%0 Dataset
%T Multi-UAV Cooperative Encirclement Training and Trajectory Simulation Results Dataset
%J National Cryosphere Desert Data Center
%I National Cryosphere Desert Data Center(www.ncdc.ac.cn)
%U http://www.ncdc.ac.cn/portal/metadata/e60ccb9a-ecf2-4e0a-963d-02d66926768f
%W NCDC
%R 10.12072/ncdc.db7477.2026
%A ma wen rui
%K Multi-UAV;Cooperative Encirclement;Deep Reinforcement Learning;MADDPG;Training Reward;Trajectory Simulation
%X This dataset supports the analysis of deep reinforcement learning training processes and trajectory-simulation visualization for cooperative multi-UAV multi-target encirclement in a constrained two-dimensional space. The data are derived from the simulation experiments reported in the paper "Deep Reinforcement Learning-Based Cooperative Encirclement Strategy for Multi-UAVs". Under the centralized-training and decentralized-execution (CTDE) framework, the paper develops a stage-reward and multi-task-driven MADDPG algorithm. The encirclement mission is decomposed into continuous stages such as approaching, tracking, surrounding, and capturing, thereby alleviating reward sparsity in multi-agent encirclement tasks. A greedy algorithm is used for target assignment, and an artificial-potential-field-inspired escape strategy is designed for the target UAVs, forming a dynamic adversarial simulation environment. The experiments were conducted in a Python, PyTorch, and OpenAI Gym multi-agent simulation environment, where six encircling UAVs were trained to cooperatively encircle two target UAVs and the trained policy was evaluated under dynamic-obstacle conditions. The submitted files are organized in metadata, data, and images directories and include one metadata workbook, two CSV data files, two PNG visualization images, and one README file. The core data entities are data/score_history.csv, data/score_history_original_row62.csv, images/MADDPG_Training_Reward.png, and images/Encirclement_Trajectory.png. The training reward CSV contains three fields: episode, reward, and moving_average_100. The episode field records episode numbers from 1 to 10,000, reward records the accumulated reward of each episode, and moving_average_100 records the 100-episode moving-average reward. Both reward and moving_average_100 are dimensionless simulation reward values. This dataset can be used to analyze training convergence in multi-agent deep reinforcement learning, the effects of stage-rewar