Speeding up Training of Linear Predictors for Multi-Antenna Frequency-Selective Channels via Meta-Learning.

Area of Science:

• Wireless Communication Engineering
• Signal Processing
• Machine Learning

Background:

• Efficient channel prediction is vital for multi-antenna frequency-selective channels, requiring minimal pilot symbols.
• Existing methods struggle with the limited data available in dynamic wireless environments.

Purpose of the Study:

• To develop novel data-driven channel prediction algorithms for multi-antenna systems.
• To integrate transfer learning and meta-learning with reduced-rank channel parametrization for faster adaptation.
• To improve prediction accuracy using fewer pilot symbols.

Main Methods:

• Proposed novel channel-prediction algorithms combining transfer and meta-learning with reduced-rank parametrization.
• Developed linear predictors optimizing past frame data for current frame adaptation.
• Introduced a novel long short-term decomposition (LSTD) for linear prediction models.
• Applied equilibrium propagation (EP) and alternating least squares (ALS) for LSTD-based prediction.

Main Results:

• Demonstrated the effectiveness of transfer and meta-learning in reducing pilot symbol requirements for channel prediction.
• Showcased the benefits of the proposed LSTD parametrization in 5G standard channel models.
• Achieved efficient channel prediction through data utilization from previous frames with distinct propagation characteristics.

Conclusions:

• Transfer and meta-learning significantly reduce pilot overhead in channel prediction for wireless systems.
• The proposed LSTD parametrization offers a promising approach for efficient and accurate channel prediction.
• The developed algorithms provide a robust data-driven strategy for next-generation wireless communication systems.