Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Personalised platformer-game generation via PCGML algorithms

Chaitidis Savvas

Full record

---

Summary

In recent years, the game development industry has shown an enormous interest in the field of Procedural Content Generation (PCG). Especially the autonomous generation of levels for video games has become a popular subject of study. Generative Adversarial Networks (GANs) are a machine learning method that has been used over the last decade and proved to be capable of generating media content and even game content. For this thesis, a new video game that combines the characteristics of a Platformer and a Dungeon Crawler has been developed in Unity from which a handcrafted dataset has been created. A GAN was successfully trained on this dataset and was therefore able to generate new level rooms. Since the core of this work was to generate rooms whose difficulty scale is adjusted based on the player’s performance, the input latent vector to the generative model was found by using an Evolutionary Strategy. Specifically, the Covariance Matrix Adaptation Evolutionary Strategy (CMA-ES) was applied which searched through the latent space of the GAN by utilizing the static characteristics of each room in form of a fitness function. The output room of this process was then given a difficulty score by evaluation and placed in the correct spot of the 3x3 grid which constitutes the whole level. The experiments of this thesis are resulted from having a variety of player models and actual players play the game and show how capable the trained GAN model is in generating novel example outputs that fit the player’s performance.

Available Files

Services