The price of decentralization in managing engineering systems through multi-agent reinforcement learning

This repository accompanies the paper and includes the imprl library used to study how multi-agent reinforcement learning decentralization affects performance when planning inspection and maintenance for engineering systems.

Repository Structure

.
├── experiments/               # published experiment configs and run instructions
│   ├── configs/
│   │   ├── k_out_of_n_infinite/
│   │   │   └── *.yaml
│   │   └── matrix_games/
│   │       └── *.yaml
│   └── ..
├── imprl/                     # published as a submodule containing the Python package
│   └── ..
├── sarsop_baseline/           # published Julia SARSOP baseline code
│   ├── main.jl
│   ├── ..
│   └── point-based-solvers/
│       └── *.toml
└── results/                   # published result tables, notebooks, and figures
    ├── data/ 
    │   └── *.csv
    ├── plot_results_main.ipynb
    ├── plot_policy_visualizations_kn_systems.ipynb
    ├── ..
    └── figures/
        └── *.png/.pdf

Installation

Clone the repository with the imprl submodule and install the Python package from imprl/:

git clone --recurse-submodules https://github.com/prateekbhustali/price-of-decentralization.git
cd price-of-decentralization

# if you already cloned without submodules
git submodule update --init --recursive

# install uv if needed
curl -LsSf https://astral.sh/uv/install.sh | sh

# create and activate a virtual environment
uv venv --python 3.11

# activate the virtual environment
source .venv/bin/activate

# installs dependencies using uv.lock (including imprl)
# sync using the imprl project + lockfile, but install into the active root .venv
uv sync --project imprl --active

See imprl/README.md for package-level installation details, optional dependency groups, and usage examples.

Results Data

The files below are already included in this repository under results/data/.

• results/data/climb_game_evals.csv: evaluation results for climb-game experiments.
• results/data/kn_infinite_results_epymarl.csv: k-out-of-n infinite-horizon EPyMARL comparison results.
• results/data/results_evals.csv: Evaluation metrics across experimental runs for each checkpoint.
• results/data/results_summary.csv: summary of results_evals.csv with best performing checkpoint per run.

Download Artifacts

You can download the following files from the Zenodo record for this project:

• model_checkpoints.tar.gz: all trained model checkpoints for all experiments.
• SARSOP_models_and_policies.tar.gz : SARSOP baseline models and policies for all environments.
• rollout_data.tar.gz: rollout data for k_out_of_n_infinite environments for all algorithms.

Use the interactive helper script from the repository root:

./download_artifacts.sh

The script downloads the Zenodo artifacts, extracts them into the expected repository locations, and removes common junk files such as .DS_Store.

Once you have the model checkpoints and SARSOP baseline models/policies downloaded, you can run the notebooks in results/ to reproduce the main results and figures in the paper.

Citation

@misc{bhustali_decentralization_2026,
      title={The price of decentralization in managing engineering systems through multi-agent reinforcement learning},
      author={Prateek Bhustali and Pablo G. Morato and Konstantinos G. Papakonstantinou and Charalampos P. Andriotis},
      year={2026},
      eprint={2603.11884},
      archivePrefix={arXiv},
      primaryClass={cs.MA},
      url={https://arxiv.org/abs/2603.11884},
}