Abstract
Predicting fading channels is a classical problem with a vast array of applications, including as an enabler of artificial intelligence (AI)-based proactive resource allocation for cellular networks. Under the assumption that the fading channel follows a stationary complex Gaussian process, as for Rayleigh and Rician fading models, the optimal predictor is linear, and it can be directly computed from the Doppler spectrum via standard linear minimum mean squared error (LMMSE) estimation. However, in practice, the Doppler spectrum is unknown, and the predictor has only access to a limited time series of estimated channels. This paper proposes to leverage meta-learning in order to mitigate the requirements in terms of training data for channel fading prediction. Specifically, it first develops an offline low-complexity solution based on linear filtering via a meta-trained quadratic regularization. Then, an online method is proposed based on gradient descent and equilibrium propagation (EP). Numerical results demonstrate the advantages of the proposed approach, showing its capacity to approach the genie-aided LMMSE solution with a small number of training data points.
Introduction
In wireless communications, "fading" refers to the rapid fluctuations in signal strength caused by the environment. To maintain high-quality connections, transmitters must predict these changes—Channel State Information (CSI)—in advance. The optimal solution is the Linear Minimum Mean Squared Error (LMMSE) estimator, but it requires perfect knowledge of the channel's Doppler spectrum, which is rarely available in practice. While deep learning can adapt to these statistics, it typically requires massive amounts of "pilot" data, which consumes valuable bandwidth. This work aims to bridge this gap by achieving the performance of optimal linear filters without prior statistical knowledge and with significantly reduced data requirements.
Methodology
The proposed framework leverages meta-learning to "learn how to learn" from diverse propagation environments. The goal is to learn a "bias vector" ($\bar{v}$) for Ridge Regression that serves as an optimized starting point for any new environment.
Offline Meta-Learning
The authors derive a closed-form solution for optimal hyperparameters. Unlike iterative methods like MAML, which require gradient descent, this approach calculates the best average predictor across historical environments in a single step, offering a low-complexity solution for offline training.
Online Meta-Learning via Equilibrium Propagation
For dynamic environments, an online method uses Equilibrium Propagation (EP). EP, inspired by biological learning, efficiently estimates gradients for the "outer" meta-learning loop. This allows the predictor to update its internal bias in real-time as it encounters new data frames, effectively accumulating knowledge from a stream of data.
Results and Discussion
The effectiveness of the meta-learning approach is evaluated using synthetic Doppler models (Jakes’ and rounded models) and 5G standard models (3GPP Spatial Channel Model CDL-C).
Efficiency in Low-Data Regimes
A key finding is the "power of one sample": the meta-learned model can produce useful predictions with just a single training sample ($L^{new}=1$). In contrast, conventional learning methods fail in this regime.
As shown in the figure above, the meta-learning approach closely matches the "genie-aided" LMMSE (which has perfect statistical knowledge), while conventional learning requires dozens of samples to reach comparable accuracy.
Online Adaptation
The online EP-based method demonstrates a steady decrease in error as it processes more frames, proving its ability to adapt to changing environments.
The results indicate that while non-linear neural networks are powerful, well-designed linear filters often outperform them when data is scarce. Meta-learning a linear filter provides a robust middle ground, maintaining low Mean Squared Error (MSE) even with significant estimation noise.
Conclusion
This paper demonstrates that meta-learning can significantly reduce the data overhead for channel fading prediction. By exploiting the underlying structure shared across different Doppler spectra, the proposed methods approach optimal performance with minimal training. Future work will address extensions to wideband frequency-selective fading, Multiple-Input Multiple-Output (MIMO) systems, and scenarios with non-stationary environments where the meta-learned bias may need more frequent updates.
Summary updated