| This thesis develops a method for identifying important input factors in large system dynamics models from an analysis based on those models' underlying structures. The identification of important input factors is commonly called factor screening and is a key step in the analysis of simulation models with many input parameters. Models under investigation are system dynamics models implemented as synchronous data flow programs, a model of computation that requires encoding the model components' dependencies in a graph format. The developed method views this graph as a stochastic process and attempts to rank the importance of inputs, or source nodes, with respect to an output, or non-source node. This ranking is accomplished primarily through the use of weighted random-walks through the graph. A comparison is made against other factor screening techniques, including fractional factorial experiments. The presented structure-based method is found to be comparably accurate to statistical factor screen experiments at magnitude order ranking. Run time of the developed method compared against a resolution III fractional factorial design is found to be similar for small models, and significantly faster for large models. |
