Computer games have become increasingly complex thus artificial game agents that are designed to learn within these environments require a significant number of high resolution sensors to facilitate good quality decision making processes. As a result, an agent’s sensor state space is often very high dimensional, which makes agent learning in real-time difficult. This PhD project will investigate approaches to dealing with the high dimensionality of an agents internal representation of a game space by both building the state representation in real-time and learning to improve the state transition process at the same time. Modular Reinforcement learning and Monte Carlo Tree Search methods will be considered in the first instance, as well as ways of compressing the representation; however as the PhD progresses other approaches including hybrid methods will be considered and devised.
First Supervisor: Charles, DK Dr
Second Supervisor: McNeill, MDJ Dr
Third Supervisor: McClean, SI Prof
Collaboration: This project does not involve collaboration with another establishment