Skill-Aligned Fairness in Multi-Agent Learning for Collaboration in Healthcare (FairSkillMARL)
Published in Submitted to NeurIPS 2026 Datasets & Benchmarks Track, 2025
Recommended citation: Promise Ekpo, B. La, T. Wiener, S. Agarwal, A. Agrawal, G. Gonzalez-Pumariega, L. Molu, A. Taylor. "Skill-Aligned Fairness in Multi-Agent Learning for Collaboration in Healthcare." Submitted, NeurIPS 2026 Datasets & Benchmarks Track. https://arxiv.org/abs/2508.18708
A skill-aligned fairness framework for multi-agent reinforcement learning, with a composite fairness metric balancing workload equality and skill-task alignment. FairSkillMARL achieves 40-60% higher skill-task alignment than workload-only baselines, with a 70-85% reduction in workload-range disparity while maintaining comparable success rates (83% vs. 80-86%). We also introduce MARLHospital, a MARL benchmark supporting heterogeneous agents with skill heterogeneity, energy levels, and customizable team compositions.