Deep Learning for Polar Codes over Flat Fading Channels

This paper proposes a deep-neural-networks scheme for decoding polar coded short packets. We consider packet transmission over frequency-flat quasi-static Rayleigh fading channels, where the channel coefficient is constant over a packet but changes packet-by-packet. Potential applications of the proposed technique are machine-type communications, messaging services, smart metering networks, and other wireless sensor networks requiring high reliability and low-latency. Computer simulations results confirm that even with simple codebook construction for an additive white Gaussian noise (AWGN) channel without fading, the proposed technique closes to the theoretical outage and achieves the coding gain in fading channel. Analyses of the learning epochs and training signal-to-noise power ratio (SNR) selections are also presented to demonstrate the effectiveness of the technique.

I. Introduction

Achieving ultra-reliable and low-latency communication (URLLC) is one of the major challenges facing future wireless communications [1]. One way to minimize the latency is by using short packets (i.e., codes with size in the range of 50 to 1000 bits However, it causes a severe degradation in channel coding gain. On the other hand, ensuring reliability usually require more resources, e.g., using strong channel codes with more redundancy or adding retransmission techniques, which increases the latency.