Alone We Can Do So Little, Together We Can Do So Much: A Combinatorial Approach for Generating Game Content

Abstract

In this paper we present a procedural content generator using Non-negative Matrix Factorisation (NMF). We use representative levels from five dissimilar content generators to train NMF models that learn patterns about the various components of the game. The constructed models are then used to automatically generate content that resembles the training data as well as to generate novel content through exploring new combinations of patterns. We describe the methodology followed and we show that the generator proposed has a more powerful capability than each of generator taken individually. The generator's output is compared to the other generators using a number of expressivity metrics. The results show that the proposed generator is able to resemble each individual generator as well as demonstrating ability to cover a wider and more novel content space.