Learned Compress-and-Forward Relaying for Primitive Relay Channels

A learning-based compress-and-forward (CF) relaying scheme is proposed for the primitive relay channel, which integrates a task-oriented neural Wyner-Ziv compressor at the relay and a neural demodulator at the destination. The learned CF relaying strategy exhibits characteristics of the optimal asymptotic CF strategy, such as binning of the quantized indices at the relay, while its performance approaches the capacity of a primitive relay channel with Gaussian inputs.