Machine Learning Based Threshold Tuning for Optimal Massive MIMO Performance

Sammanfattning

This thesis explores how machine learning can be applied to dynamically set the optimal threshold that determines if a user should be granted Sounding Reference Signals (SRS) resources at cell level. The goal of the thesis is to investigate whether this approach can improve the downlink performance on cell level compared to a static set threshold. Collected network measurements were used to train and evaluate different machine learning models for estimating spectral efficiency across different threshold configurations. Among the two evaluated approaches, XGBoost demonstrated the highest performance. The analysis revealed that features related to long-term network behavior, including the mean User Equipment (UE) SRS Signal-to-Noise Ratio (SINR), the mean path loss of the uplink and the aggregated throughput, contributed the most to the predictions, while short-term variability, the variance of the same parameters, showed less influence. The proposed adaptive threshold selection approach performed better than fixed threshold configurations in 48.3% of the evaluated cases, with a net value of around 0.15 bps/Hz, suggesting that dynamic optimization can improve network performance. At the same time, the thesis identified challenges associated with data quality, overlap between threshold configurations, and disproportion of the data set, which limited the stability and generalization of the model. The findings indicate that machine learning based dynamic threshold optimization is a promising direction for wireless network management and motivates further research on improved data collection and model development.