- Author
- Droguett, E. L.
- Title
- Methodology for the Treatment of Model Uncertainty.
- Coporate
- Maryland Univ., College Park
- Report
- THESIS, 1999, 334 p.
- Keywords
- model studies | uncertainty | methodology | fire models | risks | temperature | fire plumes
- Identifiers
- Bayesian approach; models' information as point estimates; non-homogeneous set of performance data; multiple models; arbitrary posterior distribution; models evidence as parametric probability distributions; uncertainty assessment of discrete quantities; assessment of confidence, applicability and dependence; surveying existing methodologies; experimental data for point source fire
- Abstract
- The development of a conceptual, unified, framework and methodology for treating model and parameter uncertainties is the subject of this work. Firstly, a discussion on the philosophical grounds of notions such as reality, modeling, models, and their relation is presented. On this, a characterization of the modeling process is presented. The concept of uncertainty, addressing controversial topics such as type and sources of uncertainty, are investigated arguing that uncertainty is fundamentally a characterization of lack of knowledge and as such all uncertainty are of the same type. A discussion about the roles of a model structure and model parameters is presented, in which it is argued that a distinction is for convenience and a function of the stage in the modeling process. From the foregoing discussion, a Bayesian framework for an integrated assessment of model and parameter uncertainties is developed. The methodology has as its central point the treatment of model as source of information regarding the unknown of interest. It allows for the assessment of the model characteristics affecting its performance, such as bias and precision. It also permits the assessment of possible dependencies among multiple models. Furthermore, the proposed framework makes possible the use of not only information from models (e.g., point estimates, qualitative assessments), but also evidence about the models themselves (performance data, confidence in the model, applicability of the model). The methodology is then applied in the context of fire risk models where several examples with real data are studied. These examples demonstrate how the framework and specific techniques developed in this study can address cases involving multiple models. use of perfonnance data to update the predictive capabilities of a model, and the case where a model is applied in a context other than one for which it is designed.