End-to-End Deep Learning in Wireless Communication Systems: A Tutorial Review

Abstract

The physical layer (PHY) in wireless communication systems has traditionally relied on model-based methods that are often optimized individually as independent blocks to perform tasks such as modulation, coding, and channel estimation. However, these approaches face challenges when it comes to capturing real-world nonlinearities, hardware imperfections, and increasing complexity in modern networks. This paper surveys advancements in applying deep learning (DL) for end-to-end PHY optimization by incorporating the autoencoder (AE) model as a powerful end-to-end DL framework to enable joint transmitter and receiver optimization and address challenges like dynamic channel conditions and scalability. We review cutting-edge DL models; their applications in PHY tasks such as modulation, error correction, and channel estimation; and their deployment in real-world scenarios, including point-to-point communication, multiple access, and interference channels. This work highlights the benefits of learning-based approaches over traditional methods, offering a comprehensive resource for researchers and engineers looking to innovate in next-generation wireless systems. Key insights and future directions are discussed to bridge the gap between theory and practical implementation.

Figures (19)

• Figure 1: Communication chain in a model-based system.
• Figure 2: Structure of an end-to-end DL based commmunications system modeled as an autoencoder.
• Figure 3: Diagramatic representation of section structure.
• Figure 4: Diagram of basic autoencoder structure
• Figure 5: Illustration of different AE types, including a Turbo Autoencoder, in which parallel concatenation and interleaving are implicitly realized through neural encoder blocks, as well as the proposed Autoencoder introduced in our prior work mode22.