Abstract for: Evaluation of Alternative Dynamic Behavior Representations for Automated Model Output Classification and Clustering
Automated behavior mode identification and clustering are potentially valuable additions to the analysis toolset of a system dynamics (SD) modeler. The key component for such tools is the feature vector construction; selecting a set of features to represent the dynamic behaviors to be classified or clustered. In this study, we evaluate a set of alternative feature vectors in clustering basic behavior modes encountered in SD practice. As the first case, coefficients of the polynomials fitted to the dynamic behavior are used as the features. In the second case, a given set of curves are fitted to the dynamic behavior, and the degree of fit to these curves are used as the features. The third case constructs feature vectors based on the changes in the signs of slope and curvature of the behavior. In other words, the feature vector represents the original behavior as a sequence of atomic behavior modes. In our preliminary evaluation, the third approach outperformed the former two. Later, we propose a set of extensions to the third approach in order to improve its performance while dealing with oscillatory behaviors. The modified version of the third approach is evaluated to perform better than the original one in clustering both non-oscillatory and oscillatory dynamic behaviors.