Generative Decoding of Compressed CSI for MIMO Precoding Design

TL;DR

This work proposes an ML-based, decoder-only solution for compressed CSI that reduces the data collection overhead by using a site-specific digital twin to generate synthetic CSI data for training and introduces two training schemes for the generative decoder.

Abstract

Massive MIMO systems can enhance spectral and energy efficiency, but they require accurate channel state information (CSI), which becomes costly as the number of antennas increases. While machine learning (ML) autoencoders show promise for CSI reconstruction and reducing feedback overhead, they introduce new challenges with standardization, interoperability, and backward compatibility. Also, the significant data collection needed for training makes real-world deployment difficult. To overcome these drawbacks, we propose an ML-based, decoder-only solution for compressed CSI. Our approach uses a standardized encoder for CSI compression on the user side and a site-specific generative decoder at the base station to refine the compressed CSI using environmental knowledge. We introduce two training schemes for the generative decoder: An end-to-end method and a two-stage method, both utilizing a goal-oriented loss function. Furthermore, we reduce the data collection overhead by using a site-specific digital twin to generate synthetic CSI data for training. Our simulations highlight the effectiveness of this solution across various feedback overhead regimes.

Figures (4)

• Figure 1: This figure presents the key ideas of the proposed generative decoder solution. The estimated channel is compressed using a standardized encoder on the UE, and a generative decoder refines the reported CSI at the BS. A site-specific digital twin is also leveraged to generate synthetic data for offline training, and the trained generative decoder can be deployed in the real-world system with high performance.
• Figure 2: This figure presents a bird's-eye view of the adopted Boston scenario. The BS is oriented toward the negative y-axis. The service area, represented by the user grid, is primarily a non-line-of-sight (NLoS) region
• Figure 3: This figure shows the performance of the generative decoder. The results show its effectiveness by comparing it against codebook-based CSI feedback and a reconstruction loss. The results also emphasize the importance of a site-specific dataset, as shown by the comparison with generic datasets.
• Figure 4: The figure presents results for the two-stage generative decoder, which was trained with a goal-oriented loss function on the digital twin dataset and tested in the target scenario. Despite performance degradation from the digital twin's imperfections, our proposed approach still shows a significant gain over standard Type-II codebooks.