Learning Coordinated Preference for Multi-Objective Multi-Agent Reinforcement Learning
The work proposes Preference Coordinated Multi-agent Policy Optimization (PCMA) to improve cooperative multi-agent reinforcement learning. It learns coordinated agent-specific preferences that allow agents to make complementary trade-offs, moving beyond simple scalar reward optimization. This approach models the problem as finding a team equilibrium where preference diversity leads to measurable first-order improvements in the overall team objective.