Reinforcement learning-empowered resource allocation with multi-head attention mechanism in V2X networks

Abstract

Intelligent transport systems (ITS) offer safe and autonomous service in vehicular applications. The vehicle to everything (V2X) network aids in performing communication between any vehicle to other entities such as networks, pedestrians or other objects. However, the allocation of power in the V2X network is still seen as a challenging task in recent resource allocation approaches. So, multi-head attention mechanism with reinforcement learning (MHAMRL) is utilized in resource allocation. This work considers real traffic scenes in highway traffic model and wireless transmission model. Specifically, in the mode 4 cellular V2X, every individual vehicle is considered as a resource which does not rely on the base station for resource allocation. Vehicle users are classified into V2I or V2V links based on the varied service requirements of V2X. The combination of multi-head attention mechanism sequences the signal with minimal noises which diminishes the energy consumption and improves channel gain. In the velocity range of 20-25 m/s, the proposed approach achieves a sum rate of 53 Mb/s, surpassing the 50 Mb/s achieved by the existing multi-agent deep reinforcement learning-based attention mechanism (AMARL) algorithm.