Transfer Learning for Channel Quality Prediction

The ability to predict the quality of a wireless channel is fundamental to enable anticipatory networking tasks. Traditional channel quality prediction (CQP) problems encompass predicting future conditions based on past samples of the same channel quality. In this paper we study the CQP problem across different wireless channels. To this extent, we consider a reference scenario including multiple 4G cells each of which operates multiple concurrent frequency carriers. We propose a framework based on transfer learning to predict the channel quality of a given frequency carrier when null or minimal information is available on the very same frequency carrier. For the transfer learning task we use convolutional neural networks (CNNs) and long short term memory networks (LSTMs) and compare their performance against traditional statistical methods on a dataset collected from a commercial 4G mobile radio network. The performance evaluation carried out on the reference dataset demonstrates the validity of the proposed approach which outperforms classical machine learning methods in all the tested conditions achieving an absolute 3% of error on average.