Studying the Effects of Training Data on Machine Learning-Based Procedural Content Generation

Abstract

The exploration of Procedural Content Generation via Machine Learning (PCGML) has been growing in recent years. However, while the number of PCGML techniques and methods for evaluating PCG techniques have been increasing, little work has been done in determining how the quality and quantity of the training data provided to these techniques effects the models or the output. Therefore, little is known about how much training data would actually be needed to deploy certain PCGML techniques in practice. In this paper we explore this question by studying the quality and diversity of the output of two well-known PCGML techniques (multi-dimensional Markov chains and Long Short-term Memory Recurrent Neural Networks) in generating Super Mario Bros. levels while varying the amount and quality of the training data.