Submodular Multi-Agent Policy Learning for Online Distributed Task Allocation in Open Multi-Agent Systems
SubMAPG is a centralized training, decentralized execution framework for multi-agent systems with non-additive task allocation problems under partition constraints. It uses a novel continuous relaxation called Partition Multilinear Extension (PME) and submodular difference rewards to derive unbiased credit assignment signals. This allows agents to learn optimal policies that respect the constraint of each agent performing only one action per round.