Abstract

This thesis contributes to the development of sophisticated machine learning-based AI for video games. To address the problem of limited control over the behavior of machine-learning AI and simplify its development process, the concept of formalized game spaces combined with a hierarchical structure is presented. Three different AI variants were developed as a proof of concept to accomplish a common exploration task in the turn-based strategy game Tiny Empires. The first AI just used deep reinforcement learning, while both the second and third AI were provided with additional game space data. Moreover, the third AI was structured hierarchically to allow for greater control through modularization of tasks. All three variants were compared in terms of the training progression, task performance, and learned behavior to determine if this approach is useful for machine learning in game AI.

Results/Implementation/Project Description

Conclusion

[ PDF (optional) ] 

[ Slides Kickoff/Final (optional)]