A methodology for statistical sensitivity analysis of system dynamics models

Abstract

Simulation models consist of many di erent components that analysts cannot estimate in perfect precision and results are always subject to some level of uncertainty. So, sensitivity analysis of system dynamics models should be conducted in order to reach more reliable conclusions. The selection of sensitivity type should be appropriate for the purpose of the model. Speci cally, system dynamics modeling is a behavior pattern-oriented methodology and sensitivity analysis of such models should consider the changes in behavior patterns rather than the numerical values of variables. In this thesis, a new approach, called behavior pattern sensitivity analysis, is suggested for dealing with the parameter uncertainty. In this approach pattern characteristics of output behavior are subject to statistical analysis. Changes in pattern measures, such as in ection point, period or amplitude, as a result of di erent parameter values are measured in order to determine possible sensitivity points of the model. In this thesis, two statistical analysis procedures are utilized for measuring sensitivities to model parameters. Firstly, sensitivity data is evaluated with regression method, which is a convenient way of multi-variate analysis. Secondly, ANOVA of clusters (Kleijnen and Helton, 1999a) is used on the same sensitivity data to make comparison with the regression results. This approach is applied to three di erent system dynamics models, which are the project management model by Taylor and Ford (2006), a generic supply line model and the inventory workforce model by Sterman (2000).