IMP-MARL: a Suite of Environments for Large-scale Infrastructure Management Planning via MARL

IMP-MARL offers a platform for benchmarking the scalability of cooperative MARL methods in real-world engineering applications.

In IMP-MARL, you can:

• Implement your own infrastructure management planning (IMP) environment or execute an available IMP environment.
• Train IMP policies through state-of-the-art MARL methods. The environments can be integrated with typical ecosystems via wrappers.
• Compute expert-based heuristic policies

Additionally, you will be able to:

• Retrieve the results of a benchmark campaign, where MARL methods are assessed in terms of scalability.
• Reproduce our experiments.

This repository has been developed and is maintained by Pascal Leroy & Pablo G. Morato.

Please consider opening an issue or a pull request to help us improve this repository.

Main requirements:

To work with our environments, one only needs to install Numpy.

However, to reproduce our results, more packages are required and installation instructions are provided here.

Tutorials

• Create your own IMP environment scenario
• IMP's API explained
• Train agents like in the paper and/or reproduce the results
• Retrieve directly the results
• Train your own MARL agents with PyMarl

Sets of environments available:

• (Correlated and uncorrelated) k-out-of-n system with components subject to fatigue deterioration.
• Offshore wind structural system with components subject to fatigue deterioration.

Note: A campaign cost can be activated in any environment.

Available wrappers with examples:

• Ready: PyMarl: Multi and single agent wrappers.
• Ready: Gymnasium: Single-agent wrapper.
• Ready: PettingZoo : Multi-agent wrapper.
• Ready: Rllib example: Single-agent training with RLLib and Gymnasium wrapper.
• WIP: MARLlib example: TBD
• WIP: TorchRL example: TBD

PyMarl algorithms available:

To train agents with PyMarl and one of the following algorithms, instructions are available here:

• QMIX: QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning
• QVMIX: QVMix and QVMix-Max: Extending the Deep Quality-Value Family of Algorithms to Cooperative Multi-Agent Reinforcement Learning
• QPLEX: QPLEX: Duplex Dueling Multi-Agent Q-Learning
• COMA: Counterfactual Multi-Agent Policy Gradients
• FACMAC: Factored Multi-Agent Centralised Policy Gradients
• VDN: Value-Decomposition Networks For Cooperative Multi-Agent Learning
• IQL: Independent Q-Learning

The main code is derived from PyMarl original implementation.

Expert-knowledge baselines available:

• Expert-based heuristic strategies

Run an IMP environment

env = Struct({'n_comp': 3,
               'discount_reward': 0.95,
               'k_comp': 2,
               'env_correlation': False,
               'campaign_cost': False})

obs, rewards_sum, done = env.reset(), 0, False
while not done:
    actions = {f"agent_{i}": random.randint(0,2) for i in range(3)}
    obs, rewards, done, insp_outcomes = env.step(actions) 

Citation

If you use IMP-MARL in your work, please consider citing our paper:

@misc{leroy2023impmarl,
      title={IMP-MARL: a Suite of Environments for Large-scale Infrastructure Management Planning via MARL}, 
      author={Pascal Leroy and Pablo G. Morato and Jonathan Pisane and Athanasios Kolios and Damien Ernst},
      year={2023},
      eprint={2306.11551},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}