Loss Function Fusion for Learning-Based CSI Extrapolation Enhancement

To reduce the overhead of channel state information (CSI) measurement and feedback, the extrapolation of the unmeasured CSI has drawn much attention, where deep learning is utilized to model the complex CSI correlations for improving the extrapolation accuracy. In this letter, we analyze the inconsistency of the gradient directions between minimizing the extrapolation error and maximizing the system rate. To address this issue, the linearly combined loss function of the extrapolation error and system rate is proposed to optimize the extrapolation model. Considering various dynamic ranges, the tailored successive halving algorithm is designed to efficiently search a near-optimal combination coefficient for the fused loss function. Simulation results show that our proposed loss function could achieve higher system rate performance over its conventional counterparts.