Machine learning-based Precoder Type Selection in Massive MIMO Networks

Sammanfattning

Background. Massive MIMO underpins modern 4G/5G networks by boosting spectral efficiency through beamforming. Precoding can follow either a codebook or a reciprocity principle, and the superior choice depends on the instantaneous radio context. Objective. This thesis investigates whether machine learning can learn the context and select the more efficient precoder type from production-network data. Methods. Precoder selection is cast as a two-model regression problem. One Random-Forest regressor is trained on transmissions that used codebook precoding, and a second regressor on those that used reciprocity. At run-time each HARQ process is evaluated by both models; the scheme with the higher predicted spectral efficiency is chosen. Performance is estimated offline on historical data. Results. Trained on 400,000 HARQ processes, the two models achieve low cross-validated and test nRMSE and unbiased residuals. The learned selector overrides the legacy rule in 52.5% of transmissions and increases aggregate efficiency by +6.4 %. On the switched cases the mean uplift is +16.7%; the 95th percentile reaches +44.2%. Feature-importance analysis matches domain intuition. Conclusions. While most transmissions change little, the selector consistently finds a large subset where an alternative precoder yields substantial gains. The approach is data-driven, interpretable and lightweight enough for real-time 5G deployment, though live trials are advised to validate causal uplift and tune safety margins.