Reinforcement Learning for Durable Algorithmic Recourse

Algorithmic recourse gives individuals rejected by an automated decision system counterfactual explanations — recommendations meant to improve their odds on a future application. Prior work uses reinforcement learning to keep such recommendations robust when the applicant population reacts strategically, but doesn't yet account for how the recommendations themselves shift new applicants' scores, or for keeping recommended scores stable across time.

This thesis extends that simulation framework to include these effects and adapts the reinforcement learning solution to stay robust under the added complexity.