Prototyping QoE-Aware Rate Adaptation in Cellular Networks with Commercial Applications

Szilveszter Nádas; Lars Ernström; Dan Druta; Igor Pruzhansky; David Lindero; Jonathan Lynam; Eric Petajan

Source-linked AI summary

Prototyping QoE-Aware Rate Adaptation in Cellular Networks with Commercial Applications

Szilveszter Nádas, Lars Ernström, Dan Druta, Igor Pruzhansky, David Lindero, Jonathan Lynam, Eric Petajan

arXiv:2609.09490v1Sep 8, 2026cs.NIcs.MMeess.IV

TL;DR

QoE-aware cellular allocation lacks the application, measurement, and API capabilities needed for commercial deployment. This paper prototypes an evolutionary alternative using composite spatial complexity, external QoE measurement, and bounded congestion-signal rate steering. The resulting design provides QoE-aware allocation with commercial applications while defining an evolution path toward production systems.

Problem
Commercial real-time video applications lack QoE-targeted encoding, runtime spatial-complexity estimation, rich application–network APIs, and direct QoE measurement needed for full QoE-aware allocation.
Method
The paper combines external per-session QoE measurements with spectral efficiency in a composite resource-demand model and uses incremental utility-based reallocation with congestion-signal rate steering.
Results
The prototype achieves QoE-aware allocation with external measurement and network-side rate steering without modifying the streaming application.
Takeaways & Limitations
The lab prototype provides an evolutionary bridge toward full QoE-aware resource sharing, with richer in-application measurement as the first production step.
Takeaways & Limitations
Open challenges include runtime spatial-complexity-curve estimation, temporal QoE modeling, privacy-preserving information sharing, and incentives against QoE misrepresentation.

Abstract

from arXiv · show

Prior work has shown that QoE-aware resource sharing for real-time interactive video can support up to three times more simultaneous sessions at acceptable quality compared to rate-fair allocation. However, the required capabilities (QoE-targeted encoding, runtime spatial complexity estimation, and rich application-network APIs) are not yet available in commercial deployments. In this paper, we take an evolutionary approach: we design a system that delivers QoE-aware resource allocation using only capabilities that can be assembled in a lab today. We extend the utility-based allocation framework to the radio resource domain by introducing composite spatial complexity, which combines a session's video spatial complexity with its time-variant spectral efficiency into a single resource demand function. To operate with commercial real-time video streaming applications that use rate-based congestion control and lack capability to measure QoE, we use external tooling for QoE measurements. We develop an incremental reallocation algorithm with per-interval limits that encode both the congestion control algorithm's speed constraint and that spatial complexity estimates are reliable only near the current rate. The resulting prototype combines external QoE measurements with congestion-signal-based rate steering and does not require modification to commercial applications. We chart an evolution path from this prototype toward full QoE-aware resource sharing, mapping emerging standards (IETF SCONE, CAMARA, Media over QUIC) to the progressive capabilities they enable.

I. INTRODUCTION

QoE-aware allocation promises substantial gains over rate-fair management, but commercial applications lack the capabilities needed for the ideal system. The paper therefore develops a lab-ready evolutionary approach using external QoE measurement and network-side rate steering.

I. INTRODUCTION: QoE matters to operators because experienced application–network quality affects customer retention and premium positioning as mobile video traffic grows.Traditional coverage and speed metrics do not capture this jointly delivered user experience.
I. INTRODUCTION: Up to 3× more simultaneous streams at acceptable quality were reported when allocation exploits differences in session spatial complexity.Prior work also showed gains over static rate allocation and equal-QoE strategies.
I. INTRODUCTION: Commercial deployments lack QoE-targeted encoding, frequent two-way application–network APIs, runtime SCC estimation, and cooperative trustworthy applications.Commercial applications instead use bitrate-domain congestion control and adaptation mechanisms.
I. INTRODUCTION: Congestion signals can steer commercial sessions toward target bitrates without modifying the application.Scalable congestion control and Active Rate Management provide the underlying rate-steering mechanisms.
I. INTRODUCTION: The prototype obtains per-session spatial QoE and bitrate measurements externally, combines them with spectral efficiency and utilization, and computes rate targets.This approach works with commercial applications available today.

II. QOE-AWARE RESOURCE ALLOCATION WITH SPECTRAL EFFICIENCY

The ideal allocator extends utility-based sharing to cellular radio resources by combining video spatial complexity with time-varying spectral efficiency. It allocates radio resources according to marginal utility rather than bitrate alone.

II. QOE-AWARE RESOURCE ALLOCATION WITH SPECTRAL EFFICIENCY: The framework extends fixed-capacity QoE-aware sharing to radio resources by incorporating time-variant spectral efficiency.This adds channel-dependent variation that fixed Mbps models do not capture.
II. QOE-AWARE RESOURCE ALLOCATION WITH SPECTRAL EFFICIENCY: The ideal architecture assumes frequent session spatial-complexity and spectral-efficiency information plus the total resources available for RT+I traffic.These inputs are delivered to a QoE controller through rich APIs.
II. QOE-AWARE RESOURCE ALLOCATION WITH SPECTRAL EFFICIENCY: Composite spatial complexity combines video spatial complexity with channel spectral efficiency into a time-variant resource demand for achieving a given VMAF.Spatial complexity gives the required rate, while spectral efficiency determines rate per unit radio resource.
II. QOE-AWARE RESOURCE ALLOCATION WITH SPECTRAL EFFICIENCY: Algorithm 1 distributes available RT+I radio resources to maximize total utility using marginal utility per unit radio resource.The utility function represents the operator’s resource-sharing policy.
II. QOE-AWARE RESOURCE ALLOCATION WITH SPECTRAL EFFICIENCY: The allocator iteratively selects the session with the highest marginal utility and updates its VMAF and radio-resource allocation.The algorithm stops when no session can increase or no free resources remain.

III. INCREMENTAL QOE-AWARE RATE ADAPTATION

The prototype adapts QoE-aware allocation to commercial congestion-controlled applications by combining external measurements with bounded incremental reallocation. Its limits preserve reliable local SCC information and respect the rate changes applications can achieve.

III. INCREMENTAL QOE-AWARE RATE ADAPTATION: The prototype updates targets about every 1 s, while a per-RTT scheduler uses congestion signals to steer L4S congestion control toward each target.External AMVOTS measurements substitute for in-application QoE reporting.
III. INCREMENTAL QOE-AWARE RATE ADAPTATION: A per-interval VMAF limit keeps changes near the single measured rate–VMAF point because only local SCC information is available.For example, ΔVmax can be 5 VMAF points per interval.
III. INCREMENTAL QOE-AWARE RATE ADAPTATION: Allocation-ratio bounds reflect cautious upward probing and faster congestion-driven rate cuts, with the first VMAF change always permitted.With ΔVmax = 5, convergence takes a few intervals and deliberately favors QoE stability over rapid adaptation.
III. INCREMENTAL QOE-AWARE RATE ADAPTATION: Sessions at unacceptable minimum quality are released from QoE management and returned to best-effort service, with cooldown preventing oscillation.New sessions compete for admission through the same feasibility and resource-allocation mechanism.
III. INCREMENTAL QOE-AWARE RATE ADAPTATION: The incremental algorithm reallocates resources across measured sessions, choosing utility-improving increases when resources are available and decreases when the budget is exceeded.Its output is a target VMAF and radio-resource allocation for each session.

A. Spatial complexity estimation

The prototype estimates spatial complexity locally because it observes only one rate–VMAF point per interval, while incremental reallocation keeps changes within a tractable neighborhood. Even with complete curve information, limiting QoE changes can stabilize users and reduce required knowledge.

A. Spatial complexity estimation: The prototype receives one measured rate–VMAF pair per session and interval, so it lacks the full spatial complexity curve and its local slope.Measured spectral efficiency adds one point on the composite curve, but the runtime estimation method remains unspecified.
A. Spatial complexity estimation: Incremental adjustments trace nearby operating points that can support local slope estimation, while per-interval limits confine each step to a small neighborhood.The exact estimation algorithm is left for future work.
A. Spatial complexity estimation: Algorithm 2 requires the spatial complexity curve within ±5 VMAF of the current operating point to compute resource costs.This requirement makes a local slope estimate necessary.
A. Spatial complexity estimation: Limiting QoE change remains desirable even with rich curve information because it stabilizes user experience and reduces the required curve knowledge to a small neighborhood.The paper identifies the actual curve-estimation algorithm as an open challenge outside its scope.

B. Privacy

Privacy concerns arise because QoE reporting can expose content characteristics when streams have known fingerprints. The paper proposes adding one-sided noise that preserves quality targets but slightly reduces total utility.

B. Privacy: QoE reporting is less identifying for unique real-time interactive content but can reveal stream information when recorded content has a known QoE fingerprint.The stated privacy concern is tied to recognizable content rather than QoE reporting in general.
B. Privacy: One-sided noise can inflate reported complexity or encoder quality targets, preserving quality targets at the cost of a small reduction in total utility.The noise always represents higher complexity.

C. Trust and incentives

QoE-aware allocation creates incentives for applications to misreport information to obtain resources. The paper discusses trusted SDK support as a complementary technical response, while the algorithm enforces bounded incremental changes and admission limits.

C. Trust and incentives: QoE-aware allocation can incentivize applications to misrepresent QoE, spatial complexity, or compliance to obtain more resources.The paper identifies designing economic incentives against this behavior as a hard problem.
C. Trust and incentives: A trusted SDK could handle encoding, transport, and API communication so reported metrics remain reliable and guidance is followed.The SDK is presented as a complementary technical approach to incentive design.
C. Trust and incentives: The reallocation algorithm admits sessions toward Vmin and permits normal increases only when the next step respects Vmax and the per-interval QoE-change bound.The first step can be free, after which the change is ρ-bounded.
C. Trust and incentives: Decreases are similarly bounded, except that a session at or below Vmin can release all resources when its current allocation is positive.The algorithm marks blocked decreases as unavailable and assigns them negative-infinite utility.

V. BEYOND THE PROTOTYPE

Moving from the prototype to the ideal system requires progressively richer capabilities. The paper frames this evolution as addressing broader impairments through ongoing standardization.

V. BEYOND THE PROTOTYPE: The path from prototype to ideal system requires progressively richer capabilities that address a broader set of impairments and are enabled by ongoing standardization.This passage characterizes the deployment path at a high level rather than enumerating individual standards or stages.

A. Evolution stages

The evolution begins with a prototype using external QoE measurements and congestion-signal rate steering, then progresses toward richer guidance and application-integrated QoE control. It also expands from congestion mitigation to handling radio, transport, and content-driven impairments.

Production transition: Production deployment requires moving QoE measurement from external AMVOTS tooling into the streaming application itself.The prototype currently relies on external measurements and congestion signals, while the controller treats per-session QoE as a typed input across RT+I content.
Persistent design principle: The incremental approach remains useful even in evolved systems because local QoE changes can improve session quality without requiring the full spatial complexity curve.Restricting changes to a neighborhood around the current operating point requires only the local slope.
Prototype stage: The prototype reallocates resources using measured QoE and spectral efficiency, naturally deprioritizing sessions with severely degraded channels.This frees resources for sessions that can use them effectively.
Rate guidance: Rate guidance enables faster responses to signal-quality changes and handovers than rate-based congestion-control ramp-up.It addresses prolonged quality drops after network impairments.
QoE guidance: QoE guidance and QoE-targeted encoding address content-driven complexity changes without rate overshoot, reducing ABR misprediction at its source.QoE guidance trades burstiness in QoE for burstiness in rate and resources, which current networks are not designed to handle.
Impairment handling: Application-reported QoE enables root-cause attribution by correlating degradation with radio or transport impairments for targeted optimization.Real-time and prediction capabilities must evolve to identify, attribute, and mitigate impairments across different layers and timescales.

C. Standards and protocols enabling the evolution

Emerging standards provide complementary building blocks for the evolution from congestion-signal steering to closed-loop QoE-aware control. SCONE supplies network-to-application rate advice, while CAMARA and related interfaces support application QoE reporting and network measurements.

SCONE: SCONE provides on-path QUIC throughput advice that can deliver per-session rate targets and identify flows without out-of-band mapping.It directly enables the rate-guidance stage by replacing or supplementing congestion-signal steering.
Media over QUIC: Media over QUIC supports real-time media delivery with codec flexibility and native rate adaptation through codec parameter changes.Its QUIC foundation creates a natural integration point for SCONE throughput advice.
CAMARA: CAMARA enables application-to-network QoE reporting through APIs without requiring on-path protocol changes.Aggregated device data can be correlated with network measurements for SCC estimation and RAN scheduling.
Open RAN: Open RAN interfaces can expose spectral efficiency and radio resource utilization to QoE controllers implemented as rApps or xApps.These measurements are required by the paper’s radio-domain allocation algorithms.
Closed-loop evolution: Combining SCONE throughput advice with CAMARA QoE reporting creates the richer two-way interaction needed for advanced evolution stages.A trusted SDK could bundle these protocol interactions into one application integration point.

VI. CONCLUSION

The paper shows that QoE-aware allocation can be prototyped with commercial applications despite limited observability and control. It extends allocation to radio resources, uses constrained incremental reallocation, and maps a path toward production while identifying remaining challenges.

Conclusion: The incremental algorithm treats limited observability and control as first-class constraints by bounding QoE movement near reliable SCC estimates and resource changes within CCA capabilities.This is a fundamentally different design from approximating the ideal algorithm.
Conclusion: Composite spatial complexity extends QoE-aware allocation to radio resources by combining spatial complexity with per-session spectral efficiency.The radio-domain formulation captures variation that rate-domain models cannot.
Conclusion: The prototype achieves QoE-aware allocation without modifying streaming applications by combining external QoE measurement with network-side congestion-signal rate steering.Moving to production requires replacing AMVOTS with in-application QoE measurement.
Conclusion: SCONE, CAMARA, and Media over QUIC provide standardization building blocks that can progressively relax the prototype’s per-interval limits toward ideal allocation.These standards respectively support rate guidance, application QoE reporting, and codec-level rate adaptation.
Conclusion: Open challenges include runtime SCC estimation, temporal session-level QoE modeling, privacy-preserving information sharing, and incentives against QoE misrepresentation.These challenges constrain the transition from the lab prototype to broader deployment.