Automatic Game Parameter Tuning using General Video Game Agents

Automatic Game Design is a subfield of Game Artificial Intelligence that aims to study the usage of AI algorithms for assisting in game design tasks. This dissertation presents a research work in this field, focusing on applying an evolutionary algorithm to video game parameterization. The task we are interested in is player experience. N-Tuple Bandit Evolutionary Algorithm (NTBEA) is an evolutionary algorithm that was recently proposed and successfully applied in game parameterization in a simple domain, which is the first experiment included in this project. To further investigating its ability in evolving game parameters, We applied NTBEA to evolve parameter sets for three General Video Game AI (GVGAI) games, because GVGAI has variety supplies of video games in different types and the framework has already been prepared for parameterization. 9 positive increasing functions were picked as target functions as representations of the player expected score trends. Our initial assumption was that the evolved games should provide the game environments that allow players to obtain score in the same trend as one of these functions. The experiment results confirm this for some functions, and prove that the NTBEA is very much capable of evolving GVGAI games to satisfy this task.