Deep Reinforcement Learning with Multi Head Attention Mechanism for Resource Allocation in Platoon based C-V2X Networks

Abstract

The rapid expansion of wireless services designed for local area communication has undergone the necessity for effective resource allocation within cellular vehicle-to-everything (C-V2X) networks. These networks, which operate using fifth-generation (5G) technology, bring about improved system performance by facilitating peer-to-peer communication and resource sharing among nearby devices and mobile users. This paper presents an innovative strategy known as Dynamic Resource Reservation with Deep Reinforcement Learning (DR2-DRL), which employs a multi-agent deep reinforcement learning frame-work to tackle the challenge of packet collisions in vehicular networks. The primary objective of DR2-DRL is to empower vehicles in making intelligent choices while opting for radio resources. To enhance the efficiency of training, an optimized attention-based mechanism is integrated, enabling vehicles to selectively concentrate on pertinent information obtained from observations and actions of neighboring vehicles. Remarkably, this algorithm is particularly well-suited for C-V2X, as it enables independent resource selection without relying on global information. Extensive simulations illustrate that DR2-DRL outperforms alternative decentralized approaches, underscoring its ability to scale and remain robust within dynamic vehicular networks.