Co-operative Edge Intelligence for C-V2X Communication using Federated Reinforcement Learning

This paper examines the application of federated reinforcement learning (FRL) to enable resource-constrained vehicular edge nodes to learn their communication parameters from a central parameter server (PS). In cellular vehicle-to-everything communication (C-V2X), non independently-and-identically-distributed (non-i.i.d.) data samples impose additional communication requirements and increase the training time for model convergence. By exploring correlations between local model updates and the global model aggregation distributions, we accelerate this convergence using FRL. In the proposed method, Q-values undergo weight adaptation at each training round to update the global model. Local gradient vectors at vehicles and global gradient vectors at the PS measure the contribution of vehicle local models. Furthermore, the Q-values are quantified via nonlinear mapping that reinforces positive rewards, leading to dynamic measurements of local model contributions. Using FRL, policy-based and value-based learning methods reduce the number of communication rounds by upto 40%.