The Right Variety: Improving Expressive Range Analysis with Metric Selection Methods

Expressive Range Analysis serves as a critical tool for evaluating procedural content generation, yet the field has historically lacked a standardized, objective methodology for selecting the metrics used to visualize generative spaces. To address this, a rigorous framework has been established to replace intuition-based selection with three quantitative criteria: Fitness Independence, Mutual Independence, and Minimal Alternative Metric Independence. These heuristics ensure that chosen metrics provide the most informative representation of a system’s potential output, preventing the arbitrary reuse of redundant or misleading data points.

The efficacy of this approach was demonstrated through an analysis of 9,015 Mario levels, utilizing a diverse set of 18 structural and agent-based metrics. The findings indicate that pairing structural metrics with agent-evaluated metrics yields the most robust visualizations, as these distinct categories are less prone to mutual correlation. By systematically applying these selection criteria, researchers can avoid the common pitfalls of skewed results caused by outliers and ensure that the resulting plots accurately reflect the underlying possibility space of a generative system.

While this methodology significantly enhances the precision of generative space analysis, it remains a computationally intensive process that currently relies on monotonic correlation methods. Future advancements in this field will likely focus on identifying metrics that maintain performance across varied game domains and incorporating non-linear relationship modeling to further refine the accuracy of these visualizations. By moving toward this more disciplined analytical standard, designers can achieve a more nuanced understanding of the diversity and expressive capacity of their procedural generation systems.