Confidence-aware conditional value-at-risk-constrained multi-agent reinforcement learning in a synthetic benchmark for abstract underwater deception resource allocation

Description

This dataset and code package supports a study on confidence-aware CVaR-constrained multi-agent reinforcement learning for risk-sensitive underwater deception resource allocation under threat-identification uncertainty. The package provides a synthetic abstract-simulation benchmark with 100,000 episode-level records. It includes five compared methods: Heuristic, Risk-neutral MARL, Fixed-CVaR MARL, Constrained MARL, and CA-CVaR-MARL. Four uncertainty settings are included: Low, Medium, High, and OOD biased misclassification. The data contain performance, risk, confidence, ablation, robustness, computational-cost, and mechanism-analysis results. This dataset does not contain real underwater acoustic measurements, operational data, platform parameters, or deployable-system evidence. It is intended for reproducible analysis and methodological validation of confidence-aware tail-risk control in abstract multi-agent resource allocation.

Steps to reproduce

To reproduce the dataset and analysis, first install the required Python packages: ```bash pip install numpy pandas scipy matplotlib seaborn ``` After extracting the package, check the main configuration file: ```bash configs/simulation_config.json ``` This file defines the synthetic benchmark settings, including the number of agents, threat classes, resource categories, uncertainty levels, random seeds, and test episodes. The synthetic episode-level dataset can then be regenerated by running: ```bash python code/generate_synthetic_dataset.py ``` The main output is saved as: ```bash data/episode_results.csv ``` Next, run the data-quality audit: ```bash python code/data_quality_report.py ``` This produces: ```bash data/data_quality_report.txt ``` The report checks missing values, sample sizes, confidence-score bounds, binary indicators, non-negative risk costs, and valid inference-time values. To reproduce the robustness analysis and statistical tests, run: ```bash python code/reviewer_revision_analysis.py ``` This generates the confidence-robustness summaries and high-uncertainty statistical-test results, including: ```bash data/confidence_robustness_summary.csv data/statistical_tests_high_uncertainty.csv ``` Finally, reproduce the publication-style figures with: ```bash python code/make_publication_figures.py ``` The generated figures summarize the threat-identification uncertainty settings, performance across uncertainty levels, risk-return trade-off, and ablation/mechanism analyses. All outputs are based on synthetic abstract simulation and should not be interpreted as real underwater acoustic measurements, operational data, platform parameters, or deployable-system evidence.

Institutions

• Wuhan University of Technology

  Hubei, Wuhan

Categories

Licence