Dataset for the manuscript of Analysis on constructing the training data to train neural networks for channel estimation

Current neural network solutions for channel estimation are frequently tested by training and testing on one example channel or similar channels. However, data-driven algorithms often degrade significantly on other channels which they are not trained on, because they cannot extrapolate their training knowledge. Online training can fine-tune the offline-trained neural networks to compensate for this degradation, but its feasibility is challenged by the tremendous computational resources required. To solve this degradation, we propose design criteria to generate training datasets which will ensure neural networks robustly generalize to different channels. This design criteria also constrains the channels that the trained neural networks can generalize to. Based on the proposed criteria, we further propose a benchmark design to provide maximum generalization. In this way, the offline-trained neural networks still predict the complete channel matrix to achieve both the frequency and time interpolation, even on these channels which they are not trained on. To show the general applicability, neural networks with different levels of complexity are employed to demonstrate the generalization achieved. From the simulations, the trained neural networks achieve robust generalization to wireless channels with both fixed channel profiles and variable delay spreads. We also investigate the noise effect of the training dataset and the system scalability. This paper indicates one step towards artificial general intelligence (AGI) from new perspectives.