Statistical analysis of adaptive rate and power control in urban FSO drone networks using RL-DQN-A2C

This work introduces a new Reinforcement Learning (RL) framework for optimizing Free-Space Optical (FSO) communication systems in drone-assisted smart city networks by combining Deep Q-Network (DQN) and Advantage Actor-Critic (A2C) techniques. With the help of real-time environmental feedback, the suggested hybrid RL-DQN-A2C model dynamically adjusts three important transmission parameters in real time, beam divergence, transmit power, and modulation order. This makes it possible to adapt robustly to extremely dynamic and unpredictable atmospheric conditions by doing away with the need for pre-collected datasets or preset statistical models. The FSO channel includes both air turbulence and pointing errors, which are modeled by Rayleigh and Gamma-Gamma distributions, respectively, to guarantee accurate modeling. Four traditional adaptation schemes, Optimal Rate Adaptation (ORA), Channel Inversion with Fixed Rate (CIFR), Truncated Inversion with Fixed Rate (TIFR), and Optimal Power and Rate Adaptation (OPRA), are used to compare the efficacy of the suggested approach. Evaluation of performance is done both with and without the RL-based optimization technique integrated. To solve connection issues in urban settings, the framework also investigates the deployment of Autonomous Aerial Vehicles (AAVs) as nimble, clever substitutes for conventional mobile base stations. Beam divergence, AAV positioning and orientation variations, link range, and coverage distance are just a few of the system and environmental elements that are thoroughly examined in this study in order to determine how they affect system performance. The IM/DD nonnegativity constraint is analytically integrated and evaluated to verify physical correctness, showing less than 1.5 % deviation in ASE at high SNR, while the classical Shannon capacity expression is used for capacity normalization to provide a consistent and computationally efficient benchmark. With an average inference latency of less than 8 ms and a power consumption of 12.6 W, the suggested model is also evaluated for real-time feasibility using software-in-the-loop (SIL) simulations, demonstrating its appropriateness for embedded UAV deployment. The suggested RL-DQN-A2C framework is further benchmarked by comparing it against a number of sophisticated RL models, such as Deep Meta-RL, Proximal Policy Optimization (PPO), Soft Actor-Critic (SAC), and Rainbow DQN. The findings of experiments verify that RL-DQN-A2C performs better in terms of achievable spectral efficiency (ASE), convergence stability, and flexibility in response to dynamic FSO channel errors. The efficiency of RL-DQN-A2C in optimizing adaptive transmission for drone-assisted FSO communication in urban smart city contexts is highlighted by its notable achievement of the greatest ASE of 128 bits/s/Hz, as well as its faster convergence and improved robustness. The RL-DQN-A2C framework considerably improves the ASE over all benchmark systems, according to simulation data. In particular, the OPRA method outperformed its baseline by 34.74 %, achieving the greatest ASE of 129 bits/s/Hz. Significant gains were also noted for TIFR of 31.11 %, CIFR of 16 %, (from 50 to 59 bits/s/Hz), and ORA with 32.76 %. These improvements are statistically significant (p<0.01), according to a one-way repeated-measures ANOVA, and Post hoc testing show that OPRA is superior (p<0.001). These findings highlight how the suggested RL-based method could significantly increase the spectrum efficiency and durability of drone-assisted FSO communication systems in smart city settings.