A Learning Approach to the Optimization of Massive MIMO Systems

This talk explores the use of deep learning for optimizing channel sensing and downlink precoding for both the time-domain duplex (TDD) and the frequency-domain duplex (FDD) massive MIMO systems. For the TDD system employing hybrid analog and digital beamforming, we show that the channel sensing and the downlink precoding matrices can be designed directly from the received pilots using deep learning without the intermediate channel estimation step. For the FDD cellular system with limited feedback, we further show that deep learning can be used for efficient and distributed channel estimation, quantization, feedback, and downlink multiuser precoding. The proposed methodology is generalizable, and requires significantly less training overhead than the conventional channel estimation based approaches.