Multi-Agent Reinforcement Learning for Resource Management in V2X and Sparse Update Transmission with Residual Correction for Heterogeneous Federated Learning

Abstract

Resource management in NR-V2X platooning systems, focusing on the trade off between Age of Information (AoI), communication demand, and limited onboard energy. To address this, a novel Self-sufficient Energy-Aware Multi-agent Deep De terministic Policy Gradient (SEAM-DDPG) framework is proposed. The approach enables platoons to operate in a strategic sleep mode for energy conservation and in corporates a one-step forward model to facilitate proactive decision-making under the centralized training and decentralized execution paradigm. Numerical studies show that the proposed algorithm greatly enhances system sustainability while ensuring communication efficiency. Specifically, the proactive SEAM-DDPG increases average energy reserves by up to 94.59%, extends system lifetime by 42.83%, and reduces transmit power by 30.76% relative to existing techniques. In our other paper, we introduce ReSparseFL, a communication-efficient federated learning approach to tackle both the communication and statistical heterogeneity is sues. It combines structured sparse update transmission, in which clients transmit only the most important parts of their model updates, greatly reducing communica tion costs. To mitigate the distortion caused by sparsification, ReSparseFL employs a residual correction scheme, which collects and reintroduces information lost during sparsification over training iterations. The server aggregates these sparse updates using an index-based reconstruction strategy. Extensive evaluation across multiple datasets, model architectures, and varying degrees of data heterogeneity demonstrates that ReSparseFL achieves improved training stability and competitive accuracy while significantly lowering communication costs compared to standard federated learning approaches.