Deep Adaptive Rate Allocation in Volatile Heterogeneous Wireless Networks

Gregorio Maglione; Veselin Rakocevic; Markus Amend; Touraj Soleymani

cs.ITMar 21, 2026

Deep Adaptive Rate Allocation in Volatile Heterogeneous Wireless Networks

Authors: Gregorio Maglione, Veselin Rakocevic, Markus Amend, Touraj Soleymani

Organizations: Department of Engineering, City St George’s, University of London, London EC1V 0HB, United Kingdom · Deutsche Telekom, Darmstadt, Germany

Abstract

Modern multi-access 5G+ networks provide mobile terminals with additional capacity, improving network stability and performance. However, in highly mobile environments such as vehicular networks, supporting multi-access connectivity remains challenging. The rapid fluctuations of wireless link quality often outpace the responsiveness of existing multipath schedulers and transport-layer protocols. This paper addresses this challenge by integrating Transformer-based path state forecasting with a new multipath splitting scheduler called Deep Adaptive Rate Allocation (DARA). The proposed scheduler employs a deep reinforcement learning engine to dynamically compute optimal congestion window fractions on available paths, determining data allocation among them. A six-component normalised reward function with weight-mediated conflict resolution drives a DQN policy that eliminates the observation-reaction lag inherent in reactive schedulers. Performance evaluation uses a Mininet-based Multipath Datagram Congestion Control Protocol testbed with traces from mobile users in vehicular environments. Experimental results demonstrate that DARA achieves better file transfer time reductions compared to learning-based schedulers under moderate-volatility traces. For buffered video streaming, resolution improvements are maintained across all tested conditions. Under controlled burst scenarios with sub-second buffer constraints, DARA achieves substantial rebuffering improvements whilst state-of-the-art schedulers exhibit near-continuous stalling.

Figures & tables

Fig. 1: Scheduler response to antiphase path capacity variations. Fixed-rule schedulers maintain fixed allocation regardless of path capacity. Reactive schedulers exhibit observation-reaction lag

\tau

. DARA eliminates this lag through its predictive horizon, preemptively adjusting CWND fractions to match anticipated capacity variations.

Panel of Figure 2 — Fig. 2: DARA’s Predictive Rate Control Architecture. The Transformer predicts CWND and SRTT evolution for each path. The DQN policy uses these predictions to compute CWND fractions that rate-limit each path. In this example, DARA predicts Path 1 degradation and pre-emptively limits it to $\phi_{1}=0.3$ , whilst allowing the stable path full utilisation at $\phi_{2}=1.0$ , enabling sequential packet arrival and minimal head-of-line blocking.

Block	FF Dim	Heads	Activation	Dropout
1	360	5	GELU	10%
2	360	5	GELU	10%
3	360	5	GELU	15%
Dropout increases with depth to regularise longer-range prediction targets.

TABLE I: Deployed Transformer block specifications (3-layer, selected via depth analysis in Section IV-C ).

Prediction model
Component	Value
Transformer depth	3 layers
Attention heads	5
Feedforward dimension	360
Input window	8 $\times$ 100 ms steps
Prediction horizons	5 $\times$ 100 ms (100–500 ms)

TABLE II: Final DARA deployed configuration.

Fig. 3: Path throughput normal distribution.

∗ Excluded: unclamped inverse-transform artefact.
Architecture	NRMSE	Params	Latency (ms)	Reward
Transformer (3L)	0.0009	305K	1.8	1.00
Linear	0.0022	2K	0.0	0.97
LSTM (2L)	0.0031	247K	0.2	0.97
MLP (2L)	0.0376	291K	0.1	— ∗

TABLE III: Predictor architecture comparison. NRMSE: normalised root mean square error; Reward: normalised downstream DQN performance (Transformer = 1.00).

Configuration	Actions	Reward	Preemptive (%)
2-Level	4	$17.1\pm 3.9$	53.4
3-Level	9	$101.5\pm 5.4$	56.9
5-Level	25	$155.0\pm 3.8$	63.6
7-Level	49	$152.6\pm 16.8$	62.4

TABLE IV: Action granularity analysis. Reward: mean

\pm

std over 3 runs; Preemptive: fraction of congestion events handled anticipatorily.

Fig. 4: Mean reward with 95% confidence intervals across all methods.

Method	$R$	Pre. (%)
Random	$1.0$	0.0
Reactive	$20.0$	0.0
Static 30	$-56.0$	0.0
Static 65	$-7.2$	0.0
Static 100	$-10.4$	0.0
No Transformer	$14.7$	0.0

TABLE V: Ablation results.

R

: mean reward; Pre.: preemptive rate.

Comparison	$\Delta R$	$d$	$p$	95% CI
vs. Random	156.9	23.5	$<10^{-7}$	[151.7, 163.4]
vs. Reactive	137.9	12.2	$<10^{-7}$	[132.6, 144.0]
vs. Static 65	165.1	16.3	$<10^{-7}$	[159.6, 171.7]
vs. Static 100	168.2	13.7	$<10^{-8}$	[162.9, 174.7]
vs. No Transformer	143.1	28.4	$<10^{-6}$	[138.4, 149.5]
vs. Ground-Truth DQN	40.9	5.4	$<10^{-5}$	[32.5, 49.5]

TABLE VI: Statistical comparison of DARA against baselines.

\Delta R

: mean reward difference;

d

: Cohen’s

d

;

p

: Welch’s

t

-test; CI: bootstrap 95% confidence interval.

Fig. 5: Asymmetric burst pattern showing one complete 2.5-second cycle. Path 1 exhibits 700 ms bursts peaking at 10 Mbps, collapsing to 0.2 Mbps baseline; Path 2 maintains stable 1.8 Mbps throughout.

Fig. 6: Normalised performance metrics for controlled burst scenario. Values $>1.0$ indicate improvement over single-path baseline. DARA achieves consistent leadership across applications, with advantages amplifying as latency sensitivity increases (FTP $\to$ YouTube $\to$ Live).

Fig. 7: Per-scheduler time-series showing delay (top, per-packet raw and 100-packet moving average) and throughput (bottom, per-path) for FTP under the asymmetric burst pattern. DARA (bottom panel) maintains stable low delay with sustained burst exploitation, whilst reactive schedulers exhibit periodic delay spikes aligned with burst-trough transitions.

Fig. 8: TCP file transfer performance distributions across five cellular traces (FTP application protocol; diamond markers indicate the mean; ratio versus CPF ,

{>}1.0

indicates improvement). DARA achieves the highest median throughput with the tightest interquartile range; OTIAS achieves the lowest delay at the cost of substantially reduced throughput.

Fig. 9: QUIC file transfer performance distributions across five cellular traces (cURL application; diamond markers indicate the mean; ratio versus CPF ,

{>}1.0

indicates improvement). Throughput gains are attenuated relative to TCP file transfer owing to nested congestion control interactions between tunnel-layer BBR and QUIC’s internal algorithm. Peekaboo and BLEST exhibit systematic collapse on delay and jitter axes, reflecting the combined effect of algorithmic limitations and MP-QUIC per-stream ordering constraints rather than algorithmic quality alone.

Fig. 10: YouTube QoE distributions (diamond markers indicate the mean; ratio versus CPF ).

Fig. 11: Live streaming performance distributions (diamond markers indicate the mean; ratio versus CPF ).

Wireless Resource Allocation

Wireless resource allocation assigns spectrum, power, beams, and computing capacity to meet throughput, latency, fairness, and energy goals.

See all 174 papers

Generalizable Multi-Task Learning for Wireless Networks Using Prompt Decision Transformers
Jun 3, 2026Fatih Temiz, Shavbo Salehi, Melike Erol-KantarciReinforcement Learning Transformer
School of Electrical Engineering and Computer Science, University of Ottawa, Ottawa, Canada
Adaptive and Resilient Dual-Layer Resource Slicing for Hovering Aerial Backhaul Networks
Sep 26, 2026Chuan-Chi Lai, Jen-Hsiang LiNetwork Slicing Wireless Communications
Department of Communications Engineering, National Chung Cheng University, Minxiong Township, Chiayi County 621301, Taiwan · Advanced Institute of Manufacturing with High-tech Innovations (AIM-HI), National Chung Cheng University, Minxiong Township, Chiayi County 621301, Taiwan · Department of Information Engineering and Computer Science, Feng Chia University, Taichung 407102, Taiwan
Multi-Agent Reinforcement Learning for Online Traffic Scheduling in Time-Sensitive Application
Aug 5, 2026Marcos Carvalho, Fatih Temiz, Shavbo Salehi +2Multi-Agent Coordination Edge Computing
Universidade Federal de Minas Gerais, Brazil · School of Electrical Engineering and Computer Science, University of Ottawa, Ottawa, Canada

Jun 3, 2026J/K move · Enter open · S save

Generalizable Multi-Task Learning for Wireless Networks Using Prompt Decision Transformers

Fatih Temiz, Shavbo Salehi, Melike Erol-Kantarci

School of Electrical Engineering and Computer Science, University of Ottawa, Ottawa, Canada

Reinforcement Learning Transformer Decision Transformer Multi-Task Learning Offline RL Multi-Task RL+1

Open paper AskPDF arXiv

Future wireless networks demand rapid adaptation to highly heterogeneous environments and dynamic task configurations, necessitating a shift from conventional rule-based and optimization-driven radio resource management (RRM) toward artificial intelligence (AI)-driven RRM. AI-driven approaches can learn complex nonlinear relationships, generalize across diverse network conditions and enable real-time, scalable and autonomous decision-making. Among RRM techniques, coordinated multipoint (CoMP) transmission is pivotal for mitigating inter-cell interference and enhancing cell-edge performance, thereby improving quality of experience (QoE) in dense deployments. However, optimal multi-cell selection remains a complex combinatorial challenge as it requires jointly optimizing over many possible serving-cell combinations under dynamic traffic and channel conditions. Despite their success, conventional deep reinforcement learning (DRL) methods such as proximal policy optimization (PPO) suffer from poor sample efficiency, limited generalization, and costly retraining when state and action spaces change. To address these bottlenecks, we propose a Prompt Decision Transformer (PromptDT) based multi-task learning framework capable of learning across diverse network configurations and reformulating multi-cell selection as a sequence modeling problem. By leveraging offline trajectories and task-specific prompts, PromptDT enables scalable learning across diverse network configurations, including varying base stations and user equipment counts, and scheduler policies. Experimental results demonstrate that PromptDT improves QoE by up to 49% in multi-task settings compared to baselines, with performance scaling positively alongside model capacity. Moreover, PromptDT generalizes effectively to unseen tasks, achieving robust few-shot adaptation to new network configurations without retraining or fine-tuning.

Abstract

Wireless Resource Allocation

Explore similar work

Generalizable Multi-Task Learning for Wireless Networks Using Prompt Decision Transformers

Adaptive and Resilient Dual-Layer Resource Slicing for Hovering Aerial Backhaul Networks

Multi-Agent Reinforcement Learning for Online Traffic Scheduling in Time-Sensitive Application