Adaptive Spectrum Allocation for Low Latency Vehicular Networks Using Deep Reinforcement Learning

Abstract

The impending deployment of autonomous vehicles has added new requirements about wireless communication in highly dynamic vehicular networks, particularly in terms of reliability and low latency. The allocation of the spectrum is still a difficult problem, as traffic and mobility conditions change continuously, which implies the variation in vehicle density, channel quality, interference, and communication requirements. These variations are not easily handled by the conventional static, rule-based and non-adaptive allocation schemes, affecting the efficient use of the spectrum, communication latency, and packet delivery reliability. To overcome these challenges, this study introduces an Adaptive Deep Reinforcement Learning (ADRL) based dynamic spectrum allocation framework for AVNs. The proposed framework uses a Deep Q-Network (DQN) that, based on observation of the current conditions of the vehicular network (vehicle density, channel quality, level of interference, load of the traffic flow), automatically determines the appropriate spectrum resources. A simulation-based vehicular communication scenario is created to evaluate under various traffic densities and channel-quality levels with dynamic mobility and channel conditions. Experimental analysis takes into account the following factors: spectral efficiency, spectrum utilization, communication latency, packet delivery ratio, reliability, and learning convergence. Results show that the proposed ADRL framework can ensure efficient spectrum allocation and reliable communication in a fast-growing network density and degraded channel environment and has stable convergence characteristics in its training behavior. The results show that adaptive deep reinforcement learning is an effective method to spectrum management in the next generation of autonomous vehicular communication networks that is robust, resource-efficient, and has low latency.