A Group And Individual Aware Framework For Fair Reinforcement Learning

Alexandra Cimpean, Catholijn M Jonker, Pieter Jules Karel Libin, Ann Nowe

Keywords: reinforcement learning, automated decision support, fairness framework, trustworthy AI

Abstract: Real-world sequential decision problems can be approached using a reinforcement learning approach. When these problems impact fairness across groups or individuals, considering fairness-aware techniques is crucial. Therefore, we require algorithms that can make suitable trade-offs between performance and the desired fairness notions. As the desired performance-fairness trade-off is difficult to specify a priori, we propose a framework where multiple trade-offs can be explored. As such, insights provided by the reinforcement learning algorithm, regarding the obtainable performance-fairness trade-offs, can be used by stakeholders to select the best policy for the problem at hand. To capture the appropriate fairness notions, we define an extended Markov decision process, $f$MDP, that explicitly encodes individuals and groups. Given this $f$MDP, we formalise fairness notions in the context of sequential decision problems. We formulate a fairness framework, that allows us to compute fairness notions over time. We evaluate our framework in two scenarios, each with distinct fairness requirements. The first is a job hiring setting, where strong teams must be composed, while providing equal treatment to the applicants. The second setting concerns fraud detection, where fraudulent transactions must be detected, while ensuring the burden for customers is distributed fairly.

Type Of Paper: Full paper (max page 8)

Anonymous Submission: Anonymized submission.

Submission Number: 22