Reinforcement learning for fair and ethical AI systems

Overview

This project studies how reinforcement learning (RL) can help make automated decisions fairer. Instead of fixing fairness after training, fairness is built into the learning process to create more equal outcomes for different groups. The focus is on important areas like hiring, healthcare, and finance, where biased AI can cause real harm.

The aim is to reduce unfair bias while keeping the system accurate, helping create AI that is both effective and socially responsible. You will learn about advanced RL methods, fairness measures, and how to design experiments, combining technical skills with important ethical issues.

Research activities

You will design fairness-aware RL algorithms, focusing on multi-objective optimization. You will design from scratch or improve SOTA algorithms, run experiments to evaluate performance and fairness using real-world datasets (e.g., hiring, healthcare), and compare their methods against standard approaches.

The work combines algorithm design, experimentation, and analysis, with the potential for research publication if sufficient novelty is achieved.

Outcomes

The project aims to develop and evaluate reinforcement learning methods that incorporate fairness constraints. Expected outcomes include:

• a working algorithm
• experimental results on real datasets
• insights into trade-offs between fairness and performance.

If the work demonstrates novelty, it may lead to a research publication.

Skills and experience

Students should have a foundation in Python programming and familiarity with machine learning concepts. Basic understanding of reinforcement learning principles and experience with ML libraries like PyTorch or TensorFlow will be helpful for implementing and experimenting with algorithms.

Scholarships

You may be eligible to apply for a research scholarship.

Explore our research scholarships

Keywords

• reinforcement learning
• fairness
• algorithmic bias mitigation
• ethical AI systems

Contact

Contact the supervisor for more information.