Learning to Target Incentives for Cooperation in Networked Games: Code and Data

This repository provides the source code, trained model checkpoints, generated datasets, and machine-readable evaluation results associated with the study **“Learning to Target Incentives for Cooperation in Networked Games.”**

The project studies how targeted incentives can promote cooperation in networked evolutionary games with pairwise and higher-order interactions. It combines dynamical simulations, statistical analyses, robustness experiments, centrality-based controls, and a hypergraph neural network that learns to rank agents according to their causal rescue leverage.

 Contents

The archive includes:

- A simulator for networked games with pairwise and higher-order interactions.
- Code for basin-stability, intervention-threshold, failure-depth, robustness, and sensitivity analyses.
- A hypergraph neural network for state-aware incentive targeting.
- Trained GNN and matched nodewise-MLP checkpoints.
- Fixed-pulse rollout-greedy, centrality, targeted-control, and neural baseline implementations.
- Topology-disjoint, scale-free, and empirical-network evaluation code.
- Machine-readable results in CSV, JSON, NPZ, and PKL formats.
- Statistical-analysis and figure-generation scripts.
- Reproducibility instructions, dependency versions, data-provenance information, and SHA-256 checksums.

The main generated result files include `results.csv`, `raw_uc.npz`, model-evaluation JSON files, robustness and stringency summaries, cached evaluation records, and trained PyTorch checkpoints.

Reproducibility

The archive contains fixed random seeds, experimental protocols, cached results, and trained checkpoints so that the principal quantitative claims can be inspected without repeating every computationally expensive experiment.

A complete description of the relationship between scripts, outputs, analyses, and reported results is provided in `REPRODUCIBILITY.md`. File integrity can be checked using `MANIFEST.sha256`.

Some supplementary analyses are reproduced by rerunning the corresponding seeded scripts rather than by loading archived per-run plotting arrays.

 Empirical contact-network data

The empirical evaluation uses the `contact-high-school` temporal simplicial dataset distributed by Austin R. Benson.

The original third-party files are not redistributed in this archive because the source page does not specify a repository-wide redistribution license. Download and deterministic preprocessing instructions are provided in `DATA_SOURCES.md`.

The archive includes the resulting machine-readable empirical evaluation metrics, but users who wish to rerun the preprocessing and evaluation must obtain the original dataset from its official source.

Software environment

The experiments were developed with Python 3.12 and use NumPy, SciPy, pandas, Matplotlib, and PyTorch. Tested dependency versions are listed in `requirements.txt`.

Large simulation sweeps and full neural-network retraining may require substantial CPU or GPU resources. Existing checkpoints and result records are included for verification.

License

Project-authored source code and documentation are released under the MIT License. Third-party datasets remain subject to their original licenses and terms of use.

 Citation

If you use this software or its generated results, please cite both this Zenodo record and the associated paper. Citation information for the paper should be added after publication.