«
Jacob Asofsky
Software Development
Utilizing Variable-Population Genetic Algorithms for Balance Testing
My thesis is an exploration into genetic algorithms and their use in balance testing for games . With games moving towards live-service models with frequent updates , keeping a balance in-game is crucial to keeping player interest and assuring there isn ’ t one singular way to play the game . However , balance testing typically requires significant human testing , which can be slow and costly .
With my thesis , I built a system powered by these genetic algorithms that allows visualization of item disparity in an existing game , simply by leveraging the game ’ s existing artificial intelligence ( AI ).
The genetic algorithms are variablepopulation , which means the pool of players in the game increases and decreases based on their success during the game loop . This means the more successful AIs have more sway over the outcome , visualizing item disparities
earlier than a human could feasibly catch them . Since there are no human players , rendering can be removed to allow an uncapped simulation speed , condensing hours of gameplay testing into mere seconds .
To demonstrate these principles , I built a spaceship dogfighting game where teams of ships configure themselves with modifications to fight one another . The modules they pick control everything from thrust to weapon systems , and each AI has to use what they picked to battle . Using the genetic algorithm , approximately 30 hours of gameplay can be executed in a real-time minute , generating usage charts of which items were desired and by how much . As a tool , this could be a major help in large online titles to ensure no patches would negatively impact the in-game balance .
70 SOFTWARE DEVELOPMENT