AI-based CSI acquisition for very large antenna arrays with limited overhead in 6G

Purpose and goal

Downlink reference signal overhead will become a bottleneck for CSI acquisition based on UE feedback. In this thesis, methods using AI/ML techniques to reduce reference signal overhead for CSI acquisition with a large number of antennas will be investigated.

Expected effects and result

One or more AI/ML models will be designed for CSI acquisition based on partial measurement of a full channel. The proposed AI/ML model will be benchmarked against classical methods, such as compressed sensing. The performance will be evaluated in system level simulations. The AI/ML model is supposed to greatly reduce the DL reference signal overhead for CSI acquisition.

Planned approach and implementation

Training data will be generated using Ericsson´s in-house simulator, based on 3GPP standard channel models. Training will be performed on AWS. Performance evaluation will be conducted in Ericsson´s in-house system level simulator.