Downlink Beamforming Prediction in MISO System Using Meta Learning and Unsupervised Learning

Various machine learning methods have been applied in downlink beamforming to yield performance gain. However, a prevalent limitation among these approaches is their reliance on substantial quantities of labeled data with high training complexity, especially when the number of users and antenna increases. In this paper, we propose two methodologies designed to mitigate these limitations. Our methods depart from conventional practices by training models using unlabeled data while concurrently curtailing the data volume required for effective training. Moreover, to reduce training complexity, we adopt an approach, based on the WMMSE and MLBF algorithm, that decomposes the beamforming vector prediction into lower-dimensional components. To enhance the adaptivity, we incorporate Rayleigh and Rician channels during the training. And we conducted experiments to train models under different amounts of data and assess model performance under conditions where testing distributions align with or diverge from the training distribution. Both methodologies show better performances than WMMSE at high SNR under the same distribution. Furthermore, models trained through unsupervised learning showcase heightened generality when confronted with novel environments characterized by divergent data distributions. Additionally, our meta learning approach yields commendable performance even with limited data availability.