Learning Latent Wireless Dynamics from Channel State Information

TL;DR

The work addresses predicting wireless propagation dynamics by learning latent representations of CSI through a channel-chart-inspired embedding and a velocity-conditioned predictor, yielding a Joint-Embedding Predictive Architecture (JEPA) that forecasts future latent states without reconstructing full channels. The method combines a context encoder, a target encoder updated via EMA, and an autoregressive predictor trained with a latent $L_2$ loss, augmented by a two-stage curriculum learning strategy. On the DICHASUS indoor dataset, JEPA achieves notable improvements in long-horizon prediction accuracy and channel-chart quality, with pre-training on a globally robust channel distance yielding the best results and robustness to moderate velocity-noise. This approach offers a data-efficient, self-supervised path to leverage latent dynamics for wireless control tasks, such as resource allocation and scheduling, in future networks.

Abstract

In this work, we propose a novel data-driven machine learning (ML) technique to model and predict the dynamics of the wireless propagation environment in latent space. Leveraging the idea of channel charting, which learns compressed representations of high-dimensional channel state information (CSI), we incorporate a predictive component to capture the dynamics of the wireless system. Hence, we jointly learn a channel encoder that maps the estimated CSI to an appropriate latent space, and a predictor that models the relationships between such representations. Accordingly, our problem boils down to training a joint-embedding predictive architecture (JEPA) that simulates the latent dynamics of a wireless network from CSI. We present numerical evaluations on measured data and show that the proposed JEPA displays a two-fold increase in accuracy over benchmarks, for longer look-ahead prediction tasks.

Figures (5)

• Figure 1: Common architectures for self-supervised learning.
• Figure 2: System model and our proposed method.
• Figure 3: Impact of pre-training the encoder on the learned channel charts.
• Figure 4: The user's dynamics as modelled by the predictor in the channel charts.
• Figure 5: Downstream task numerical results.