Surrogate Models of Self-Organising Systems

Surrogate models have been heavily used in artificial intelligence to speed up machine learning and optimisation processes. Yet, they also offer other groundbreaking perspectives, including reduction of computational costs at runtime and built-in explainability. These perspectives are of great relevance to self-organising systems research, considering the potentially high costs for computing their components' high degrees of freedom, the arising complex system dynamics and emergent system properties. Surrogate modelling strives to tackle these challenges by approximating the original models through statistical analysis, resulting in replacement models that are significantly easier to compute or interpret. In this paper, we consider existing use cases of according surrogate modelling approaches and highlight currently faced tradeoffs. Finally, we present a research agenda that shows how complexity reduction may be used to overcome these tradeoffs and provide a basis for interdisciplinary research endeavours in this field.