Open Collections

Spelchan, Billy Dwaine

University of British Columbia

Procedural generation of levels for video games has been around for decades but tends to not produce results as good as human-crafted maps. Procedural Content Generation via Machine Learning (PCGML) is a recent sub-category where machine learning approaches have been applied to the generation of levels. By training a model to replicate human-crafted designs it is anticipated that the generated levels will be of a quality similar to that produced by a designer. We believe that such techniques can also be used to aid designers in the creation of levels. To explore this, we have created a mixed initiative editor that uses a variety of metrics to provide useful information to the designer. A suggestion feature that uses a simple protocol allows for creation of several suggestion generators for the testing of different PCGML methods. We worked out several methods of providing suggestions and wrote generators using genetic algorithms, autoencoders, and Long Short Term Memory (LSTM) neural networks. A small user study was then conducted online (due to Covid-19 restrictions) which resulted in a rudimentary confirmation that such an editor was beneficial to the designer. The levels that users created during the study were compared against professional levels and procedurally generated levels using various metrics. We developed a few new metrics to help detect generation issues regarding tile placement. We further explore the suggestion mechanisms that we developed to see if the suggestions can be applied to other procedurally generated content to improve the quality of that content. We conclude that PCGML generators, and our new methods of generating levels, are useful and are a good fit for mixed initiative editors.

Text

2022-04-19

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/81243

Computer Science

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/