Analysis of agent-based simulation models through metamodeling

In this thesis, we propose a three-stage analysis procedure to interpret input-output relationships of agent-based simulation models. In the first stage, we use metamodels of agent-based models to approximate the input-output dynamics by constructing a functional relationship. For this purpose, we employ Support Vector Regression and Random Forest techniques from machine learning domain. Based on the results of an experimental analysis on Segregation and Traffic Basic models, we observe that Support Vector Regression stands as a promising technique only for output prediction while Random Forest is more appropriate when the main aim is to clearly depict input-output dynamics with a minimal loss of accuracy. In the second stage, we focus on efficient training of Random Forest metamodels. We observe that a sequential sampling strategy considering feedbacks from the metamodel generates more accurate metamodels compared to metamodels trained on randomly selected input-output couples. Besides, the iterative training process of metamodels also gives valuable information about the dynamics of the model by depicting boundary points between different types of model behaviors and counterintuitive outcomes. In the last stage, we devise a rule extraction technique from Random Forest metamodels based on a set partitioning problem formulation. Finally, we apply the three-stage analysis technique to an agent-based influenza epidemic model to analyze the effectiveness of different combinations of intervention strategies in the presence of various transmissibility scenarios of influenza.