The place of causal analysis in the analysis of simulation data

Abstract

This talk briefly reviews selected basic concepts and principles of structural approach to causal analysis, and outlines how they could be harnessed for analyzing and summarizing the data from simulations of complex dynamic systems, and for exploratory analysis of simulation models through machine learning. We illustrate the proposed method in the context of human behaviour modeling on a sample scenario from the EDA project A-0938-RT-GC EUSAS. The method revolves around the twin concepts of a causal partition of a variable of interest, and a causal summary of a simulation run. We broadly define a causal summary as a partition of the significant values of the analyzed variables (in our case the simulated motives fear and anger of human beings) into separate contributions by various causing factors, such as social influence or external events.We demonstrate that such causal summaries can be processed by machine learning techniques (e.g. clustering and classification) and facilitate meaningful interpretations of the emergent behaviours of complex agent-based models.