3
Strategy A — Reward Shaping
3.4 Phase 2: Multi-Flow Fairness Evaluation 3.4.1 Design Space and Rationale The three fairness strategies are not chosen ad hoc; they are a deliberate axis-aligned decomposition of the minimal-modification design space around Aurora’s training pipeline. Each strategy alters exactly one component while every other component— network architecture, action space, PPO hyperparameters, episode length, training link distribution— is held identical to the unmodified baseline. The three axes are: the reward function (Strategy A adds a fair-share penalty term), the observation space (Strategy B appends two competition-aware features), and a single reward coefficient (Strategy C scales the loss penalty). This orthogonality lets us attribute any fairness change to a single intervention and makes the three strategies cleanly comparable; it also defines the scope of the study— we deliberately exclude heavier modifications such as multi-agent self-play training or policy-architecture changes, which lie outside the post-hoc minimal-edit regime. Sweep ranges follow a geometric grid around the unmodified baseline value, which is a standard convention for first-pass hyperparameter exploration when no prior literature estimate is available. Strategy A sweeps λ ∈ {0.5, 1.0, 2.0, 5.0} (roughly 2× steps starting from a value where the penalty term is comparable to a single-MI reward unit). Strategy C sweeps the loss coefficient ∈ {4000, 8000, 16000} (2×, 4×, 8× the Aurora baseline of 2000). Strategy B has no scalar to sweep so it is evaluated at a single configuration. The reported numbers in this paper use the sweepbest setting for each strategy (λ=2.0, loss=8000); the over-penalization regime at the high end of each sweep is itself a finding, discussed in Section 5.3.
where Ĉ is estimated as send rate×send ratio. No inter-agent communication is required. We sweep λ ∈ {0.5, 1.0, 2.0, 5.0}, training a separate model for each value and evaluating in heterogeneous competition (ego: modified; background: unmodified Aurora).
Strategy A adds a self-restraint penalty when a flow’s throughput exceeds its estimated fair share: r′ = r − λ · max 0, xi − Ĉ/N (2)
3.4.4
Strategy B — State Augmentation
Strategy B retains the original reward but extends the 30-dimensional observation with two additional features: (i) estimated number of competing flows (inferred from throughput variance and loss patterns relative to solo-flow behavior), and (ii) ego throughput as a fraction of estimated total capacity. The augmented observation has 32 dimensions. 3.4.5
Strategy C — Loss-Sensitivity Tuning
Aurora’s base reward uses a loss coefficient of 2000. Strategy C increases this coefficient to values in {4000, 8000, 16000}, making the agent more sensitive to packet loss. In a shared-bottleneck environment, elevated loss signals congestion from competing flows; higher sensitivity should trigger AIMD-like backoff, incidentally improving fairness. No architectural changes are required. 3.4.6
Evaluation Protocol
All strategies share the same protocol: ego = strategy-trained policy; background = unmodified Aurora; flows = 2; episodes = 50; stagger = 50 steps. Primary metric: mean Jain’s J with standard deviation and minimum J. Secondary metrics: per-flow and aggregate throughput. 3.4.7
Hyperparameter Selection Results
Table 3 reports the full sweep under the protocol above; this same protocol serves as the insweep selection criterion. We select the highest mean J at lowest standard deviation σ, subject to aggregate-throughput preservation (3.51–3.56 Mbps across all rows, within ∼1% of the all-Aurora baseline). The sweep-best configurations carried forward to the extended evaluation (Section 4) are λ=2.0 for Strategy A and loss=8000 for Strategy C; Strategy B has no scalar to tune. Interior monotonicity (Strategy A: J rises 0.806 → 0.854 → 0.876 across λ=0.5→1.0→2.0; Strategy C: J rises 0.820 → 0.873 from loss=4000 to 8000) supports the selected configurations as genuine sweet spots
3.4.2 Baseline Measurement We run N unmodified Aurora flows on a shared bottleneck and measure Jain’s J over 50 episodes of 400 MIs each. We use a staggered start: flow 0 operates alone for 50 steps before flows 1. . .N −1 enter at their initial rates. Without the stagger, both flows observe symmetric congestion and back off together, yielding artificially high J ≈ 0.88. With the stagger, the incumbent establishes a rate the newcomer must compete against which is a more realistic fairness evaluation. 3
Table 2: Phase 1 single-flow reproduction over the 25-second 20/40 Mbps bandwidth-switching trace. “Aurora (paper)” reports the values in [1] on a similar trace. Metric
TCP CUBIC
Aurora
Aurora (paper)
18.4 65.5% 9.68
26.3 93.8% 1.83
∼25–27 ∼92% ∼2
Mean throughput (Mbps) Link utilization Mean |gap| (Mbps)
flow trace where the active set evolves {1}→{1, 2}→{1, 2, 3}→{1, 2, 3, 4}→{1, 3, 4} over the episode.
Table 3: Phase 2 hyperparameter sweep. Mean J, σ, and per-flow throughput over 50 episodes of 2-flow staggered competition (ego vs. unmodified Aurora). Bold rows are the selected configurations carried forward; aggregate throughput is preserved across all rows (3.51–3.56 Mbps). Str.
Config
J
σ
Ego
BG
Base
2×Aurora
0.901
0.088
—
—
A A A A
λ=0.5 λ=1.0 λ=2.0 λ=5.0
0.806 0.854 0.876 0.638
0.128 0.120 0.109 0.067
2.6 2.4 2.3 3.1
0.9 1.2 1.3 0.4
B
state aug
0.863
0.118
2.3
1.2
C C C
loss=4000 loss=8000 loss=16000
0.820 0.873 0.832
0.126 0.114 0.125
2.5 1.3 1.1
1.1 2.2 2.4
4
Extended Evaluation
This section reports results for the two extended scenarios introduced in Section 3.5. Strategy C is the most TCP-friendly mechanism by a wide margin in mixed traffic; Strategy B is the most stable through dynamic flow-set changes; Strategy A is the worst in the mixed-traffic case. The mixed-traffic table uses 50 episodes; the dynamic table is a single deterministic trace. 4.1 Mixed Aurora + TCP CUBIC We pair each ego policy with one TCP CUBIC flow on the shared bottleneck. Table 4 reports the 50episode mean. Baseline Aurora dominates CUBIC by a factor of 9.5× in throughput, in line with the prediction in Section 5.3. Strategy C narrows the ratio to 5.2× by ceding bandwidth: its high loss-coefficient response approximates CUBIC’s AIMD-style backoff. Contrary to the conjecture in Section 5.4 that Strategy C would be exploitable by an AIMD-aggressive flow, the empirical result is the opposite — Strategy A (calibrated against another Aurora) overshoots on CUBIC and worsens the imbalance to 13.0×. We complement Jain’s index with the Harm metric (introduced following reviewer feedback; see also recent fairness-aware CC literature), defined as the fractional throughput reduction CUBIC suffers due to the ego flow’s presence: ! x̂solo − x̂mixed g g Harm(f →g) = max 0, , (3) x̂solo g
rather than noisy maxima. The high-end degradation (λ=5.0: J=0.638; loss=16000: J=0.832) is the over-penalization regime, mechanistically analysed in Section 5.3. All sweep models are trained at a single seed (seed=42).
3.5 Extended Scenarios Building on the 2-flow homogeneous protocol described in the previous subsection, we add two scenarios from the proposal’s Step 3 evaluation plan. Each runs against a frozen Aurora background unless noted; the mixed-traffic scenario uses 50 episodes, the dynamic scenario uses a single deterministic trace. One backward-compatible simulator extension supports these (all 129 unit tests pass with default arguments): Active mask. MultiFlowNetwork.step(active=...) where x̂solo is CUBIC’s mean throughput on the same accepts a per-flow boolean mask; muted flows emit g link without competition. Harm captures the counterno packets and receive zero metrics, allowing flows factual damage that Jain’s index cannot: even if two to enter and leave mid-episode without perturbing flows have equal shares, a Harm of 0 means neither queue dynamics for active flows. flow was actually hurt. Strategy C achieves a mean The two extended scenarios are: Harm of 0.122—CUBIC loses only 12% of its solo (1) Mixed Aurora–CUBIC: ego ∈ {baseline, A, throughput on average—versus 0.327 for Strategy A, B, C}; background = one TCP CUBIC flow. confirming that Strategy C is the only mechanism (2) Dynamic entry/exit: a single 4that genuinely avoids penalising existing TCP traffic. 4
Table 4: Mixed Aurora + TCP CUBIC, 50 Episodes (Tput in Mbps; Ratio = Ego/CUBIC; Harm = fractional CUBIC throughput loss vs. solo baseline) J ±σ
Ego
CUBIC
Ratio
Harm±σ
0.621 ± 0.091 0.583 ± 0.044 0.600 ± 0.074 0.722 ± 0.139
3.03 3.24 3.17 2.35
0.32 0.25 0.28 0.45
9.5× 13.0× 11.4× 5.2×
0.243 ± 0.314 0.327 ± 0.311 0.284 ± 0.320 0.122 ± 0.255
Ego Baseline Aurora A (λ=2.0) B (state-aug) C (loss=8000)
This does not invalidate the comparison. The baseline J ≈ 0.9 with episodes dipping to 0.67 is a legitimate starting point, and all strategy improvements are measured relative to it.
Table 5: Dynamic Entry/Exit: Steady-State J Per Phase (Single Seed) Baseline Strategy A Strategy B Strategy C
P0(1)
P1(2)
P2(3)
P3(4)
P4(3)
1.00 1.00 1.00 1.00
0.99 0.89 0.97 0.96
0.92 0.86 0.93 0.86
0.91 0.83 0.92 0.88
0.92 0.86 0.95 0.83
5.2
On Reward Shaping and the Sweet Spot Hypothesis Strategy A confirms that reward shaping can steer RL congestion control toward fairer outcomes. The monotonic improvement from λ=0.5 to λ=2.0 (J: 0.806 → 0.876, σ: 0.128 → 0.109) demonstrates that the penalty causes the ego to cap its sending rate near the estimated fair share, leaving room for the background flow. Each doubling of λ produces a consistent, measurable improvement in both mean fairness and variance. The collapse at λ=5.0 (J=0.638) reveals a critical interaction with the staggered-start evaluation protocol. At evaluation time, the 50-step solo phase gives the ego an uncontested ramp with no penalty applied. The λ=5.0 policy optimized to avoid overshooting fair share during training’s multi-flow phase may have converged to an aggressive early-ramp strategy that exploits this window and then holds its position. The background Aurora cannot recover, producing a more extreme imbalance than any moderate-λ configuration. This over-penalization failure mode generalizes: any RL agent trained with a self-restraint penalty in a staggered environment may learn to front-load aggression in the uncontested window. Fairness rewards must be paired with temporal consistency, either applying the penalty from the first timestep or retroactively based on the eventual competitive context.
4.2 Dynamic Flow Entry / Exit A single 4-flow trace exercises five active-set phases. Table 5 reports steady-state J in each phase (mean of the last 20 steps before the next event). In P1 (2 flows) the symmetric all-Aurora baseline peaks at J=0.99, edging out Strategy B; from P2 onward Strategy B leads or ties the baseline and maintains the highest J through the flow-departure phase. The two extra observation features (count of competing flows, ego throughput fraction) make phase changes directly visible to the policy, while Strategy A lags in every phase and Strategy C degrades after a flow departs. 5
Discussion
5.1
On the Reproducibility of Astraea’s Unfairness Finding Our baseline measurement (J ≈ 0.901 for two unmodified Aurora flows) does not reproduce Astraea’s reported J<0.5. Our reproduction was trained with random loss sampled from 0–5%, producing a loss penalty that ranges up to −100 reward units per monitor interval. Under this regime, Aurora learns to retreat quickly from congestion signals; when two such agents compete, both back off symmetrically, fair collapse rather than one-flow dominance. Astraea’s scenario likely requires a lower-loss training distribution (producing a more aggressive incumbent) or network parameters that allow one established flow to hold its rate while a newcomer cannot penetrate the backlogged queue. The staggered-start protocol partially reproduces this dynamic (minimum J = 0.672 across 50 episodes), but our Aurora’s symmetric conservatism limits the depth of unfairness.
5.3
On Implicit Fairness Through State Augmentation Strategy B achieves J=0.863 (σ=0.118) without any explicit fairness incentive. This suggests that information asymmetry, rather than a misaligned reward, is a significant contributor to RL unfairness. An agent that knows it is competing and can observe its throughput fraction can learn cooperative behavior from the original throughput-maximizing 5
objective, because hoarding bandwidth in a contested environment produces diminishing marginal returns once the queue is saturated. The practical advantage of Strategy B is that it requires no change to the reward function. Operators reluctant to alter the training objective can add observational features and retrain, achieving most of the fairness benefit without fairness-specific reward engineering.
9.5× (baseline) to 5.2×. This contradicts the conjecture in Section 5.4 that Strategy C’s conservative backoff would be exploited by an AIMD-aggressive flow. The mechanism is the opposite of the one we predicted: with loss coef=8000, the ego responds to shared-queue drops faster than the unmodified Aurora, so it concedes the contested bandwidth before CUBIC’s slower window growth can claim it. Strategies A and B, calibrated for Aurora-on-Aurora competition, instead overshoot in the wrong direction — treating CUBIC’s retreat as an opportunity to reclaim more capacity. The practical implication: for any near-term deployment of Aurora-class controllers on the public Internet, Strategy C is the only mechanism studied that does not strictly disadvantage existing TCP traffic.
5.4 On Loss Sensitivity as a Fairness Proxy Strategy C (loss=8000) achieves J=0.873, comparable to Strategy A λ=2.0, but with a qualitatively different per-flow outcome: 1.3 Mbps for the ego vs. 2.2 Mbps for the background. This is the only configuration in which the background flow consistently outperforms the ego, suggesting AIMD-like behavior: the ego interprets shared-queue loss as a signal to reduce below its fair share, while the background Aurora (less loss-sensitive) holds its rate. In mixed-traffic scenarios where Aurora competes with TCP CUBIC, Strategy C’s conservative backoff may be exploited by a CUBIC flow that does not reduce its window at the same loss level. Strategy A λ=2.0, by contrast, uses a throughput estimate to compute a fair-share threshold, which is a more principled approach less vulnerable to exploitation. The over-penalization effect also appears in Strategy C: loss=16000 underperforms loss=8000 for the same structural reason as λ=5.0.
5.7
On State Augmentation as the DynamicScenario Winner
Section 4.2 shows that Strategy B leads or ties the all-Aurora baseline in every phase except P1 (2 flows), where the symmetric baseline peaks at J=0.99 and B reaches 0.97. The gap is within noise, but the pattern is clear: when flows are few and symmetric the baseline’s uniformity is hard to beat; as the active set grows (P2–P3) or shrinks (P4), B’s two extra observation features (count of competing flows, ego throughput fraction) give the policy a direct read on the changed fair share, whereas A and C must infer it indirectly from latency or loss. This is the practical counterpart to the implicit-fairness theme of Section 5.3: information design is sometimes a more direct route to the desired behaviour than reward design.
5.5 Aggregate Throughput Preservation Aggregate throughput remains stable at 3.51– 3.56 Mbps regardless of fairness configuration. This Pareto property (fairness improvement without efficiency loss) is the ideal outcome for a sharedbottleneck CC system. The link is a hard capacity constraint; redistribution cannot create or destroy bandwidth. The strategies cause the ego to give up bandwidth it was holding above its fair share: bandwidth that was either sitting in the queue or causing unnecessary loss. The implication is that deploying any of these strategies in a real network would improve competing flows’ experience without degrading aggregate utilization, in contrast to per-flow rate limiters that impose efficiency costs in exchange for fairness guarantees.
5.8
Limitations
Several of the limitations identified in earlier drafts of this work are now addressed by the extended evaluation in Section 4: mixed Aurora–CUBIC competition (Section 4.1) and dynamic flow entry/exit (Section 4.2). The remaining limitations apply. First, a fluid simulator rather than full packet-level simulation understates timing-noise variance. Second, varying stagger length and the number of competing flows would strengthen the claims. Third, all experiments use a fixed-architecture background flow, sampling only one point in a heterogeneous deployment space.
5.6
On Strategy C’s Coexistence with TCP CUBIC Section 4.1 provides the strongest single result of this study: Strategy C is the only mechanism that genuinely improves Aurora’s coexistence with TCP CUBIC, narrowing the throughput ratio from
6
Conclusion
We have implemented and evaluated three fairness strategies for Aurora, a deep RL congestion controller trained in single-flow environments. Modest reward shaping achieves the best Jain’s fairness 6
index (J=0.876) with the lowest variance; state augmentation (Strategy B) provides competitive fairness (J=0.863) without modifying the reward function; and loss-sensitivity tuning achieves comparable fairness through conservative backoff. All three strategies preserve aggregate throughput within 1% of the unmodified Aurora baseline. We identify an overpenalization failure mode in Strategy A and Strategy C in which stronger fairness incentives produce worse outcomes due to interaction with the staggered evaluation protocol. The broader lesson is that RL-based congestion control fairness is a training-time problem that cannot be fully addressed at deployment time. The three strategies studied here represent complementary points in a design space bounded by reward engineering (A), information design (B), and penalty calibration (C), and together suggest that practical improvements to RL-CC fairness are achievable without abandoning the RL framework.
global Internet. IEEE J. Sel. Areas Commun., 13(8):1465–1480, 1995. [4] H.-C. Yen et al. On the fairness of Internet congestion control over WiFi with deep reinforcement learning. arXiv preprint, 2024. [5] N. Cardwell, Y. Cheng, C. S. Gunn, S. H. Yeganeh, and V. Jacobson. BBR: Congestionbased congestion control. ACM Queue, 14(5), 2016. [6] R. Jain, D.-M. Chiu, and W. Hawe. A quantitative measure of fairness and discrimination for resource allocation in shared computer systems. DEC Technical Report TR-301, 1984. [7] H. Mao, R. Netravali, and M. Alizadeh. Neural adaptive video streaming with Pensieve. In Proc. ACM SIGCOMM 2017, 2017. [8] N. Jay et al. Internet congestion control via deep reinforcement learning (Custard). arXiv:1810.03259, 2018.
References [1] N. Jay, N. Rotman, B. Godfrey, M. Schapira, and A. Tamar. A deep reinforcement learning perspective on Internet congestion control. In Proc. 36th ICML, 2019.
[9] S. Ha, I. Rhee, and L. Xu. CUBIC: A new TCPfriendly high-speed TCP variant. ACM SIGOPS Oper. Syst. Rev., 42(5):64–74, 2008.
[2] X. Liao et al. Towards fair and efficient learningbased congestion control. In Proc. EuroSys 2024, 2024.
[10] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv:1707.06347, 2017.
[3] L. S. Brakmo and L. L. Peterson. TCP Vegas: End to end congestion avoidance on a
7