Transfer learning for 5G communication scenarios under different mobile speed

Abstract

The cellular network has become a cutting-edge topic due to the rapid advance of intelligent applications and the increasing user demand. With the help of machine learning, the result of message recovery could be acceptable when the wireless channel does not widely change. Compared with traditional machine learning, domain adaptation, a case in transfer learning, can be a more appropriate approach to reduce the influence caused by the Doppler effect and multi-path propagation to increase the reliability of communication. In this case, although the source data and target data might not be independent and identically distributed because of the changes in communication scenarios, a model could be implemented to recover the received signals with a low symbol-error rate (SER), and the training time and the computation complexity can be reduced to satisfy the requirement of low processing duration compared with learning from scratch.