Free Space Optical (FSO) communication systems offer high bandwidth and immunity to electromagnetic interference, making them attractive for next-generation wireless communications. However, they are highly susceptible to atmospheric conditions such as fog, dust, and precipitation, which can severely degrade signal quality. In this paper, we propose a novel hybrid artificial intelligence framework that enhances the robustness and adaptability of FSO systems. Our approach integrates deep computer vision and reinforcement learning (RL) techniques to dynamically respond to changing weather conditions. First, we utilize a convolutional neural network to classify environmental weather conditions from captured images. This image-based classification enables precise identification of channel impairments, providing crucial state information for decision-making. Next, we employ a Q-learning agent that receives the classified weather condition and selects optimal actions to adjust the transmission power and data rate. The goal is to maintain signal quality by maximizing the Signal-to-Noise Ratio, minimizing Bit Error Rate, and reducing power consumption. Extensive experiments were conducted using a custom dataset of weather images and a simulated FSO channel model. Results demonstrate that the proposed system can accurately classify seven distinct weather types and significantly improve FSO performance through intelligent control.

Abstract Inter‐cell interference and smart jammer attacks significantly impair the performance of non‐orthogonal multiple access (NOMA) networks. This issue is particularly critical when considering strategic interactions with malicious actors. To address this challenge, the power allocation problem is framed in a two‐cell NOMA network as a sequential game. In this game, each base station acts as a leader, choosing a power allocation strategy, while the smart jammer acts as a follower, reacting optimally to the base stations' choices. To address this multi‐agent scenario, four multi‐agent reinforcement learning algorithms are proposed: Q‐learning based unselfish (QLU), deep QLU, hot booting deep QLU, and decreased state deep QLU. A game‐theoretic analysis that demonstrates the algorithms' convergence to the optimal network‐wide strategy with high probability is provided. Simulation results further confirm the superiority of our proposed algorithms compared to the Q‐learning‐based selfish NOMA power allocation method.

ABSTRACT In this study, an innovative architecture is proposed to enhance the performance of semantic communication networks by leveraging deep learning and joint source‐channel coding. A fundamental challenge in this field is the strong dependence of conventional networks on a fixed signal‐to‐noise ratio (SNR) during training, which leads to performance degradation under varying channel conditions. To address this limitation, we introduce a novel attention‐based approach that enables dynamic adaptation to different SNR levels, ensuring more stable and optimized communication performance. The proposed model learns more generalized features that exhibit greater resilience to channel variations. To evaluate its effectiveness, extensive simulations were conducted, comparing the performance of the proposed architecture with DeepSC, a state‐of‐the‐art benchmark model in the field. While the baseline model, trained at a single SNR, experiences performance drops under mismatched conditions, the proposed model, trained across a range of SNRs, achieves improvement of 16.2%, 30.8%, 42.8%, and 53.8% for 1, 2, 3, and 4‐gram precisions, respectively, in bilingual evaluation understudy score and an 11.4% increase in sentence similarity across challenging low‐SNR conditions. Furthermore, the model maintains robust performance with 48% less training data, highlighting its efficiency and data efficiency under practical constraints. These gains confirm the model's superior adaptability and high‐quality data reconstruction under diverse conditions. The results of this study underscore the significant benefits of attention‐based architectures in semantic communication, particularly in environments with unpredictable channel variations, and highlight their potential for reliable deployment in real‐world applications.

Free-space optical (FSO) communications provide higher bandwidth and immunity to electromagnetic interference compared to radio communications. To increase FSO system capacity, space-division multiplexing (SDM) with orbital angular momentum (OAM) modes has been explored. However, atmospheric turbulence induces power leakage among modes, leading to significant intermodal interference and performance degradation. Efficient power allocation and mode selection are essential to mitigate these effects. This paper addresses the joint mode selection and power allocation (JMSPA) problem for optimizing data transmission in OAM-multiplexed FSO systems in the presence of atmospheric turbulence. The proposed algorithm extends the Generalized Linear Fractional Programming (GLFP) approach by leveraging its polyblock approximation and projection framework while introducing key modifications to handle the mixed-integer nonlinear fractional structure of JMSPA. Unlike standard GLFP, JMSPA employs a specialized projection step based on a bisection-based search, enabling efficient resolution of the nonlinear rate constraints within the max-min weighted rate (WR) problem. Additionally, instead of a single polyblock approximation, JMSPA constructs distinct polyblocks for each mode subset, effectively managing the mode selection constraints and ensuring a globally optimal solution. The performance of the proposed algorithms is evaluated through channel simulations under different turbulence levels. Results demonstrate significant link performance improvements over uniform power allocation, with notable gains in stronger turbulence conditions.

In this study, two novel multi‐hop relay‐assisted hybrid FSO/RF communication systems are presented and compared. In these structures, RF and FSO links, at each hop, are parallel and send data simultaneously. This is the first time that in a multi‐hop hybrid FSO/RF structure, detect and forward protocol is used. In the first structure, at each hop, received signals with a higher signal‐to‐noise ratio are selected. However, in the second structure, at each hop, received FSO and RF signals are separately detected and forwarded and selection is done only at the last hop. Considering FSO link in negative exponential atmospheric turbulence and RF link in Rayleigh fading, for the first time, closed‐form expressions are derived for outage probability ( ) and bit error rate of the proposed structures. MATLAB simulations are provided to verify derived expressions. Results indicate that the structure with the selection at each hop has better performance than the structure with a selection at the last hop.