Fairness-Aware DRL-Based Power Allocation for Parallel FSO/RF-NOMA in HAP-Assisted Networks

High-altitude platform (HAP)-based networks have emerged as a practical solution to complement satellite systems by providing flexible and scalable connectivity. In this paper, we present an HAP-assisted network that integrates parallel free-space optical (FSO) and radio frequency (RF) transmissions with non-orthogonal multiple access (NOMA) to support communication in both dense and sparsely populated environments. FSO links are used to deliver high-capacity backhaul between the HAP and multiple access points (APs). RF transmission supports user access, where local users are served via the APs and remote users are directly connected to the HAP. To address fairness among users experiencing heterogeneous channel conditions, the transmit power allocation problem is formulated as a min–max outage probability optimization. This problem is inherently non-convex due to inter-user interference and the stochastic nature of the wireless channels. To overcome this challenge, a deep reinforcement learning (DRL)-based approach is developed, where multiple actor networks and a centralized critic are jointly employed to learn efficient power allocation policies. The proposed approach enables near-optimized performance while maintaining low computational complexity during online operation. Numerical results show that the proposed DRL method closely tracks successive convex approximation (SCA)-based optimization and differential evolution (DE)-based global search, while outperforming equal power allocation and gradient ascent. The results further confirm the robustness of the proposed framework under different transmit-power levels, RF fading conditions, decoding thresholds, HAP zenith angles, network sizes, and atmospheric-severity regimes.