Automated Analysis of Regularities Between Model Parameters and Output Using Support Vector Regression in Conjunction with Decision Trees

Opening the black-box of nonlinear relationships between model inputs and outputs significantly contributes to the understanding about the dynamic problem being studied. Considering the weaknesses and disadvantages of human-guided and systematic techniques o ered in the literature, this paper presents a model analysis and exploration tool for agent-based models. The tool first approximates the input-output relationship by developing a metamodel, a simplified representation of the original agent-based model. For this purpose, it utilizes support vector regression, which is capable of approximating highly nonlinear systems accurately. Following the metamodel fitting, the tool incorporates a tree-based method to extract the knowledge embedded in the metamodel. The resulting tree is then expressed as a set of IF-THEN rules that have high comprehensibility compared to complex metamodel function. We utilize the tool for the exploration of Tra ic Basic model and the results transparently show the relationship between model inputs and output. Furthermore, rules extracted from the metamodel point out some counterintuitive results about the model which are not easily inferred from the raw input-output data. We also discuss potential uses of our tool, and provide the R script which makes the analysis repeatable on other agent-based models.