An Enhanced Loss Function for 6G Space-Time Domain Predictive Channel Models

Facing the sixth-generation (6G) technical vision, the machine learning (ML) based channel modeling has drawn a lot of attention. The prediction accuracy of ML-based channel modeling methods depends on the dataset size, model parameter settings, and loss functions. The design of loss function is critical to improve the accuracy of predictive channel models. This paper proposes a novel space-time-aware weighted loss function that reflects the correlation between space-time domain channel characteristics. The proposed loss function is compared to widely used loss functions, e.g., mean square error (MSE), Huber loss, and log-cosh loss. Based on the channel measurements, the performance of the proposed loss function is evaluated. According to the experiment results, the predictive channel model with the proposed loss function achieves the highest accuracy compared to the channel model with conventional loss functions. The proposed loss function can be used to design more accurate ML-based predictive channel models for 6G wireless channels.