Autoencoder-bank based design for adaptive channel-blind robust transmission

Abstract The idea of employing deep autoencoders (AEs) has been recently proposed to capture the end-to-end performance in physical layer communication systems. However, most current methods for applying AEs are developed based on assumption that there exists an explicit channel model training matches actual online transmission. variation indeed imposes a major limitation AE-based In this paper, without relying model, we propose adaptive scheme increase reliability system over different conditions. Specifically, partition coefficient values into sub-intervals, train AE each offline phase, and constitute bank AEs. Then, condition phase average block error rate (BLER), optimal pair encoder decoder is selected data To gain knowledge about conditions, assume realistic scenario which instantaneous not known, blindly estimate it at Rx, i.e., any pilot symbols. Our simulation results confirm superiority existing terms power consumption. For instance, when target BLER equal $$10^{-4}$$ 10 - 4 , our algorithm with 5 pairs can achieve 1.2 dB compared non-adaptive scheme.