Experimental Dataset for Graph-Relational Curriculum Multi-Agent Reinforcement Learning with Bilateral Value Decomposition in Multi-UAV Cooperative Air Combat

This dataset is the complete experimental record of the paper "Graph-Relational Curriculum Multi-Agent Reinforcement Learning with Bilateral Value Decomposition for Multi-UAV Cooperative Air Combat", produced on the pure-Python MARL training and 3-D visualization scenario module of the Netherwind Air Combat Simulation Platform (JSBSim F-16-class 6-DOF flight dynamics). It contains: (1) 44 final deterministic re-evaluation CSV summaries—the source of every number in the paper's tables, each run evaluated over 100 paired episodes under common random numbers; (2) 182 SWA (stochastic weight averaged) PyTorch checkpoints actually evaluated in the paper, covering the full BCA model (15 seeds), seven baselines (IPPO, MAPPO, HAPPO, MADDPG, SAC, MAT, Random), ablation variants (w/o GRE, w/o BVD, w/o PBC, with reward shaping), hyperparameter fairness checks, sensitivity analyses, opponent-strength robustness tests, and 5v5 scalability runs; (3) 595 training-log CSVs organized into 38 experiment groups (per-episode return, win rate, losses, HP/ammunition statistics, and periodic evaluations); (4) 31 recorded combat state sequences (.npz: aircraft positions/attitudes, missile states, launch events) underlying the 3-D emergent-tactics figure (Fig. 10).

Together with the public code repository (github.com/debriscraft/NetherWind), the data-to-paper mapping table in the bundled README, and the exact commands in reproduce_paper.md, every number, table, and figure in the article can be independently reproduced.