Adaptive sampling with feature elimination for agent-based models

In this thesis, we propose an efficient and user-friendly metamodeling procedure simultaneously incorporating adaptive sampling with feature elimination in order to successfully capture the relationships between the parameters and the output of simulation models in the presence of insignificant model parameters with respect to an output of interest. In this procedure, Random Forest metamodel is utilized. While adaptive sampling efficiently yields the metamodel with a high-quality training set, feature elimination promotes the performance and efficiency of the metamodeling procedure by reducing complexity and dimensionality due to insignificant parameters. In this respect, the proposed metamodeling procedure is also potentially applicable to high-dimensional simulation models. For illustrative purposes, the proposed procedure is applied to an agent-based segregation model. It is observed that adaptive sampling strategy can generate metamodels with higher predictive performance compared to input-oriented and random sampling strategies. Yet, the metamodel accuracy is considerably influenced by the mtry parameter of Random Forest metamodel and the presence of the insignificant simulation model parameters. However, experimental results show that the proposed procedure successfully alleviates the drawbacks of insignificant simulation parameters on the metamodel and sampling by eliminating them, and reduces the dependency of the metamodel performance to the mtry parameter. As a result, the proposed metamodeling procedure stands out as a robust and efficient way to produce a metamodel that successfully approximates the dynamics embedded in the simulation model. The proposed procedure can be applied by researchers to reduce the time required for comprehensive model analyses, to gain awareness about the model behaviors, and to detect the model parameters conditioning the model behavior.