Machine Learning-Based Predictive Medium Access Control Protocols for Vehicular Networks

Abstract

The design of robust Medium Access Control (MAC) protocols for vehicular networks remains a critical challenge due to dynamic topologies, intermittent channel conditions, and stringent latency requirements. This paper presents a machine learning-based predictive MAC framework that leverages historical and real-time channel state information to optimize contention window adaptation, priority scheduling, and collision avoidance. A hybrid architecture integrating recurrent neural networks (RNNs) with reinforcement learning (RL) agents is developed to predict temporal traffic patterns and dynamically adjust channel access parameters. The RNN module processes time-series data from vehicular nodes to forecast short-term network congestion, while the RL agent optimizes transmission policies through a discounted reward mechanism based on collision probability and throughput maximization. Extensive simulations under urban and highway scenarios demonstrate a 27% reduction in end-to-end latency and 33% improvement in packet delivery ratio compared to IEEE 802.11p and conventional CSMA/CA. However, the model exhibits a 15–18% performance degradation in non-line- of-sight (NLOS) environments with multipath fading, attributed to imperfect channel estimation. Further analysis reveals quadratic computational complexity relative to node density, limiting scalability beyond 150 nodes per coverage area. The proposed framework is shown to achieve a Nash equilibrium in transmission scheduling under stationary traffic, though convergence time increases exponentially with velocity variance. These results highlight the potential of cross- layer integration of machine learning into MAC protocols while underscoring the need for distributed computation to address real-time constraints.