QoS-Aware Intelligent Model Selection for Autonomous Vehicles under Dynamic Network and Resource Conditions

Abstract

Autonomous-vehicle perception systems increasingly operate across heterogeneous compute and communication environments, where the most accurate model is not necessarily the model that provides the best end-to-end service quality. A high-capacity perception model may improve recognition quality while increasing inference time, CPU consumption, memory demand, and sensitivity to network conditions. This paper proposes a Quality-of-Service (QoS)-Aware Intelligent Model Selection framework that predicts the expected QoS of candidate perception models from model, network, and hardware context and dynamically selects a feasible model according to application priorities. The framework extends a network-aware autonomous-vehicle benchmarking foundation with a QoS prediction layer, multi-objective utility function, Pareto filtering, and a hysteresis-based switching controller. Three tabular prediction methods—linear regression, random forest, and gradient boosting—are compared. A synthetic dataset of 1,800 observations is generated from six candidate model profiles, five network conditions, three hardware conditions, and repeated measurements. The synthetic evaluation indicates that nonlinear predictors outperform the linear baseline and that QoS-aware selection can reduce mean response time by approximately 31.8–42.1% relative to an accuracy-only baseline in the modeled scenarios. Pareto and ablation analyses further show that latency and resource terms materially influence the selected model. Because no real AV testbed experiment has yet been conducted, these numerical results are explicitly treated as synthetic validation rather than empirical evidence. The paper therefore provides a reproducible research design and a publication-oriented methodology whose final claims should be revalidated with measured AV executions.

Introduction

This text presents a QoS-aware model-selection framework for autonomous vehicles (AVs). The central idea is that selecting an AV perception model should not depend only on accuracy. Instead, the system should consider response time, network conditions, hardware/resource availability, and application priorities because these factors jointly affect end-to-end vehicle performance.

The work builds on earlier AV benchmarking research, particularly Pylot, which demonstrated the importance of balancing accuracy and latency, and subsequent work on distributed QoS measurement. The proposed research extends these ideas by making model selection context-dependent and adaptive. Rather than always choosing the most accurate model, the system predicts the expected QoS of candidate models under the current operating conditions and selects the most suitable one.

Main Research Questions

The study investigates four questions:

1. Can model and environmental features predict response-time QoS accurately?
2. Can predicted QoS be transformed into a transparent multi-objective selection rule?
3. Can context-aware selection reduce response time compared with accuracy-only selection while maintaining acceptable accuracy?
4. How does model selection change when the relative importance of latency, accuracy, and resource consumption changes?

Proposed Framework

The framework contains six major layers:

• Application/request context
• Candidate model pool
• QoS and resource monitoring
• QoS prediction
• Multi-objective model selection
• Execution and feedback

The system predicts the expected performance of each candidate model, removes infeasible or dominated choices using Pareto filtering, and then applies a weighted utility function based on accuracy, latency, and resource efficiency.

A switching threshold and minimum dwell time are also introduced to prevent unstable behavior in which the system repeatedly switches models because of small fluctuations in predicted QoS.

Experimental Design

Since a physical AV testbed was unavailable, the study uses a synthetic dataset of 1,800 observations. The experiments vary:

• 6 candidate models
• 5 network conditions
• 3 hardware conditions
• 20 repetitions for each combination

The six candidate models range from Model-A, which is fastest and most resource-efficient, to Model-F, which provides the highest accuracy but also the greatest latency and CPU demand.

The profiles illustrate the fundamental trade-off: higher accuracy generally comes with increased latency and resource consumption.

Overall Contribution

The proposed work combines QoS prediction, Pareto optimization, multi-objective ranking, and dynamic model switching into a closed-loop architecture for AV perception. Its main contribution is not simply introducing adaptive model selection, but applying it specifically to context-aware AV QoS optimization, where the best model can change according to network, hardware, and application conditions.

In short, the study argues that the "best" AV model is not necessarily the most accurate model; it is the model that provides the best overall trade-off between accuracy, latency, and resource usage under the current operating conditions.

Conclusion

This paper presented a QoS-aware intelligent model-selection framework for autonomous-vehicle perception services. The framework extends a benchmarking-oriented view of AV systems by adding QoS prediction, multi-objective ranking, Pareto filtering, and stable dynamic switching. The approach explicitly recognizes that model accuracy, latency, and resource consumption form a competing objective space. A synthetic experiment with 1,800 observations demonstrated the complete analytical pipeline. Ensemble predictors achieved lower response-time prediction error than the linear baseline, and the proposed selection policy reduced modeled response time by up to 42.1% relative to an accuracy-only baseline. Pareto and ablation analyses further illustrated the contribution of the individual decision components. The principal limitation is also the most important qualification: the numerical results are synthetic. The paper should therefore be treated as a methodology/prototype-validation manuscript until the framework is executed on measured AV workloads. The immediate next step is to connect the selector to the existing gRPC-based benchmarking pipeline, run repeated experiments over controlled network and hardware conditions, and replace the synthetic observations with empirical measurements. Such validation would enable stronger claims regarding statistical significance, generalization, and real-world safety relevance. 1) Note on validation. All numerical observations reported in this paper are synthetically generated for methodological validation and are not measurements from a physical autonomous vehicle. Future work should repeat the protocol using measured AV execution, network and resource data. 2) Validation note: All numerical observations in this study are synthetically generated for methodological validation. They are not measurements obtained from a physical autonomous vehicle or production perception stack. Accordingly, the reported findings are limited to the synthetic study design and should not be interpreted as empirical evidence of real-world AV performance. 3) Validation note: All numerical observations in this study are synthetically generated for methodological validation and are not measurements from a physical autonomous vehicle.

References

[1] I. Gog, S. Kalra, P. Schafhalter, M. A. Wright, J. E. Gonzalez, and I. Stoica, “Pylot: A Modular Platform for Exploring Latency-Accuracy Tradeoffs in Autonomous Vehicles,” in Proc. IEEE International Conference on Robotics and Automation (ICRA), 2021, pp. 8806–8813. DOI: 10.1109/ICRA48506.2021.9561747. [2] A. Geiger, P. Lenz, and R. Urtasun, “Are We Ready for Autonomous Driving? The KITTI Vision Benchmark Suite,” in Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012, pp. 3354–3361. [3] M. Cordts et al., “The Cityscapes Dataset for Semantic Urban Scene Understanding,” in Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 3213–3223. [4] A. Dosovitskiy et al., “CARLA: An Open Urban Driving Simulator,” in Proc. Conference on Robot Learning (CoRL), 2017, pp. 1–16. [5] Z. Zheng, H. Ma, M. R. Lyu, and I. King, “WS-DREAM: A Distributed Reliability Assessment Mechanism for Web Services,” in Proc. IEEE/IFIP International Conference on Dependable Systems and Networks (DSN), 2010. [6] M. Frei, R. German, and A. Djanatliev, “An Edge Resource Selection Scheme Considering QoS and Computational Parameters,” in Proc. International Conference on Computer Communications and Networks (ICCCN), 2024. DOI: 10.1109/ICCCN61486.2024.10637567. [7] S. Sarkar, A. Trivedi, R. Bansal, and A. Sahu, “QoS Aware Mixed-Criticality Task Scheduling in Vehicular Edge Cloud System,” arXiv:2407.14793, 2024. [8] K. Maruvada, H. H. Hassan, A. Kattepur, and G. Bouloukakis, “Adaptive Model Selection using Meta Models and Drift Adaptation,” in Proc. IEEE COMSNETS, 2026, pp. 1361–1366. DOI: 10.1109/COMSNETS67989.2026.11417773. [9] R. Farahani, Z. Azimi, M. Colosi, and S. Dustdar, “LMEdge: QoS-Aware LLM Inference Orchestration on Edge Clusters,” arXiv:2607.17175, 2026. [10] S. Paruthi, “Benchmarking Dataset for Latency-Accuracy Tradeoffs in Autonomous Vehicle Systems,” M.Tech. Research Thesis, Indraprastha Institute of Information Technology Delhi, 2025. The thesis provides the direct benchmarking foundation used by the present research.

Copyright

Copyright © 2026 Sandeep Kumar Singh, Atul Kumar Mishra. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.