This dataset supports the paper "Cardinality-Adaptive Multi-Agent Reinforcement Learning for Variable-Scale Cooperative Aerial Games"

Abstract

This dataset supports the paper "Cardinality-Adaptive Multi-Agent Reinforcement Learning for Variable-Scale Cooperative Aerial Games" (submitted to Applied Intelligence). It contains the complete training and evaluation records of CA-MAPPO, a cardinality-adaptive multi-agent proximal policy optimization algorithm for variable-scale cooperative aerial-game formations, generated with the JSBSim six-degree-of-freedom flight-dynamics engine in the loop.The dataset includes: (1) per-iteration training logs (episode return, periodic win rate, policy/value losses, entropy, KL divergence; JSONL) from three-seed main training runs and the gated three-stage curriculum (single-scale consolidation at 2v2, mixed-scale main stage on {2v2, 3v3, 5v5}, and ace-weighted final consolidation), covering trained scales and strictly held-out zero-shot scales {4v4, 6v6}; (2) paired evaluation results (win/draw/loss counts with Wilson confidence intervals; JSON) for CA-MAPPO, MAPPO, QMIX and ablation variants, organized by opponent level (L1–L4) × team size under a common-random-number protocol with 20 episodes per cell, plus zero-shot extrapolation evaluations at 10v10, 14v14 and 20v20; (3) ablation study data (without geometric attention bias, without set-attention critic, single-scale training, without per-scale advantage normalization, without win-rate-gated curriculum, UPDeT-faithful adaptation, and double-budget controls); (4) full-episode step-by-step simulation records (per-step state, hybrid action, reward, termination flag, and flight parameters for every agent); (5) automated maneuver-detector statistics behind the emergent-tactics analysis and self-attention weight maps (NPZ) behind the attention-analysis figures.The dataset contains raw data only — no figures, no plotted curves, no trained model weights.