Learning Coordinated Preference for Multi-Objective Multi-Agent Reinforcement Learning

The paper introduces Preference Coordinated Multi-agent Policy Optimization (PCMA), a novel approach to cooperative multi-objective multi-agent reinforcement learning (MOMARL) that learns agent-specific preferences to facilitate trade-offs among agents with diverse roles and observations. Theoretical foundations are provided by framing cooperative MOMARL as a team-optimal game, demonstrating that preference diversity can enhance overall team performance. Experimental results across various cooperative MOMA environments, including a traffic-control scenario, indicate significant improvements in both performance and coordination of trade-offs, which is crucial for practitioners working on multi-agent systems with conflicting objectives.