SetDQN: An Agent Number Invariant Reinforcement Learning Model Using Set Transformers
Gabor Paczolay, Istvan Harmati · 2025
Many robotic tasks consist of several agents that are working on the same task with the same objective. In these tasks, it can be useful if the reinforcement learning architecture is not only suited to one configuration, but is invariant to agent numbers. Two utilization of this system would be a scalable realization and resilience to agent failure. In this paper, a novel multi-agent architecture is proposed with agent number invariance, utilizing Set Transformers. It is benchmarked on a robotic task where the agents are electromagnetic transmitters and they need to cover as large area as possible, where we achieve good results for the three-agent environment pretrained on four agents.