Neural Network Optimal UW-OFDM

We propose a novel end-to-end learning scheme for wireless communication systems employing the unique word (UW)-OFDM signaling scheme. The work is motivated by the recent advances of machine learning in channel equalization and data estimation. Our idea is to design a non-systematically encoded UW-OFDM system optimal for neural network (NN) based estimators. To this order, we introduce model-based neural network architectures that optimize the transmitter and receiver sides, i.e. the UW-OFDM symbol generation and the NN data estimation together for minimal bit error ratio (BER). The proposed model is evaluated in a simulation environment, and compared with NN-based and traditional estimators.