2026-10-11T12:49:05Z https://www.scidb.cn/oai

10.57760/sciencedb.00zq1 2026-08-20T16:25:17Z

2026-08-20 Heterogeneous Air-Combat Formation MARL Dataset: Full RAC-MAPPO Training and Paired-Seed Evaluation Logs from a JSBSim 6-DoF Testbed doi:10.57760/sciencedb.00zq1 en This dataset supports the paper "Role-Attentive Centralised Multi-AgentReinforcement Learning for Heterogeneous Air-Combat Formations". It containsthe complete training and evaluation logs produced on a JSBSim sixdegree-of-freedom testbed (data only; no figures, no model weights):(1) training logs and 100-episode paired-seed evaluation records of RAC-MAPPOand all baselines (MAPPO, IPPO, HAPPO, MAT, QMIX, ROMA, RODE, HASAC) againsta four-level scripted adversary ladder (L1-L4);(2) full ablation runs (team reward, curriculum, set-attention critic, roleembedding, parameter sharing, Pk-aware fire gate);(3) generalisation (unseen geometry/equipment), scalability (3v2/6v6zero-shot, fine-tuning, from-scratch), self-play control, and point-massfidelity-control experiments;(4) step-by-step combat trajectory records with flight parameters (altitude,Mach, angle of attack, load factor, specific energy).All evaluations follow a paired-seed protocol; every number in the paper'stables and figures can be independently recomputed from this dataset.Format: JSONL/JSON, 909 files in 163 experiment directories. air combat; multi-agent reinforcement learning; heterogeneous formation; flight dynamics logs; paired-seed evaluation Zhou Changxin PUBLIC https://creativecommons.org/licenses/by/4.0/ dataset Science Data Bank