AdaFairMARL: Enforcing Adaptive Fairness Constraints in Multi Agent Reinforcement Learning

Cooperative multi agent reinforcement learning (MARL) can produce highly unequal workloads, particularly when agents differ in capability. In safety critical teams, such imbalance can overburden individual agents and delay important actions. Existing fairness approaches in MARL primarily modify rewards through fixed penalties, inequality aversion, or social welfare objectives. These methods can encourage fair behavior, but they provide no direct framework for specifying and enforcing a desired fairness level. We introduce AdaFairMARL, a constrained MARL framework that enforces a user specified workload fairness level beyond reward shaping.