Utilizing Unsupervised Learning for Improving ML Channel State Feedback in Cellular Networks
Bryse Flowers, Adarsh Sawant, Runxin Wang, Dustin Zhang · 2023
Recent studies have shown that Machine Learning (ML) techniques can be used to compress Channel State Feedback (CSF) in order to reduce overhead; however, it is unclear if a single ML model will be adequate to generalize across all cellular deployments. This work explores a potential enhancement to ML CSF by studying whether multiple ML CSF models can achieve superior performance over a single model (with a fixed model size and architecture). The proposed approach uses unsupervised learning to partition wireless channels before training ML CSF models and is shown to provide a performance improvement, measured through end-to-end Normalized Mean Squared Error (NMSE) of the reconstructed CSF, in every investigated scenario for ML CSF and achieves a 1.9dB better overall NMSE than using a single ML CSF model. A separate potential approach for utilizing multiple ML CSF models is evaluated for reference by directly using channel model labels to partition the wireless channel before training ML CSF models. It is shown that this reference approach does not always lead to a performance improvement, highlighting the need for incorporating unsupervised learning in the creation of multiple ML CSF models.