Deep learning for noncoherent OTFS modulation

In this paper, we consider an emerging modulation scheme named as orthogonal time frequency space (OTFS) modulation to effectively suffer from high-mobility time-varying channels. Thus, it has ability to overcome inter-carrier interference in the orthogonal frequency division multiplexing (OFDM) system. The classical OTFS requires accurate channel state information (CSI) to exactly detect the signal. However, when the channels are high-mobility, it is challenging to estimate CSI perfectly. This paper proposes a novel deep learning-aided noncoherent autoencoder OTFS (NAE-OTFS) system which models both the transmitter and receiver as a deep neural network encoder and decoder of an AE architecture. This design enables a joint optimization of transmitter and receiver via an end-to-end training procedure. By doing so, our proposed NAE-OTFS can detect data bits without the CSI estimation requirement at the transmitter as well as receiver. Besides, the proposed scheme fully exploits multi-path diversity to improve the detection performance. Simulation results show that our proposal achieves a superior BER performance over the baselines which are unable to harness the multi-path diversity. Moreover, our scheme offers lower training overhead than learning OFDM-based baselines, since it is trained only with a single training signal to noise ratio (SNR) while still performs well in a range of other testing SNRs.