Uncertainty Quantification and Data Fusion Based on Dempster-Shafer Theory

Yanyan He

2013-01-01

Abstract:Quantifying uncertainty in modeling and simulation is crucial since the parameters of the physical system are inherently non-deterministic and knowledge of the system embodied in the model is incomplete or inadequate. The most well-developed nonadditive-measure theory – the DempsterShafer theory of evidence – is explored for uncertainty quantification and propagation. For “uncertainty quantification,” we propose the MinMax method to construct belief functions to represent uncertainty in the information (data set) involving the inseparably mixed type of uncertainties. Using the principle of minimum uncertainty and the concepts of entropy and specificity, the MinMax method specifies a partition of a finite interval on the real line and assigns belief masses to the uniform subintervals. The method is illustrated in a simple example and applied to the total uncertainty quantification in flight plan of two actual flights. For “uncertainty propagation,” we construct belief/probability density functions for the output or the statistics of the output given the belief/probability density functions for the uncertain input variables. Different approaches are introduced for aleatory uncertainty propagation, epistemic uncertainty propagation, and mixed type of uncertainty propagation. The impact of the uncertain input parameters on the model output is studied using these approaches in a simple example of aerodynamic flow: quasi-one-dimensional nozzle flow. In the situation that multiple models are available for the same quantity of interest, the combination rules in the Dempster-Shafer theory can be utilized to integrate the predictions from the different models. In the present work, we propose a robust and comprehensive procedure to combine multiple bodies of evidence. It is robust in that it can combine multiple bodies of evidence, consistent or otherwise. It is comprehensive in the sense that it examines the bodies of evidence strongly conflicted with others, reconstructs the basic belief mass functions by discounting, and then fuses all the bodies of evidence using an optimally parametrized combination rule. The proposed combination procedure is applied to radiotherapy dose response outcome analysis.