Visualising Generative Spaces Using Convolutional Neural Network Embeddings
In this EXAG 2022 presentation, Oliver Withington discusses how interest in procedural content generation (PCG) for games has increased both commercially and among researchers, so has the need for methodologies for understanding the generative spaces of PCG systems. They introduce and evaluate a novel approach for visualising the generative spaces of systems for generating game levels, using embeddings extracted from a trained convolutional neural network. They evaluate the approach in terms of its ability to produce 2D visualisations that correlate with the behavioural characteristics of the levels. The results across two alternative game domains, Super Mario and Boxoban, indicate that this approach is powerful in certain settings and worthy of further investigation. However the implementation used in this work was also inconsistent in comparison to the benchmarks, as well as being prone to intermittent failure. They conclude that this method is worthy of further evaluation, but that future implementations of it require significant refinement.