Dynamic programming with Q-learning based reinforcement learning optimization for multi unmanned aerial vehicles swarm path planning

Path planning for multi-Unmanned Aerial Vehicle (UAV) swarms presents critical challenges due to the dynamic mission environments and diverse cost metrics involved. Traditional algorithms often struggle with high-dimensionality and dense obstacles, limiting their adaptability and performance. To tackle these challenges, we integrated the Dynamic Programming with Q-Learning (Dyna-Q) method from reinforcement learning and proposed a novel artificial intelligence-based metaheuristic algorithm (MA) called the Dyna-Q-based Improved Artemisinin Optimization (DQAO), and introduced two strategies to improve Dyna-Q: the Real-Sim Interaction Strategy (RIS) and the Distribution Guided Decision Strategy (DDS), RIS transforms the agent based on action strategies with the most extreme outcomes in the environment model, enabling the interaction between virtual and real experiences, and the DDS analyzes the overall distribution of action-state pairs stored in the model and uses it as a basis to optimize the search process. This adaptive mechanism allows DQAO to dynamically adjust its update policies, resulting in improved optimization performance and robustness. Extensive experiments on the Congress on Evolutionary Computation 2005 (CEC2005) and Congress on Evolutionary Computation 2017 (CEC2017) benchmark suites demonstrate that DQAO achieves superior accuracy, convergence speed, and path safety compared to several state-of-the-art algorithms, especially in complex and obstacle-rich scenarios.