Deployment-Aware Multi-Objective Optimization of Machine Learning Models for Predictive Congestion Control in Multi-Hop Vehicular Networks

Abstract

Multi-hop vehicular wireless networks are highly susceptible to channel congestion due to dynamic topologies, heterogeneous traffic densities, and contention-based medium access. This paper proposes a deployment-aware predictive congestion control framework for vehicular communications, where packet transmission decisions are guided by supervised machine learning models optimized under predictive and computational constraints. Its novelty lies in coupling packet-level congestion prediction with deployment-aware model selection, allowing transmission-control models to be selected according to predictive quality, measured native-inference latency, and serialized model footprint. A packet-level dataset was constructed from realistic vehicle trajectories generated with SUMO using road networks extracted from OpenStreetMap for three topologically diverse cities: Liverpool, Rio de Janeiro, and Houston. These trajectories support vehicular network evaluation under heterogeneous urban conditions, enabling the assessment of predictive transmission decisions across different road topologies. Several supervised learning models were considered, including linear classifiers, neural networks, and ensemble machine learning architectures such as Random Forest, Weighted Soft Voting, and Stacking. Hyperparameter tuning was formulated as a multi-objective optimization problem and solved using the NSGA-II evolutionary algorithm, jointly maximizing predictive performance while minimizing inference latency and serialized model size. Pareto-efficient models were exported to ONNX format and executed through a native inference pipeline to evaluate their suitability for computationally constrained vehicular communication environments. Results across heterogeneous urban scenarios show that the proposed framework improves vehicular communication performance in terms of packet delivery reliability and communication delay, while preserving competitive machine learning performance under deployment-oriented constraints.