SU-MIMO Port Selection Using Convolutional Neural Networks

Sammanfattning

Background: The exponential increase in user equipment (UE) units within mobile networks necessitates more efficient Massive MIMOalgorithms. To address this demand, integrating artificial intelligence (AI) into various network aspects is gaining traction. Goal: This thesis explores the feasibility of employing a lightweight convolutional neural network (CNN) to optimize port selection in single-usermultiple-input multiple-output (SU-MIMO) networks. Port selection, a critical component of all forms of MIMO networks, determines theoptimal ports on a UE for data transmission. The objective is to enhance selection speed, reduce computational complexity,and minimize memory consumption. Method: The methodology involves a quasi-experiment where a CNN model, trained on data transfer logs between a basestation and a UE, specifically a mobile phone, is compared with a self-implemented version of the port selection algorithm utilised in Ericssonbase stations. The evaluation criteria include time-, computational-, and spatial complexity. The accuracy of the port selection capabilities of themodels is also recorded. Results: Despite the complexity of the CNN models, the results indicate subpar performance and low test accuracies.This suggests that achieving satisfactory performance would either necessitate an increased model complexity and size or that a convolutionalneural network is not the correct choice for replacing the algorithm. Conclusion: In conclusion, the thesis finds that a lightweight CNN may not be the optimal solution for port selectionoptimization in SU-MIMO networks. However, it suggests potential avenues for further research to explore alternative approaches to this task.