Visualising Generative Spaces Using Convolutional Neural Network Embeddings (2210.17464v1)

Abstract: As academic interest in procedural content generation (PCG) for games has increased, so has the need for methodologies for comparing and contrasting the output spaces of alternative PCG systems. In this paper we introduce and evaluate a novel approach for visualising the generative spaces of level generation systems, using embeddings extracted from a trained convolutional neural network. We evaluate the approach in terms of its ability to produce 2D visualisations of encoded game levels that correlate with their behavioural characteristics. The results across two alternative game domains, Super Mario and Boxoban, indicate that this approach is powerful in certain settings and that it has the potential to supersede alternative methods for visually comparing generative spaces. However its performance was also inconsistent across the domains investigated in this work, as well as it being susceptible to intermittent failure. We conclude that this method is worthy of further evaluation, but that future implementations of it would benefit from significant refinement.