Bayesian treatment of model uncertainty for partially applicable models

Area of Science:

• Decision analysis
• Bayesian statistics
• Model uncertainty quantification

Background:

• Assessing model uncertainty is crucial for decision-making.
• Existing Bayesian frameworks handle multiple models and experimental data.
• Subjective expert beliefs on model credibility are often overlooked.

Purpose of the Study:

• To extend a Bayesian framework for model uncertainty assessment.
• To incorporate subjective expert beliefs on model credibility and applicability.
• To improve uncertainty quantification for unknown variables.

Main Methods:

• Specialization of a previously presented Bayesian framework for model uncertainty.
• Formalism treating models as information sources.
• Extension to include subjective information on model credibility and domain applicability.

Main Results:

• A methodology to assess and incorporate expert beliefs into uncertainty quantification.
• The framework accommodates predictions from multiple models and experimental validation.
• Demonstration of the approach using a fire risk modeling example.

Conclusions:

• Expert beliefs on model credibility can be formally integrated into uncertainty assessments.
• The extended Bayesian framework provides a more comprehensive approach to model uncertainty.
• This method is valuable for applications where model applicability is uncertain, such as fire risk analysis.