go back

Disentangling Interaction using Maximum Entropy Reinforcement Learning in Multi-Agent Systems

David Rother, Jan Peters, Thomas Weisswange, "Disentangling Interaction using Maximum Entropy Reinforcement Learning in Multi-Agent Systems", 26th European Conference on Artificial Intelligence (ECAI 2023), pp. 1994-2001, 2023.


Research on multi-agent interaction involving multiple humans is still in its infancy. Most recent approaches have focused on environments with collaboration-focused human behavior, or providing only a small, defined set of situations. When deploying robots in general human-inhabited environments in the future it will be unlikely that all intentions can be guaranteed to fit a pre-defined model of collaboration while one might still expect a robot to behave collaboratively. Existing approaches are unlikely to effectively create such behaviors in such "co-existence" environments. To tackle this issue, we introduce a novel framework that decomposes interaction and task-solving into separate learning problems. Policies are learned with maximum entropy reinforcement learning to support their recombination, allowing us to create interaction-impact-aware agents and scaling the cost of training agents linearly with the number of agents and available tasks. For the combination of the different action distributions, we propose a weighting function covering their alignment with the original task to guarantee fulfillment of a robot's own goals. We demonstrate that our framework addresses the scaling problem while solving a given task and considering collaboration opportunities in a co-existence particle environment. Our work introduces a new learning paradigm that opens the path to more complex multi-robot, multi-human interactions.

Download Bibtex file Download PDF