Detecting Emergent Behavior in Complex Systems: A Machine Learning Approach

The live identification of emergent behavior in complex systems with little a-priori information is a challenging task and existing approaches are either applicable to a small subset of models or do not scale well. In contrast, post-mortem approaches that have a more in-depth understanding of the characteristics of emergent properties often struggle with analyzing a large amount of data to extract relationships between the variables, events, and entities whose interaction eventually leads to emergent behavior. Machine
learning approaches have been promoted as potential replacements of existing approaches, due to their ability to analyze large amounts of data without a-priori knowledge of existing relationships. In this paper, we present a first step towards the use of supervised learning approaches to identify and predict emergent behavior. Our hybrid approach unifies live and post-mortem perspectives by relying on a visual inspection of the simulation run and the simulation data set to identify a set of features that are more likely to generate emergent behavior (post-mortem) which are then used by a machine learning
module to predict emergent behavior (live). Our analysis shows the potential of such approaches but also highlights challenges and future avenues of research.