Parameter-Efficient Online Fine-Tuning of ML-Based Hybrid Beamforming With LoRA

We propose low-rank adaptation (LoRA) for machine learning-aided hybrid beamforming (HBF) in episodically dynamic millimeter-wave multiple-input multiple-output (MIMO) systems. This approach introduces low-rank trainable matrices and uses a small buffer with recent channel samples, making it ideal for real-time adjustments. Evaluated for a large MIMO HBF system across both an environment-specific channel using ray tracing and clustered delay line channel models, simulation results show that rank-2 LoRA achieves efficient retraining with only 6% of the original network’s parameters and 128 samples, improving average achievable information rate (AIR) by over 45% compared to the pre-trained model in both scenarios. The method significantly outperforms transfer-learning with full-model online fine-tuning and model-agnostic meta-learning with its “almost-no-inner-loop” variant.