AdaFair-MARL: Enforcing Adaptive Fairness Constraints in Multi-Agent Reinforcement Learning
Published in Submitted to IEEE Conference on Decision and Control (CDC) 2025, 2025
Recommended citation: Promise Ekpo, S. Agarwal, F. Grimm, J. Liu, L. Molu, A. Taylor. "AdaFair-MARL: Enforcing Adaptive Fairness Constraints in Multi-Agent Reinforcement Learning." Submitted, IEEE CDC, 2025. https://arxiv.org/pdf/2511.14135
An adaptive fairness framework (Fair-GNE) for multi-agent reinforcement learning that automatically learns penalty parameters via Lagrangian dual ascent, eliminating manual hyperparameter tuning for fairness in heterogeneous teams. It improves workload balance over fixed-penalty baselines (JFI: 0.89 vs. 0.33, p < 0.01, Cohen’s d = 8.99) while maintaining 86% task success, with theoretical convergence guarantees through KKT conditions (95-100% KKT satisfaction, 88-100% constraint satisfaction across fairness thresholds).