General level generation

Abstract

This thesis explores techniques and metrics for creating game level generators that can work for different games with minimal changes. We focus on methods that generate content by searching and that generate generators by searching. For the first approach, we explore search-based techniques to find a set of playable and diverse levels. We introduce the Constrained MAP-Elites algorithm to generate levels using a set of generalizable diversity and quality metrics. We test the technique on five different games: Super Mario Bros and four games in General Video Game AI corpus (Zelda, Solarfox, Plants, and Portals). The results show that on simple games with few mechanics such as Zelda and Super Mario Bros, the algorithm is able to find a set of diverse and playable levels; less so for more complex games. Next, we approach the problem of general level generation from a different angle: instead of generating the level, we generate the level generator itself. We introduce two different representations to represent level generators: neural network weights and a generator description language called Marahel. Neural networks performed better than using Marahel scripts as all the trained generators were able to generate almost 100% playable levels, while the Marahel generators struggled to achieve playability. On the other hand, the Marahel generators are more understandable and can be easily modified by humans compared to the neural network ones. Overall, both approaches can help us to achieve our goal of general level generation from two different perspectives. Searching for content takes a long time but can produce a large corpus of diverse and playable levels. This makes this approach suitable for offline generation. Searching for generators takes a long time, but the found generators are very fast, although they do not guarantee playability. The speed of these generators allows for multiple re-sampling, making them suitable for online generation.