A Comprehensive Review of Resource Allocation via Sparsity-Aware Orthogonal Initialization of Deep Neural Networks in Free Space Optical Communications

Abstract

Free Space Optical (FSO) communication has emerged as a promising high-speed wireless communication technology due to its large bandwidth, low deployment cost, and high security. However, atmospheric turbulence, beam divergence, and channel fading significantly degrade system performance, necessitating efficient resource allocation strategies. Traditional optimization techniques often struggle with the non-linear and dynamic nature of FSO channels. Recent advancements in deep learning have introduced intelligent approaches for solving resource allocation problems in FSO systems. In particular, sparsity-aware orthogonal initialization (SAO) of deep neural networks has gained attention for improving training efficiency and reducing computational complexity. SAO enables the construction of sparse yet maximally connected neural networks using orthogonal weight initialization, ensuring better convergence and stability during training. Additionally, deep learning-based resource allocation methods, such as primal-dual learning frameworks, have demonstrated superior performance in handling stochastic optimization problems in FSO communications, including power allocation and relay selection. This review provides a comprehensive analysis of recent methods and architectures integrating SAO-based neural networks and deep learning techniques for efficient resource allocation in FSO systems. It highlights key advancements, challenges, and future research directions, emphasizing the role of sparse neural architectures in achieving energy-efficient and high-performance optical communication systems.