Unmanned aerial vehicle (UAV) systems offer significant advantages in terms of rapid decision-making processes, precise operational execution, and robust performance in complex environments. With the evolution of security paradigms and advancements in equipment systems, collaborative multi-UAV operations have become essential for modern protection frameworks and represent a key developmental direction. To address the need for coordinated multi-objective operations within complex network environments—and to overcome the limitation in current research where single operational loops can only handle individual objectives—this study proposes an innovative intelligent loop recommendation method. By formulating the operational loop recommendation problem as a network flow shortest-path model, we implement an exact algorithm to generate operational loops that enable coordinated multi-UAV operations. Additionally, we develop a learning-inspired algorithm (LIA) incorporating Pareto optimization strategies and specialized learning mechanisms to effectively resolve multi-objective conflicts in UAV task allocation. This research integrates complex network theory with operational loop optimization concepts, providing new technical support for intelligent mission management systems.

Grid-based maritime search and rescue (SAR) planning requires sequential decisions from incomplete spatial observations. Increasing map size also expands the search space and can intensify reward sparsity. This paper proposes CCAL-DQN, an LSTM deep Q-network with a cropping-concatenation (CC) observation representation and a deterministic boundary sign mask for simulated SAR path planning under partial observability. The method formulates the task as a partially observable Markov decision process and encodes observation features before action-value estimation. Deterministic boundary-aware cropping and concatenation keep the observation input dimension fixed, with nonlocal samples included near map boundaries. In the original evaluation, the method was trained on $9\times 9$ maps and evaluated without fine-tuning on fixed $20\times 20$ maps with an 80-step mission budget. The reported score is the map-wise ratio of policy-collected value to the incumbent returned by Gurobi within 60 s on the same map. CCAL-DQN achieves a higher mean relative score than the classical ACO, PSO, and GA algorithms under the tested conditions. It also has the shortest measured post-training evaluation time, excluding offline training cost. In the reported DRL comparison, LSTM-DQN obtains 0.5275 compared with 0.4232 for DRQN. Cropping-concatenation obtains 0.66 compared with 0.64 for full observation. Additional implementation and CC sign comparisons yielded higher mean collection fractions for the LSTM-based and inversion-enabled variants, respectively, under their separately specified protocols. A supplementary analysis using DDQN, DRQN, and PPO indicates that the effect of global context augmentation depends on the backbone. These results support fixed-size observation representations for algorithmic search planning in the tested simulated-grid setting.