Knowledge-Empowered Distributed Learning Platform in Internet of Unmanned Aerial Agents to Support NR-V2X Communication

NR-V2X Mode 2 is introduced by the third generation partnership project (3GPP) to support vehicle-to-everything (V2X) communication. In NR-V2X Mode 2, vehicles select resources for the exchange of cooperative awareness messages (CAM) in a decentralized manner based on their local observation using semi-persistent scheduling. Resources are distributed over the 2-D frequency and time domain, following the long-term evolution frame structure. Since vehicles select resources based on their local observations and due to spectrum scarcity, this may lead to contention. Hence, selecting a resource is challenging, and as each vehicle strives to select a resource, it becomes a consensus problem. To resolve resource contention, in this article, we propose a knowledge-empowered distributed multiagent deep reinforcement learning (K-MADRL) approach. Based on traffic flow information, long short-term memory (LSTM) is employed to deploy Unmanned Internet of Aerial Agents (UIAAs) to collect vehicle state information. UIAAs gather vehicle state knowledge and train the local deep reinforcement learning (DRL) model. The locally trained model at the UIAA is shared and aggregated at the gNB for the global model update. The trained policy is then sent to the vehicles over system synchronization blocks for distributed execution. Moreover, the vehicles select the resource based on the joint action, i.e., by anticipating the actions of the neighboring vehicles. Our scheme is compared with other methods, such as DRL, optimization techniques, the SPS method, and random allocation methods, used in the NR-V2X environment. The results of the simulations show that our scheme outperforms the other methods.