Resource Allocation via Model-Free Deep Learning in Free Space Optical Communications

This paper investigates the general problem of resource allocation for mitigating channel fading effects in Free Space Optical (FSO) communications. The is modeled as constrained stochastic optimization framework, which covers a variety FSO scenarios involving power adaptation, relay selection and their joint allocation. Under this we propose two algorithms that solve problems. We first present Stochastic Dual Gradient (SDG) algorithm shown to exactly by exploiting strong duality but whose implementation necessarily requires explicit accurate system models. As an alternative Primal-Dual Deep Learning (PDDL) based on SDG algorithm, parameterizes policy with Neural Networks (DNNs) optimizes latter via primal-dual method. parameterized incurs only small loss optimality due representational DNNs, can be moreover implemented without knowledge A wide set numerical experiments are performed corroborate proposed demonstrate superior performance computational efficiency compared baseline methods both continuous binary settings.