PAPR Reduction Scheme for Deep Learning-Based Communication Systems Using Autoencoders

Melika Vahdat, Koosha Pourtahmasi Roshandeh, Masoud Ardakani, Hai Jiang

Abstract: Deep neural networks (DNN) have gained considerable attention in the communication literature during the past few years. In particular, as a well-known DNN architecture, autoencoders (AE) are used to model the end-to-end communication systems achieving a reasonable performance in terms of block error rate (BLER). However, autoencoders significantly suffer from high peak-to-average-power-ratio (PAPR), resulting in power amplifier saturation. This paper proposes a novel DNN architecture for reducing PAPR in autoencoder-based communication systems. Simulation results verify that the proposed scheme outperforms the conventional PAPR reduction method, i.e., loss function-based PAPR reduction approach, in terms of both bit error rate (BER) and PAPR.