Machine learning for channel quality prediction: From concept to experimental validation

We focus on prediction of channel quality between any two devices using Deep Neural Network (DNN) from information already known to mobile networks. The DNN-based prediction reduces a cost of a common pilot-based channel quality measurement in scenarios with many ad-hoc communicating devices. However, collecting a sufficient number of high-quality and well-distributed training samples in real-world is not feasible. Hence, in this paper, we develop and validate a concept of DNN-based channel quality prediction between any two devices based on a low-complexity and easy-to-create digital twin. The digital twin serves for a generation of a large synthetic training dataset for channel quality prediction. As the low-complexity digital twin cannot capture all real-world aspects of the channels, we enhance the digital twin with real-world measured and artificially augmented inputs via transfer learning. The proposed concept is implemented and validated in software defined mobile network. We demonstrate that the proposed concept predicts the channel quality with a very high accuracy (mean average error of only 0.66 dB) in a real-world complex indoor scenario. Such error is sufficient for practical applications of the developed channel quality prediction concept and the error is few times lower than the error achievable by state-of-the-art solutions.