Code and Artifact Supplement for: "Explanations as a Reward Signal: Closed-Loop Active XAI for Hierarchical Reinforcement Learning in Contested Airspace"
Year: 2026
Type: software
Abstract: CODE AND ARTIFACT SUPPLEMENT FOR:"EXPLANATIONS AS A REWARD SIGNAL: CLOSED-LOOP ACTIVE XAI FORHIERARCHICAL REINFORCEMENT LEARNING IN CONTESTED AIRSPACE" This repository contains the complete code, pre-trained model weights, evaluationscenarios, and statistical analysis scripts required to reproduce all findingsreported in the manuscript: "Explanations as a Reward Signal: Closed-Loop Active XAI for Hierarchical Reinforcement Learning in Contested Airspace" İ. Ardıç, N. Yumuşak Sakarya University · Turkish Aerospace Submitted to IEEE Access (2026). CONTENTS • Training/ — Strategic DQN, Tactical PPO, and Maneuver SAC training scripts with fixed seeds (1024 / 2048 / 4096). • Model_3_Oracle/, Model_4_Blind/, Model_5_ActiveXAI/ — pre-trained model weights and evaluation harnesses for the privileged-oracle baseline, the blind-ablation baseline, and the proposed Active XAI agent. • scenarios/ — 20,000 procedurally generated training scenarios and 1,000 deterministic Grand-Test scenarios with fixed evaluation seeds. • scripts/ — statistical analysis (chi-square, TOST equivalence test), faithfulness measurement for both Strategic DQN and Tactical PPO, SHAP latency benchmark, diagnostic-figure generation, and the lambda / tau sensitivity sweeps. • results/ — pre-computed result files (JSON) for reproducibility cross-checking. • requirements.txt — pinned Python dependencies (Python 3.14, PyTorch 2.9, Stable-Baselines3 2.7, JSBSim 1.3, SHAP 0.49). • README.md — full reproduction guide. REPRODUCIBILITY All seed values are fixed. Per-episode seeds use the deterministic mapping"42 + episode_id". Running scripts/reproduce_table_VI.py reproduces Table VIof the manuscript exactly; scripts/analyze_statistics.py reproduces thechi-square (p = 0.9313) and TOST (p = 0.0062, +/- 3% margin) results;scripts/generate_diagnostic_figure.py reproduces Figure 8 deterministicallyfrom Model_5_ActiveXAI/results_model5.json. ACCESS POLICY This archive is shared under RESTRICTED ACCESS in light of dual-useconsiderations associated with high-fidelity autonomous flight-controlartifacts (discussed in Section VII-F of the manuscript). The editorial teamand peer reviewers can request direct access credentials from thecorresponding author for the review process. Following publication, thematerials will remain in restricted-access mode and will be made available toqualified researchers for non-commercial academic reproducibility uponreasonable request to the corresponding author ([email protected]). CITATION If you use this artifact, please cite the IEEE Access article (forthcoming)and this Zenodo record: Ardıç, İ. and Yumuşak, N. (2026). "Code and Artifact Supplement for: 'Explanations as a Reward Signal: Closed-Loop Active XAI for Hierarchical Reinforcement Learning in Contested Airspace'". Zenodo. DOI: 10.5281/zenodo.20606445. URL: https://doi.org/10.5281/zenodo.20606445 FUNDING & AFFILIATION This research was conducted independently as part of doctoral studies atSakarya University. The corresponding author is employed by Turkish Aerospace(TUSAŞ); however, this work was carried out without direct funding orinfluence from the employer. The authors declare no conflict of interest.
Authors: Ardıç, İbrahim, Yumusak, Nejat
Institution: Sakarya University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7164026194