Analysing the effect of latent space mutation strategies for PCGML

Procedural Content Generation Via Machine Learning (PCGML), describes an evolving area of research, that incorporates the use of machine learning models in a generative context. These innovative approaches accelerate the creation of content while concurrently reducing the need for human intervention. In particular, autoencoder models have been steadily employed in collaboration with supporting algorithms to facilitate the generation of novel game content. The standard autoencoder provides a means of replicating given data by learning a lower dimensional embedding, called the latent representation. Applying various operations to these representations can yield uniquely diverse content. This paper offers a versatile and robust framework for game level generation using a standard autoencoder for level generation and a denoising autoencoder for level repair and enhancement. Additionally, the integration of clustering techniques effectively identifies the core components that make up the various types of levels in a Rogue-like game domain. The essence of this approach lies in harnessing the knowledge within the clusters to guide the mutation phase of the level’s latent spaces. This refined methodology provides significant insights on level analysis, generation and repair, highlighting how autoencoders can be used as a basis for game development.