Using causal discovery to analyze emergence in agent-based models

Analyzing agent-based models is a complex task. Agent-based models typically contain complex non-linear interactions between agents and generate emergent properties that cannot easily be explained. They are most commonly analyzed using sensitivity analysis techniques. While these techniques help understanding agent-based models better, they are not a one-size-fits-all solution. This paper explores the novel use of causal discovery algorithms from the field of causality as an additional means to analyze agent-based models. We propose the AbACaD methodology: Agent-based model Analysis using Causal Discovery. In this methodology, emergence in agent-based models is analyzed using causal discovery in combination with both machine learning and sensitivity analysis techniques. AbACaD combines different causal discovery algorithms, using a novel causal graph merging algorithm, to generate a causal graph based on agent-based simulation outcomes. This graph represents the causal relationships between the model parameters and the output variables of the model, and is then exploited to improve the understanding of emergent properties in the model. To demonstrate the effectiveness of AbACaD, it is applied to two models: the El Farol bar model, and an airport security and efficiency model. New emergent properties, such as the moment agents change their strategy in the El Farol bar model were identified. Furthermore, we found queue length to be an important factor in the number of casualties in an improvised explosive device (IED) attack. These emergent properties were well identified using AbACaD, but are hard to identify with traditional analysis techniques alone.