IMP-MARL: a Suite of Environments for Large-scale Infrastructure Management Planning via MARL
Saved in:
| Title: | IMP-MARL: a Suite of Environments for Large-scale Infrastructure Management Planning via MARL |
|---|---|
| Authors: | Leroy, Pascal, Morato, Pablo G., Pisane, Jonathan, Kolios, Athanasios, Ernst, Damien |
| Source: | Advances in Neural Information Processing Systems (2023-12); Thirty-seventh Conference on Neural Information Processing Systems, La Nouvelle-Orléans, United States [US], 10/12/2023 |
| Publisher Information: | Morgan Kaufmann Publishers, 2023. |
| Publication Year: | 2023 |
| Subject Terms: | Computer Science - Learning, Computer Science - Multiagent Systems, cs.SY, eess.SY, Engineering, computing & technology, Computer science, Ingénierie, informatique & technologie, Sciences informatiques |
| Description: | We introduce IMP-MARL, an open-source suite of multi-agent reinforcementlearning (MARL) environments for large-scale Infrastructure Management Planning(IMP), offering a platform for benchmarking the scalability of cooperative MARLmethods in real-world engineering applications. In IMP, a multi-componentengineering system is subject to a risk of failure due to its components'damage condition. Specifically, each agent plans inspections and repairs for aspecific system component, aiming to minimise maintenance costs whilecooperating to minimise system failure risk. With IMP-MARL, we release severalenvironments including one related to offshore wind structural systems, in aneffort to meet today's needs to improve management strategies to supportsustainable and reliable energy systems. Supported by IMP practical engineeringenvironments featuring up to 100 agents, we conduct a benchmark campaign, wherethe scalability and performance of state-of-the-art cooperative MARL methodsare compared against expert-based heuristic policies. The results reveal thatcentralised training with decentralised execution methods scale better with thenumber of agents than fully centralised or decentralised RL approaches, whilealso outperforming expert-based heuristic policies in most IMP environments.Based on our findings, we additionally outline remaining cooperation andscalability challenges that future MARL methods should still address. ThroughIMP-MARL, we encourage the implementation of new environments and the furtherdevelopment of MARL methods. |
| Document Type: | conference paper http://purl.org/coar/resource_type/c_5794 conferenceObject peer reviewed |
| Language: | English |
| Relation: | https://github.com/moratodpg/imp_marl; urn:issn:1049-5258 |
| Access URL: | https://orbi.uliege.be/handle/2268/304451 |
| Rights: | open access http://purl.org/coar/access_right/c_abf2 info:eu-repo/semantics/openAccess |
| Accession Number: | edsorb.304451 |
| Database: | ORBi |
| Abstract: | We introduce IMP-MARL, an open-source suite of multi-agent reinforcementlearning (MARL) environments for large-scale Infrastructure Management Planning(IMP), offering a platform for benchmarking the scalability of cooperative MARLmethods in real-world engineering applications. In IMP, a multi-componentengineering system is subject to a risk of failure due to its components'damage condition. Specifically, each agent plans inspections and repairs for aspecific system component, aiming to minimise maintenance costs whilecooperating to minimise system failure risk. With IMP-MARL, we release severalenvironments including one related to offshore wind structural systems, in aneffort to meet today's needs to improve management strategies to supportsustainable and reliable energy systems. Supported by IMP practical engineeringenvironments featuring up to 100 agents, we conduct a benchmark campaign, wherethe scalability and performance of state-of-the-art cooperative MARL methodsare compared against expert-based heuristic policies. The results reveal thatcentralised training with decentralised execution methods scale better with thenumber of agents than fully centralised or decentralised RL approaches, whilealso outperforming expert-based heuristic policies in most IMP environments.Based on our findings, we additionally outline remaining cooperation andscalability challenges that future MARL methods should still address. ThroughIMP-MARL, we encourage the implementation of new environments and the furtherdevelopment of MARL methods. |
|---|