Abstract for:Classification of Fundamental System Dynamics Model Outputs

Several computer-based tools have been developed to automate certain tasks in System Dynamics (SD) modeling cycle such as hypothesis testing, validation, and policy analysis. Introduction of these tools is motivated by the dramatic increase in the number of experiments and the number of resulting dynamic patterns due to the increase in the number of model parameters. However, the analysis of model outputs require expert visual judgment to determine how the system qualitatively evolves over time. Most of the computer-based tools use a Hidden Markov Model based classification algorithm, ISTS, which was developed to replace human expert judgment. However, ISTS fails to provide correct classification under certain challenging conditions. This study introduces standard time series classification techniques to classify SD model outputs. Although there is no statistical significance, experimental results show that Symbolic Representation for Multivariate Time Series (SMTS) and 1-nearest neighbor algorithms with Dynamic Time Warping (DTW) and Longest Common Subsequence (LCSS) outperform ISTS in terms of classification accuracy for non-noisy data. However, ISTS cannot handle noisy data and its classification performance significantly deteriorates. As a good performing approach for both noisy and non-noisy data, SMTS is planned to be implemented in BATS, a recent SD model analysis tool.