Abstract for:A Comparative Analysis of Alternative Designs for Automated Clustering SD Model Output
In the ideal modeling cycle, the analyst is expected to conduct simulation experiments for sensitivity analysis, validation, uncertainty analysis, scenario analysis, and policy design and analysis purposes. Let alone large-scale modeling projects, the set of experiments that should ideally be conducted even for a modestly sized simulation model turns out to be very large. However, in theory, this large set of experiments can be conducted in a fully automated manner. In that respect, the pattern recognition field appears to be the place to look for solutions. Considering the potential benefits of pattern recognition in general, and clustering approaches in specific, we provide a brief review of clustering algorithms and their basic components. We also provide the results of a set of experiments for comparing alternative designs for a SD-tailored clustering algorithm. Results on a sample dataset that represents the most commonly encountered dynamic patterns in SD studies show that a preprocessing step where the model output is scaled to a predefined range through standardization or normalization yields to significant accuracy improvements with almost no computational cost. The atomic behavior sequence (ABS) representation is evaluated to be an appropriate and promising way to capture the key pattern features for clustering purposes.