Ml-Based Codebook-Free CSI Feedback: Feature, Architecture, and Loss Design

Abstract

This paper introduces a novel codebook-free machine learning approach for enhancing channel state information (CSI) feedback in 5G and beyond networks. The proposed solution leverages a neural network with a streamlined yet effective architecture, achieving high compression ratios while maintaining minimal reconstruction losses. The model's performance is rigorously evaluated in a realistic urban environment, through ray tracing data, demonstrating its robustness against uncertainties even at high compression levels. Notably, the study highlights the advantages of using the time-domain CSI over frequency-domain data as input features, while also highlighting the importance of computing loss and assessing the performance in the frequencydomain. Comparative analysis reveals the proposed model's superiority over existing state-of-the-art (SotA) models, achieving less than 0.5 % CSI reconstruction error at 256 -fold compression, effectively underscoring its potential for practical deployment in next-generation wireless networks. This work represents a significant step forward in the development of efficient and reliable CSI feedback mechanisms, paving the way for more resilient and high-performance communication systems.