Ethically-Aligned Multiobjective Reinforcement Learning
Reinforcement Learning (RL) relies on a reward function as the only feedback in an interaction to learn complex tasks. Multi-objective RL (MORL) uses multiple reward functions. The aim of this thesis will be to produce new MORL algorithms that are capable to produce optimal and ethically aligned actions. The project will delve into ethically-aligned design to develop these new breeds of algorithms. The candidate will work closely with Prof. Hussein Abbass (http://www.husseinabbass.net/) team at UNSW-Canberra and will contribute MORL algorithms that could work in distributed environments; especially for the shepherding problem. The project is suitable for a candidate who would enjoy a balanced approach that covers theory, experiment, programming, and ethics.