LEARNING-BASED TWO-WAY FEEDBACK CHANNEL CODING: A COMPARATIVE ANALYSIS OF NEURAL NETWORK ARCHITECTURES

Abstract

The integration of deep learning techniques with classical channel coding theory has ushered in a new generation of communication systems capable of end-to-end optimisation under realistic channel conditions. This article presents a systematic comparative investigation of five distinct neural network paradigms—convolutional neural network (CNN) encoder-decoder pairs, long short-term memory (LSTM) recurrent architectures, attention-based transformer-inspired models, turbo autoencoders, and a novel hybrid CNN-LSTM framework—applied to the two-way feedback channel coding problem. By leveraging backward channel feedback, encoder representations are iteratively refined across multiple transmission rounds, significantly improving bit error rate (BER) performance beyond the single-shot, open-loop regime. Simulations conducted over additive white Gaussian noise (AWGN) and Rayleigh fading channels demonstrate that the proposed hybrid architecture achieves a coding gain of approximately 5.3 dB relative to an uncoded baseline at a BER of 10⁻´, while exhibiting favourable trade-offs between computational latency and error-correction capability. These results establish the viability of learning-based feedback coding as a foundation for sixth-generation (6G) physical-layer design.