Unmanned aerial vehicle (UAV) ad hoc networks are increasingly vital for civilian and military applications due to their rapid deployment and collaborative capabilities. However, their dynamic nature and diverse quality of service (QoS) demands, especially for time-sensitive multi-priority services (e.g., command, control, payload), impose significant challenges on medium access control (MAC) protocols. The statistic priority-based multiple access (SPMA) protocol, distinguished by its priority differentiation and channel load statistics, is a promising candidate for UAV networks. A critical factor influencing SPMA performance, particularly end-to-end delay and transmission success rate, is the backoff mechanism. Existing SPMA backoff strategies predominantly rely on static function models (e.g.,, Binary Exponential Backoff- BEB, logarithmic functions) with limited optimization dimensions (often only priority). These static approaches are inadequate for adapting to the highly dynamic and heterogeneous environment of UAV ad hoc networks, characterized by fluctuating channel loads, varying priority thresholds, and stringent resource constraints. This necessitates the development of an intelligent backoff strategy that can dynamically learn optimal backoff times by comprehensively considering multiple factors, such as service priority, current channel load, and priority thresholds, to enhance network efficiency and meet diverse QoS requirements.

To address the limitations of static backoff models, this paper modeled the dynamic decision-making process of selecting backoff times in SPMA as a Markov decision process (MDP). An intelligent backoff strategy based on the double deep Q-network (DDQN) algorithm was proposed. The UAV node acts as the agent interacting with the network environment. The state space was designed to capture both the priority of the current data packet to be sent and the channel load statistics, including historical values to learn trends. The action space consisted of discrete backoff times within finite windows defined by the maximum validity periods of different priority services. A reward function was meticulously crafted to guide learning, incorporating: (1) a backoff result factor (+1 for success, -1 for failure upon re-checking conditions after backoff); (2) the priority of the service (emphasizing higher priority); (3) the difference between the priority threshold and the channel load at the end of the backoff period (promoting minimal, efficient backoff). By employing separate current and target Q-networks updated at different frequencies, the DDQN framework mitigated the overestimation issue common in DQN and enhanced stability. The agent selects actions (backoff times) using an epsilon-greedy policy, balancing exploration and exploitation. Training leverages experience replay for sample efficiency. Crucially, the "offline training, online deployment" paradigm was adopted, where the DDQN model was trained on a server, and the lightweight inference model was deployed on UAVs, minimizing onboard computational burden.

The performance of the proposed DDQN-based intelligent backoff strategy was rigorously evaluated through simulation using OPNET Modeler 14.5 for the SPMA protocol and Python for the ML components, incorporating dynamic elements such as node mobility and bursty traffic. Comparative analyses were conducted against traditional BEB and a logarithmic function-based backoff strategy under various traffic load scenarios. Key results demonstrate the superiority of the DDQN approach: (1) Significant delay reduction: The strategy achieves up to a 33.3% reduction in transmission delay for low-priority services compared to the benchmark strategies. (2) Improved success rate: It enhances the initial backoff success rate by up to 18%, indicating a higher probability of successful transmission immediately after the backoff period. (3) Enhanced adaptability: The strategy effectively improves the overall transmission success rate and demonstrates robust adaptability to variations in network scale (number of UAV nodes). These improvements are attributed to the DDQN agent's ability to learn complex, non-linear relationships between the state factors (priority, channel load) and the optimal backoff action, dynamically adjusting to network conditions far more effectively than static models.

This study successfully addresses the critical challenge of designing an efficient backoff mechanism for the SPMA protocol in dynamic UAV ad hoc networks. By formulating the backoff time selection as an MDP and leveraging the DDQN algorithm, we developed an intelligent backoff strategy that dynamically optimizes backoff times based on a comprehensive consideration of service priority, channel load statistics, and priority thresholds. The simulation results conclusively demonstrate that this strategy outperforms traditional BEB and function-based approaches, significantly reducing delay (especially for low-priority traffic), improving initial backoff success rates, and enhancing overall transmission reliability while adapting well to network size changes. The "offline training, online deployment" approach ensures practical feasibility on resource-constrained UAV platforms. This DDQN-driven intelligent backoff technology presents a promising solution for achieving low-latency and high-reliability MAC layer communication in mission-critical UAV operations, paving the way for more efficient and robust UAV ad hoc networks. Future work could explore decentralized training and robustness under extreme mobility or jamming.