文献类型： 外文期刊

作者： Hu, Chunyang ¹ ; Li, Jingchen ² ; Yang, Yusen ² ; Gu, Qiong ¹ ; Wu, Zhao ¹ ; Ning, Bin ¹ ;

作者机构： 1.Hubei Univ Arts & Sci, Sch Comp Engn, Xiangyang 441053, Hubei, Peoples R China

2.Beijing Acad Agr & Forestry Sci, Informat Technol Res Ctr, Beijing 100079, Peoples R China

关键词： Unmanned aerial vehicle; Large-scale multi-agent systems; Multi-agent reinforcement learning

期刊名称：INTERNATIONAL JOURNAL OF FUZZY SYSTEMS （ 影响因子：3.6； 五年影响因子：3.1 ）

ISSN： 1562-2479

年卷期： 2025 年

页码：

收录情况： SCI

摘要： In unmanned aerial vehicle (UAV) swarm confrontations, the optimal policies obtained through deep reinforcement learning methods face an exponential increase in computational and storage resource consumption with the number of UAVs. To achieve efficient policies in large-scale UAV swarm confrontations while keeping the amount of parameters and floating-point operations within an acceptable range, this study proposes a method based on fuzzy multi-agent reinforcement learning. This method models the confrontation in large-scale UAV swarms as a fuzzy game, establishing the corresponding group decision-making process. With the proof of the Markov property, interactions among UAVs are fuzzified into interactions among a few abstract agents, so that policies assigned to abstract agents rather than individual UAV, while the storage consumption is reduced. Through defuzzification calculations, policies of abstract agents are mapped to specific UAV behaviors, significantly reducing the computing consumption while ensuring policy effectiveness. Comparative experiments with other baseline methods show that our approach significantly reduces the required floating-point operations and parameters in UAV swarm confrontations of various numbers of UAVs, with comparable performance of the learned polices.