Adaptive NPC Behavior in Maze Chase Game Using Genetic Algorithms

head shot of Karen Suzue

Name: Karen Suzue
Major: Computer Science
Minor: Studio Art
Advisor: Drew Guarnera; Sofia Visa (second reader)

Game balancing, which involves fine-tuning in-game factors to match the capabilities and demands of the player, is a crucial aspect of game design. This is because having good game balance maximizes player enjoyment or satisfaction. While there are many factors involved in creating a balanced game, one of the most common types of game balancing is difficulty adjustment. This involves ensuring a balance between the player’s skill and the level of challenge present in game.

Static difficulty adjustment (SDA) is currently the most widely employed form of difficulty balancing in video games. In many games, players are given a fixed range of predefined difficulty modes to choose from. Such difficulty modes are derived from data generalization and extensive trial-and-error testing that is both costly and tedious. Furthermore, this approach produces unreliable results as it fails to consider the wide range of skill among players. Thus, it is often difficult for games with SDA to consistently ensure player satisfaction.

To solve these issues, we turn to dynamic difficulty adjustment (DDA) which modifies game features in real time depending on the player’s skill. An intuitive approach to implementing DDA would be through genetic algorithms (GA), a group of unsupervised learning algorithms based on biological evolution and natural selection. This study explores the use of GA for DDA with a focus on adjusting NPC behavior. Based on previous theoretical techniques, a maze-chase game and GA are developed for demonstration purposes. In this game, the player plays against a group of enemies through several runs and allows the enemy population to evolve a set of desirable traits. In the future, the resulting enemy traits can be saved into a “seeding pool” or “case bank” to be injected in later playthroughs. Results from testing shows that the GA behaves in a logical manner, rewarding faster enemies when the player wins and slower enemies when the player loses. Further research is needed, however, to implement the case-injection phase of the game.

Posted in Comments Enabled, Independent Study, Symposium 2022.

4 responses to “Adaptive NPC Behavior in Maze Chase Game Using Genetic Algorithms”

  1. Benjamin Hassan says:

    I like the evolution idea, Karen. How would this ideally be implemented in a video game?

  2. Pavithra says:

    Excellent work Karen! Congratulations

  3. Max Johnson (Fellow Gamer) says:

    Congrats! One day I hope we can use these findings to better develop Among Us single player, a sorely lacking game mode. I’ll have my people reach out to your people when I get to the difficulty curve of Minecraft 2.

  4. Jodi says:

    Karen, this is amazing. I believe even the thinking process that brought you to your conclusions will be immensely valuable applied to real life, beyond game design.
    CONGRATULATIONS! You did it!!