Architectural Design and Performance Trade-offs of Deep Learning-Based Autoencoders for Wireless Communication Systems | ||
| JES. Journal of Engineering Sciences | ||
| Articles in Press, Accepted Manuscript, Available Online from 17 November 2025 | ||
| Document Type: Research Paper | ||
| DOI: 10.21608/jesaun.2025.439559.1835 | ||
| Authors | ||
| Mohammed Abo-Zahhad1; Eman Ismail* 2; Abdelhay Ali2; Osama O Elnahas3 | ||
| 11-Department of Electronics and Communications Engineering, ,Egypt-Japan University of Science and Technology, E-JUST, Alexandra, Egypt. 2-Department of Electrical Engineering, Faculty of Engineering, Assuit University, Assuit, Egypt. | ||
| 2Department of Electrical Engineering, Faculty of Engineering, Assuit University, Assuit, Egypt | ||
| 31-Department of Electrical Engineering, Faculty of Engineering, Assuit University, Assuit, Egypt. 2-Department of Information Technology, Borg Alarab Technological University, Alexandra, Egypt. | ||
| Abstract | ||
| Deep learning-based autoencoders have shown significant potential in wireless communication systems by enabling the joint optimization of modulation and coding schemes. This study provides an analysis of the architectural design of deep learning autoencoders in such systems, emphasizing the tradeoff between model complexity and block error rate (BLER) performance. Using 7,4 autoencoders as a case study, we investigate how varying the depth (number of layers) and width (neurons per layer) of both encoder and decoder networks impacts communication reliability over Additive White Gaussian Noise (AWGN) and Rayleigh fading channels. Our findings indicate that increasing the width of single-layer autoencoders markedly decreases BLER until a saturation threshold, after which additional complexity offers minimal improvement. In contrast, deeper multilayer networks yield poorer performance despite increased complexity. We also observe that the effects of signal-to-noise ratio (SNR) and codeword length on autoencoder efficiency vary. The Rayleigh fading channel notably degrades performance, underscoring the need for more resilient architecture. Based on these insights, we propose design guidelines that strike a balance between computational cost and performance for deployment in resource-constrained wireless environments. Future research will explore advanced architectures, including transformer-based masked autoencoders and attention mechanisms, to improve fading robustness and generalization. Additionally, hardware implementation and unsupervised learning methods for real-world datasets are recommended to advance the application of deep learning autoencoders in next-generation wireless communications. This work offers critical insights into achieving an optimal balance between complexity and performance, enabling efficient, reliable, and scalable wireless communication through the application of advanced deep learning techniques. | ||
| Keywords | ||
| Autoencoders; Architectural Optimization; Additive White Gaussian Noise (AWGN); Rayleigh Fading Channels; Computational complexity | ||
|
Statistics Article View: 4 |
||
