Idea: Bridging Theoretical Fairness Definitions with Multi-Agent Coordination in the Real World

Current multi-agent reinforcement learning (MARL) theory defines fairness primarily as workload balance, assuming homogeneity across agents in real-world domains. Meanwhile, in fast-paced collaborative settings such as healthcare, there is a need to reduce cognitive overload by ensuring that tasks are allocated to agents based on their skill levels, which correlates with their efficiency at performing tasks. Hence, fairness requires skill-task alignment. Experiment results from exploring different fairness definitions, based on either workload balance, skill-task alignment, or both, reveal potential opportunities for theoretical contributions. Our findings reveal that (1) optimal fairness emerges from balancing multiple objectives rather than optimising single metrics, and (2) Stronger fairness penalties favour simpler algorithms unless properly tuned. We propose research directions to develop theoretical foundations that bridge current fairness concepts with the real-world coordination challenges of heterogeneous multi-agent systems.