Details
| Original language | English |
|---|---|
| Title of host publication | Intelligent Decision Technologies |
| Subtitle of host publication | Proceedings of the 14th KES-IDT 2022 Conference |
| Editors | Ireneusz Czarnowski, Robert J. Howlett, Robert J. Howlett, Lakhmi C. Jain |
| Place of Publication | Singapore |
| Pages | 87-97 |
| Number of pages | 11 |
| ISBN (electronic) | 978-981-19-3444-5 |
| Publication status | Published - 27 Jul 2022 |
| Event | 14th International KES Conference on Intelligent Decision Technologies, KES-IDT 2022 - Virtual, Online Duration: 20 Jun 2022 → 22 Jun 2022 |
Publication series
| Name | Smart Innovation, Systems and Technologies |
|---|---|
| Volume | 309 |
| ISSN (Print) | 2190-3018 |
| ISSN (electronic) | 2190-3026 |
Abstract
Multi-agent reinforcement learning facilitates agents learning to solve complex decision-making problems requiring collaboration. However, reinforcement learning methods are underpinned by stochastic mechanisms, making them unsuitable for safety-critical domains. To solve this issue, approaches such as assured multi-agent reinforcement learning, which utilises quantitative verification to produce formal guarantees of safety requirements during the agents learning process, have been developed. However, this approach relies on accurate knowledge about the environment to be effectively used which can be detrimental if this knowledge is inaccurate. Therefore, we developed an extension to assured multi-agent reinforcement learning called agent interaction driven adaptability, an automated process to securing reliable safety constraints, allowing inaccurate and missing knowledge to be used without detriment. Our preliminary results showcase the ability of agent interaction driven adaptability to allow safe multi-agent reinforcement learning to be utilised in safety-critical scenarios.
Keywords
- Assurance, Deep reinforcement learning, Multi-agent reinforcement learning, Quantitative verification, Safety
ASJC Scopus subject areas
- Decision Sciences(all)
- General Decision Sciences
- Computer Science(all)
- General Computer Science
Cite this
- Standard
- Harvard
- Apa
- Vancouver
- BibTeX
- RIS
Intelligent Decision Technologies : Proceedings of the 14th KES-IDT 2022 Conference. ed. / Ireneusz Czarnowski; Robert J. Howlett; Robert J. Howlett; Lakhmi C. Jain. Singapore, 2022. p. 87-97 (Smart Innovation, Systems and Technologies; Vol. 309).
Research output: Chapter in book/report/conference proceeding › Conference contribution › Research › peer review
}
TY - GEN
T1 - Assured Multi-agent Reinforcement Learning with Robust Agent-Interaction Adaptability
AU - Riley, Joshua
AU - Calinescu, Radu
AU - Paterson, Colin
AU - Kudenko, Daniel
AU - Banks, Alec
PY - 2022/7/27
Y1 - 2022/7/27
N2 - Multi-agent reinforcement learning facilitates agents learning to solve complex decision-making problems requiring collaboration. However, reinforcement learning methods are underpinned by stochastic mechanisms, making them unsuitable for safety-critical domains. To solve this issue, approaches such as assured multi-agent reinforcement learning, which utilises quantitative verification to produce formal guarantees of safety requirements during the agents learning process, have been developed. However, this approach relies on accurate knowledge about the environment to be effectively used which can be detrimental if this knowledge is inaccurate. Therefore, we developed an extension to assured multi-agent reinforcement learning called agent interaction driven adaptability, an automated process to securing reliable safety constraints, allowing inaccurate and missing knowledge to be used without detriment. Our preliminary results showcase the ability of agent interaction driven adaptability to allow safe multi-agent reinforcement learning to be utilised in safety-critical scenarios.
AB - Multi-agent reinforcement learning facilitates agents learning to solve complex decision-making problems requiring collaboration. However, reinforcement learning methods are underpinned by stochastic mechanisms, making them unsuitable for safety-critical domains. To solve this issue, approaches such as assured multi-agent reinforcement learning, which utilises quantitative verification to produce formal guarantees of safety requirements during the agents learning process, have been developed. However, this approach relies on accurate knowledge about the environment to be effectively used which can be detrimental if this knowledge is inaccurate. Therefore, we developed an extension to assured multi-agent reinforcement learning called agent interaction driven adaptability, an automated process to securing reliable safety constraints, allowing inaccurate and missing knowledge to be used without detriment. Our preliminary results showcase the ability of agent interaction driven adaptability to allow safe multi-agent reinforcement learning to be utilised in safety-critical scenarios.
KW - Assurance
KW - Deep reinforcement learning
KW - Multi-agent reinforcement learning
KW - Quantitative verification
KW - Safety
UR - http://www.scopus.com/inward/record.url?scp=85135905227&partnerID=8YFLogxK
U2 - 10.1007/978-981-19-3444-5_8
DO - 10.1007/978-981-19-3444-5_8
M3 - Conference contribution
AN - SCOPUS:85135905227
SN - 9789811934438
T3 - Smart Innovation, Systems and Technologies
SP - 87
EP - 97
BT - Intelligent Decision Technologies
A2 - Czarnowski, Ireneusz
A2 - Howlett, Robert J.
A2 - Howlett, Robert J.
A2 - Jain, Lakhmi C.
CY - Singapore
T2 - 14th International KES Conference on Intelligent Decision Technologies, KES-IDT 2022
Y2 - 20 June 2022 through 22 June 2022
ER -