New outcome-fair RMAB policy reduces bias in deadline scheduling tasks

In stochastic deadline scheduling systems—where tasks must be completed before random deadlines—the classic Restless Multi-Armed Bandit (RMAB) framework with the Whittle index policy maximizes expected cumulative reward but ignores fairness across demographic groups. Researchers Shakti Sharma and Rahul Meshram from the Electrical Engineering and Systems Science domain address this gap by introducing an outcome-fair RMAB model. They define fairness criteria based on long-term completion rate guarantees and develop an outcome-fair Whittle index policy. A key innovation is a virtual queue mechanism that dynamically adjusts priorities to enforce equal average completion rates across structurally disadvantaged classes, without requiring explicit group labels in the scheduling decisions.

The team compared their outcome-fair Whittle index policy against standard Whittle (no fairness) and input-fair Whittle index approaches using numerical simulations. Results show the outcome-fair policy achieves significantly better fairness metrics, though with a modest reduction in total profit. Importantly, this trade-off diminishes as server capacity increases—meaning the policy scales efficiently in resource-rich environments. The work, presented in an 8-page conference paper, demonstrates that fairness constraints can be integrated into real-time scheduling without catastrophic efficiency loss, opening the door for more equitable AI-driven resource allocation in cloud computing, job queues, and real-time systems.