Location
Hilton Hawaiian Village, Honolulu, Hawaii
Event Website
https://hicss.hawaii.edu/
Start Date
3-1-2024 12:00 AM
End Date
6-1-2024 12:00 AM
Description
The Agent Environment Cycle (AEC) of PettingZoo has been a major paradigm shift in the implementation of Multi-Agent Reinforcement Learning (MARL) frameworks, providing a unified and concise interface for any kind of multi-agent environment. Based on this model, we propose DRAMA, a principled approach for dynamic action space restrictions. DRAMA can be used to add statically computed physical constraints as well as a self-learning multi-agent governance: It generalizes the idea of action masking to continuous action spaces and self-learning restrictions, while being fully compatible with the AEC implementation of PettingZoo—and, by transitivity, with most major MARL frameworks. In this paper, we provide the theoretical background of restricted multi-agent systems, present an extension of PettingZoo via wrapper classes, and show the potential of our approach for various use cases. By treating dynamic restrictions as an additional player of a multi-agent system, our approach offers novel capabilities and flexibility in handling multi-agent environments and thus serves as a valuable tool for researchers and practitioners in the field.
Recommended Citation
Oesterle, Michael; Grams, Tim; and Bartelt, Christian, "DRAMA at the PettingZoo: Dynamically Restricted Action Spaces for Multi-Agent Reinforcement Learning Frameworks" (2024). Hawaii International Conference on System Sciences 2024 (HICSS-57). 13.
https://aisel.aisnet.org/hicss-57/st/sw_development/13
DRAMA at the PettingZoo: Dynamically Restricted Action Spaces for Multi-Agent Reinforcement Learning Frameworks
Hilton Hawaiian Village, Honolulu, Hawaii
The Agent Environment Cycle (AEC) of PettingZoo has been a major paradigm shift in the implementation of Multi-Agent Reinforcement Learning (MARL) frameworks, providing a unified and concise interface for any kind of multi-agent environment. Based on this model, we propose DRAMA, a principled approach for dynamic action space restrictions. DRAMA can be used to add statically computed physical constraints as well as a self-learning multi-agent governance: It generalizes the idea of action masking to continuous action spaces and self-learning restrictions, while being fully compatible with the AEC implementation of PettingZoo—and, by transitivity, with most major MARL frameworks. In this paper, we provide the theoretical background of restricted multi-agent systems, present an extension of PettingZoo via wrapper classes, and show the potential of our approach for various use cases. By treating dynamic restrictions as an additional player of a multi-agent system, our approach offers novel capabilities and flexibility in handling multi-agent environments and thus serves as a valuable tool for researchers and practitioners in the field.
https://aisel.aisnet.org/hicss-57/st/sw_development/13